diff options
Diffstat (limited to 'llama.cpp/ggml/src/ggml-opencl/kernels')
86 files changed, 13024 insertions, 0 deletions
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl new file mode 100644 index 0000000..509bf17 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl @@ -0,0 +1,190 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// add +//------------------------------------------------------------------------------ + +// general-purpose kernel for addition of two tensors +// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3 +// cons: not very efficient +kernel void kernel_add( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) + *((global float *)(src1_ptr + i10*nb10)); + } +} + +// assumption: src1 is a row +// broadcast src1 into src0 +kernel void kernel_add_row( + global float4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float4 * dst, + ulong offsetd, + int ne +) { + src0 = (global float4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] + src1[idx1]; +} + +kernel void kernel_add_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int type_src0, + int type_src1 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + + half v0, v1; + if (type_src0 == 1) { + v0 = convert_half(*((global float *)(src0_ptr + i0*nb00))); + } else { + v0 = *((global half *)(src0_ptr + i0*nb00)); + } + + if (type_src1 == 1) { + v1 = convert_half(*((global float *)(src1_ptr + i10*nb10))); + } else { + v1 = *((global half *)(src1_ptr + i10*nb10)); + } + + *((global half *)(dst_ptr + i0*nb0)) = v0 + v1; + } +} + +kernel void kernel_add_row_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global half4 * dst, + ulong offsetd, + int ne, + int type_src0, + int type_src1 +) { + dst = (global half4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + + half4 v0, v1; + if (type_src0 == 1) { + global float4* src0_f32 = (global float4*)((global char*)src0 + offset0); + v0 = convert_half4(src0_f32[gid]); + } else { + global half4* src0_f16 = (global half4*)((global char*)src0 + offset0); + v0 = src0_f16[gid]; + } + + if (type_src1 == 1) { + global float4* src1_f32 = (global float4*)((global char*)src1 + offset1); + v1 = convert_half4(src1_f32[idx1]); + } else { + global half4* src1_f16 = (global half4*)((global char*)src1 + offset1); + v1 = src1_f16[idx1]; + } + + dst[gid] = v0 + v1; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl new file mode 100644 index 0000000..e9c6d55 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl @@ -0,0 +1,42 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// add_id +//------------------------------------------------------------------------------ +kernel void kernel_add_id( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb02, + ulong nb11, + ulong nb21, + int ne0, + int ne1 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + src2 = (global char*)((global char*)src2 + offset2); + dst = (global char*)((global char*)dst + offsetd); + + int i1 = get_group_id(0); + int i2 = get_group_id(1); + + const int i11 = *((global const int *) (src2 + i1*sizeof(int) + i2*nb21)); + + const size_t nb1 = ne0 * sizeof(float); + const size_t nb2 = ne1 * nb1; + + global float * dst_row = (global float *)((global char *)dst + i1*nb1 + i2*nb2); + global float * src0_row = (global float *)((global char *)src0 + i1*nb01 + i2*nb02); + global float * src1_row = (global float *)((global char *)src1 + i11*nb11); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + dst_row[i0] = src0_row[i0] + src1_row[i0]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl new file mode 100644 index 0000000..af4adc7 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl @@ -0,0 +1,86 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define SWAP(x, y, T) { T tmp = (x); (x) = (y); (y) = tmp; } + +enum ggml_sort_order { + GGML_SORT_ORDER_ASC, + GGML_SORT_ORDER_DESC, +}; + +kernel void kernel_argsort_f32_i32( + global float * src0, + ulong offset0, + global int * dst, + ulong offsetd, + const int ne00, + const int ne00_pad, + const int order, + local int * dst_row +) { + // bitonic sort + int col = get_local_id(0); + int row = get_group_id(1); + + if (col >= ne00_pad) { + return; + } + + src0 = (global char *)((global char *)src0 + offset0); + dst = (global float *)((global char *)dst + offsetd); + + global float * x_row = src0 + row * ne00; + + // initialize indices + dst_row[col] = col; + + barrier(CLK_LOCAL_MEM_FENCE); + + for (int k = 2; k <= ne00_pad; k *= 2) { + for (int j = k / 2; j > 0; j /= 2) { + int ixj = col ^ j; + if (ixj > col) { + if ((col & k) == 0) { + if (dst_row[col] >= ne00 || + (dst_row[ixj] < ne00 && (order == GGML_SORT_ORDER_ASC ? + x_row[dst_row[col]] > x_row[dst_row[ixj]] : + x_row[dst_row[col]] < x_row[dst_row[ixj]])) + ) { + SWAP(dst_row[col], dst_row[ixj], int); + } + } else { + if (dst_row[ixj] >= ne00 || + (dst_row[col] < ne00 && (order == GGML_SORT_ORDER_ASC ? + x_row[dst_row[col]] < x_row[dst_row[ixj]] : + x_row[dst_row[col]] > x_row[dst_row[ixj]])) + ) { + SWAP(dst_row[col], dst_row[ixj], int); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + } + + // copy the result to dst without the padding + if (col < ne00) { + dst[row * ne00 + col] = dst_row[col]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl new file mode 100644 index 0000000..ae60324 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl @@ -0,0 +1,20 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// clamp +//------------------------------------------------------------------------------ +kernel void kernel_clamp( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + float min, + float max +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = src0[get_global_id(0)] < min ? + min : + (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl new file mode 100644 index 0000000..0c1b3d7 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl @@ -0,0 +1,51 @@ +kernel void kernel_concat_f32( + global const char * src0, + ulong offset0, + global const char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int dim +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + const int i3 = get_group_id(2); + const int i2 = get_group_id(1); + const int i1 = get_group_id(0); + + int o[4] = {0, 0, 0, 0}; + o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); + + global const float * x; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (global const float *)(src0 + (i3 )*nb03 + (i2 )*nb02 + (i1 )*nb01 + (i0 )*nb00); + } else { + x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10); + } + + global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + *y = *x; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl new file mode 100644 index 0000000..e339c90 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl @@ -0,0 +1,185 @@ +#ifdef USE_FP16 +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#define T_FLOAT half +#define T_FLOAT4 half4 +#define VSTORE_T_FLOAT4(data, offset, p) vstore_half4_rte(data, offset, p) +#else +#define T_FLOAT float +#define T_FLOAT4 float4 +#define VSTORE_T_FLOAT4(data, offset, p) vstore4(data, offset, p) +#endif + +#if defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#else +#define REQD_SUBGROUP_SIZE_128 +#endif + +#define T_ACCUM float4 +#define VEC_SIZE 4 + +#define BS_K 64 +#define BS_NPQ 64 +#define BS_CRS 16 + +#define TS_K 4 +#define TS_NPQ 8 + +#define WG_K (BS_K / TS_K) +#define WG_NPQ (BS_NPQ / TS_NPQ) + +#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE) +#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE) + +static inline uint splitWork(uint work_size, uint block_size){ + return (work_size + block_size - 1) / block_size; +} + +REQD_SUBGROUP_SIZE_128 +kernel void kernel_conv_2d( + global void* p_knl, + ulong off_knl, + global void* p_src, + ulong off_src, + global void* p_dst, + ulong off_dst, + local void* shared, + uint Cout, uint Cin, uint N, + uint KW, uint KH, uint W, uint H, uint OW, uint OH, + uint s0, uint s1, uint p0, uint p1, uint d0, uint d1, + uint nb01, uint nb02, uint nb03, + uint nb11, uint nb12, uint nb13, + uint nb1, uint nb2, uint nb3 +) { + global T_FLOAT* knl_data = (global T_FLOAT*) ((global char*)p_knl + off_knl); + global T_FLOAT* src_data = (global T_FLOAT*) ((global char*)p_src + off_src); + global T_FLOAT* dst_data = (global T_FLOAT*) ((global char*)p_dst + off_dst); + + const uint K = Cout; + const uint CRS = Cin*KH*KW; + const uint NPQ = N*OH*OW; + + const uint lid_k = get_local_id(0); + const uint lid_npq = get_local_id(1); + const uint tid = lid_npq * WG_K + lid_k; + + const uint B_idx_K = get_group_id(0); + const uint B_idx_NPQ = get_group_id(1); + + const uint offset_k = B_idx_K * BS_K; + const uint offset_npq = B_idx_NPQ * BS_NPQ; + + local T_FLOAT* Ash = (local T_FLOAT*)shared; + local T_FLOAT4* Bsh = (local T_FLOAT4*) &Ash[BS_K * BS_CRS]; + + T_ACCUM regC[TS_K][TS_NPQ_VEC]; + for (int i = 0; i < TS_K; ++i) { + for (int j = 0; j < TS_NPQ_VEC; ++j) { + regC[i][j] = (T_ACCUM)(0.0f); + } + } + + const uint NB_CRS = splitWork(CRS, BS_CRS); + + for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) { + const uint offset_crs = B_idx_CRS * BS_CRS; + + for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) { + const uint k_l = i / BS_CRS; + const uint crs_l = i % BS_CRS; + const uint k_g = offset_k + k_l; + const uint crs_g = offset_crs + crs_l; + + if (k_g < K && crs_g < CRS) { + const uint Cin_idx = crs_g / (KW*KH); + const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW; + const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW; + const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03; + Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx]; + } else { + Ash[k_l * BS_CRS + crs_l] = (T_FLOAT)0.0f; + } + } + + for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) { + const uint crs_l = i / BS_NPQ_VEC; + const uint npq_l_vec = i % BS_NPQ_VEC; + const uint crs_g = offset_crs + crs_l; + + T_FLOAT4 val = (T_FLOAT4)(0.0f); + if (crs_g < CRS) { + const uint Cin_idx = crs_g / (KW * KH); + const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW; + const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW; + for (int v = 0; v < VEC_SIZE; ++v) { + const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v; + if (npq_g < NPQ) { + const uint N_idx = npq_g / (OH * OW); + const uint pq_idx = npq_g % (OH * OW); + const uint OH_idx = pq_idx / OW; + const uint OW_idx = pq_idx % OW; + const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1); + const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0); + + if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) { + const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13; + ((T_FLOAT*)&val)[v] = src_data[src_idx]; + } + } + } + } + Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val; + } + + barrier(CLK_LOCAL_MEM_FENCE); + + #pragma unroll + for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) { + T_FLOAT regA[TS_K]; + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l]; + } + + for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) { + T_FLOAT4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg]; + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), convert_float4(regB), regC[k_l_reg][npq_l_vec_reg]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + const uint k_g = offset_k + lid_k * TS_K + k_l_reg; + if (k_g >= K) continue; + + for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) { + const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE; + + const uint N_idx = npq_g_base / (OH * OW); + const uint pq_idx = npq_g_base % (OH * OW); + const uint OH_idx = pq_idx / OW; + const uint OW_idx = pq_idx % OW; + + if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) { + const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3; + VSTORE_T_FLOAT4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]); + } else { + T_ACCUM res = regC[k_l_reg][npq_l_vec_reg]; + for (int v = 0; v < VEC_SIZE; ++v) { + const uint npq_g = npq_g_base + v; + if (npq_g < NPQ) { + const uint N_idx_s = npq_g / (OH*OW); + const uint pq_idx_s = npq_g % (OH*OW); + const uint OH_idx_s = pq_idx_s / OW; + const uint OW_idx_s = pq_idx_s % OW; + const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3; + dst_data[dst_idx_s] = (T_FLOAT)(((float*)&res)[v]); + } + } + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl new file mode 100644 index 0000000..cb05637 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl @@ -0,0 +1,176 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#if defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#else +#define REQD_SUBGROUP_SIZE_128 +#endif + +#define T_ACCUM float4 +#define VEC_SIZE 4 + +#define BS_K 64 +#define BS_NPQ 64 +#define BS_CRS 16 + +#define TS_K 4 +#define TS_NPQ 8 + +#define WG_K (BS_K / TS_K) +#define WG_NPQ (BS_NPQ / TS_NPQ) + +#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE) +#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE) + +static inline uint splitWork(uint work_size, uint block_size){ + return (work_size + block_size - 1) / block_size; +} + +REQD_SUBGROUP_SIZE_128 +kernel void kernel_conv_2d( + global void* p_knl, + ulong off_knl, + global void* p_src, + ulong off_src, + global void* p_dst, + ulong off_dst, + local void* shared, + uint Cout, uint Cin, uint N, + uint KW, uint KH, uint W, uint H, uint OW, uint OH, + uint s0, uint s1, uint p0, uint p1, uint d0, uint d1, + uint nb01, uint nb02, uint nb03, + uint nb11, uint nb12, uint nb13, + uint nb1, uint nb2, uint nb3 +) { + global half* knl_data = (global half*) ((global char*)p_knl + off_knl); + global float* src_data = (global float*) ((global char*)p_src + off_src); + global float* dst_data = (global float*) ((global char*)p_dst + off_dst); + + const uint K = Cout; + const uint CRS = Cin*KH*KW; + const uint NPQ = N*OH*OW; + + const uint lid_k = get_local_id(0); + const uint lid_npq = get_local_id(1); + const uint tid = lid_npq * WG_K + lid_k; + + const uint B_idx_K = get_group_id(0); + const uint B_idx_NPQ = get_group_id(1); + + const uint offset_k = B_idx_K * BS_K; + const uint offset_npq = B_idx_NPQ * BS_NPQ; + + local half* Ash = (local half*)shared; + local float4* Bsh = (local float4*) &Ash[BS_K * BS_CRS]; + + T_ACCUM regC[TS_K][TS_NPQ_VEC]; + for (int i = 0; i < TS_K; ++i) { + for (int j = 0; j < TS_NPQ_VEC; ++j) { + regC[i][j] = (T_ACCUM)(0.0f); + } + } + + const uint NB_CRS = splitWork(CRS, BS_CRS); + + for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) { + const uint offset_crs = B_idx_CRS * BS_CRS; + + for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) { + const uint k_l = i / BS_CRS; + const uint crs_l = i % BS_CRS; + const uint k_g = offset_k + k_l; + const uint crs_g = offset_crs + crs_l; + + if (k_g < K && crs_g < CRS) { + const uint Cin_idx = crs_g / (KW*KH); + const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW; + const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW; + const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03; + Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx]; + } else { + Ash[k_l * BS_CRS + crs_l] = (half)0.0f; + } + } + + for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) { + const uint crs_l = i / BS_NPQ_VEC; + const uint npq_l_vec = i % BS_NPQ_VEC; + const uint crs_g = offset_crs + crs_l; + + float4 val = (float4)(0.0f); + if (crs_g < CRS) { + const uint Cin_idx = crs_g / (KW * KH); + const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW; + const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW; + for (int v = 0; v < VEC_SIZE; ++v) { + const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v; + if (npq_g < NPQ) { + const uint N_idx = npq_g / (OH * OW); + const uint pq_idx = npq_g % (OH * OW); + const uint OH_idx = pq_idx / OW; + const uint OW_idx = pq_idx % OW; + const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1); + const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0); + + if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) { + const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13; + ((float*)&val)[v] = src_data[src_idx]; + } + } + } + } + Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val; + } + + barrier(CLK_LOCAL_MEM_FENCE); + + #pragma unroll + for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) { + half regA[TS_K]; + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l]; + } + + for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) { + float4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg]; + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), regB, regC[k_l_reg][npq_l_vec_reg]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) { + const uint k_g = offset_k + lid_k * TS_K + k_l_reg; + if (k_g >= K) continue; + + for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) { + const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE; + + const uint N_idx = npq_g_base / (OH * OW); + const uint pq_idx = npq_g_base % (OH * OW); + const uint OH_idx = pq_idx / OW; + const uint OW_idx = pq_idx % OW; + + if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) { + const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3; + vstore4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]); + } else { + T_ACCUM res = regC[k_l_reg][npq_l_vec_reg]; + for (int v = 0; v < VEC_SIZE; ++v) { + const uint npq_g = npq_g_base + v; + if (npq_g < NPQ) { + const uint N_idx_s = npq_g / (OH*OW); + const uint pq_idx_s = npq_g % (OH*OW); + const uint OH_idx_s = pq_idx_s / OW; + const uint OW_idx_s = pq_idx_s % OW; + const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3; + dst_data[dst_idx_s] = ((float*)&res)[v]; + } + } + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl new file mode 100644 index 0000000..9369351 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl @@ -0,0 +1,184 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// cpy +//------------------------------------------------------------------------------ + +kernel void kernel_cpy_f16_f16( + global half * src0, + ulong offset0, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global half*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + int i3 = n / (ne2*ne1*ne0); + int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0); + + global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + global const half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + dst_data[i00] = src[0]; + } +} + +kernel void kernel_cpy_f16_f32( + global half * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + + src0 = (global half*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + int i3 = n / (ne2*ne1*ne0); + int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0); + + global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + global half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + dst_data[i00] = src[0]; + } +} + +kernel void kernel_cpy_f32_f16( + global float * src0, + ulong offset0, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + int i3 = n / (ne2*ne1*ne0); + int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0); + + global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + + dst_data[i00] = src[0]; + } +} + +kernel void kernel_cpy_f32_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + int i3 = n / (ne2*ne1*ne0); + int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0); + int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0; + int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0); + + global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); + + dst_data[i00] = src[0]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl new file mode 100644 index 0000000..9fb4347 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl @@ -0,0 +1,366 @@ +//------------------------------------------------------------------------------ +// This file is contains kernels for data conversion. +// These kernels are used when loading the model, so its performance is less +// important. +//------------------------------------------------------------------------------ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +//------------------------------------------------------------------------------ +// block_q6_K +//------------------------------------------------------------------------------ +struct block_q6_K { + uint8_t ql[QK_K/2]; // quants, lower 4 bits + uint8_t qh[QK_K/4]; // quants, upper 2 bits + int8_t scales[QK_K/16]; // scales, quantized with 8 bits + half d; // super-block scale +}; + +//------------------------------------------------------------------------------ +// kernel_convert_block_q4_0 +// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA). +// This kernel does not deshuffle the bits. +//------------------------------------------------------------------------------ +kernel void kernel_convert_block_q4_0( + global struct block_q4_0 * src0, + global uchar * dst_q, + global half * dst_d +) { + global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0); + global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + *d = b->d; + + for (int i = 0; i < QK4_0/2; ++i) { + q[i] = b->qs[i]; + } +} + +kernel void kernel_restore_block_q4_0( + global uchar * src_q, + global half * src_d, + global struct block_q4_0 * dst +) { + global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0); + global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0); + global half * d = (global half *) src_d + get_global_id(0); + + b->d = *d; + for (int i = 0; i < QK4_0/2; ++i) { + b->qs[i] = q[i]; + } +} + +//------------------------------------------------------------------------------ +// kernel_convert_block_q4_0_noshuffle +// Flatten q4_0 weights and unshuffle the bits +//------------------------------------------------------------------------------ + +kernel void kernel_convert_block_q4_0_noshuffle( + global struct block_q4_0 * src0, + global uchar * dst_q, + global half * dst_d +) { + global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0); + global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + *d = b->d; + for (int i = 0; i < QK4_0/4; ++i) { + uchar x0 = b->qs[2*i + 0]; + uchar x1 = b->qs[2*i + 1]; + + q[i + 0 ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4); + q[i + QK4_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0); + +#ifdef ADRENO_GPU + // Workaround for adreno - must have the following printf statement for + // the kernel to work properly. Otherwise it produces incorrect result. + // convert_uchar above also seems necessary. + // Compare against a large number so that it does not print anything. + // get_sub_group_local_id() also works. + if (get_global_id(0) == 65536*4096) { + printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0)); + } +#endif + } +} + +kernel void kernel_restore_block_q4_0_noshuffle( + global uchar * src_q, + global half * src_d, + global struct block_q4_0 * dst, + uchar mask_0F, + uchar mask_F0 +) { + global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0); + global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0); + global half * d = (global half *) src_d + get_global_id(0); + + b->d = *d; + for (int i = 0; i < QK4_0/4; ++i) { + uchar x0 = q[i + 0 ] ; + uchar x1 = q[i + QK4_0/4]; + + b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4)); + b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0)); + } +} + +//------------------------------------------------------------------------------ +// block_mxfp4 +//------------------------------------------------------------------------------ +#define QK_MXFP4 32 +struct block_mxfp4 { + uchar e; // E8M0 + uchar qs[QK_MXFP4 / 2]; +}; + +//------------------------------------------------------------------------------ +// kernel_convert_block_mxfp4 +// Convert the block_mxfp4 format to 2 separate arrays (AOS -> SOA). +// This kernel does not deshuffle the bits. +//------------------------------------------------------------------------------ +kernel void kernel_convert_block_mxfp4( + global struct block_mxfp4 * src0, + global uchar * dst_q, + global uchar * dst_e +) { + global struct block_mxfp4 * b = (global struct block_mxfp4 *) src0 + get_global_id(0); + global uchar * q = (global uchar *) dst_q + QK_MXFP4 / 2 * get_global_id(0); + global uchar * e = (global uchar *) dst_e + get_global_id(0); + + *e = b->e; + + for (int i = 0; i < QK_MXFP4 / 2; ++i) { + q[i] = b->qs[i]; + } +} + +kernel void kernel_convert_block_mxfp4_trans( + global struct block_mxfp4 * src0, + __global uint4 * dst_q, + __global uchar * dst_e, + uint ne00, + uint ne01 +) { + int i00 = get_global_id(1); + uint i01 = get_global_id(0); + uint i02 = get_global_id(2); + + uint ne00_blk = ne00 / QK_MXFP4; + uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01; + uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01; + + global struct block_mxfp4 * b = src0 + src_blk_offset; + + dst_q[dst_blk_offset] = ((global uint4 *)(&(b->qs[0])))[0]; + dst_e[dst_blk_offset] = b->e; +} + +kernel void kernel_restore_block_mxfp4( + global uchar * src_q, + global half * src_e, + global struct block_mxfp4 * dst +) { + global struct block_mxfp4 * b = (global struct block_mxfp4 *) dst + get_global_id(0); + global uchar * q = (global uchar *) src_q + QK_MXFP4 / 2 * get_global_id(0); + global uchar * e = (global uchar *) src_e + get_global_id(0); + + b->e = *e; + for (int i = 0; i < QK_MXFP4 / 2; ++i) { + b->qs[i] = q[i]; + } +} + +kernel void kernel_restore_block_mxfp4_trans( + __global uint4 * src_q, + __global uchar * src_e, + global struct block_mxfp4 * dst, + uint ne00, + uint ne01 +) { + int i00 = get_global_id(1); + uint i01 = get_global_id(0); + uint i02 = get_global_id(2); + + uint ne00_blk = ne00 / QK_MXFP4; + uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01; + uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01; + + global struct block_mxfp4 * b = dst + dst_blk_offset; + + ((global uint4 *)(&(b->qs[0])))[0] = src_q[src_blk_offset]; + b->e = src_e[src_blk_offset]; +} + +//------------------------------------------------------------------------------ +// block_q8_0 +//------------------------------------------------------------------------------ +typedef struct { + half d; // delta + char qs[QK8_0]; // quants +} block_q8_0; + +kernel void kernel_convert_block_q8_0( + global block_q8_0 * src0, + global uchar * dst_q, + global half * dst_d +) { + global block_q8_0 * b = (global block_q8_0 *) src0 + get_global_id(0); + global uchar * q = (global uchar *) dst_q + QK8_0*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + *d = b->d; + + for (int i = 0; i < QK8_0; ++i) { + q[i] = b->qs[i]; + } +} + +kernel void kernel_restore_block_q8_0( + global uchar * src_q, + global half * src_d, + global block_q8_0 * dst +) { + global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0); + global uchar * q = (global uchar *) src_q + QK8_0*get_global_id(0); + global half * d = (global half *) src_d + get_global_id(0); + + b->d = *d; + for (int i = 0; i < QK8_0; ++i) { + b->qs[i] = q[i]; + } +} + +kernel void kernel_restore_block_q8_0_trans( + global uchar * src_q, + global half * src_d, + global block_q8_0 * dst, + uint ne00, + uint ne01 +){ + uint num_blk_per_row = ne00 / QK8_0; + + global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0) * num_blk_per_row; + global uchar * q = (global uchar *) src_q + get_global_id(0) * 4; // 4 8-bit packed + global half * d = (global half *) src_d + get_global_id(0); + + for (uint blk = 0; blk < num_blk_per_row; blk++) { + b->d = *d; + + for (uint i = 0; i < QK8_0; i+=4) { + b->qs[i] = q[0]; + b->qs[i+1] = q[1]; + b->qs[i+2] = q[2]; + b->qs[i+3] = q[3]; + + q += 4 * ne01; // M stride + } + + d += ne01; + + b++; + } +} + +//------------------------------------------------------------------------------ +// kernel_convert_block_q6_K +// Convert the block_q6_K format to 3 separate arrays (AOS -> SOA). +// This kernel does not deshuffle the bits. +// Each thread processes a super block. +//------------------------------------------------------------------------------ +kernel void kernel_convert_block_q6_K( + global struct block_q6_K * src0, + global uchar * dst_ql, + global uchar * dst_qh, + global char * dst_s, + global half * dst_d +) { + global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0); + global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0); + global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0); + global char * s = (global char *) dst_s + QK_K/16*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + *d = b->d; + + for (int i = 0; i < QK_K/2; ++i) { + ql[i] = b->ql[i]; + } + for (int i = 0; i < QK_K/4; ++i) { + qh[i] = b->qh[i]; + } + for (int i = 0; i < QK_K/16; ++i) { + s[i] = b->scales[i]; + } +} + +// Restore block_q6_K from flattened arrays. +// Each thread processes a super block. +kernel void kernel_restore_block_q6_K( + global uchar * dst_ql, + global uchar * dst_qh, + global char * dst_s, + global half * dst_d, + global struct block_q6_K * dst +) { + global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0); + global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0); + global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0); + global char * s = (global char *) dst_s + QK_K/16*get_global_id(0); + global half * d = (global half *) dst_d + get_global_id(0); + + b->d = *d; + + for (int i = 0; i < QK_K/2; ++i) { + b->ql[i] = ql[i]; + } + for (int i = 0; i < QK_K/4; ++i) { + b->qh[i] = qh[i]; + } + for (int i = 0; i < QK_K/16; ++i) { + b->scales[i] = s[i]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl new file mode 100644 index 0000000..36eff04 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl @@ -0,0 +1,58 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// diag_mask_inf kernels +//------------------------------------------------------------------------------ +kernel void kernel_diag_mask_inf( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int n_past +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int i02 = get_global_id(2); + int i01 = get_global_id(1); + int i00 = get_global_id(0); + + if (i00 > n_past + i01) { + dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY; + } else { + dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00]; + } +} + +kernel void kernel_diag_mask_inf_8( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd, + int ne00, + int ne01, + int n_past +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + int i = 2*get_global_id(0); + + dst[i+0] = src0[i+0]; + dst[i+1] = src0[i+1]; + int i4 = 4*i; + int i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01; + int i01 = i4/(ne00); i4 -= i01*ne00; + int i00 = i4; + for (int k = 3; k >= 0; --k) { + if (i00 + 4 + k <= n_past + i01) { + break; + } + (&dst[i+1])[k] = -INFINITY; + if (i00 + k > n_past + i01) { + (&dst[i])[k] = -INFINITY; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl new file mode 100644 index 0000000..6d9b4ad --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl @@ -0,0 +1,138 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// div +//------------------------------------------------------------------------------ +kernel void kernel_div( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) / *((global float *)(src1_ptr + i10*nb10)); + } +} + +// assumption: src1 is a row +// broadcast src1 into src0 +kernel void kernel_div_row( + global float4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float4 * dst, + ulong offsetd, + int ne +) { + src0 = (global float4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] / src1[idx1]; +} + +kernel void kernel_div_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) / *((global half *)(src1_ptr + i10*nb10)); + } +} + +kernel void kernel_div_row_f16( + global half4 * src0, + ulong offset0, + global half4 * src1, + ulong offset1, + global half4 * dst, + ulong offsetd, + int ne +) { + src0 = (global half4*)((global char*)src0 + offset0); + src1 = (global half4*)((global char*)src1 + offset1); + dst = (global half4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] / src1[idx1]; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py b/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py new file mode 100644 index 0000000..b5d1d72 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py @@ -0,0 +1,26 @@ +# + +import sys +import logging +logger = logging.getLogger("opencl-embed-kernel") + + +def main(): + logging.basicConfig(level=logging.INFO) + + if len(sys.argv) != 3: + logger.info("Usage: python embed_kernel.py <input_file> <output_file>") + sys.exit(1) + + ifile = open(sys.argv[1], "r") + ofile = open(sys.argv[2], "w") + + for i in ifile: + ofile.write('R"({})"\n'.format(i)) + + ifile.close() + ofile.close() + + +if __name__ == "__main__": + main() diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl new file mode 100644 index 0000000..126298a --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl @@ -0,0 +1,82 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// expm1 +//------------------------------------------------------------------------------ +kernel void kernel_expm1_f32_nd( + global void * p_src0_base, + ulong off_src0_abs, + global void * p_dst_base, + ulong off_dst_abs, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13 +) { + int i0 = get_global_id(0); + int i1 = get_global_id(1); + int i2 = get_global_id(2); + + if (i0 < ne10 && i1 < ne11 && i2 < ne12) { + for (int i3 = 0; i3 < ne13; ++i3) { + ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03; + global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor); + + ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13; + global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor); + + *dst_val_ptr = exp(*src_val_ptr) - 1; + } + } +} + +kernel void kernel_expm1_f16_nd( + global void * p_src0_base, + ulong off_src0_abs, + global void * p_dst_base, + ulong off_dst_abs, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13 +) { + int i0 = get_global_id(0); + int i1 = get_global_id(1); + int i2 = get_global_id(2); + + if (i0 < ne10 && i1 < ne11 && i2 < ne12) { + for (int i3 = 0; i3 < ne13; ++i3) { + ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03; + global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor); + + ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13; + global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor); + + *dst_val_ptr = exp(*src_val_ptr) - 1; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl new file mode 100644 index 0000000..9b73938 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl @@ -0,0 +1,17 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// fill +//------------------------------------------------------------------------------ +__kernel void kernel_fill_f32( + __global float *dst, + ulong offsetd, + float v, + int n + +) { + dst = (global float*)((global char*)dst + offsetd); + if(get_global_id(0) < n){ + dst[get_global_id(0)] = v; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl new file mode 100644 index 0000000..8f43c4f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl @@ -0,0 +1,370 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define ACC_TYPE float +#define ACC_TYPE4 float4 +#define DATA_TYPE half +#define DATA_TYPE4 half4 +#define CONVERT_ACC4(x) convert_float4(x) +#define CONVERT_DATA4(x) convert_half4(x) + +#define DK_VEC (DK/4) +#define DV_VEC (DV/4) +#define WG_SIZE (BLOCK_M) +#define Q1_WG_SIZE 64 + +inline float get_alibi_slope( + const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1 +) { + if (max_bias <= 0.0f) { + return 1.0f; + } + const float base = h < n_head_log2 ? m0 : m1; + const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + return pow(base, exph); +} +__kernel void flash_attn_f16( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int block_q_idx = get_group_id(0); + const int head_batch_idx = get_global_id(1); + + const int my_query_row = block_q_idx * BLOCK_M + tid; + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + if (my_query_row < n_q) { + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1; + const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_ACC4(q_ptr[i]); + } + } + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = (ACC_TYPE4)(0.0f); + } + ACC_TYPE m_i = -INFINITY; + ACC_TYPE l_i = 0.0f; + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC]; + __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC]; + + for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) { + for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) { + const int row = i / DK_VEC; + const int col = i % DK_VEC; + const int k_row_idx = k_start + row; + if (k_row_idx < n_kv) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1; + l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col]; + } + } + for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) { + const int row = i / DV_VEC; + const int col = i % DV_VEC; + const int v_row_idx = k_start + row; + if (v_row_idx < n_kv) { + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1; + l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col]; + } + } + barrier(CLK_LOCAL_MEM_FENCE); + + if (my_query_row >= n_q) { + continue; + } + + for (int j = 0; j < BLOCK_N; j += 2) { + const int k_row0 = k_start + j; + const int k_row1 = k_start + j + 1; + + ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f); + ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0); + dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1); + } + ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale; + ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale; + + if (is_causal) { + if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY; + if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY; + } + + if (k_row0 >= n_kv) score0 = -INFINITY; + if (k_row1 >= n_kv) score1 = -INFINITY; + + if (mask_base != NULL) { + const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1); + if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0]; + if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1]; + } + + if (logit_softcap > 0.0f) { + score0 = logit_softcap * tanh(score0 / logit_softcap); + score1 = logit_softcap * tanh(score1 / logit_softcap); + } + + const ACC_TYPE m_new = max(m_i, max(score0, score1)); + const ACC_TYPE p0 = exp(score0 - m_new); + const ACC_TYPE p1 = exp(score1 - m_new); + const ACC_TYPE scale_prev = exp(m_i - m_new); + + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]); + } + l_i = l_i * scale_prev + p0 + p1; + m_i = m_new; + } + } + + if (my_query_row < n_q) { + if (sinks_void != NULL) { + const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + const ACC_TYPE m_sink = sinks_ptr[head_idx]; + const ACC_TYPE m_final = max(m_i, m_sink); + + const ACC_TYPE scale_o = exp(m_i - m_final); + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] *= scale_o; + } + + l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final); + } + + const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1; + global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset); + if (l_i > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_i; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv); + } + } else { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = (DATA_TYPE4)(0.0f); + } + } + } +} + +__kernel void flash_attn_f16_q1( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int head_batch_idx = get_global_id(1); + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2; + const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_ACC4(q_ptr[i]); + } + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + const global ACC_TYPE* sinks_ptr = NULL; + if (sinks_void != NULL) { + sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + } + + ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY; + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + m_i = max(m_i, score); + } + + __local ACC_TYPE local_m[Q1_WG_SIZE]; + local_m[tid] = m_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]); + barrier(CLK_LOCAL_MEM_FENCE); + } + const ACC_TYPE m_final = local_m[0]; + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f); + ACC_TYPE l_i = 0.0f; + + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1; + const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset); + const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + const ACC_TYPE p = exp(score - m_final); + l_i += p; + #pragma unroll + for (int i = 0; i < DV_VEC; i++) { + o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]); + } + } + + __local ACC_TYPE local_l[Q1_WG_SIZE]; + __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE]; + local_l[tid] = l_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_l[tid] += local_l[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + + const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1; + global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset); + ACC_TYPE l_final = local_l[0]; + + if (sinks_ptr != NULL) { + l_final += exp(sinks_ptr[head_idx] - m_final); + } + + if (l_final > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_final; + for (int i = 0; i < DV_VEC; i++) { + local_o_comp[tid] = o_acc[i]; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_o_comp[tid] += local_o_comp[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv); + } + } + } else if (tid == 0) { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl new file mode 100644 index 0000000..a6d7479 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl @@ -0,0 +1,371 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define ACC_TYPE float +#define ACC_TYPE4 float4 +#define DATA_TYPE float +#define DATA_TYPE4 float4 +#define MASK_DATA_TYPE half +#define CONVERT_ACC4(x) (x) +#define CONVERT_DATA4(x) (x) + +#define DK_VEC (DK/4) +#define DV_VEC (DV/4) +#define WG_SIZE (BLOCK_M) +#define Q1_WG_SIZE 64 + +inline float get_alibi_slope( + const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1 +) { + if (max_bias <= 0.0f) { + return 1.0f; + } + const float base = h < n_head_log2 ? m0 : m1; + const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + return pow(base, exph); +} +__kernel void flash_attn_f32( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int block_q_idx = get_group_id(0); + const int head_batch_idx = get_global_id(1); + + const int my_query_row = block_q_idx * BLOCK_M + tid; + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + if (my_query_row < n_q) { + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1; + const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_ACC4(q_ptr[i]); + } + } + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = (ACC_TYPE4)(0.0f); + } + ACC_TYPE m_i = -INFINITY; + ACC_TYPE l_i = 0.0f; + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC]; + __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC]; + + for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) { + for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) { + const int row = i / DK_VEC; + const int col = i % DK_VEC; + const int k_row_idx = k_start + row; + if (k_row_idx < n_kv) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1; + l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col]; + } + } + for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) { + const int row = i / DV_VEC; + const int col = i % DV_VEC; + const int v_row_idx = k_start + row; + if (v_row_idx < n_kv) { + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1; + l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col]; + } + } + barrier(CLK_LOCAL_MEM_FENCE); + + if (my_query_row >= n_q) { + continue; + } + + for (int j = 0; j < BLOCK_N; j += 2) { + const int k_row0 = k_start + j; + const int k_row1 = k_start + j + 1; + + ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f); + ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0); + dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1); + } + ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale; + ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale; + + if (is_causal) { + if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY; + if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY; + } + + if (k_row0 >= n_kv) score0 = -INFINITY; + if (k_row1 >= n_kv) score1 = -INFINITY; + + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1); + if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0]; + if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1]; + } + + if (logit_softcap > 0.0f) { + score0 = logit_softcap * tanh(score0 / logit_softcap); + score1 = logit_softcap * tanh(score1 / logit_softcap); + } + + const ACC_TYPE m_new = max(m_i, max(score0, score1)); + const ACC_TYPE p0 = exp(score0 - m_new); + const ACC_TYPE p1 = exp(score1 - m_new); + const ACC_TYPE scale_prev = exp(m_i - m_new); + + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]); + } + l_i = l_i * scale_prev + p0 + p1; + m_i = m_new; + } + } + + if (my_query_row < n_q) { + if (sinks_void != NULL) { + const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + const ACC_TYPE m_sink = sinks_ptr[head_idx]; + const ACC_TYPE m_final = max(m_i, m_sink); + + const ACC_TYPE scale_o = exp(m_i - m_final); + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] *= scale_o; + } + + l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final); + } + + const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1; + global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset); + if (l_i > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_i; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv); + } + } else { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = (DATA_TYPE4)(0.0f); + } + } + } +} + +__kernel void flash_attn_f32_q1( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int head_batch_idx = get_global_id(1); + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2; + const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_ACC4(q_ptr[i]); + } + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + const global ACC_TYPE* sinks_ptr = NULL; + if (sinks_void != NULL) { + sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + } + + ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY; + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + m_i = max(m_i, score); + } + + __local ACC_TYPE local_m[Q1_WG_SIZE]; + local_m[tid] = m_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]); + barrier(CLK_LOCAL_MEM_FENCE); + } + const ACC_TYPE m_final = local_m[0]; + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f); + ACC_TYPE l_i = 0.0f; + + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1; + const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset); + const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + const ACC_TYPE p = exp(score - m_final); + l_i += p; + #pragma unroll + for (int i = 0; i < DV_VEC; i++) { + o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]); + } + } + + __local ACC_TYPE local_l[Q1_WG_SIZE]; + __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE]; + local_l[tid] = l_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_l[tid] += local_l[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + + const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1; + global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset); + ACC_TYPE l_final = local_l[0]; + + if (sinks_ptr != NULL) { + l_final += exp(sinks_ptr[head_idx] - m_final); + } + + if (l_final > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_final; + for (int i = 0; i < DV_VEC; i++) { + local_o_comp[tid] = o_acc[i]; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_o_comp[tid] += local_o_comp[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv); + } + } + } else if (tid == 0) { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl new file mode 100644 index 0000000..ec7361b --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl @@ -0,0 +1,373 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define ACC_TYPE float +#define ACC_TYPE4 float4 +#define Q_DATA_TYPE4 float4 +#define KV_DATA_TYPE4 half4 +#define O_DATA_TYPE4 float4 +#define MASK_DATA_TYPE half +#define CONVERT_Q_ACC4(x) (x) +#define CONVERT_KV_ACC4(x) convert_float4(x) +#define CONVERT_O_DATA4(x) (x) + +#define DK_VEC (DK/4) +#define DV_VEC (DV/4) +#define WG_SIZE (BLOCK_M) +#define Q1_WG_SIZE 64 + +inline float get_alibi_slope( + const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1 +) { + if (max_bias <= 0.0f) { + return 1.0f; + } + const float base = h < n_head_log2 ? m0 : m1; + const int exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + return pow(base, exph); +} +__kernel void flash_attn_f32_f16( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int block_q_idx = get_group_id(0); + const int head_batch_idx = get_global_id(1); + + const int my_query_row = block_q_idx * BLOCK_M + tid; + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + if (my_query_row < n_q) { + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1; + const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]); + } + } + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = (ACC_TYPE4)(0.0f); + } + ACC_TYPE m_i = -INFINITY; + ACC_TYPE l_i = 0.0f; + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC]; + __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC]; + + for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) { + for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) { + const int row = i / DK_VEC; + const int col = i % DK_VEC; + const int k_row_idx = k_start + row; + if (k_row_idx < n_kv) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1; + l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col]; + } + } + for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) { + const int row = i / DV_VEC; + const int col = i % DV_VEC; + const int v_row_idx = k_start + row; + if (v_row_idx < n_kv) { + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1; + l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col]; + } + } + barrier(CLK_LOCAL_MEM_FENCE); + + if (my_query_row >= n_q) { + continue; + } + + for (int j = 0; j < BLOCK_N; j += 2) { + const int k_row0 = k_start + j; + const int k_row1 = k_start + j + 1; + + ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f); + ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0); + dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1); + } + ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale; + ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale; + + if (is_causal) { + if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY; + if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY; + } + + if (k_row0 >= n_kv) score0 = -INFINITY; + if (k_row1 >= n_kv) score1 = -INFINITY; + + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1); + if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0]; + if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1]; + } + + if (logit_softcap > 0.0f) { + score0 = logit_softcap * tanh(score0 / logit_softcap); + score1 = logit_softcap * tanh(score1 / logit_softcap); + } + + const ACC_TYPE m_new = max(m_i, max(score0, score1)); + const ACC_TYPE p0 = exp(score0 - m_new); + const ACC_TYPE p1 = exp(score1 - m_new); + const ACC_TYPE scale_prev = exp(m_i - m_new); + + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]); + } + l_i = l_i * scale_prev + p0 + p1; + m_i = m_new; + } + } + + if (my_query_row < n_q) { + if (sinks_void != NULL) { + const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + const ACC_TYPE m_sink = sinks_ptr[head_idx]; + const ACC_TYPE m_final = max(m_i, m_sink); + + const ACC_TYPE scale_o = exp(m_i - m_final); + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_acc[i] *= scale_o; + } + + l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final); + } + + const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1; + global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset); + if (l_i > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_i; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv); + } + } else { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) { + o_row[i] = (O_DATA_TYPE4)(0.0f); + } + } + } +} + +__kernel void flash_attn_f32_f16_q1( + const global void * q_void, ulong q_offset, + const global void * k_void, ulong k_offset, + const global void * v_void, ulong v_offset, + global void * o_void, ulong o_offset, + const float scale, + const int n_q, + const int n_kv, + const int is_causal, + const int n_head, + const ulong q_nb1, const ulong q_nb2, const ulong q_nb3, + const ulong k_nb1, const ulong k_nb2, const ulong k_nb3, + const ulong v_nb1, const ulong v_nb2, const ulong v_nb3, + const ulong o_nb1, const ulong o_nb2, const ulong o_nb3, + const float max_bias, + const float m0, + const float m1, + const int n_head_log2, + const float logit_softcap, + const int n_head_kv, + const global void* mask_void, + const ulong mask_offset, + const ulong mask_nb1, + const ulong mask_nb2, + const ulong mask_nb3, + const int mask_ne2, + const int mask_ne3, + const global void* sinks_void, + const ulong sinks_offset +) { + const int tid = get_local_id(0); + const int head_batch_idx = get_global_id(1); + + const int batch_idx = head_batch_idx / n_head; + const int head_idx = head_batch_idx % n_head; + + const int gqa_ratio = n_head / n_head_kv; + const int head_kv_idx = head_idx / gqa_ratio; + + const global char* q_base = (const global char*)q_void + q_offset; + const global char* k_base = (const global char*)k_void + k_offset; + const global char* v_base = (const global char*)v_void + v_offset; + global char* o_base = (global char*)o_void + o_offset; + + const global char* mask_base = NULL; + if (mask_void != NULL) { + const int mask_head_idx = head_idx % mask_ne2; + const int mask_batch_idx = batch_idx % mask_ne3; + mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2; + } + + ACC_TYPE4 q_priv[DK_VEC]; + const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2; + const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset); + #pragma unroll + for (int i = 0; i < DK_VEC; ++i) { + q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]); + } + + float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1); + + const global ACC_TYPE* sinks_ptr = NULL; + if (sinks_void != NULL) { + sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset); + } + + ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY; + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + m_i = max(m_i, score); + } + + __local ACC_TYPE local_m[Q1_WG_SIZE]; + local_m[tid] = m_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]); + barrier(CLK_LOCAL_MEM_FENCE); + } + const ACC_TYPE m_final = local_m[0]; + + ACC_TYPE4 o_acc[DV_VEC]; + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f); + ACC_TYPE l_i = 0.0f; + + for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) { + const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1; + const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1; + const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset); + const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset); + ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f); + #pragma unroll + for (int k = 0; k < DK_VEC; k++) { + dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc); + } + ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale; + if (mask_base != NULL) { + const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base); + score += slope * (ACC_TYPE)mask_ptr[k_idx]; + } + if (logit_softcap > 0.0f) { + score = logit_softcap * tanh(score / logit_softcap); + } + const ACC_TYPE p = exp(score - m_final); + l_i += p; + #pragma unroll + for (int i = 0; i < DV_VEC; i++) { + o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]); + } + } + + __local ACC_TYPE local_l[Q1_WG_SIZE]; + __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE]; + local_l[tid] = l_i; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_l[tid] += local_l[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + + const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1; + global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset); + ACC_TYPE l_final = local_l[0]; + + if (sinks_ptr != NULL) { + l_final += exp(sinks_ptr[head_idx] - m_final); + } + + if (l_final > 0.0f) { + const ACC_TYPE l_inv = 1.0f / l_final; + for (int i = 0; i < DV_VEC; i++) { + local_o_comp[tid] = o_acc[i]; + barrier(CLK_LOCAL_MEM_FENCE); + #pragma unroll + for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) local_o_comp[tid] += local_o_comp[tid + s]; + barrier(CLK_LOCAL_MEM_FENCE); + } + if (tid == 0) { + o_row[i] = CONVERT_O_DATA4(local_o_comp[0] * l_inv); + } + } + } else if (tid == 0) { + #pragma unroll + for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl new file mode 100644 index 0000000..1ab426c --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl @@ -0,0 +1,89 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// gelu +//------------------------------------------------------------------------------ +#define GELU_COEF_A 0.044715f +#define GELU_QUICK_COEF -1.702f +#define SQRT_2_OVER_PI 0.79788456080286535587989211986876f +#define SQRT_2_INV 0.70710678118654752440084436210484f + +kernel void kernel_gelu( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + float x = src0[get_global_id(0)]; + + dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); +} + +kernel void kernel_gelu_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + float4 x = src0[get_global_id(0)]; + + dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); +} + +kernel void kernel_gelu_erf( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + float x = src0[get_global_id(0)]; + dst[get_global_id(0)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV)); +} + +kernel void kernel_gelu_erf_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + float4 x = src0[get_global_id(0)]; + dst[get_global_id(0)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV)); +} + +kernel void kernel_gelu_quick( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + float x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x))); +} + +kernel void kernel_gelu_quick_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + float4 x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x))); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl new file mode 100644 index 0000000..3917aa3 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl @@ -0,0 +1,162 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable + +#define QK_MXFP4 32 +#define N_SIMDGROUP 2 +#define SIMDGROUP_WIDTH 64 + +static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) { + ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b; + fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00; + fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00; + fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00; + fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0; + bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0; + bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0; + bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0; + + fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0; + fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0; + fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0; + fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0; + + sign_a.lo = (fp4x8.s0 << 12) & 0x8000; + sign_a.hi = (fp4x8.s0 << 8) & 0x8000; + sign_b.lo = (fp4x8.s0 << 4) & 0x8000; + sign_b.hi = fp4x8.s0 & 0x8000; + + fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0; + fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0; + + ushort2 fp16_packed_a_1, fp16_packed_b_1; + fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00; + fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00; + fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00; + fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0; + bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0; + bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0; + bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0; + + fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0; + fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0; + fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0; + fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0; + + sign_a.lo = (fp4x8.s1 << 12) & 0x8000; + sign_a.hi = (fp4x8.s1 << 8) & 0x8000; + sign_b.lo = (fp4x8.s1 << 4) & 0x8000; + sign_b.hi = fp4x8.s1 & 0x8000; + + fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1; + fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1; + + return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1)); +} + +static inline float e8m0_to_fp32(uchar x) { + int bits; + bits = (x == 0) ? 0x00400000 : ((uint) x << 23); + return as_float(bits); +} + + +__attribute__((qcom_reqd_sub_group_size("half"))) +__kernel void kernel_gemm_moe_mxfp4_f32( + __global uint4 * src0_q, + __global uchar * src0_e, + __read_only image1d_buffer_t src1, + __global ushort4 * src2, + __global float * dst, + ulong offsetd, + int ne00, + int ne01, + int tile_size +) { + uint i01 = get_global_id(0); + uint i20 = get_global_id(2); + uint sgid = get_local_id(1); + uint slid = get_sub_group_local_id(); + + ushort4 router = src2[i20]; + ushort expert_id = router.x; + ushort i11 = router.y; + ushort i1 = router.z; + ushort tile_id = router.w; + + if (tile_id * tile_size + i01 >= ne01) { // handle edge case when ne01 is not multiple of tile_size + return; + } + + uint expert_offset = expert_id * ne00 * ne01 / 32; + uint tile_offset = expert_offset + tile_id * tile_size + i01; + + __private float sum = 0.0f; // each thread calculate partial sum of one output + + // loop along ne00 in block granularity, skip 4 blocks every iter + for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) { + // load one block of q + uint4 regQ = src0_q[tile_offset + ib00 * ne01]; + // convert 8 fp4 to fp16 + half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0)); + + uint offset = i11 * ne00 / 4 + ib00 * 8; + float4 shared_y4; + shared_y4 = read_imagef(src1, (offset + 0)); + float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 4)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1)); + + shared_y4 = read_imagef(src1, (offset + 1)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 5)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2)); + + shared_y4 = read_imagef(src1, (offset + 2)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 6)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3)); + + shared_y4 = read_imagef(src1, (offset + 3)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 7)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + uchar regE = src0_e[tile_offset + ib00 * ne01]; + sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3)); + } + + // reduction in local memory, assumes #subgroups=4 + __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)]; + if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum; + // if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum; + // if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum; + barrier(CLK_LOCAL_MEM_FENCE); + if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid]; + // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid]; + // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid]; + + // 1 outputs per thread in subgroup 0 + if (sgid == 0) { + dst = dst + (offsetd >> 2); + dst[i01 + tile_id * tile_size + i1 * ne01] = sum; + } + +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl new file mode 100644 index 0000000..b4b1e51 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl @@ -0,0 +1,156 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable + +#define QK_MXFP4 32 +#define N_SIMDGROUP 4 +#define SIMDGROUP_WIDTH 64 + +static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) { + ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b; + fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00; + fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00; + fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00; + fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0; + bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0; + bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0; + bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0; + + fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0; + fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0; + fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0; + fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0; + + sign_a.lo = (fp4x8.s0 << 12) & 0x8000; + sign_a.hi = (fp4x8.s0 << 8) & 0x8000; + sign_b.lo = (fp4x8.s0 << 4) & 0x8000; + sign_b.hi = fp4x8.s0 & 0x8000; + + fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0; + fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0; + + ushort2 fp16_packed_a_1, fp16_packed_b_1; + fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00; + fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00; + fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00; + fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0; + bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0; + bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0; + bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0; + + fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0; + fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0; + fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0; + fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0; + + sign_a.lo = (fp4x8.s1 << 12) & 0x8000; + sign_a.hi = (fp4x8.s1 << 8) & 0x8000; + sign_b.lo = (fp4x8.s1 << 4) & 0x8000; + sign_b.hi = fp4x8.s1 & 0x8000; + + fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1; + fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1; + + return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1)); +} + +static inline float e8m0_to_fp32(uchar x) { + int bits; + bits = (x == 0) ? 0x00400000 : ((uint) x << 23); + return as_float(bits); +} + + +__attribute__((qcom_reqd_sub_group_size("half"))) +__kernel void kernel_gemv_moe_mxfp4_f32( + __global uint4 * src0_q, + __global uchar * src0_e, + __read_only image1d_buffer_t src1, + __global uint * src2, + __global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne11 +) { + uint i01 = get_global_id(0); + uint i20 = get_global_id(2); + uint sgid = get_local_id(1); + uint slid = get_sub_group_local_id(); + + uint i11 = i20 % ne11; + + uint expert_id = src2[i20]; + uint expert_offset = expert_id * ne00 * ne01 / 32; + + __private float sum = 0.0f; // each thread calculate partial sum of one output + + // loop along ne00 in block granularity, skip 4 blocks every iter + for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) { + + // load one block of q + uint4 regQ = src0_q[expert_offset + ib00 * ne01 + i01]; + + uint offset = i11 * ne00 / 4 + ib00 * 8; + + half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0)); + + float4 shared_y4; + shared_y4 = read_imagef(src1, (offset + 0)); + float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 4)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1)); + + shared_y4 = read_imagef(src1, (offset + 1)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 5)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2)); + + shared_y4 = read_imagef(src1, (offset + 2)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 6)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + + fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3)); + + shared_y4 = read_imagef(src1, (offset + 3)); + acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6); + + shared_y4 = read_imagef(src1, (offset + 7)); + acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7); + + uchar regE = src0_e[ib00 * ne01 + i01 + expert_offset]; + sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3)); + } + + // reduction in local memory, assumes #subgroups=4 + __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)]; + if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum; + if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum; + if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum; + barrier(CLK_LOCAL_MEM_FENCE); + if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid]; + if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid]; + if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid]; + + // 1 outputs per thread in subgroup 0 + if (sgid == 0) { + dst = dst + (offsetd >> 2); + dst[i01 + i20 * ne01] = sum; + } + +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl new file mode 100644 index 0000000..ee5c79f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl @@ -0,0 +1,268 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable + +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#endif + +// assume +#define QK4_0 32 +#define N_SIMDGROUP 4 + +#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \ + float shared_y; \ + shared_y = sub_group_broadcast(y.s0, 0); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 0); \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 0); \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 0); \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 0); \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 0); \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 0); \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 0); \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s0, 1); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 1); \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 1); \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 1); \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 1); \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 1); \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 1); \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 1); \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \ + shared_y = sub_group_broadcast(y.s0, 2); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 2); \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 2); \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 2); \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 2); \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 2); \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 2); \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 2); \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s0, 3); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 3); \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 3); \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 3); \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 3); \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 3); \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 3); \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 3); \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \ + float8 shared_y; \ + shared_y = sub_group_broadcast(y, 0); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + shared_y = sub_group_broadcast(y, 1); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \ + shared_y = sub_group_broadcast(y, 2); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + shared_y = sub_group_broadcast(y, 3); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +__kernel void kernel_gemv_noshuffle( + __read_only image1d_buffer_t src0_q, // quantized A + global half2 * src0_d, // A scales + __read_only image1d_buffer_t src1, // B + ulong offset1, // offset to B (0) + global float * dst, // C + ulong offsetd, // offset to C (0) + uint K, // K + int ne01, // M + int ne02, // 1 + int ne10, // K + int ne12, // 1 + int ne0, // M + int ne1, // N + int r2, // 1 + int r3) +{ + uint groupId = get_local_id(1); + uint gid = get_global_id(0); + ushort slid = get_sub_group_local_id(); + + __private uint4 regA; + __private half2 regS; + __private float8 regB; + + __private float2 totalSum = (float2)(0.0f); + + // loop along K in block granularity, skip 4 blocks every iter + for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) { + regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows + // first 4 fibers in each wave load 8 B values to its private scope + if (slid < 4) { + regB.s0123 = read_imagef(src1, (slid * 2 + k * 8)); + regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8)); + } + + // load half weights for two blocks in consecutive rows + regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x; + regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x; + regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x; + regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x; +#ifdef VECTOR_SUB_GROUP_BROADCAT + dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB); +#else + dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB); +#endif // VECTOR_SUB_GROUP_BROADCAT + + regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x; + regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x; + regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x; + regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x; +#ifdef VECTOR_SUB_GROUP_BROADCAT + dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB); +#else + dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB); +#endif // VECTOR_SUB_GROUP_BROADCAT + } + + // reduction in local memory, assumes #wave=4 + __local float2 reduceLM[SIMDGROUP_WIDTH * 3]; + if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum; + if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum; + if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum; + barrier(CLK_LOCAL_MEM_FENCE); + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid]; + + // 2 outputs per fiber in wave 0 + if (groupId == 0) { + dst = (global float*)((global char*)dst + offsetd); + vstore2(totalSum, 0, &(dst[gid * 2])); + } + +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl new file mode 100644 index 0000000..469d3ed --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl @@ -0,0 +1,274 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable + +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#endif + +// assume +#define QK4_0 32 +#define N_SIMDGROUP 4 + +#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \ + float shared_y; \ + shared_y = sub_group_broadcast(y.s0, 0); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 0); \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 0); \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 0); \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 0); \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 0); \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 0); \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 0); \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s0, 1); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 1); \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 1); \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 1); \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 1); \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 1); \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 1); \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 1); \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \ + shared_y = sub_group_broadcast(y.s0, 2); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 2); \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 2); \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 2); \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 2); \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 2); \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 2); \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 2); \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s0, 3); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 3); \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 3); \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 3); \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s4, 3); \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 3); \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 3); \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 3); \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \ + float8 shared_y; \ + shared_y = sub_group_broadcast(y, 0); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + shared_y = sub_group_broadcast(y, 1); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + + +#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \ + shared_y = sub_group_broadcast(y, 2); \ + total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + shared_y = sub_group_broadcast(y, 3); \ + total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \ + total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \ + total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \ + total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \ + total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \ + total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \ + total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \ + total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \ + total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \ + total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \ + total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \ + total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \ + total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \ + total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \ + total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ + total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +__kernel void kernel_gemv_noshuffle( + __read_only image1d_buffer_t src0_q, // quantized A + global half2 * src0_d, // A scales + __read_only image1d_buffer_t src1, // B + ulong offset1, // offset to B (0) + global float * dst, // C + ulong offsetd, // offset to C (0) + int ne00, // K + int ne01, // M + int ne02, // 1 + int ne10, // K + int ne12, // 1 + int ne0, // M + int ne1, // N + int r2, // 1 + int r3) +{ + uint groupId = get_local_id(1); + uint gid = get_global_id(0); + ushort slid = get_sub_group_local_id(); + + uint K = ne00; + uint M = ne01; + + uint LINE_STRIDE_A = M / 2; + uint BLOCK_STRIDE_A = N_SIMDGROUP * M; + + __private uint4 regA; + __private half2 regS; + __private float8 regB; + + __private float2 totalSum = (float2)(0.0f); + + // loop along K in block granularity, skip 4 blocks every iter + for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) { + regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows + // first 4 fibers in each wave load 8 B values to its private scope + if (slid < 4) { + regB.s0123 = read_imagef(src1, (slid * 2 + k * 8)); + regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8)); + } + + // load half weights for two blocks in consecutive rows + regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x; + regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x; + regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x; + regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x; +#ifdef VECTOR_SUB_GROUP_BROADCAT + dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB); +#else + dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB); +#endif // VECTOR_SUB_GROUP_BROADCAT + + regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x; + regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x; + regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x; + regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x; +#ifdef VECTOR_SUB_GROUP_BROADCAT + dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB); +#else + dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB); +#endif // VECTOR_SUB_GROUP_BROADCAT + } + + // reduction in local memory, assumes #wave=4 + __local float2 reduceLM[SIMDGROUP_WIDTH * 3]; + if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum; + if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum; + if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum; + barrier(CLK_LOCAL_MEM_FENCE); + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid]; + + // 2 outputs per fiber in wave 0 + if (groupId == 0) { + dst = (global float*)((global char*)dst + offsetd); + vstore2(totalSum, 0, &(dst[gid * 2])); + } + +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl new file mode 100644 index 0000000..f944ef3 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl @@ -0,0 +1,195 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable + +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#endif + +#define QK8_0 32 +#define N_SIMDGROUP 4 + +#define dequantizeBlockAccum_ns_sgbroadcast_1(total_sums, bits8, scale, y) \ + float shared_y; \ + char elem; \ + \ + shared_y = sub_group_broadcast(y.s0, 0); \ + elem = (char)(bits8.s0 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 0); \ + elem = (char)((bits8.s0 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 0); \ + elem = (char)((bits8.s0 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 0); \ + elem = (char)((bits8.s0 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s4, 0); \ + elem = (char)(bits8.s1 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 0); \ + elem = (char)((bits8.s1 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 0); \ + elem = (char)((bits8.s1 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 0); \ + elem = (char)((bits8.s1 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s0, 1); \ + elem = (char)(bits8.s2 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 1); \ + elem = (char)((bits8.s2 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 1); \ + elem = (char)((bits8.s2 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 1); \ + elem = (char)((bits8.s2 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s4, 1); \ + elem = (char)(bits8.s3 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 1); \ + elem = (char)((bits8.s3 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 1); \ + elem = (char)((bits8.s3 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 1); \ + elem = (char)((bits8.s3 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s0, 2); \ + elem = (char)(bits8.s4 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 2); \ + elem = (char)((bits8.s4 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 2); \ + elem = (char)((bits8.s4 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 2); \ + elem = (char)((bits8.s4 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s4, 2); \ + elem = (char)(bits8.s5 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 2); \ + elem = (char)((bits8.s5 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 2); \ + elem = (char)((bits8.s5 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 2); \ + elem = (char)((bits8.s5 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s0, 3); \ + elem = (char)(bits8.s6 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s1, 3); \ + elem = (char)((bits8.s6 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s2, 3); \ + elem = (char)((bits8.s6 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s3, 3); \ + elem = (char)((bits8.s6 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + \ + shared_y = sub_group_broadcast(y.s4, 3); \ + elem = (char)(bits8.s7 & 0x000000FF); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s5, 3); \ + elem = (char)((bits8.s7 & 0x0000FF00) >> 8); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s6, 3); \ + elem = (char)((bits8.s7 & 0x00FF0000) >> 16); \ + total_sums += convert_int(elem) * scale * shared_y; \ + shared_y = sub_group_broadcast(y.s7, 3); \ + elem = (char)((bits8.s7 & 0xFF000000) >> 24); \ + total_sums += convert_int(elem) * scale * shared_y; \ + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +__kernel void kernel_gemv_noshuffle( + __read_only image1d_buffer_t src0_q, // quantized A + global half * src0_d, // A scales + __read_only image1d_buffer_t src1, // B + ulong offset1, // offset to B (0) + global float * dst, // C + ulong offsetd, // offset to C + int ne00, // K + int ne01, // M + int ne02, // 1 + int ne10, // K + int ne12, // 1 + int ne0, // M + int ne1, // N + int r2, // 1 + int r3) +{ + uint groupId = get_local_id(1); + uint gid = get_global_id(0); + ushort slid = get_sub_group_local_id(); + + uint K = ne00; + uint M = ne01; + + uint LINE_STRIDE_A = M; + uint BLOCK_STRIDE_A = 8 * M; // 32 / 4 = 8 + + __private uint8 regA; + __private half regS; + __private float8 regB; + + __private float totalSum = (float)(0.0f); + + // loop along K in block granularity, skip 4 blocks every iter + #pragma unroll 1 /* tell compiler not to unroll */ + for (uint k = groupId; k < (K / QK8_0); k += N_SIMDGROUP) { + regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of one rows + // first 4 fibers in each wave load 8 B values to its private scope + if (slid < 4) { + regB.s0123 = read_imagef(src1, (slid * 2 + k * 8)); + regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8)); + } + + // load weights for one block in consecutive rows + regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x; + regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x; + regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x; + regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x; + regA.s4 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x; + regA.s5 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x; + regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x; + regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x; + + dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB); + } + + // reduction in local memory, assumes #wave=4 + __local float reduceLM[SIMDGROUP_WIDTH * 3]; + if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum; + if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum; + if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum; + barrier(CLK_LOCAL_MEM_FENCE); + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid]; + if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid]; + + // 1 outputs per fiber in wave 0 + if (groupId == 0) { + dst = (global float*)((global char*)dst + offsetd); + dst[gid] = totalSum; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl new file mode 100644 index 0000000..c2962ed --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl @@ -0,0 +1,187 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +#define QK4_0 32 + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + + +//------------------------------------------------------------------------------ +// dequantize_q4_0_f32, dequantize_q4_0_f16 +//------------------------------------------------------------------------------ +void dequantize_q4_0_f32(global struct block_q4_0 * xb, short il, float16 * reg) { + global ushort * qs = ((global ushort *)xb + 1); + float d1 = il ? (xb->d / 16.h) : xb->d; + float d2 = d1 / 256.f; + float md = -8.h * xb->d; + ushort mask0 = il ? 0x00F0 : 0x000F; + ushort mask1 = mask0 << 8; + + reg->s0 = d1 * (qs[0] & mask0) + md; + reg->s1 = d2 * (qs[0] & mask1) + md; + + reg->s2 = d1 * (qs[1] & mask0) + md; + reg->s3 = d2 * (qs[1] & mask1) + md; + + reg->s4 = d1 * (qs[2] & mask0) + md; + reg->s5 = d2 * (qs[2] & mask1) + md; + + reg->s6 = d1 * (qs[3] & mask0) + md; + reg->s7 = d2 * (qs[3] & mask1) + md; + + reg->s8 = d1 * (qs[4] & mask0) + md; + reg->s9 = d2 * (qs[4] & mask1) + md; + + reg->sa = d1 * (qs[5] & mask0) + md; + reg->sb = d2 * (qs[5] & mask1) + md; + + reg->sc = d1 * (qs[6] & mask0) + md; + reg->sd = d2 * (qs[6] & mask1) + md; + + reg->se = d1 * (qs[7] & mask0) + md; + reg->sf = d2 * (qs[7] & mask1) + md; +} + + +//------------------------------------------------------------------------------ +// get_rows +//------------------------------------------------------------------------------ +kernel void kernel_get_rows_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int i10 = get_group_id(0); + int i11 = get_group_id(1); + int i12 = get_group_id(2); + + int r = ((global int *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0]; + + int i02 = i11; + int i03 = i12; + + for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) { + if (ind >= ne00) { + return; + } + ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] = + ((global float *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind]; + } +} + +kernel void kernel_get_rows_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int i10 = get_group_id(0); + int i11 = get_group_id(1); + int i12 = get_group_id(2); + + int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0]; + + int i02 = i11; + int i03 = i12; + + for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) { + if (ind >= ne00) { + return; + } + ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] = + ((global half *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind]; + } +} + +kernel void kernel_get_rows_q4_0( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + const int NL = 2; + + int i10 = get_group_id(0); + int i11 = get_group_id(1); + int i12 = get_group_id(2); + + int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0]; + + int i02 = i11; + int i03 = i12; + + for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) { + float16 temp; + if (ind >= ne00) { + return; + } + dequantize_q4_0_f32( + ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03)) + ind/NL, ind%NL, &temp); + *(((global float16 *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1)) + ind) = temp; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl new file mode 100644 index 0000000..059a4bb --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl @@ -0,0 +1,378 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define GELU_COEF_A 0.044715f +#define GELU_QUICK_COEF -1.702f +#define SQRT_2_OVER_PI 0.79788456080286535587989211986876f +#define SQRT_2_INV 0.70710678118654752440084436210484f + +//------------------------------------------------------------------------------ +// geglu +//------------------------------------------------------------------------------ +kernel void kernel_geglu( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const float x0 = src0_row[i0]; + const float x1 = src1_row[i0]; + + const float gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0))); + + dst_row[i0] = gelu*x1; + } +} + +kernel void kernel_geglu_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const half x0 = src0_row[i0]; + const half x1 = src1_row[i0]; + + const half gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0))); + + dst_row[i0] = gelu*x1; + } +} + +//------------------------------------------------------------------------------ +// reglu +//------------------------------------------------------------------------------ +kernel void kernel_reglu( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const float x0 = src0_row[i0]; + const float x1 = src1_row[i0]; + + dst_row[i0] = x0*x1*(x0 > 0.0f); + } +} + +kernel void kernel_reglu_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const half x0 = src0_row[i0]; + const half x1 = src1_row[i0]; + + dst_row[i0] = x0*x1*(x0 > 0.0f); + } +} + +//------------------------------------------------------------------------------ +// swiglu +//------------------------------------------------------------------------------ +kernel void kernel_swiglu( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const float x0 = src0_row[i0]; + const float x1 = src1_row[i0]; + + const float silu = x0 / (1.0f + exp(-x0)); + + dst_row[i0] = silu*x1; + } +} + +kernel void kernel_swiglu_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const half x0 = src0_row[i0]; + const half x1 = src1_row[i0]; + + const half silu = x0 / (1.0f + exp(-x0)); + + dst_row[i0] = silu*x1; + } +} + +//------------------------------------------------------------------------------ +// swiglu_oai +//------------------------------------------------------------------------------ +kernel void kernel_swiglu_oai( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off, + float limit, + float alpha +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + float x0 = src0_row[i0]; + float x1 = src1_row[i0]; + + x0 = min(x0, limit); + x1 = max(min(x1, limit), -limit); + + float out_glu = x0 / (1.0f + exp(-x0 * alpha)); + out_glu = out_glu * (1.0f + x1); + + dst_row[i0] = out_glu; + } +} + +//------------------------------------------------------------------------------ +// geglu_erf +//------------------------------------------------------------------------------ +kernel void kernel_geglu_erf( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const float x0 = src0_row[i0]; + const float x1 = src1_row[i0]; + + const float gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV)); + + dst_row[i0] = gelu_erf*x1; + } +} + +kernel void kernel_geglu_erf_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const half x0 = src0_row[i0]; + const half x1 = src1_row[i0]; + + const half gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV)); + + dst_row[i0] = gelu_erf*x1; + } +} + +//------------------------------------------------------------------------------ +// geglu_quick +//------------------------------------------------------------------------------ +kernel void kernel_geglu_quick( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global float * dst_row = (global float *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const float x0 = src0_row[i0]; + const float x1 = src1_row[i0]; + + const float gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0))); + + dst_row[i0] = gelu_quick*x1; + } +} + +kernel void kernel_geglu_quick_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb01, + ulong nb11, + int ne0, + ulong nb1, + int ne00_off, + int ne10_off +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off; + global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off; + global half * dst_row = (global half *) ((global char *) dst + get_group_id(0)*nb1); + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const half x0 = src0_row[i0]; + const half x1 = src1_row[i0]; + + const half gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0))); + + dst_row[i0] = gelu_quick*x1; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl new file mode 100644 index 0000000..8e4fa0e --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl @@ -0,0 +1,121 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +// Workgroup must be a subgroup +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_32 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_group_norm( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne, + int group_size, + float eps +) { + src0 = (global float *)((global char *)src0 + offset0); + dst = (global float *)((global char *)dst + offsetd); + + int start = get_group_id(0) * group_size; + int end = start + group_size; + + start += get_local_id(0); + + if (end >= ne) { + end = ne; + } + + float tmp = 0.0f; + + for (int j = start; j < end; j += get_local_size(0)) { + tmp += src0[j]; + } + + tmp = sub_group_reduce_add(tmp); + + const float mean = tmp / group_size; + tmp = 0.0f; + + for (int j = start; j < end; j += get_local_size(0)) { + float xi = src0[j] - mean; + dst[j] = xi; + tmp += xi * xi; + } + + tmp = sub_group_reduce_add(tmp); + + const float variance = tmp / group_size; + const float scale = 1.0f/sqrt(variance + eps); + for (int j = start; j < end; j += get_local_size(0)) { + dst[j] *= scale; + } +} + +//------------------------------------------------------------------------------ +// group_norm_mul_add +//------------------------------------------------------------------------------ +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_32 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_group_norm_mul_add( + global float * src0, ulong offset0, + global float * src1, ulong offset1, + global float * src2, ulong offset2, + global float * dst, ulong offsetd, + int ne, + int group_size, + float eps +) { + src0 = (global float *)((global char *)src0 + offset0); + src1 = (global float *)((global char *)src1 + offset1); + src2 = (global float *)((global char *)src2 + offset2); + dst = (global float *)((global char *)dst + offsetd); + + int start = get_group_id(0) * group_size; + int end = start + group_size; + if (end > ne) { + end = ne; + } + + float sum = 0.0f; + float sum_sq = 0.0f; + + for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) { + float val = src0[j]; + sum += val; + sum_sq += val*val; + } + + sum = sub_group_reduce_add(sum); + sum_sq = sub_group_reduce_add(sum_sq); + + const float mean = sum / group_size; + const float var = sum_sq / group_size - mean * mean; + const float scale = rsqrt(var + eps); + + for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) { + dst[j] = ((src0[j] - mean) * scale) * src1[j] + src2[j]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl new file mode 100644 index 0000000..cf6cdaa --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl @@ -0,0 +1,57 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_im2col_f16( + global float * src1, + ulong offset1, + global half * dst, + ulong offsetd, + ulong batch_offset, + ulong delta_offset, + long IW, + long IH, + long IC, + long OW, + long OH, + long KW, + long KH, + long pelements, + long CHW, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1 +) { + long i = get_global_id(0); + if (i >= pelements) { + return; + } + + src1 = (global float*)((global char*)src1 + offset1); + dst = (global half*)((global char*)dst + offsetd); + + long ksize = OW * KH; + long kx = i / ksize; + long kd = kx * ksize; + long ky = (i - kd) / OW; + long ix = i % OW; + + long oh = get_group_id(1); + long batch = get_group_id(2) / IC; + long ic = get_group_id(2) % IC; + + long iiw = ix * s0 + kx * d0 - p0; + long iih = oh * s1 + ky * d1 - p1; + + long offset_dst = + ((batch * OH + oh) * OW + ix) * CHW + + (ic * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + long offset_src = ic * delta_offset + batch * batch_offset; + dst[offset_dst] = src1[offset_src + iih * IW + iiw]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl new file mode 100644 index 0000000..1ecdb23 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl @@ -0,0 +1,57 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_im2col_f32( + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + ulong batch_offset, + ulong delta_offset, + long IW, + long IH, + long IC, + long OW, + long OH, + long KW, + long KH, + long pelements, + long CHW, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1 +) { + long i = get_global_id(0); + if (i >= pelements) { + return; + } + + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + long ksize = OW * KH; + long kx = i / ksize; + long kd = kx * ksize; + long ky = (i - kd) / OW; + long ix = i % OW; + + long oh = get_group_id(1); + long batch = get_group_id(2) / IC; + long ic = get_group_id(2) % IC; + + long iiw = ix * s0 + kx * d0 - p0; + long iih = oh * s1 + ky * d1 - p1; + + long offset_dst = + ((batch * OH + oh) * OW + ix) * CHW + + (ic * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + long offset_src = ic * delta_offset + batch * batch_offset; + dst[offset_dst] = src1[offset_src + iih * IW + iiw]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl new file mode 100644 index 0000000..5c3e8bc --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl @@ -0,0 +1,39 @@ + +kernel void kernel_mean_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global float *)((global char *)src0 + offset0); + dst = (global float *)((global char *)dst + offsetd); + + int i3 = get_global_id(2); + int i2 = get_global_id(1); + int i1 = get_global_id(0); + + if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) { + return; + } + + global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03); + global float * dst_row = (global float *) ((global char *) dst + i1*nb1 + i2*nb2 + i3*nb3); + + float row_sum = 0; + + for (int i0 = 0; i0 < ne00; i0++) { + row_sum += src_row[i0]; + } + + dst_row[0] = row_sum / ne00; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl new file mode 100644 index 0000000..b12a592 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl @@ -0,0 +1,152 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// mul +//------------------------------------------------------------------------------ +kernel void kernel_mul( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10)); + } +} + +// assumption: src1 is a row +// broadcast src1 into src0 +kernel void kernel_mul_row( + global float4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float4 * dst, + ulong offsetd, + int ne +) { + src0 = (global float4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] * src1[idx1]; +} + +kernel void kernel_mul_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10)); + } +} + +kernel void kernel_mul_row_f16( + global half4 * src0, + ulong offset0, + global half4 * src1, + ulong offset1, + global half4 * dst, + ulong offsetd, + int ne +) { + src0 = (global half4*)((global char*)src0 + offset0); + src1 = (global half4*)((global char*)src1 + offset1); + dst = (global half4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] * src1[idx1]; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl new file mode 100644 index 0000000..ecb577b --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl @@ -0,0 +1,139 @@ +// src0_q, src0_d, src1 are transposed as a preprocessing step +// 4-bit weights are transposed in groups of 4 (unsigned short int) +// consider weights originally "next to each other", now "on top of each other" +// each fiber computes a 8x4 tile of output elements +// using unshuffled weights + +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable + +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_128 +#endif + +kernel void kernel_mul_mat_Ab_Bi_8x4( + global const ushort * src0_q, // quantized A + global const half * src0_d, // A scales + __read_only image1d_buffer_t src1, // B (1d image) + global float * dst, // C + int m, // M + int n, // N with padding + int k, // K + int n_no_padding // N without padding +) { + + int m_4 = m >> 2; + int n_4 = n >> 2; + + int gy = get_global_id(0); + int gx = get_global_id(1); + int gx_2 = gx << 2; + + half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0; // 8x4 output elements + half8 B; // registers for activations + half4 dequantized_weights; // registers for dequantized weights + __global const ushort* weight_ptr = src0_q + gx_2; // pointer for weights + __global const half* scale_ptr = src0_d + gx_2; // pointer for scales + + for(int i=0; i<k; i+=4){ //loop through K dimension + + B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4)); + B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1); + + // keep (i/4) and (i/32) in parenthesis, rounds down + // load 4 consecutive groups of 4 weights + ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m)); // (i/4) because weights grouped in 4s + + // load 4 consecutive scales + half4 scale = vload4(0, scale_ptr + (i/32)*(m));// (i/32) because 1 scale per 32 elements + + // j=0 + dequantized_weights.s0 = ((bits4.s0 & (0x000F)) - 8) * scale.s0; // dequantize a row of the 16 weights + dequantized_weights.s1 = ((bits4.s1 & (0x000F)) - 8) * scale.s1; + dequantized_weights.s2 = ((bits4.s2 & (0x000F)) - 8) * scale.s2; + dequantized_weights.s3 = ((bits4.s3 & (0x000F)) - 8) * scale.s3; + c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate + c1 += B * dequantized_weights.s1; + c2 += B * dequantized_weights.s2; + c3 += B * dequantized_weights.s3; + + // j=1 + B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4)); + B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1); + dequantized_weights.s0 = (((bits4.s0 & (0x00F0)) >> 4) - 8) * scale.s0; // dequantize a row of the 16 weights + dequantized_weights.s1 = (((bits4.s1 & (0x00F0)) >> 4) - 8) * scale.s1; + dequantized_weights.s2 = (((bits4.s2 & (0x00F0)) >> 4) - 8) * scale.s2; + dequantized_weights.s3 = (((bits4.s3 & (0x00F0)) >> 4) - 8) * scale.s3; + c0 += B * dequantized_weights.s0; //vector-scalar multiplication to accumulate + c1 += B * dequantized_weights.s1; + c2 += B * dequantized_weights.s2; + c3 += B * dequantized_weights.s3; + + // j=2 + B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4)); + B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1); + dequantized_weights.s0 = (((bits4.s0 & (0x0F00)) >> 8) - 8) * scale.s0; // dequantize a row of the 16 weights + dequantized_weights.s1 = (((bits4.s1 & (0x0F00)) >> 8) - 8) * scale.s1; + dequantized_weights.s2 = (((bits4.s2 & (0x0F00)) >> 8) - 8) * scale.s2; + dequantized_weights.s3 = (((bits4.s3 & (0x0F00)) >> 8) - 8) * scale.s3; + c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate + c1 += B * dequantized_weights.s1; + c2 += B * dequantized_weights.s2; + c3 += B * dequantized_weights.s3; + + // j=3 + B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4)); + B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1); + dequantized_weights.s0 = (((bits4.s0 & (0xF000)) >> 12) - 8) * scale.s0; // dequantize a row of the 16 weights + dequantized_weights.s1 = (((bits4.s1 & (0xF000)) >> 12) - 8) * scale.s1; + dequantized_weights.s2 = (((bits4.s2 & (0xF000)) >> 12) - 8) * scale.s2; + dequantized_weights.s3 = (((bits4.s3 & (0xF000)) >> 12) - 8) * scale.s3; + c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate + c1 += B * dequantized_weights.s1; + c2 += B * dequantized_weights.s2; + c3 += B * dequantized_weights.s3; + } + + int idx = (gy<<3)*m + (gx<<2); // vectorized store 16 elements + + // conditional check if store is to a valid location. Required when N is not a multiple of 8 + // if statements allow registers to be reused for each store + // provides a performance boost due to reduced register footprint, which increases number of concurrent waves + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl new file mode 100644 index 0000000..73a8884 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl @@ -0,0 +1,130 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#if defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#else +#define REQD_SUBGROUP_SIZE_128 +#endif + +#define OPWM 64 +#define OPWN 64 +#define CPWK 8 +#define OPTM 4 +#define OPTN 8 + +#define WG_M (OPWM / OPTM) +#define WG_N (OPWN / OPTN) +#define VEC_K (CPWK / 4) + +REQD_SUBGROUP_SIZE_128 +__kernel void mul_mat_f16_f32( + const int M, const int N, const int K, + __global const void* A_void, ulong A_offset, + __global const void* B_void, ulong B_offset, + __global void* C_void, ulong C_offset) { + + __global const half* A = (__global const half* )((__global const char*)A_void + A_offset); + __global const float* B = (__global const float*)((__global const char*)B_void + B_offset); + __global float* C = (__global float*)((__global char*)C_void + C_offset); + + const int lidm = get_local_id(0); + const int lidn = get_local_id(1); + const int lid = lidn * WG_M + lidm; + + const int offsetM = get_group_id(0) * OPWM; + const int offsetN = get_group_id(1) * OPWN; + + __local half4 Alocal[OPWM][VEC_K]; + __local float4 Blocal[OPWN][VEC_K]; + + float sum[OPTM][OPTN]; + + for (int wm = 0; wm < OPTM; wm++) { + for (int wn = 0; wn < OPTN; wn++) { + sum[wm][wn] = 0.0f; + } + } + + const int numTiles = (K + CPWK - 1) / CPWK; + + const int load_row_a = lid % OPWM; + const int load_vec_k_a = lid / OPWM; + const int global_row_a = offsetM + load_row_a; + + const int load_row_b = lid % OPWN; + const int load_vec_k_b = lid / OPWN; + const int global_row_b = offsetN + load_row_b; + + for (int t = 0; t < numTiles; t++) { + const int k_start = t * CPWK; + const int k_vec_start_a = k_start + load_vec_k_a * 4; + const int k_vec_start_b = k_start + load_vec_k_b * 4; + + if (global_row_a < M && k_vec_start_a < K) { + if (k_vec_start_a + 3 < K) { + Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a); + } else { + half4 tempA = (half4)(0.0h); + if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a]; + if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1]; + if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2]; + Alocal[load_row_a][load_vec_k_a] = tempA; + } + } else { + Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h); + } + + if (global_row_b < N && k_vec_start_b < K) { + if (k_vec_start_b + 3 < K) { + Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b); + } else { + float4 tempB = (float4)(0.0f); + if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b]; + if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1]; + if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2]; + Blocal[load_row_b][load_vec_k_b] = tempB; + } + } else { + Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f); + } + + barrier(CLK_LOCAL_MEM_FENCE); + + #pragma unroll + for (int k_vec = 0; k_vec < VEC_K; k_vec++) { + float4 a_fvecs[OPTM]; + int current_row_a = lidm; + for (int wm = 0; wm < OPTM; wm++) { + a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]); + current_row_a += WG_M; + } + + float4 b_fvecs[OPTN]; + int current_row_b = lidn; + for (int wn = 0; wn < OPTN; wn++) { + b_fvecs[wn] = Blocal[current_row_b][k_vec]; + current_row_b += WG_N; + } + + for (int wm = 0; wm < OPTM; wm++) { + for (int wn = 0; wn < OPTN; wn++) { + sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + for (int wm = 0; wm < OPTM; wm++) { + int globalRow = offsetM + lidm + wm * WG_M; + if (globalRow < M) { + for (int wn = 0; wn < OPTN; wn++) { + int globalCol = offsetN + lidn + wn * WG_N; + if (globalCol < N) { + C[globalCol * M + globalRow] = sum[wm][wn]; + } + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl new file mode 100644 index 0000000..ac0274b --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl @@ -0,0 +1,273 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_khr_subgroups : enable + +#define LM_FIRST_256B 0 +#define LM_SECOND_256B 64 +#define LM_THIRD_256B 128 +#define LM_FOURTH_256B 192 + + +inline float16 mm_load_a( + image1d_buffer_t matrix_A, + uint subMatrixAStartInElements, + int nb01, + int line_stride_matrix_A_in_bytes +) { + __private float8 regA; + size_t sub_block_id_m = get_local_id(0); + +#ifdef KQV + uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * nb01/4); +#else // KQ + uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * line_stride_matrix_A_in_bytes/4); +#endif + + regA.s0123 = read_imagef(matrix_A, a_texCoord/4); + regA.s4567 = read_imagef(matrix_A, (a_texCoord+4)/4); + + return convert_float16(as_half16(regA)); +} + +inline float4 alu_32( + float16 regA, + __local float4* matrix_B_vec +) { + + __private float4 rC = 0; + int i = get_sub_group_id() * 64; + + rC += regA.s0 * matrix_B_vec[i]; + rC += regA.s1 * matrix_B_vec[i + 16]; + rC += regA.s4 * matrix_B_vec[i + 1]; + rC += regA.s5 * matrix_B_vec[i + 17]; + rC += regA.s8 * matrix_B_vec[i + 2]; + rC += regA.s9 * matrix_B_vec[i + 18]; + rC += regA.sc * matrix_B_vec[i + 3]; + rC += regA.sd * matrix_B_vec[i + 19]; + + i += 32; + + rC += regA.s2 * matrix_B_vec[i]; + rC += regA.s3 * matrix_B_vec[i + 16]; + rC += regA.s6 * matrix_B_vec[i + 1]; + rC += regA.s7 * matrix_B_vec[i + 17]; + rC += regA.sa * matrix_B_vec[i + 2]; + rC += regA.sb * matrix_B_vec[i + 18]; + rC += regA.se * matrix_B_vec[i + 3]; + rC += regA.sf * matrix_B_vec[i + 19]; + + return rC; +} + +inline float16 alu_16( + float16 regA, + __local float* matrix_B_local +) { + float16 out; + __local float4* matrix_B_vec = (__local float4*)matrix_B_local; + + out.s0123 = alu_32(regA, matrix_B_vec); + out.s4567 = alu_32(regA, matrix_B_vec + 4); + out.s89ab = alu_32(regA, matrix_B_vec + 8); + out.scdef = alu_32(regA, matrix_B_vec + 12); + + return out; +} + +inline void mm_mad( + __local float* matrix_B_local, + float16 regA, + float8 regB, + uint b_localOffsetInWords, + float16* regC0_ptr, + float16* regC1_ptr +) { + int offset = b_localOffsetInWords + get_sub_group_id() * 256; + + matrix_B_local[offset + LM_FIRST_256B] = regB.s0; + matrix_B_local[offset + LM_SECOND_256B] = regB.s1; + matrix_B_local[offset + LM_THIRD_256B] = regB.s2; + matrix_B_local[offset + LM_FOURTH_256B] = regB.s3; + + float16 add0 = alu_16(regA, matrix_B_local); + *regC0_ptr += add0; + + matrix_B_local[offset + LM_FIRST_256B] = regB.s4; + matrix_B_local[offset + LM_SECOND_256B] = regB.s5; + matrix_B_local[offset + LM_THIRD_256B] = regB.s6; + matrix_B_local[offset + LM_FOURTH_256B] = regB.s7; + + float16 add1 = alu_16(regA, matrix_B_local); + *regC1_ptr += add1; +} + +inline void mm_store_c_N( + __write_only image1d_buffer_t matrix_C, + float16 regC0, + float16 regC1, + uint subMatrixCStartInElements, + int line_stride_matrix_C_in_bytes, + int mask +) { + size_t sub_block_id_m = get_local_id(0); + + uint strideInWords = line_stride_matrix_C_in_bytes/4; + uint c_coordInWords_0 = (subMatrixCStartInElements + sub_block_id_m); + + uint c_coordInWords_1 = c_coordInWords_0 + 1 * strideInWords; + uint c_coordInWords_2 = c_coordInWords_0 + 2 * strideInWords; + uint c_coordInWords_3 = c_coordInWords_0 + 3 * strideInWords; + uint c_coordInWords_4 = c_coordInWords_0 + 4 * strideInWords; + uint c_coordInWords_5 = c_coordInWords_0 + 5 * strideInWords; + uint c_coordInWords_6 = c_coordInWords_0 + 6 * strideInWords; + uint c_coordInWords_7 = c_coordInWords_0 + 7 * strideInWords; + uint c_coordInWords_8 = c_coordInWords_0 + 8 * strideInWords; + uint c_coordInWords_9 = c_coordInWords_0 + 9 * strideInWords; + uint c_coordInWords_10 = c_coordInWords_0 + 10 * strideInWords; + uint c_coordInWords_11 = c_coordInWords_0 + 11 * strideInWords; + uint c_coordInWords_12 = c_coordInWords_0 + 12 * strideInWords; + uint c_coordInWords_13 = c_coordInWords_0 + 13 * strideInWords; + uint c_coordInWords_14 = c_coordInWords_0 + 14 * strideInWords; + uint c_coordInWords_15 = c_coordInWords_0 + 15 * strideInWords; + uint c_coordInWords_16 = c_coordInWords_0 + 16 * strideInWords; + uint c_coordInWords_17 = c_coordInWords_0 + 17 * strideInWords; + uint c_coordInWords_18 = c_coordInWords_0 + 18 * strideInWords; + uint c_coordInWords_19 = c_coordInWords_0 + 19 * strideInWords; + uint c_coordInWords_20 = c_coordInWords_0 + 20 * strideInWords; + uint c_coordInWords_21 = c_coordInWords_0 + 21 * strideInWords; + uint c_coordInWords_22 = c_coordInWords_0 + 22 * strideInWords; + uint c_coordInWords_23 = c_coordInWords_0 + 23 * strideInWords; + uint c_coordInWords_24 = c_coordInWords_0 + 24 * strideInWords; + uint c_coordInWords_25 = c_coordInWords_0 + 25 * strideInWords; + uint c_coordInWords_26 = c_coordInWords_0 + 26 * strideInWords; + uint c_coordInWords_27 = c_coordInWords_0 + 27 * strideInWords; + uint c_coordInWords_28 = c_coordInWords_0 + 28 * strideInWords; + uint c_coordInWords_29 = c_coordInWords_0 + 29 * strideInWords; + uint c_coordInWords_30 = c_coordInWords_0 + 30 * strideInWords; + uint c_coordInWords_31 = c_coordInWords_0 + 31 * strideInWords; + + if (mask > 0) { write_imagef(matrix_C, c_coordInWords_0, regC0.s0); } + if (mask > 1) { write_imagef(matrix_C, c_coordInWords_1, regC0.s1); } + if (mask > 2) { write_imagef(matrix_C, c_coordInWords_2, regC0.s2); } + if (mask > 3) { write_imagef(matrix_C, c_coordInWords_3, regC0.s3); } + if (mask > 4) { write_imagef(matrix_C, c_coordInWords_4, regC0.s4); } + if (mask > 5) { write_imagef(matrix_C, c_coordInWords_5, regC0.s5); } + if (mask > 6) { write_imagef(matrix_C, c_coordInWords_6, regC0.s6); } + if (mask > 7) { write_imagef(matrix_C, c_coordInWords_7, regC0.s7); } + if (mask > 8) { write_imagef(matrix_C, c_coordInWords_8, regC0.s8); } + if (mask > 9) { write_imagef(matrix_C, c_coordInWords_9, regC0.s9); } + if (mask > 10) { write_imagef(matrix_C, c_coordInWords_10, regC0.sa); } + if (mask > 11) { write_imagef(matrix_C, c_coordInWords_11, regC0.sb); } + if (mask > 12) { write_imagef(matrix_C, c_coordInWords_12, regC0.sc); } + if (mask > 13) { write_imagef(matrix_C, c_coordInWords_13, regC0.sd); } + if (mask > 14) { write_imagef(matrix_C, c_coordInWords_14, regC0.se); } + if (mask > 15) { write_imagef(matrix_C, c_coordInWords_15, regC0.sf); } + if (mask > 16) { write_imagef(matrix_C, c_coordInWords_16, regC1.s0); } + if (mask > 17) { write_imagef(matrix_C, c_coordInWords_17, regC1.s1); } + if (mask > 18) { write_imagef(matrix_C, c_coordInWords_18, regC1.s2); } + if (mask > 19) { write_imagef(matrix_C, c_coordInWords_19, regC1.s3); } + if (mask > 20) { write_imagef(matrix_C, c_coordInWords_20, regC1.s4); } + if (mask > 21) { write_imagef(matrix_C, c_coordInWords_21, regC1.s5); } + if (mask > 22) { write_imagef(matrix_C, c_coordInWords_22, regC1.s6); } + if (mask > 23) { write_imagef(matrix_C, c_coordInWords_23, regC1.s7); } + if (mask > 24) { write_imagef(matrix_C, c_coordInWords_24, regC1.s8); } + if (mask > 25) { write_imagef(matrix_C, c_coordInWords_25, regC1.s9); } + if (mask > 26) { write_imagef(matrix_C, c_coordInWords_26, regC1.sa); } + if (mask > 27) { write_imagef(matrix_C, c_coordInWords_27, regC1.sb); } + if (mask > 28) { write_imagef(matrix_C, c_coordInWords_28, regC1.sc); } + if (mask > 29) { write_imagef(matrix_C, c_coordInWords_29, regC1.sd); } + if (mask > 30) { write_imagef(matrix_C, c_coordInWords_30, regC1.se); } + if (mask > 31) { write_imagef(matrix_C, c_coordInWords_31, regC1.sf); } +} + +#define TILESIZE_K 16 +#define TILESIZE_M 64 +#define TILESIZE_N 32 +#ifdef KQV +__kernel void mul_mm_f16_f32_kqv( +#else +__kernel void mul_mm_f16_f32_kq( +#endif + __read_only image1d_buffer_t matrix_A, + int offset0, + __global float* matrix_B, + int offset1, + __write_only image1d_buffer_t matrix_C, + int offsetd, + int M, int K, int N, + int D_A, + int D_B, + int nb01 +) { + + uint block_id_m = get_global_id(1); + uint block_id_n = get_global_id(2) % ((N+TILESIZE_N-1)/TILESIZE_N); + uint block_id_d = get_global_id(2) / ((N+TILESIZE_N-1)/TILESIZE_N); + + __private float16 regA; + __private float8 regB; + __private float16 regC0; + __private float16 regC1; + + const uint col = block_id_m * TILESIZE_M; + const uint row = block_id_n * TILESIZE_N; + const uint depth_A = block_id_d / (D_B/D_A); + const uint depth_B = block_id_d; + +#ifdef KQV + int line_stride_matrix_A_in_bytes = nb01 * M; + int line_stride_matrix_B_in_bytes = K * N * 4; +#else + int line_stride_matrix_A_in_bytes = K * D_A * 2; + int line_stride_matrix_B_in_bytes = K * D_B * 4; +#endif + + int line_stride_matrix_C_in_bytes = M * 4; + + const uint strideAinElements = line_stride_matrix_A_in_bytes / 2; + const uint strideBinElements = line_stride_matrix_B_in_bytes / 4; + + size_t sub_block_id_m = get_local_id(0); + + uint b_localOffsetInWords = (sub_block_id_m/16)*16 + + ((((sub_block_id_m)>>0)&1)<<2) + + ((((sub_block_id_m)>>1)&1)<<3) + + ((((sub_block_id_m)>>2)&1)<<0) + + ((((sub_block_id_m)>>3)&1)<<1); + + uint2 b_globalOffsetInWords_xy = {((sub_block_id_m%4)*4), (sub_block_id_m>>2)}; + uint b_globalOffsetInWords00, b_globalOffsetInWords16; +#ifdef KQV + b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*K; + b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * K); + uint subMatrixAStartInElements = depth_A * strideAinElements + col * nb01 / 2; + uint subMatrixBStartInElements = depth_B * strideBinElements + row * K; +#else + b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*line_stride_matrix_B_in_bytes/4; + b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * line_stride_matrix_B_in_bytes/4); + uint subMatrixAStartInElements = col * strideAinElements + depth_A * K; + uint subMatrixBStartInElements = row * strideBinElements + depth_B * K; +#endif + + __local float matrix_B_local[1024]; + + for (uint step=0; step < K; step+=TILESIZE_K) { + size_t sub_block_id_m = get_local_id(0); + regA = mm_load_a(matrix_A, subMatrixAStartInElements, nb01, line_stride_matrix_A_in_bytes); + + uint b_coordInWords00 = subMatrixBStartInElements + b_globalOffsetInWords00; + uint b_coordInWords16 = subMatrixBStartInElements + b_globalOffsetInWords16; + + regB.s0123 = vload4(b_coordInWords00/4, matrix_B); + regB.s4567 = vload4(b_coordInWords16/4, matrix_B); + + mm_mad(matrix_B_local, regA, regB, b_localOffsetInWords, ®C0, ®C1); + + subMatrixAStartInElements += TILESIZE_K; + subMatrixBStartInElements += TILESIZE_K; + } + + uint subMatrixCStartInElements = depth_B * N * M + row * M + col; + mm_store_c_N(matrix_C, regC0, regC1, subMatrixCStartInElements, line_stride_matrix_C_in_bytes, (N-block_id_n*32)); +} + diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl new file mode 100644 index 0000000..6982f8f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl @@ -0,0 +1,146 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define LOAD_VEC_A 4 +#define LOAD_VEC_B 4 + +#define BM 64 +#define BN 64 +#define BK 16 +#define TM 4 +#define TN 8 + +kernel void kernel_mul_mm_f16_f32_l4_lm( + global half4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float * dst, + ulong offsetd, + + int ne00, + int ne01, + int ne02, + int ne11, + int ne12, + + int stride_a, + int stride_b, + int stride_d, + + int batch_stride_a, + int batch_stride_b, + int batch_stride_d, + + int r2, + int r3 +) { + src0 = (global half4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + local half buf_a[BM * BK]; + local float buf_b[BN * BK]; + + const int batch_idx = get_global_id(2); + + const int i13 = batch_idx / ne12; + const int i12 = batch_idx % ne12; + + const int i03 = i13 / r3; + const int i02 = i12 / r2; + + const int batch_idx_a = i03 * ne02 + i02; + + const int ir = get_group_id(0); + const int ic = get_group_id(1); + + const int tid = get_local_id(0); + const int th_r = tid % (BM / TM); + const int th_c = tid / (BM / TM); + + const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A); + const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A); + const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B); + const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B); + + const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK; + const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK; + + int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A; + int pos_b = (batch_idx * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B; + + float sums[TM * TN]; + half cache_a[TM]; + float cache_b[TN]; + + for (int i = 0; i < TM * TN; i++) { + sums[i] = 0.0f; + } + + for (int block = 0; block < ne00; block += BK) { + for (int l = 0; l < BM; l += loadstride_a) { + if (ir*BM + loadc_a + l < ne01) { + const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a; + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3; + } else { + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0h; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0h; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0h; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0h; + } + } + + for (int l = 0; l < BN; l += loadstride_b) { + if (ic*BN + loadc_b + l < ne11) { + const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b; + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3; + } else { + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0h; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0h; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0h; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0h; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + pos_a += BK / LOAD_VEC_A; + pos_b += BK / LOAD_VEC_B; + + for (int i = 0; i < BK; i++) { + for (int j = 0; j < TM; j++) { + cache_a[j] = buf_a[(i) * BM + th_r * TM + j]; + } + for (int j = 0; j < TN; j++) { + cache_b[j] = buf_b[(i) * BN + th_c * TN + j]; + } + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + const int sums_idx = cc*TM + cr; + sums[sums_idx] = mad(convert_float(cache_a[cr]), cache_b[cc], sums[sums_idx]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + const int dr = ir * BM + th_r * TM; + const int dc = ic * BN + th_c * TN; + + const int offsets = batch_idx * batch_stride_d; + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + if (dr + cr < ne01 && dc + cc < ne11) { + dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr]; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl new file mode 100644 index 0000000..d7d5ba6 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl @@ -0,0 +1,147 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define LOAD_VEC_A 4 +#define LOAD_VEC_B 4 + +#define BM 64 +#define BN 64 +#define BK 16 +#define TM 4 +#define TN 8 + +kernel void kernel_mul_mm_f32_f32_l4_lm( + global float4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float * dst, + ulong offsetd, + + int ne00, + int ne01, + int ne02, + int ne11, + int ne12, + + int stride_a, + int stride_b, + int stride_d, + + int batch_stride_a, + int batch_stride_b, + int batch_stride_d, + + int r2, + int r3 +) { + src0 = (global float4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + local float buf_a[BM * BK]; + local float buf_b[BN * BK]; + + const int batch_idx = get_global_id(2); + + const int i13 = batch_idx / ne12; + const int i12 = batch_idx % ne12; + + const int i03 = i13 / r3; + const int i02 = i12 / r2; + + const int batch_idx_a = i03 * ne02 + i02; + + const int ir = get_group_id(0); + const int ic = get_group_id(1); + + const int tid = get_local_id(0); + const int th_r = tid % (BM / TM); + const int th_c = tid / (BM / TM); + + const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A); + const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A); + const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B); + const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B); + + const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK; + const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK; + + int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A; + int pos_b = (batch_idx * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B; + + float sums[TM * TN]; + float cache_a[TM]; + float cache_b[TN]; + + for (int i = 0; i < TM * TN; i++) { + sums[i] = 0.0f; + } + + for (int block = 0; block < ne00; block += BK) { + for (int l = 0; l < BM; l += loadstride_a) { + if (ir*BM + loadc_a + l < ne01) { + const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a; + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3; + } else { + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f; + } + } + + for (int l = 0; l < BN; l += loadstride_b) { + if (ic*BN + loadc_b + l < ne11) { + const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b; + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3; + } else { + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + pos_a += BK / LOAD_VEC_A; + pos_b += BK / LOAD_VEC_B; + + for (int i = 0; i < BK; i++) { + for (int j = 0; j < TM; j++) { + cache_a[j] = buf_a[(i) * BM + th_r * TM + j]; + } + + for (int j = 0; j < TN; j++) { + cache_b[j] = buf_b[(i) * BN + th_c * TN + j]; + } + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + const int sums_idx = cc*TM + cr; + sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + const int dr = ir * BM + th_r * TM; + const int dc = ic * BN + th_c * TN; + + const int offsets = batch_idx * batch_stride_d; + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + if (dr + cr < ne01 && dc + cc < ne11) { + dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr]; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl new file mode 100644 index 0000000..3602c92 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl @@ -0,0 +1,158 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define LOAD_VEC_A 2 +#define LOAD_VEC_B 4 + +#define BM 64 +#define BN 64 +#define BK 32 +#define TM 4 +#define TN 8 + +kernel void kernel_mul_mm_q6_k_f32_l4_lm( + global uchar * src0_ql, + global uchar * src0_qh, + global char * src0_s, + global half * src0_d, + global float4 * src1, + ulong offset1, + global float * dst, + ulong offsetd, + + int ne00, + int ne01, + int ne02, + int ne11, + int ne12, + + int stride_a, + int stride_b, + int stride_d, + + int batch_stride_a, + int batch_stride_b, + int batch_stride_d, + + int r2, + int r3 +) { + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float *)((global char*)dst + offsetd); + + local float buf_a[BM * BK]; + local float buf_b[BN * BK]; + + const int batch_idx = get_global_id(2); + + const int i13 = batch_idx / ne12; + const int i12 = batch_idx % ne12; + + const int i03 = i13 / r3; + const int i02 = i12 / r2; + + const int batch_idx_a = i03 * ne02 + i02; + + const int ir = get_group_id(0); + const int ic = get_group_id(1); + + const int tid = get_local_id(0); + const int th_r = tid % (BM / TM); + const int th_c = tid / (BM / TM); + + const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A); + const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A); + const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B); + const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B); + + const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK; + const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK; + + int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A; + int pos_b = (batch_idx * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B; + + float sums[TM * TN]; + float cache_a[TM]; + float cache_b[TN]; + + for (int i = 0; i < TM * TN; i++) { + sums[i] = 0.0f; + } + + for (int block = 0; block < ne00; block += BK) { + for (int l = 0; l < BM; l += loadstride_a) { + if (ir*BM + loadc_a + l < ne01) { + int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a; + + int ib = idx / 128; // 2 values per idx + int iqs = idx % 128; // 0..127 + + int n = iqs / 64; // 0,1 + int b = (iqs % 64) / 32; // 0,1 + int is_b = (iqs % 16) / 8; // 0,1 + int qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6 + int is = 8 * n + qhshift + is_b; // 0..15 + int qsi = n * 64 + (iqs % 32) * 2; // 0,2,4..126 + int qhi = n * 32 + (iqs % 16) * 2; // 0,2,4..62 + + float dscale = (float)src0_d[ib] * (float)src0_s[ib*16 + is]; + + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 0] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 0] >> qhshift) & 3) << 4)) - 32); + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 1] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 1] >> qhshift) & 3) << 4)) - 32); + } else { + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f; + } + } + + for (int l = 0; l < BN; l += loadstride_b) { + if (ic*BN + loadc_b + l < ne11) { + int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b; + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3; + } else { + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + pos_a += BK / LOAD_VEC_A; + pos_b += BK / LOAD_VEC_B; + + for (int i = 0; i < BK; i++) { + for (int j = 0; j < TM; j++) { + cache_a[j] = buf_a[(i) * BM + th_r * TM + j]; + } + + for (int j = 0; j < TN; j++) { + cache_b[j] = buf_b[(i) * BN + th_c * TN + j]; + } + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + const int sums_idx = cc*TM + cr; + sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + const int dr = ir * BM + th_r * TM; + const int dc = ic * BN + th_c * TN; + + const int offsets = batch_idx * batch_stride_d; + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + if (dr + cr < ne01 && dc + cc < ne11) { + dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr]; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl new file mode 100644 index 0000000..51ce212 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl @@ -0,0 +1,129 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable + +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_128 +#endif + +kernel void kernel_mul_mm_q8_0_f32_8x4( + global const uint * src0_q, + global const half * src0_d, + __read_only image1d_buffer_t src1, + global float * dst, + int k, + int m, + int n, + int n_no_padding, + ulong offsetd +) { + + int m_4 = m >> 2; + int n_4 = n >> 2; + + int gy = get_global_id(0); + int gx = get_global_id(1); + int gx_2 = gx << 2; + dst = (global float *)((global char*)dst + offsetd); + + + half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0; + half8 B; + half4 deq; + + __global const uint* wptr = src0_q + gx_2; + __global const half* sptr = src0_d + gx_2; + + for (int i = 0; i < k; i += 4) { + uint4 pack4 = vload4(0, wptr + (i / 4) * m); + half4 scale = vload4(0, sptr + (i / 32) * m); + + char4 p0 = as_char4(pack4.s0); + char4 p1 = as_char4(pack4.s1); + char4 p2 = as_char4(pack4.s2); + char4 p3 = as_char4(pack4.s3); + + // ------------------- j = 0 (k = i+0) ------------------- + B.s0123 = read_imageh(src1, gy * 2 + (i + 0) * n_4); + B.s4567 = read_imageh(src1, gy * 2 + (i + 0) * n_4 + 1); + + half4 wj0 = convert_half4((char4)(p0.s0, p1.s0, p2.s0, p3.s0)) * scale; + + c0 += B * wj0.s0; + c1 += B * wj0.s1; + c2 += B * wj0.s2; + c3 += B * wj0.s3; + + // ------------------- j = 1 (k = i+1) ------------------- + B.s0123 = read_imageh(src1, gy * 2 + (i + 1) * n_4); + B.s4567 = read_imageh(src1, gy * 2 + (i + 1) * n_4 + 1); + + half4 wj1 = convert_half4((char4)(p0.s1, p1.s1, p2.s1, p3.s1)) * scale; + + c0 += B * wj1.s0; + c1 += B * wj1.s1; + c2 += B * wj1.s2; + c3 += B * wj1.s3; + + // ------------------- j = 2 (k = i+2) ------------------- + B.s0123 = read_imageh(src1, gy * 2 + (i + 2) * n_4); + B.s4567 = read_imageh(src1, gy * 2 + (i + 2) * n_4 + 1); + + half4 wj2 = convert_half4((char4)(p0.s2, p1.s2, p2.s2, p3.s2)) * scale; + + c0 += B * wj2.s0; + c1 += B * wj2.s1; + c2 += B * wj2.s2; + c3 += B * wj2.s3; + + // ------------------- j = 3 (k = i+3) ------------------- + B.s0123 = read_imageh(src1, gy * 2 + (i + 3) * n_4); + B.s4567 = read_imageh(src1, gy * 2 + (i + 3) * n_4 + 1); + + half4 wj3 = convert_half4((char4)(p0.s3, p1.s3, p2.s3, p3.s3)) * scale; + + c0 += B * wj3.s0; + c1 += B * wj3.s1; + c2 += B * wj3.s2; + c3 += B * wj3.s3; + } + + int idx = (gy << 3) * m + (gx << 2); + + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx); + idx += m; + } + if(idx+3 < m*n_no_padding){ + vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl new file mode 100644 index 0000000..147b66f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl @@ -0,0 +1,154 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#define LOAD_VEC_A 4 +#define LOAD_VEC_B 4 + +#define BM 64 +#define BN 64 +#define BK 32 +#define TM 4 +#define TN 8 + +kernel void kernel_mul_mm_q8_0_f32_l4_lm( + global char4 * src0_q, + global half * src0_d, + global float4 * src1, + ulong offset1, + global float * dst, + ulong offsetd, + + int ne00, + int ne01, + int ne02, + int ne11, + int ne12, + + int stride_a, + int stride_b, + int stride_d, + + int batch_stride_a, + int batch_stride_b, + int batch_stride_d, + + int r2, + int r3 +) { + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float *)((global char*)dst + offsetd); + + local float buf_a[BM * BK]; + local float buf_b[BN * BK]; + + const int batch_idx = get_global_id(2); + + const int i13 = batch_idx / ne12; + const int i12 = batch_idx % ne12; + + const int i03 = i13 / r3; + const int i02 = i12 / r2; + + const int batch_idx_a = i03 * ne02 + i02; + + const int ir = get_group_id(0); + const int ic = get_group_id(1); + + const int tid = get_local_id(0); + const int th_r = tid % (BM / TM); + const int th_c = tid / (BM / TM); + + const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A); + const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A); + const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B); + const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B); + + const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK; + const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK; + + int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A; + int pos_b = (batch_idx * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B; + + float sums[TM * TN]; + float cache_a[TM]; + float cache_b[TN]; + + for (int i = 0; i < TM * TN; i++) { + sums[i] = 0.0f; + } + + for (int block = 0; block < ne00; block += BK) { + for (int l = 0; l < BM; l += loadstride_a) { + if (ir*BM + loadc_a + l < ne01) { + int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a; + int ib = idx / 8; + int iqs = idx % 8; + + float d = (float)src0_d[ib]; + global char4 * qs = src0_q + ib*8 + iqs; + char4 q = *qs; + float4 v = convert_float4(q)*d; + + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v.s0; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v.s1; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v.s2; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v.s3; + } else { + buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f; + buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f; + } + } + + for (int l = 0; l < BN; l += loadstride_b) { + if (ic*BN + loadc_b + l < ne11) { + int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b; + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3; + } else { + buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f; + buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f; + } + } + + barrier(CLK_LOCAL_MEM_FENCE); + + pos_a += BK / LOAD_VEC_A; + pos_b += BK / LOAD_VEC_B; + + for (int i = 0; i < BK; i++) { + for (int j = 0; j < TM; j++) { + cache_a[j] = buf_a[(i) * BM + th_r * TM + j]; + } + + for (int j = 0; j < TN; j++) { + cache_b[j] = buf_b[(i) * BN + th_c * TN + j]; + } + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + const int sums_idx = cc*TM + cr; + sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]); + } + } + } + barrier(CLK_LOCAL_MEM_FENCE); + } + + const int dr = ir * BM + th_r * TM; + const int dc = ic * BN + th_c * TN; + + const int offsets = batch_idx * batch_stride_d; + + for (int cc = 0; cc < TN; cc++) { + for (int cr = 0; cr < TM; cr++) { + if (dr + cr < ne01 && dc + cc < ne11) { + dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr]; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl new file mode 100644 index 0000000..9393b54 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl @@ -0,0 +1,118 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define N_F16_F16 4 + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_f16_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3) +{ + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int r0 = get_group_id(0); + int rb = get_group_id(1)*N_F16_F16; + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + + global half * x = (global half *) (src0 + offset_src0); + + if (ne00 < 128) { + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global half * y = (global half *) (src1 + offset_src1); + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) { + sumf += (half) x[i] * (half) y[i]; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + global half4 * x4 = (global half4 *)x; + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global half * y = (global half *) (src1 + offset_src1); + global half4 * y4 = (global half4 *) y; + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) { + sumf += (half) x4[i].s0 * y4[i].s0; + sumf += (half) x4[i].s1 * y4[i].s1; + sumf += (half) x4[i].s2 * y4[i].s2; + sumf += (half) x4[i].s3 * y4[i].s3; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) { + all_sum += (half) x[i] * y[i]; + } + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl new file mode 100644 index 0000000..e52d3c6 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl @@ -0,0 +1,118 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define N_F16_F32 4 + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_f16_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int r0 = get_group_id(0); + int rb = get_group_id(1)*N_F16_F32; + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + + global half * x = (global half *) (src0 + offset_src0); + + if (ne00 < 128) { + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global float * y = (global float *) (src1 + offset_src1); + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) { + sumf += convert_float(x[i]) * y[i]; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + global half4 * x4 = (global half4 *)x; + for (int row = 0; row < N_F16_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global float * y = (global float *) (src1 + offset_src1); + global float4 * y4 = (global float4 *) y; + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) { + sumf += convert_float(x4[i].s0) * y4[i].s0; + sumf += convert_float(x4[i].s1) * y4[i].s1; + sumf += convert_float(x4[i].s2) * y4[i].s2; + sumf += convert_float(x4[i].s3) * y4[i].s3; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) { + all_sum += (float) x[i] * y[i]; + } + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl new file mode 100644 index 0000000..28d3021 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl @@ -0,0 +1,94 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_f16_f32_1row( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global half * x = (global half *) (src0 + offset_src0); + global float * y = (global float *) (src1 + offset_src1); + + float sumf = 0; + if (ne00 < 128) { + for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) { + sumf += (float) x[i] * (float) y[i]; + } + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } else { + global half4 * x4 = (global half4 *) x; + global float4 * y4 = (global float4 *) y; + for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) { + sumf += (float) x4[i].s0 * y4[i].s0; + sumf += (float) x4[i].s1 * y4[i].s1; + sumf += (float) x4[i].s2 * y4[i].s2; + sumf += (float) x4[i].s3 * y4[i].s3; + } + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) { + all_sum += (float) x[i] * y[i]; + } + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl new file mode 100644 index 0000000..cdf8197 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl @@ -0,0 +1,84 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +// Assumes row size (ne00) is a multiple of 4 +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_f16_f32_l4( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int nrows = ne11; + int r0 = get_group_id(0); + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + + global half4 * x4 = (global half4 *) (src0 + offset_src0); + + for (int r1 = 0; r1 < nrows; ++r1) { + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global float4 * y4 = (global float4 *) (src1 + offset_src1); + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) { + sumf += convert_float(x4[i].s0) * y4[i].s0; + sumf += convert_float(x4[i].s1) * y4[i].s1; + sumf += convert_float(x4[i].s2) * y4[i].s2; + sumf += convert_float(x4[i].s3) * y4[i].s3; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl new file mode 100644 index 0000000..ec71b87 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl @@ -0,0 +1,118 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define N_F32_F32 4 + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_f32_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int r0 = get_group_id(0); + int rb = get_group_id(1)*N_F32_F32; + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + + global float * x = (global float *) (src0 + offset_src0); + + if (ne00 < 128) { + for (int row = 0; row < N_F32_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global float * y = (global float *) (src1 + offset_src1); + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) { + sumf += (float) x[i] * (float) y[i]; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + global float4 * x4 = (global float4 *)x; + for (int row = 0; row < N_F32_F32; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global float * y = (global float *) (src1 + offset_src1); + global float4 * y4 = (global float4 *) y; + + float sumf = 0; + for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) { + sumf += (float) x4[i].s0 * y4[i].s0; + sumf += (float) x4[i].s1 * y4[i].s1; + sumf += (float) x4[i].s2 * y4[i].s2; + sumf += (float) x4[i].s3 * y4[i].s3; + } + + float all_sum = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) { + all_sum += (float) x[i] * y[i]; + } + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl new file mode 100644 index 0000000..d50bd1f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl @@ -0,0 +1,189 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK_MXFP4 32 +typedef struct { + uchar e; // E8M0 + uchar qs[QK_MXFP4/2]; +} block_mxfp4; + +constant static float kvalues_mxfp4_f[16] = { + 0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f +}; + +static inline float e8m0_to_fp32(uchar x) { + int bits; + + if (x == 0) { + bits = 0x00400000; + } else { + bits = (uint) x << 23; + } + + return as_float(bits); +} + +#ifdef INTEL_GPU +#define N_R0_MXFP4 2 // number of rows each subgroup works on +#define N_SG_MXFP4 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_MXFP4 2 +#define N_SG_MXFP4 2 +#define N_SIMDWIDTH 64 +#endif + +inline void mul_mv_mxfp4_f32( + global char * src0, + global char * src1, + global char * dst, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3, + local char * shmem +) { + local float * shmem_f32 = (local float *) shmem; + int nb = ne00/QK_MXFP4; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = 0; + + int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4; + + uint i12 = im%ne12; + uint i13 = im/ne12; + + ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0); + global float * y = (global float *) (src1 + offset_src1); + + const short ix = get_sub_group_local_id()/2; // 0...15 + const short it = get_sub_group_local_id()%2; // 0 or 1 + + shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16]; + barrier(CLK_LOCAL_MEM_FENCE); + + float4 yl[4]; + float sumf[N_R0_MXFP4] = {0.f}; + + global float * yb = y + ix * QK_MXFP4 + it * 8; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + global float4 * y4 = (global float4 *)yb; + yl[0] = y4[0]; + yl[1] = y4[4]; + yl[2] = y4[1]; + yl[3] = y4[5]; + + for (short row = 0; row < N_R0_MXFP4; row++) { + global block_mxfp4 * xb = x + row*nb + ib; + global uchar * q2 = (global uchar *)(xb->qs + 8*it); + + float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] & 0x0F], shmem_f32[q2[1] & 0x0F], shmem_f32[q2[2] & 0x0F], shmem_f32[q2[3] & 0x0F]); + float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4 ], shmem_f32[q2[1] >> 4 ], shmem_f32[q2[2] >> 4 ], shmem_f32[q2[3] >> 4 ]); + float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] & 0x0F], shmem_f32[q2[5] & 0x0F], shmem_f32[q2[6] & 0x0F], shmem_f32[q2[7] & 0x0F]); + float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4 ], shmem_f32[q2[5] >> 4 ], shmem_f32[q2[6] >> 4 ], shmem_f32[q2[7] >> 4 ]); + + acc1 = (acc1 + acc3) + (acc2 + acc4); + + sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3)); + } + + yb += (N_SIMDWIDTH/2) * QK_MXFP4; + } + + global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0; + + for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) { + float sum_all = sub_group_reduce_add(sumf[row]); + if (get_sub_group_local_id() == 0) { + dst_f32[first_row + row] = sum_all; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_id_mxfp4_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne11, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne20, + int ne21, + ulong nb21, + int ne0, + int ne1, + int r2, + int r3, + local char * shmem +) { + src0 = (global char *)((global char *)src0 + offset0); + src1 = (global char *)((global char *)src1 + offset1); + src2 = (global char *)((global char *)src2 + offset2); + dst = (global char *)((global char *)dst + offsetd); + + const int iid1 = get_group_id(2)/ne20; + const int idx = get_group_id(2)%ne20; + + int i02 = ((global int *) (src2 + iid1*nb21))[idx]; + + int i11 = idx % ne11; + int i12 = iid1; + + int i1 = idx; + int i2 = i12; + + global char * src0_cur = src0 + i02*nb02; + global char * src1_cur = src1 + i11*nb11 + i12*nb12; + + global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float); + + mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur, + ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl new file mode 100644 index 0000000..f65e86e --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl @@ -0,0 +1,176 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK_MXFP4 32 + +static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) { + ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b; + fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00; + fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00; + fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00; + fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800; + bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800; + bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800; + bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800; + + fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo; + fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi; + fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo; + fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi; + + sign_a.lo = (fp4x4 << 12) & 0x8000; + sign_a.hi = (fp4x4 << 8) & 0x8000; + sign_b.lo = (fp4x4 << 4) & 0x8000; + sign_b.hi = fp4x4 & 0x8000; + + fp16_packed_a = sign_a + bias_a + fp16_packed_a; + fp16_packed_b = sign_b + bias_b + fp16_packed_b; + + return as_half4((ushort4)(fp16_packed_a, fp16_packed_b)); +} + +static inline float e8m0_to_fp32(uchar x) { + int bits; + bits = (x == 0) ? 0x00400000 : ((uint) x << 23); + return as_float(bits); +} + +#ifdef INTEL_GPU +#define N_R0_MXFP4 2 // number of rows each subgroup works on +#define N_SG_MXFP4 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_MXFP4 4 +#define N_SG_MXFP4 1 +#define N_SIMDWIDTH 64 +#define SRC0Q_IMG +#endif + +kernel void kernel_mul_mv_id_mxfp4_f32_flat( +#ifdef SRC0Q_IMG + __read_only image1d_buffer_t src0_q, +#else + global uchar * src0_q, +#endif + global uchar * src0_e, + global uchar * src1, + ulong offset1, + global uchar * src2, + ulong offset2, + global uchar * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne11, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne20, + int ne21, + ulong nb21, + int ne0, + int ne1, + int r2, + int r3 +) { + dst = dst + offsetd; + + const int iid1 = get_group_id(2) / ne20; + const int idx = get_group_id(2) % ne20; + + uint i02 = ((global uint *) (src2 + offset2 + iid1 * nb21))[idx]; + + int i11 = idx % ne11; + + int nb = ne00 / QK_MXFP4; + + uint src0_off = i02*nb02; + src0_off /= 17; // 17 = sizeof(block_mxfp4) + + src0_e = src0_e + src0_off; + + dst = dst + (idx * ne0 + iid1 * ne1 * ne0) * sizeof(float); + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + + int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4; + + uint offset_src0 = first_row*nb01; + offset_src0 /= 17; // 17 = sizeof(block_mxfp4) +#ifdef SRC0Q_IMG + ulong offset_q = src0_off + offset_src0; +#else + src0_q = src0_q + src0_off*16; + global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0; +#endif + global uchar * x_e = src0_e + offset_src0; + + const short ix = get_sub_group_local_id() >> 1; + const short it = get_sub_group_local_id() & 1; + + float sumf[N_R0_MXFP4] = {0.f}; + + src1 = src1 + offset1 + i11 * nb11 + iid1 * nb12; + global float * y = (global float *) (src1 + r1 * nb11); + global float * yb = y + ix * QK_MXFP4 + it * 8; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH / 2) { + global float4 * y4 = (global float4 *)yb; + + #pragma unroll + for (short row = 0; row < N_R0_MXFP4; row++) { + uchar xb_e = x_e[row * nb + ib]; +#ifdef SRC0Q_IMG + ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy); +#else + ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it)); +#endif + + half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0); + half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1); + float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2); + acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3); + + fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2); + fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3); + acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2); + acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3); + + sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3)); + } + + yb += (N_SIMDWIDTH / 2) * QK_MXFP4; + } + + global float * dst_f32 = (global float *)dst + (ulong)r1 * ne0; + + for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) { + float sum_all = sub_group_reduce_add(sumf[row]); + if (get_sub_group_local_id() == 0) { + dst_f32[first_row + row] = sum_all; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl new file mode 100644 index 0000000..7ccf41e --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl @@ -0,0 +1,283 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +// This function requires the original shuffled weights. +// As a reminder, the original weights are shuffled so that (q[0], q[16]) are +// packed together in a byte, so are (q[1], q[17]) and so on. +inline float block_q_4_0_dot_y_flat( + global uchar * x, + global half * dh, + float sumy, + float16 yl, + int il +) { + float d = *dh; + global ushort * qs = ((global ushort *)x + il/2); + float acc = 0.f; + + acc += yl.s0 * (qs[0] & 0x000F); + acc += yl.s1 * (qs[0] & 0x0F00); + acc += yl.s8 * (qs[0] & 0x00F0); + acc += yl.s9 * (qs[0] & 0xF000); + + acc += yl.s2 * (qs[1] & 0x000F); + acc += yl.s3 * (qs[1] & 0x0F00); + acc += yl.sa * (qs[1] & 0x00F0); + acc += yl.sb * (qs[1] & 0xF000); + + acc += yl.s4 * (qs[2] & 0x000F); + acc += yl.s5 * (qs[2] & 0x0F00); + acc += yl.sc * (qs[2] & 0x00F0); + acc += yl.sd * (qs[2] & 0xF000); + + acc += yl.s6 * (qs[3] & 0x000F); + acc += yl.s7 * (qs[3] & 0x0F00); + acc += yl.se * (qs[3] & 0x00F0); + acc += yl.sf * (qs[3] & 0xF000); + + return d * (sumy * -8.f + acc); +} + +// +// This variant outputs 8 values. +// +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 8 // each SIMD group works on 8 rows +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_DST 8 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif + +inline void mul_vec_q_n_f32_8x_flat( + global char * src0_q, + global half * src0_d, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + const ulong nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = 0; + + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + // The number of scales is the same as the number of blocks. + ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + // Each block contains QK4_0/2 uchars, hence offset for qs is as follows. + ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2; + + global uchar * x = (global uchar *) src0_q + offset0_q; + global half * d = (global half *) src0_d + offset0_d; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float16 yl; + float8 sumf = 0.f; + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix*QK4_0 + il; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0.f; + + sumy += yb[0]; + sumy += yb[1]; + sumy += yb[2]; + sumy += yb[3]; + sumy += yb[4]; + sumy += yb[5]; + sumy += yb[6]; + sumy += yb[7]; + + sumy += yb[16]; + sumy += yb[17]; + sumy += yb[18]; + sumy += yb[19]; + sumy += yb[20]; + sumy += yb[21]; + sumy += yb[22]; + sumy += yb[23]; + + yl.s0 = yb[0]; + yl.s1 = yb[1]/256.f; + + yl.s2 = yb[2]; + yl.s3 = yb[3]/256.f; + + yl.s4 = yb[4]; + yl.s5 = yb[5]/256.f; + + yl.s6 = yb[6]; + yl.s7 = yb[7]/256.f; + + yl.s8 = yb[16]/16.f; + yl.s9 = yb[17]/4096.f; + + yl.sa = yb[18]/16.f; + yl.sb = yb[19]/4096.f; + + yl.sc = yb[20]/16.f; + yl.sd = yb[21]/4096.f; + + yl.se = yb[22]/16.f; + yl.sf = yb[23]/4096.f; + + sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il); + sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il); + sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il); + sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il); + + sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il); + sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il); + sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il); + sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il); + + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + float8 tot = (float8)( + sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3), + sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5), + sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + + if (first_row + 4 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4; + } + if (first_row + 5 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5; + } + if (first_row + 6 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6; + } + if (first_row + 7 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_id_q4_0_f32_8x_flat( + global char * src0_q, + global half * src0_d, + global float * src1, + ulong offset1, + global char * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + ulong nb00, + ulong nb02, + int ne10, + int ne11, + int ne12, + ulong nb11, + ulong nb12, + int ne20, + int ne21, + ulong nb21, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float *)((global char *)src1 + offset1); + src2 = (global char *)((global char *)src2 + offset2); + dst = (global float *)((global char *)dst + offsetd); + + const int iid1 = get_group_id(2)/ne20; + const int idx = get_group_id(2)%ne20; + + const int i02 = ((global int *)(src2 + iid1*nb21))[idx]; + + const int i11 = idx%ne11; + const int i12 = iid1; + + const int i1 = idx; + const int i2 = i12; + + global char * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2); + global half * src0_d_cur = src0_d + (i02*nb02/nb00); + global float * src1_cur = (global float *)((global char *) src1 + i11*nb11 + i12*nb12); + global float * dst_cur = dst + i1*ne0 + i2*ne1*ne0; + + mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl new file mode 100644 index 0000000..f37e83e --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl @@ -0,0 +1,140 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK8_0 32 +typedef struct { + half d; // delta + char qs[QK8_0]; // quants +} block_q8_0; + +#define NB_Q8_0 8 + +#ifdef INTEL_GPU +#define N_R0_Q8_0 4 // number of rows each subgroup works on +#define N_SG_Q8_0 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_Q8_0 4 +#define N_SG_Q8_0 2 +#define N_SIMDWIDTH 64 +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_id_q8_0_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + ulong nb01, + ulong nb02, + int ne11, + int ne12, + ulong nb11, + ulong nb12, + int ne20, + int ne21, + ulong nb21, + int ne0, + int ne1 +) { + src0 = (global char *)((global char *)src0 + offset0); + src1 = (global char *)((global char *)src1 + offset1); + src2 = (global char *)((global char *)src2 + offset2); + dst = (global char *)((global char *)dst + offsetd); + + int iid1 = get_group_id(2)/ne20; + int idx = get_group_id(2)%ne20; + + int i02 = ((global int *) (src2 + iid1*nb21))[idx]; + + int i11_ = idx % ne11; + int i12_ = iid1; + + int i1 = idx; + int i2 = i12_; + + global char * src0_cur = src0 + i02*nb02; + global char * src1_cur = src1 + i11_*nb11 + i12_*nb12; + + global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float); + + int nb = ne00/QK8_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + + int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0; + + ulong offset_src1 = r1*nb11; + global float * y = (global float *) (src1_cur + offset_src1); + + // pointers to src0 rows + global block_q8_0 * ax[N_R0_Q8_0]; + for (int row = 0; row < N_R0_Q8_0; ++row) { + ulong offset_src0 = (first_row + row)*nb01; + ax[row] = (global block_q8_0 *) ((global char *) src0_cur + offset_src0); + } + + float yl[NB_Q8_0]; + float sumf[N_R0_Q8_0] = { 0.f }; + + const short ix = get_sub_group_local_id()/4; + const short il = get_sub_group_local_id()%4; + + global float * yb = y + ix*QK8_0 + il*NB_Q8_0; + + // each thread handles NB_Q8_0 quants at a time + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) { + for (short i = 0; i < NB_Q8_0; ++i) { + yl[i] = yb[i]; + } + + for (short row = 0; row < N_R0_Q8_0; row++) { + global char * qs = ax[row][ib].qs + il*NB_Q8_0; + float sumq = 0.f; + for (short iq = 0; iq < NB_Q8_0; ++iq) { + sumq += qs[iq] * yl[iq]; + } + sumf[row] += sumq*ax[row][ib].d; + } + + yb += N_SIMDWIDTH*NB_Q8_0; + } + + global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0; + + for (int row = 0; row < N_R0_Q8_0; ++row) { + float tot = sub_group_reduce_add(sumf[row]); + + if (get_sub_group_local_id() == 0 && first_row + row < ne01) { + dst_f32[first_row + row] = tot; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl new file mode 100644 index 0000000..fd3a071 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl @@ -0,0 +1,222 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK8_0 32 +typedef struct { + half d; // delta + char qs[QK8_0]; // quants +} block_q8_0; + +#define NB_Q8_0 8 + +#ifdef INTEL_GPU +#define N_R0_Q8_0 4 // number of rows each subgroup works on +#define N_SG_Q8_0 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_Q8_0 4 +#define N_SG_Q8_0 2 +#define N_SIMDWIDTH 64 +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_id_q8_0_f32_flat( + global char * src0_q, + global half * src0_d, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + ulong nb01, + ulong nb02, + int ne11, + int ne12, + ulong nb11, + ulong nb12, + int ne20, + int ne21, + ulong nb21, + int ne0, + int ne1 +) { + src1 = (global char *)((global char *)src1 + offset1); + src2 = (global char *)((global char *)src2 + offset2); + dst = (global char *)((global char *)dst + offsetd); + + int iid1 = (int)get_group_id(2)/ne20; + int idx = (int)get_group_id(2)%ne20; + + int i02 = ((global int *) (src2 + iid1*nb21))[idx]; + + int i11_ = idx % ne11; + int i12_ = iid1; + + int i1 = idx; + int i2 = i12_; + + // 34 == sizeof(block_q8_0) + uint src0_off = i02*nb02; + src0_off /= 34; + + global char * src0_q_cur = src0_q + src0_off*sizeof(char)*QK8_0; + global half * src0_d_cur = src0_d + src0_off; + global char * src1_cur = src1 + i11_*nb11 + i12_*nb12; + + global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float); + + int nb = ne00/QK8_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + + int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0; + + ulong offset_src1 = r1*nb11; + global float * y = (global float *) (src1_cur + offset_src1); + + // pointers to src0 rows + uint offset_src0_base = first_row*nb01; + + global char * ax0, * ax1, * ax2, * ax3; + global half * ad0, * ad1, * ad2, * ad3; + uint offset_src0; + + offset_src0 = offset_src0_base + 0*nb01; + offset_src0 = offset_src0/34; + ax0 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0); + ad0 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 1*nb01; + offset_src0 = offset_src0/34; + ax1 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0); + ad1 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 2*nb01; + offset_src0 = offset_src0/34; + ax2 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0); + ad2 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 3*nb01; + offset_src0 = offset_src0/34; + ax3 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0); + ad3 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half)); + + const short ix = get_sub_group_local_id()/4; + const short il = get_sub_group_local_id()%4; + + global float * yb = y + ix*QK8_0 + il*NB_Q8_0; + + float8 yl; + float8 qv; + float4 sumf = 0.f; + float sumq = 0.f; + global char * qs; + + // each thread handles NB_Q8_0 quants at a time + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) { + yl = vload8(0, yb); + + qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s0 += sumq*ad0[ib]; + + qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s1 += sumq*ad1[ib]; + + qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s2 += sumq*ad2[ib]; + + qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s3 += sumq*ad3[ib]; + + yb += N_SIMDWIDTH*NB_Q8_0; + } + + global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0; + + float4 tot = (float4)( + sub_group_reduce_add(sumf.s0), + sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), + sub_group_reduce_add(sumf.s3) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst_f32[first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst_f32[first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst_f32[first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst_f32[first_row + 3] = tot.s3; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl new file mode 100644 index 0000000..9a4d4b9 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl @@ -0,0 +1,144 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK_MXFP4 32 +typedef struct { + uchar e; // E8M0 + uchar qs[QK_MXFP4/2]; +} block_mxfp4; + +constant static float kvalues_mxfp4_f[16] = { + 0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f +}; + +static inline float e8m0_to_fp32(uchar x) { + int bits; + + if (x == 0) { + bits = 0x00400000; + } else { + bits = (uint) x << 23; + } + + return as_float(bits); +} + +#ifdef INTEL_GPU +#define N_R0_MXFP4 2 // number of rows each subgroup works on +#define N_SG_MXFP4 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_MXFP4 2 +#define N_SG_MXFP4 2 +#define N_SIMDWIDTH 64 +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_mxfp4_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3, + local char * shmem +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + local float * shmem_f32 = (local float *) shmem; + int nb = ne00/QK_MXFP4; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4; + + uint i12 = im%ne12; + uint i13 = im/ne12; + + ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + ulong offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0); + global float * y = (global float *) (src1 + offset_src1); + + const short ix = get_sub_group_local_id()/2; // 0...15 + const short it = get_sub_group_local_id()%2; // 0 or 1 + + shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16]; + barrier(CLK_LOCAL_MEM_FENCE); + + float4 yl[4]; + float sumf[N_R0_MXFP4] = {0.f}; + + global float * yb = y + ix * QK_MXFP4 + it * 8; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + global float4 * y4 = (global float4 *)yb; + yl[0] = y4[0]; + yl[1] = y4[4]; + yl[2] = y4[1]; + yl[3] = y4[5]; + + for (short row = 0; row < N_R0_MXFP4; row++) { + global block_mxfp4 * xb = x + row*nb + ib; + global uchar * q2 = (global uchar *)(xb->qs + 8*it); + + float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] & 0x0F], shmem_f32[q2[1] & 0x0F], shmem_f32[q2[2] & 0x0F], shmem_f32[q2[3] & 0x0F]); + float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4 ], shmem_f32[q2[1] >> 4 ], shmem_f32[q2[2] >> 4 ], shmem_f32[q2[3] >> 4 ]); + float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] & 0x0F], shmem_f32[q2[5] & 0x0F], shmem_f32[q2[6] & 0x0F], shmem_f32[q2[7] & 0x0F]); + float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4 ], shmem_f32[q2[5] >> 4 ], shmem_f32[q2[6] >> 4 ], shmem_f32[q2[7] >> 4 ]); + + acc1 = (acc1 + acc3) + (acc2 + acc4); + + sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3)); + } + + yb += (N_SIMDWIDTH/2) * QK_MXFP4; + } + + global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0; + + for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) { + float sum_all = sub_group_reduce_add(sumf[row]); + if (get_sub_group_local_id() == 0) { + dst_f32[first_row + row] = sum_all; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl new file mode 100644 index 0000000..3d5a923 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl @@ -0,0 +1,167 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK_MXFP4 32 + +static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) { + ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b; + fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00; + fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00; + fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00; + fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00; + + bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800; + bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800; + bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800; + bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800; + + fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo; + fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi; + fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo; + fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi; + + sign_a.lo = (fp4x4 << 12) & 0x8000; + sign_a.hi = (fp4x4 << 8) & 0x8000; + sign_b.lo = (fp4x4 << 4) & 0x8000; + sign_b.hi = fp4x4 & 0x8000; + + fp16_packed_a = sign_a + bias_a + fp16_packed_a; + fp16_packed_b = sign_b + bias_b + fp16_packed_b; + + return as_half4((ushort4)(fp16_packed_a, fp16_packed_b)); +} + +static inline float e8m0_to_fp32(uchar x) { + int bits; + bits = (x == 0) ? 0x00400000 : ((uint) x << 23); + return as_float(bits); +} + +#ifdef INTEL_GPU +#define N_R0_MXFP4 2 // number of rows each subgroup works on +#define N_SG_MXFP4 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_MXFP4 2 +#define N_SG_MXFP4 2 +#define N_SIMDWIDTH 64 +#define SRC0Q_IMG +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_mxfp4_f32_flat( +#ifdef SRC0Q_IMG + __read_only image1d_buffer_t src0_q, +#else + global uchar * src0_q, +#endif + global uchar * src0_e, + global uchar * src1, + ulong offset1, + global uchar * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = src1 + offset1; + dst = dst + offsetd; + + int nb = ne00 / QK_MXFP4; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4; + + uint i12 = im % ne12; + uint i13 = im / ne12; + + uint offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + // 17 = sizeof(block_mxfp4) + offset_src0 /= 17; +#ifdef SRC0Q_IMG + ulong offset_q = offset_src0; +#else + global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0; +#endif + global uchar * x_e = src0_e + offset_src0; + + ulong offset_src1 = r1 * nb11 + i12 * nb12 + i13 * nb13; + global float * y = (global float *)(src1 + offset_src1); + + const short ix = get_sub_group_local_id() >> 1; // 0...15 + const short it = get_sub_group_local_id() & 1; // 0 or 1 + + float sumf[N_R0_MXFP4] = {0.f}; + + global float * yb = y + ix * QK_MXFP4 + it * 8; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + global float4 * y4 = (global float4 *)yb; + + #pragma unroll + for (short row = 0; row < N_R0_MXFP4; row++) { + uchar xb_e = x_e[row * nb + ib]; +#ifdef SRC0Q_IMG + ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy); +#else + ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it)); +#endif + + half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0); + half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1); + float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2); + acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3); + + fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2); + fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3); + acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2); + acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3); + + sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3)); + } + + yb += (N_SIMDWIDTH/2) * QK_MXFP4; + } + + global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0; + + for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) { + float sum_all = sub_group_reduce_add(sumf[row]); + if (get_sub_group_local_id() == 0) { + dst_f32[first_row + row] = sum_all; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl new file mode 100644 index 0000000..52141e0 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl @@ -0,0 +1,192 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +//------------------------------------------------------------------------------ +// mul_vec_q_n_f32 +//------------------------------------------------------------------------------ +// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i]) +// il indicates where the q4 quants begin (0 or QK4_0/4) +// we assume that the yl's have been multiplied with the appropriate scale factor +// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096) +inline float block_q_4_0_dot_y( + global struct block_q4_0 * qb_curr, + float sumy, + private float * yl, + int il +) { + float d = qb_curr->d; + float2 acc = 0.f; + global ushort * qs = ((global ushort *)qb_curr + 1 + il/2); + for (int i = 0; i < 8; i+=2) { + acc.s0 += yl[i + 0] * (qs[i / 2] & 0x000F) + + yl[i + 1] * (qs[i / 2] & 0x0F00); + acc.s1 += yl[i + 8] * (qs[i / 2] & 0x00F0) + + yl[i + 9] * (qs[i / 2] & 0xF000); + } + return d * (sumy * -8.f + acc.s0 + acc.s1); +} + +#ifdef INTEL_GPU +#define N_DST 4 // each SIMD group works on 4 rows +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // assuming SIMD group size is 16 +#elif defined (ADRENO_GPU) +#define N_DST 4 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif + +inline void mul_vec_q_n_f32( + global void * src0, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + + const ulong nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global + // id of a SIMD group in the grid. + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + + global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float yl[16]; // src1 vector cache + float sumf[N_DST]={0.f}; + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix * QK4_0 + il; + + // each thread in a SIMD group deals with half a block. + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0; + for (int i = 0; i < 8; i += 2) { + sumy += yb[i] + yb[i+1]; + yl[i+0] = yb[i+ 0]; + yl[i+1] = yb[i+ 1]/256.f; + sumy += yb[i+16] + yb[i+17]; + yl[i+8] = yb[i+16]/16.f; + yl[i+9] = yb[i+17]/4096.f; + } + + for (int row = 0; row < N_DST; row++) { + sumf[row] += block_q_4_0_dot_y(x+ib+row*nb, sumy, yl, il); + } + + // One thread in a SIMD group (i.e., subgroup) handles a half block, + // hence then entire SIMD group handles SIMDWIDTH/2 blocks. + // y points to the activation matrix (of type float). Therefore for + // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because + // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of + // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size. + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + // The above does not work for Adreno - it produces incorrect results for + // row = 1, 2, 3 and only row = 0 gives the correct result. + // If N_DST is changed, the below array must be initialized accordingly. + // This also seems to perform better on Intel. + float tot[N_DST] = { + sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]), + sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])}; + for (int row = 0; row < N_DST; ++row) { + if (get_sub_group_local_id() == 0 && first_row + row < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row]; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_q4_0_f32( + global void * src0, + ulong offset0, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl new file mode 100644 index 0000000..3eebab8 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl @@ -0,0 +1,307 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +inline float mm_block_q_4_0_dot_y_flat( + global uchar * x, + global half * dh, + float sumy, + float16 yl, + int il +) { + float d = *dh; + global ushort * qs = ((global ushort *)x + il/2); + float acc = 0.f; + + acc += yl.s0 * (qs[0] & 0x000F); + acc += yl.s1 * (qs[0] & 0x0F00); + acc += yl.s8 * (qs[0] & 0x00F0); + acc += yl.s9 * (qs[0] & 0xF000); + + acc += yl.s2 * (qs[1] & 0x000F); + acc += yl.s3 * (qs[1] & 0x0F00); + acc += yl.sa * (qs[1] & 0x00F0); + acc += yl.sb * (qs[1] & 0xF000); + + acc += yl.s4 * (qs[2] & 0x000F); + acc += yl.s5 * (qs[2] & 0x0F00); + acc += yl.sc * (qs[2] & 0x00F0); + acc += yl.sd * (qs[2] & 0xF000); + + acc += yl.s6 * (qs[3] & 0x000F); + acc += yl.s7 * (qs[3] & 0x0F00); + acc += yl.se * (qs[3] & 0x00F0); + acc += yl.sf * (qs[3] & 0xF000); + + return d * (sumy * -8.f + acc); +} + +#ifdef INTEL_GPU +#define N_DST 16 // each SIMD group works on 8 rows (in weights matrix) +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // assuming SIMD group size is 16 +#elif defined (ADRENO_GPU) +#define N_DST 16 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif +// +// This variant performs 1d blocking with 16x output. +// Eeach simdgroup outputs 16 values on `n0` dim (row in the output matrix). +// +inline void mul_mat_q_n_f32_1d_16x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + const int nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of + // a SIMD group in the grid. Each SIMD group produces N_DST values in the + // result, hence uses nb blocks, i.e., the offset becomes first_row*nb. + // Currently with llama2 7B, im is always 0. + // TODO: how to handle im/gqa*(nb*ne0)? + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + // The number of scales is the same as the number of blocks. + ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + // Each block contains QK4_0/2 uchars, hence offset for qs is as follows. + ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2; + + global uchar * x = (global uchar *) src0_q + offset0_q; + global half * d = (global half *) src0_d + offset0_d; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float16 yl; + float16 sumf = (float16)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, + 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f); + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix*QK4_0 + il; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0.f; + + sumy += yb[0]; + sumy += yb[1]; + sumy += yb[2]; + sumy += yb[3]; + sumy += yb[4]; + sumy += yb[5]; + sumy += yb[6]; + sumy += yb[7]; + + sumy += yb[16]; + sumy += yb[17]; + sumy += yb[18]; + sumy += yb[19]; + sumy += yb[20]; + sumy += yb[21]; + sumy += yb[22]; + sumy += yb[23]; + + yl.s0 = yb[0]; + yl.s1 = yb[1]/256.f; + + yl.s2 = yb[2]; + yl.s3 = yb[3]/256.f; + + yl.s4 = yb[4]; + yl.s5 = yb[5]/256.f; + + yl.s6 = yb[6]; + yl.s7 = yb[7]/256.f; + + yl.s8 = yb[16]/16.f; + yl.s9 = yb[17]/4096.f; + + yl.sa = yb[18]/16.f; + yl.sb = yb[19]/4096.f; + + yl.sc = yb[20]/16.f; + yl.sd = yb[21]/4096.f; + + yl.se = yb[22]/16.f; + yl.sf = yb[23]/4096.f; + + sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il); + sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il); + sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il); + sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il); + + sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il); + sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il); + sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il); + sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il); + + sumf.s8 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 8*nb*QK4_0/2, d + ib + 8*nb, sumy, yl, il); + sumf.s9 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 9*nb*QK4_0/2, d + ib + 9*nb, sumy, yl, il); + sumf.sa += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 10*nb*QK4_0/2, d + ib + 10*nb, sumy, yl, il); + sumf.sb += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 11*nb*QK4_0/2, d + ib + 11*nb, sumy, yl, il); + + sumf.sc += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 12*nb*QK4_0/2, d + ib + 12*nb, sumy, yl, il); + sumf.sd += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 13*nb*QK4_0/2, d + ib + 13*nb, sumy, yl, il); + sumf.se += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 14*nb*QK4_0/2, d + ib + 14*nb, sumy, yl, il); + sumf.sf += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 15*nb*QK4_0/2, d + ib + 15*nb, sumy, yl, il); + + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + float16 tot = (float16)( + sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3), + sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5), + sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7), + + sub_group_reduce_add(sumf.s8), sub_group_reduce_add(sumf.s9), + sub_group_reduce_add(sumf.sa), sub_group_reduce_add(sumf.sb), + sub_group_reduce_add(sumf.sc), sub_group_reduce_add(sumf.sd), + sub_group_reduce_add(sumf.se), sub_group_reduce_add(sumf.sf) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + + if (first_row + 4 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4; + } + if (first_row + 5 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5; + } + if (first_row + 6 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6; + } + if (first_row + 7 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7; + } + + if (first_row + 8 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 8] = tot.s8; + } + if (first_row + 9 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 9] = tot.s9; + } + if (first_row + 10 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 10] = tot.sa; + } + if (first_row + 11 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 11] = tot.sb; + } + + if (first_row + 12 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 12] = tot.sc; + } + if (first_row + 13 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 13] = tot.sd; + } + if (first_row + 14 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 14] = tot.se; + } + if (first_row + 15 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 15] = tot.sf; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_q4_0_f32_1d_16x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + mul_mat_q_n_f32_1d_16x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl new file mode 100644 index 0000000..38024d0 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl @@ -0,0 +1,265 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +inline float mm_block_q_4_0_dot_y_flat( + global uchar * x, + global half * dh, + float sumy, + float16 yl, + int il +) { + float d = *dh; + global ushort * qs = ((global ushort *)x + il/2); + float acc = 0.f; + + acc += yl.s0 * (qs[0] & 0x000F); + acc += yl.s1 * (qs[0] & 0x0F00); + acc += yl.s8 * (qs[0] & 0x00F0); + acc += yl.s9 * (qs[0] & 0xF000); + + acc += yl.s2 * (qs[1] & 0x000F); + acc += yl.s3 * (qs[1] & 0x0F00); + acc += yl.sa * (qs[1] & 0x00F0); + acc += yl.sb * (qs[1] & 0xF000); + + acc += yl.s4 * (qs[2] & 0x000F); + acc += yl.s5 * (qs[2] & 0x0F00); + acc += yl.sc * (qs[2] & 0x00F0); + acc += yl.sd * (qs[2] & 0xF000); + + acc += yl.s6 * (qs[3] & 0x000F); + acc += yl.s7 * (qs[3] & 0x0F00); + acc += yl.se * (qs[3] & 0x00F0); + acc += yl.sf * (qs[3] & 0xF000); + + return d * (sumy * -8.f + acc); +} + +#ifdef INTEL_GPU +#define N_DST 8 // each SIMD group works on 8 rows (in weights matrix) +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // assuming SIMD group size is 16 +#elif defined (ADRENO_GPU) +#define N_DST 8 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif +// +// This variant performs 1d blocking with 8x output. +// Eeach simdgroup outputs 8 values on `n0` dim (row in the output matrix). +// +inline void mul_mat_q_n_f32_1d_8x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + const int nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of + // a SIMD group in the grid. Each SIMD group produces N_DST values in the + // result, hence uses nb blocks, i.e., the offset becomes first_row*nb. + // Currently with llama2 7B, im is always 0. + // TODO: how to handle im/gqa*(nb*ne0)? + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + // The number of scales is the same as the number of blocks. + ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + // Each block contains QK4_0/2 uchars, hence offset for qs is as follows. + ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2; + + global uchar * x = (global uchar *) src0_q + offset0_q; + global half * d = (global half *) src0_d + offset0_d; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float16 yl; + float8 sumf = (float8)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f); + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix*QK4_0 + il; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0.f; + + sumy += yb[0]; + sumy += yb[1]; + sumy += yb[2]; + sumy += yb[3]; + sumy += yb[4]; + sumy += yb[5]; + sumy += yb[6]; + sumy += yb[7]; + + sumy += yb[16]; + sumy += yb[17]; + sumy += yb[18]; + sumy += yb[19]; + sumy += yb[20]; + sumy += yb[21]; + sumy += yb[22]; + sumy += yb[23]; + + yl.s0 = yb[0]; + yl.s1 = yb[1]/256.f; + + yl.s2 = yb[2]; + yl.s3 = yb[3]/256.f; + + yl.s4 = yb[4]; + yl.s5 = yb[5]/256.f; + + yl.s6 = yb[6]; + yl.s7 = yb[7]/256.f; + + yl.s8 = yb[16]/16.f; + yl.s9 = yb[17]/4096.f; + + yl.sa = yb[18]/16.f; + yl.sb = yb[19]/4096.f; + + yl.sc = yb[20]/16.f; + yl.sd = yb[21]/4096.f; + + yl.se = yb[22]/16.f; + yl.sf = yb[23]/4096.f; + + sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il); + sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il); + sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il); + sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il); + + sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il); + sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il); + sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il); + sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il); + + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + float8 tot = (float8)( + sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3), + sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5), + sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + + if (first_row + 4 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4; + } + if (first_row + 5 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5; + } + if (first_row + 6 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6; + } + if (first_row + 7 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_q4_0_f32_1d_8x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + mul_mat_q_n_f32_1d_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl new file mode 100644 index 0000000..aed1ce7 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl @@ -0,0 +1,272 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +// This function requires the original shuffled weights. +// As a reminder, the original weights are shuffled so that (q[0], q[16]) are +// packed together in a byte, so are (q[1], q[17]) and so on. +inline float block_q_4_0_dot_y_flat( + global uchar * x, + global half * dh, + float sumy, + float16 yl, + int il +) { + float d = *dh; + global ushort * qs = ((global ushort *)x + il/2); + float acc = 0.f; + + acc += yl.s0 * (qs[0] & 0x000F); + acc += yl.s1 * (qs[0] & 0x0F00); + acc += yl.s8 * (qs[0] & 0x00F0); + acc += yl.s9 * (qs[0] & 0xF000); + + acc += yl.s2 * (qs[1] & 0x000F); + acc += yl.s3 * (qs[1] & 0x0F00); + acc += yl.sa * (qs[1] & 0x00F0); + acc += yl.sb * (qs[1] & 0xF000); + + acc += yl.s4 * (qs[2] & 0x000F); + acc += yl.s5 * (qs[2] & 0x0F00); + acc += yl.sc * (qs[2] & 0x00F0); + acc += yl.sd * (qs[2] & 0xF000); + + acc += yl.s6 * (qs[3] & 0x000F); + acc += yl.s7 * (qs[3] & 0x0F00); + acc += yl.se * (qs[3] & 0x00F0); + acc += yl.sf * (qs[3] & 0xF000); + + return d * (sumy * -8.f + acc); +} + +// +// This variant outputs 8 values. +// +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 8 // each SIMD group works on 8 rows +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // assuming SIMD group size is 32 +#elif defined (ADRENO_GPU) +#define N_DST 8 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif + +inline void mul_vec_q_n_f32_8x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + const ulong nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of + // a SIMD group in the grid. Each SIMD group produces N_DST values in the + // result, hence uses nb blocks, i.e., the offset becomes first_row*nb. + // Currently with llama2 7B, im is always 0. + // TODO: how to handle im/gqa*(nb*ne0)? + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + // The number of scales is the same as the number of blocks. + ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + // Each block contains QK4_0/2 uchars, hence offset for qs is as follows. + ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2; + + global uchar * x = (global uchar *) src0_q + offset0_q; + global half * d = (global half *) src0_d + offset0_d; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float16 yl; + float8 sumf = 0.f; + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix*QK4_0 + il; + + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0.f; + + sumy += yb[0]; + sumy += yb[1]; + sumy += yb[2]; + sumy += yb[3]; + sumy += yb[4]; + sumy += yb[5]; + sumy += yb[6]; + sumy += yb[7]; + + sumy += yb[16]; + sumy += yb[17]; + sumy += yb[18]; + sumy += yb[19]; + sumy += yb[20]; + sumy += yb[21]; + sumy += yb[22]; + sumy += yb[23]; + + yl.s0 = yb[0]; + yl.s1 = yb[1]/256.f; + + yl.s2 = yb[2]; + yl.s3 = yb[3]/256.f; + + yl.s4 = yb[4]; + yl.s5 = yb[5]/256.f; + + yl.s6 = yb[6]; + yl.s7 = yb[7]/256.f; + + yl.s8 = yb[16]/16.f; + yl.s9 = yb[17]/4096.f; + + yl.sa = yb[18]/16.f; + yl.sb = yb[19]/4096.f; + + yl.sc = yb[20]/16.f; + yl.sd = yb[21]/4096.f; + + yl.se = yb[22]/16.f; + yl.sf = yb[23]/4096.f; + + sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il); + sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il); + sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il); + sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il); + + sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il); + sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il); + sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il); + sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il); + + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + float8 tot = (float8)( + sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3), + sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5), + sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + + if (first_row + 4 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4; + } + if (first_row + 5 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5; + } + if (first_row + 6 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6; + } + if (first_row + 7 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_q4_0_f32_8x_flat( + global uchar * src0_q, + global half * src0_d, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + mul_vec_q_n_f32_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl new file mode 100644 index 0000000..9295521 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl @@ -0,0 +1,254 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q4_0 +//------------------------------------------------------------------------------ +struct block_q4_0 +{ + half d; + uint8_t qs[QK4_0 / 2]; +}; + +// +// This variant unrolls the loops and uses vector types instead of pointers. +// It improves performance on Adreno but not so much on Intel. +// +inline float block_q_4_0_dot_y_v( + global struct block_q4_0 * qb_curr, + float sumy, + float16 yl, + int il +) { + float d = qb_curr->d; + float acc = 0.f; + global ushort * qs = ((global ushort *)qb_curr + 1 + il/2); + + acc += yl.s0 * (qs[0] & 0x000F); + acc += yl.s1 * (qs[0] & 0x0F00); + acc += yl.s8 * (qs[0] & 0x00F0); + acc += yl.s9 * (qs[0] & 0xF000); + + acc += yl.s2 * (qs[1] & 0x000F); + acc += yl.s3 * (qs[1] & 0x0F00); + acc += yl.sa * (qs[1] & 0x00F0); + acc += yl.sb * (qs[1] & 0xF000); + + acc += yl.s4 * (qs[2] & 0x000F); + acc += yl.s5 * (qs[2] & 0x0F00); + acc += yl.sc * (qs[2] & 0x00F0); + acc += yl.sd * (qs[2] & 0xF000); + + acc += yl.s6 * (qs[3] & 0x000F); + acc += yl.s7 * (qs[3] & 0x0F00); + acc += yl.se * (qs[3] & 0x00F0); + acc += yl.sf * (qs[3] & 0xF000); + + return d * (sumy * -8.f + acc); +} + +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 4 // each SIMD group works on 4 rows +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // assuming SIMD group size is 16 +#elif defined (ADRENO_GPU) +#define N_DST 4 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif + +inline void mul_vec_q_n_f32_v( + global void * src0, + global float * src1, + global float * dst, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + const ulong nb = ne00/QK4_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global + // id of a SIMD group in the grid. + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + + global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0; + global float * y = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float16 yl; // src1 vector cache + float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f); + + int ix = get_sub_group_local_id()/2; + int il = 8*(get_sub_group_local_id()%2); + + global float * yb = y + ix * QK4_0 + il; + + // each thread in a SIMD group deals with half a block. + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) { + float sumy = 0; + + sumy += yb[0]; + sumy += yb[1]; + sumy += yb[2]; + sumy += yb[3]; + sumy += yb[4]; + sumy += yb[5]; + sumy += yb[6]; + sumy += yb[7]; + + sumy += yb[16]; + sumy += yb[17]; + sumy += yb[18]; + sumy += yb[19]; + sumy += yb[20]; + sumy += yb[21]; + sumy += yb[22]; + sumy += yb[23]; + + + yl.s0 = yb[0]; + yl.s1 = yb[1]/256.f; + + yl.s2 = yb[2]; + yl.s3 = yb[3]/256.f; + + yl.s4 = yb[4]; + yl.s5 = yb[5]/256.f; + + yl.s6 = yb[6]; + yl.s7 = yb[7]/256.f; + + yl.s8 = yb[16]/16.f; + yl.s9 = yb[17]/4096.f; + + yl.sa = yb[18]/16.f; + yl.sb = yb[19]/4096.f; + + yl.sc = yb[20]/16.f; + yl.sd = yb[21]/4096.f; + + yl.se = yb[22]/16.f; + yl.sf = yb[23]/4096.f; + + sumf.s0 += block_q_4_0_dot_y_v(x+ib+0*nb, sumy, yl, il); + sumf.s1 += block_q_4_0_dot_y_v(x+ib+1*nb, sumy, yl, il); + sumf.s2 += block_q_4_0_dot_y_v(x+ib+2*nb, sumy, yl, il); + sumf.s3 += block_q_4_0_dot_y_v(x+ib+3*nb, sumy, yl, il); + + // One thread in a SIMD group (i.e., subgroup) handles a half block, + // hence then entire SIMD group handles SIMDWIDTH/2 blocks. + // y points to the activation matrix (of type float). Therefore for + // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because + // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of + // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size. + yb += QK4_0 * (N_SIMDWIDTH/2); + } + + // The above does not work for Adreno - it produces incorrect results for + // row = 1, 2, 3 and only row = 0 gives the correct result. + // If N_DST is changed, the below array must be initialized accordingly. + // This also seems to perform better on Intel. + float4 tot = (float4)( + sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + } +} + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mat_q4_0_f32_v( + global void * src0, + ulong offset0, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + mul_vec_q_n_f32_v(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl new file mode 100644 index 0000000..71ab989 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl @@ -0,0 +1,180 @@ +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +//------------------------------------------------------------------------------ +// block_q4_K +//------------------------------------------------------------------------------ +#define QK_K 256 +#define K_SCALE_SIZE 12 + +// 8 blocks of 32 elements each +// weight is represented as x = a * q + b +typedef struct { + half d; // super-block scale for quantized scales + half dmin; // super-block scale for quantized mins + + uchar scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits + uchar qs[QK_K/2]; // 4-bit quants +} block_q4_K; + +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 4 // number of rows each SIMD group works on +#define N_SIMDGROUP 1 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // SIMD group size +#elif defined (ADRENO_GPU) +#define N_DST 4 +#define N_SIMDGROUP 1 +#define N_SIMDWIDTH 64 +#endif + +#undef BLOCK_STRIDE +// number of (super) blocks each subgroup processes +// each thread in a subgroup processes a block (32 weights) +#define BLOCK_STRIDE (N_SIMDWIDTH/8) + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q4_K_f32( + global char * src0, + int offset0, + global char * src1, + int offset1, + global char * dst, + int offsetd, + int ne00, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + ushort kmask1 = 0x3f3f; + ushort kmask2 = 0x0f0f; + ushort kmask3 = 0xc0c0; + + int ix = get_sub_group_local_id()/8; // super block index + int it = get_sub_group_local_id()%8; // block index (inside super block) + int iq = it/4; // 0 or 1 - first or second half of the super block + int ir = it%4; // 0...3 - block index in the half super block + + int nb = ne00/QK_K; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST; + + int i12 = im%ne12; + int i13 = im/ne12; + + int offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + int offset_src1 = r1*nb11 + (i12 )*nb12 + (i13 )*nb13; + + global block_q4_K * x = (global block_q4_K *) (src0 + offset_src0); + global float * y = (global float *) (src1 + offset_src1); + + float yl[16]; + float yh[16]; + float sumf[N_DST] = {0.f}; + float all_sum; + + global float * y4 = y + ix * QK_K + 64 * iq + 8 * ir; + + ushort sc16[4]; + uchar * sc8 = (uchar *)sc16; + + for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) { + float4 sumy = {0.f, 0.f, 0.f, 0.f}; + for (int i = 0; i < 8; ++i) { + yl[i+0] = y4[i+0]; + sumy.s0 += yl[i+0]; + + yl[i+8] = y4[i+32]; + sumy.s1 += yl[i+8]; + + yh[i+0] = y4[i+128]; + sumy.s2 += yh[i+0]; + + yh[i+8] = y4[i+160]; + sumy.s3 += yh[i+8]; + } + + global ushort * sc = (global ushort *)x[ib].scales + iq; + global ushort * q1 = (global ushort *)x[ib].qs + 16 * iq + 4 * ir; + global half * dh = &x[ib].d; + + for (int row = 0; row < N_DST; row++) { + sc16[0] = sc[0] & kmask1; + sc16[1] = sc[2] & kmask1; + sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2); + sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2); + + global ushort * q2 = q1 + 32; + + float4 acc1 = {0.f, 0.f, 0.f, 0.f}; + float4 acc2 = {0.f, 0.f, 0.f, 0.f}; + for (int i = 0; i < 8; i += 2) { + acc1.s0 += yl[i+0] * (q1[i/2] & 0x000F); + acc1.s1 += yl[i+1] * (q1[i/2] & 0x0F00); + acc1.s2 += yl[i+8] * (q1[i/2] & 0x00F0); + acc1.s3 += yl[i+9] * (q1[i/2] & 0xF000); + acc2.s0 += yh[i+0] * (q2[i/2] & 0x000F); + acc2.s1 += yh[i+1] * (q2[i/2] & 0x0F00); + acc2.s2 += yh[i+8] * (q2[i/2] & 0x00F0); + acc2.s3 += yh[i+9] * (q2[i/2] & 0xF000); + } + + float dall = dh[0]; + float dmin = dh[1]; + sumf[row] += dall * ((acc1.s0 + 1.f/256.f * acc1.s1) * sc8[0] + + (acc1.s2 + 1.f/256.f * acc1.s3) * sc8[1] * 1.f/16.f + + (acc2.s0 + 1.f/256.f * acc2.s1) * sc8[4] + + (acc2.s2 + 1.f/256.f * acc2.s3) * sc8[5] * 1.f/16.f) - + dmin * (sumy.s0 * sc8[2] + sumy.s1 * sc8[3] + sumy.s2 * sc8[6] + sumy.s3 * sc8[7]); + + q1 += nb01/2; + sc += nb01/2; + dh += nb01/2; + } + + y4 += BLOCK_STRIDE * QK_K; + } + + global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0; + + for (int row = 0; row < N_DST; ++row) { + all_sum = sub_group_reduce_add(sumf[row]); + if (first_row + row < ne01) { + if (get_sub_group_local_id() == 0) { + dst_f32[first_row + row] = all_sum; + } + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl new file mode 100644 index 0000000..819e519 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl @@ -0,0 +1,194 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK4_0 32 +#define QR4_0 2 +#define QK4_1 32 +#define QR4_1 2 +#define QK5_0 32 +#define QR5_0 2 +#define QK5_1 32 +#define QR5_1 2 +#define QK8_0 32 +#define QR8_0 1 +#define QK_K 256 +#define K_QUANTS_PER_ITERATION 2 + +typedef char int8_t; +typedef uchar uint8_t; +typedef short int16_t; +typedef ushort uint16_t; +typedef int int32_t; +typedef uint uint32_t; + +//------------------------------------------------------------------------------ +// block_q6_K +//------------------------------------------------------------------------------ +// 6-bit quantization +// weight is represented as x = a * q +// 16 blocks of 16 elements each +// Effectively 6.5625 bits per weight +typedef struct { + uint8_t ql[QK_K/2]; // quants, lower 4 bits + uint8_t qh[QK_K/4]; // quants, upper 2 bits + int8_t scales[QK_K/16]; // scales, quantized with 8 bits + half d; // super-block scale +} block_q6_K; + +//------------------------------------------------------------------------------ +// kernel_mul_mv_q6_K_f32 +//------------------------------------------------------------------------------ + +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 1 // number of rows each SIMD group works on +#define N_SIMDGROUP 2 // number of SIMD groups in a thread group +#define N_SIMDWIDTH 16 // SIMD group size +#elif defined (ADRENO_GPU) +#define N_DST 1 +#define N_SIMDGROUP 2 +#define N_SIMDWIDTH 64 +#endif + +#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q6_K_f32( + global void * src0, + ulong offset0, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + uchar kmask1 = 0x03; + uchar kmask2 = 0x0C; + uchar kmask3 = 0x30; + uchar kmask4 = 0xC0; + + int nb = ne00/QK_K; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int row = N_SIMDGROUP * r0 + get_sub_group_id(); + + if (row >= ne01) { + return; + } + + int i12 = im%ne12; + int i13 = im/ne12; + + ulong offset_src0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + + global block_q6_K * x = (global block_q6_K *) src0 + row*nb + offset_src0; + global float * yy = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + float sumf = 0; + + // For Q6_K quantization, 16 values forms a subblock, 16 subblock forms a + // block. Values in a subblock shares a scale that is quantized with 8 bits; + // the entire block shares a single floating point scale. + // For work distribution, each thread processes a subblock (16 weights), hence + // 16 threads process a (super) block -- a subgroup thus handles SIMDWIDTH/16 + // (super) blocks -- this is the block stride. + // The 16 threads that process a (super) block are split into 2 portions, each has + // 8 threads; each portion works on 8 subblocks. + // For subgroup of 16 threads, the entire subgroup works on a single (super) block + // before moving to the next (super) block. Thread0 - thread7 work on the + // first 8 subblocks; thread8 - thread15 works on the last 8 subblocks. + // Thread0 - thread3 work on subblocks 0, 2, 4, 6; thread4 - thread7 work on + // subblocks 1, 3, 5, 7. Each thread does not work on an entire subblock, but + // works on a total of 16 weight values. + int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0 + int ix = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1 + int ip = tid/8; // first or second half of (super) block (0 or 1) + int il = tid%8; // each half has 8 parts, one per scale + int n = 4; // 4 scales at a time (and 4 sums) + int l0 = n*il; // offset into half-block, 0..28 + int is = 8*ip + l0/16; // 0, 1, 8, 9 + + int y_offset = 128*ip + l0; + int q_offset_l = 64*ip + l0; + int q_offset_h = 32*ip + l0; + + for (int i = ix; i < nb; i += BLOCK_STRIDE) { + + global uint8_t * q1 = x[i].ql + q_offset_l; + global uint8_t * q2 = q1 + QK_K/8; + global uint8_t * qh = x[i].qh + q_offset_h; + global int8_t * sc = x[i].scales + is; + + global float * y = yy + i * QK_K + y_offset; + + float dall = x[i].d; + + float4 sums = {0.f, 0.f, 0.f, 0.f}; + + sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & kmask1) << 4)) - 32.f); + sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & kmask2) << 2)) - 32.f); + sums.s2 += y[0+64] * ((float)((q1[0] >> 4) | ((qh[0] & kmask3) << 0)) - 32.f); + sums.s3 += y[0+96] * ((float)((q2[0] >> 4) | ((qh[0] & kmask4) >> 2)) - 32.f); + + sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & kmask1) << 4)) - 32.f); + sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & kmask2) << 2)) - 32.f); + sums.s2 += y[1+64] * ((float)((q1[1] >> 4) | ((qh[1] & kmask3) << 0)) - 32.f); + sums.s3 += y[1+96] * ((float)((q2[1] >> 4) | ((qh[1] & kmask4) >> 2)) - 32.f); + + sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & kmask1) << 4)) - 32.f); + sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & kmask2) << 2)) - 32.f); + sums.s2 += y[2+64] * ((float)((q1[2] >> 4) | ((qh[2] & kmask3) << 0)) - 32.f); + sums.s3 += y[2+96] * ((float)((q2[2] >> 4) | ((qh[2] & kmask4) >> 2)) - 32.f); + + sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & kmask1) << 4)) - 32.f); + sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & kmask2) << 2)) - 32.f); + sums.s2 += y[3+64] * ((float)((q1[3] >> 4) | ((qh[3] & kmask3) << 0)) - 32.f); + sums.s3 += y[3+96] * ((float)((q2[3] >> 4) | ((qh[3] & kmask4) >> 2)) - 32.f); + + sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]); + } + + float tot = sub_group_reduce_add(sumf); + if (get_sub_group_local_id() == 0) { + dst[r1*ne0 + im*ne0*ne1 + row] = tot; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl new file mode 100644 index 0000000..86fe09c --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl @@ -0,0 +1,194 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +//------------------------------------------------------------------------------ +// kernel_mul_mv_q6_K_f32_flat +//------------------------------------------------------------------------------ +#define Q6_K_MASK1 0x03 +#define Q6_K_MASK2 0x0C +#define Q6_K_MASK3 0x30 +#define Q6_K_MASK4 0xC0 + +#define QK_K 256 + +inline float block_q_6_K_dot_y_flat( + global uchar * blk_ql, + global uchar * blk_qh, + global char * blk_scales, + global half * blk_d, + global float * yy, + int ib, + int ip, + int is, + int l0 +) { + int y_offset = 128*ip + l0; + int q_offset_l = 64*ip + l0; + int q_offset_h = 32*ip + l0; + + global uchar * q1 = blk_ql + ib*128 + q_offset_l; + global uchar * q2 = q1 + QK_K/8; + global uchar * qh = blk_qh + ib*64 + q_offset_h; + global char * sc = blk_scales + ib*16 + is; + + global float * y = yy + ib * QK_K + y_offset; + + float dall = blk_d[ib]; + + float sumf = 0; + float4 sums = {0.f, 0.f, 0.f, 0.f}; + + sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[0+64] * ((float)((q1[0] >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[0+96] * ((float)((q2[0] >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[1+64] * ((float)((q1[1] >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[1+96] * ((float)((q2[1] >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[2+64] * ((float)((q1[2] >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[2+96] * ((float)((q2[2] >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f); + + sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f); + sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f); + sums.s2 += y[3+64] * ((float)((q1[3] >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f); + sums.s3 += y[3+96] * ((float)((q2[3] >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f); + + sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]); + + return sumf; +} + +#undef N_DST +#undef N_SIMDGROUP +#undef N_SIMDWIDTH + +#ifdef INTEL_GPU +#define N_DST 4 +#define N_SIMDGROUP 2 +#define N_SIMDWIDTH 16 +#elif defined (ADRENO_GPU) +#define N_DST 4 +#define N_SIMDGROUP 2 +#define N_SIMDWIDTH 64 +#endif + +#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q6_K_f32_flat( + global uchar * src0_ql, + global uchar * src0_qh, + global char * src0_s, + global half * src0_d, + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne10, + int ne12, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + int nb = ne00/QK_K; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int i12 = im%ne12; + int i13 = im/ne12; + + int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST; + + ulong offset_src0 = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02); + ulong offset_src0_ql = offset_src0 * 128; + ulong offset_src0_qh = offset_src0 * 64; + ulong offset_src0_s = offset_src0 * 16; + ulong offset_src0_d = offset_src0; + + global uchar * blk_ql = (global uchar *) src0_ql + offset_src0_ql; + global uchar * blk_qh = (global uchar *) src0_qh + offset_src0_qh; + global char * blk_scales = (global char *) src0_s + offset_src0_s; + global half * blk_d = (global half *) src0_d + offset_src0_d; + global float * yy = (global float *) src1 + r1*ne10 + im*ne00*ne1; + + int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0 + int ix = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1 + int ip = tid/8; // first or second half of (super) block (0 or 1) + int il = tid%8; // each half has 8 parts, one per scale + int n = 4; // 4 scales at a time (and 4 sums) + int l0 = n*il; // offset into half-block, 0..28 + int is = 8*ip + l0/16; // 0, 1, 8, 9 + + float4 sumf = 0; + + for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) { + if (first_row + 0 < ne01) { + sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0); + } + if (first_row + 1 < ne01) { + sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0); + } + if (first_row + 2 < ne01) { + sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0); + } + if (first_row + 3 < ne01) { + sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0); + } + } + + float4 tot = (float4)( + sub_group_reduce_add(sumf.s0), + sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), + sub_group_reduce_add(sumf.s3) + ); + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl new file mode 100644 index 0000000..7e88c74 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl @@ -0,0 +1,125 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK8_0 32 +typedef struct { + half d; // delta + char qs[QK8_0]; // quants +} block_q8_0; + +#define NB_Q8_0 8 + +#ifdef INTEL_GPU +#define N_R0_Q8_0 4 // number of rows each subgroup works on +#define N_SG_Q8_0 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_Q8_0 4 +#define N_SG_Q8_0 2 +#define N_SIMDWIDTH 64 +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q8_0_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src0 = (global char*)((global char*)src0 + offset0); + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + int nb = ne00/QK8_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0; + + uint i12 = im%ne12; + uint i13 = im/ne12; + + ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13; + global float * y = (global float *) (src1 + offset_src1); + + // pointers to src0 rows + global block_q8_0 * ax[N_R0_Q8_0]; + for (int row = 0; row < N_R0_Q8_0; ++row) { + ulong offset_src0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + ax[row] = (global block_q8_0 *) ((global char *) src0 + offset_src0); + } + + float yl[NB_Q8_0]; + float sumf[N_R0_Q8_0] = { 0.f }; + + const short ix = get_sub_group_local_id()/4; + const short il = get_sub_group_local_id()%4; + + global float * yb = y + ix*QK8_0 + il*NB_Q8_0; + + // each thread handles NB_Q8_0 quants at a time + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) { + for (short i = 0; i < NB_Q8_0; ++i) { + yl[i] = yb[i]; + } + + for (short row = 0; row < N_R0_Q8_0; row++) { + global char * qs = ax[row][ib].qs + il*NB_Q8_0; + float sumq = 0.f; + for (short iq = 0; iq < NB_Q8_0; ++iq) { + sumq += qs[iq] * yl[iq]; + } + sumf[row] += sumq*ax[row][ib].d; + } + + yb += N_SIMDWIDTH*NB_Q8_0; + } + + global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0; + + for (int row = 0; row < N_R0_Q8_0; ++row) { + float tot = sub_group_reduce_add(sumf[row]); + + if (get_sub_group_local_id() == 0 && first_row + row < ne01) { + dst_f32[first_row + row] = tot; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl new file mode 100644 index 0000000..71d159f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl @@ -0,0 +1,202 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#define QK8_0 32 +typedef struct { + half d; // delta + char qs[QK8_0]; // quants +} block_q8_0; + +#define NB_Q8_0 8 + +#ifdef INTEL_GPU +#define N_R0_Q8_0 4 // number of rows each subgroup works on +#define N_SG_Q8_0 2 // number of subgroups in a work group +#define N_SIMDWIDTH 16 // subgroup size +#elif defined (ADRENO_GPU) +#define N_R0_Q8_0 4 +#define N_SG_Q8_0 2 +#define N_SIMDWIDTH 64 +#endif + +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_16 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_mul_mv_q8_0_f32_flat( + global char * src0_q, + global half * src0_d, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + int ne1, + int r2, + int r3 +) { + src1 = (global char*)((global char*)src1 + offset1); + dst = (global char*)((global char*)dst + offsetd); + + int nb = ne00/QK8_0; + + int r0 = get_group_id(0); + int r1 = get_group_id(1); + int im = get_group_id(2); + + int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0; + + uint i12 = im%ne12; + uint i13 = im/ne12; + + ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13; + global float * y = (global float *) (src1 + offset_src1); + + // pointers to src0 rows + uint offset_src0_base = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03; + + global char * ax0, * ax1, * ax2, * ax3; + global half * ad0, * ad1, * ad2, * ad3; + uint offset_src0; + + offset_src0 = offset_src0_base + 0*nb01; + offset_src0 = offset_src0/34; + ax0 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0); + ad0 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 1*nb01; + offset_src0 = offset_src0/34; + ax1 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0); + ad1 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 2*nb01; + offset_src0 = offset_src0/34; + ax2 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0); + ad2 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half)); + + offset_src0 = offset_src0_base + 3*nb01; + offset_src0 = offset_src0/34; + ax3 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0); + ad3 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half)); + + const short ix = get_sub_group_local_id()/4; + const short il = get_sub_group_local_id()%4; + + global float * yb = y + ix*QK8_0 + il*NB_Q8_0; + + float8 yl; + float8 qv; + float4 sumf = 0.f; + float sumq = 0.f; + global char * qs; + + // each thread handles NB_Q8_0 quants at a time + for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) { + yl = vload8(0, yb); + + qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s0 += sumq*ad0[ib]; + + qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s1 += sumq*ad1[ib]; + + qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s2 += sumq*ad2[ib]; + + qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0; + qv = convert_float8(vload8(0, qs)); + sumq = 0; + sumq += qv.s0*yl.s0; + sumq += qv.s1*yl.s1; + sumq += qv.s2*yl.s2; + sumq += qv.s3*yl.s3; + sumq += qv.s4*yl.s4; + sumq += qv.s5*yl.s5; + sumq += qv.s6*yl.s6; + sumq += qv.s7*yl.s7; + sumf.s3 += sumq*ad3[ib]; + + yb += N_SIMDWIDTH*NB_Q8_0; + } + + global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0; + + float4 tot = (float4)( + sub_group_reduce_add(sumf.s0), + sub_group_reduce_add(sumf.s1), + sub_group_reduce_add(sumf.s2), + sub_group_reduce_add(sumf.s3) + ); + + if (get_sub_group_local_id() == 0) { + if (first_row + 0 < ne01) { + dst_f32[first_row + 0] = tot.s0; + } + if (first_row + 1 < ne01) { + dst_f32[first_row + 1] = tot.s1; + } + if (first_row + 2 < ne01) { + dst_f32[first_row + 2] = tot.s2; + } + if (first_row + 3 < ne01) { + dst_f32[first_row + 3] = tot.s3; + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl new file mode 100644 index 0000000..170f822 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl @@ -0,0 +1,161 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +//------------------------------------------------------------------------------ +// norm +//------------------------------------------------------------------------------ +kernel void kernel_norm( + global void * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb01, + ulong nb02, + ulong nb03, + float eps, + local float * sum +) { + src0 = (global void*)((global char*)src0 + offset0); + dst = (global void*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); + + // MEAN + // parallel sum + sum[get_local_id(0)] = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + sum[get_local_id(0)] += x[i00]; + } + // reduce + barrier(CLK_LOCAL_MEM_FENCE); + for (uint i = get_local_size(0)/2; i > 0; i /= 2) { + if (get_local_id(0) < i) { + sum[get_local_id(0)] += sum[get_local_id(0) + i]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + float mean = sum[0] / ne00; + + // recenter and VARIANCE + barrier(CLK_LOCAL_MEM_FENCE); + global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + sum[get_local_id(0)] = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + y[i00] = x[i00] - mean; + sum[get_local_id(0)] += y[i00] * y[i00]; + } + + // reduce + barrier(CLK_LOCAL_MEM_FENCE); + for (uint i = get_local_size(0)/2; i > 0; i /= 2) { + if (get_local_id(0) < i) { + sum[get_local_id(0)] += sum[get_local_id(0) + i]; + } + barrier(CLK_LOCAL_MEM_FENCE); + } + float variance = sum[0] / ne00; + + float scale = 1.0f/sqrt(variance + eps); + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + y[i00] = y[i00] * scale; + } +} + +//------------------------------------------------------------------------------ +// norm_mul_add +//------------------------------------------------------------------------------ +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_32 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_norm_mul_add( + global char * src0_ptr, ulong src0_offset, + global char * src1_ptr, ulong src1_offset, + global char * src2_ptr, ulong src2_offset, + global char * dst_ptr, ulong dst_offset, + int ne00, int ne01, int ne02, int ne03, + ulong nb01, ulong nb02, ulong nb03, + int ne10, int ne11, int ne12, int ne13, + ulong nb11, ulong nb12, ulong nb13, + int ne20, int ne21, int ne22, int ne23, + ulong nb21, ulong nb22, ulong nb23, + ulong nbd1, ulong nbd2, ulong nbd3, + float eps, + local float2 * sums +) { + const int i03 = get_group_id(2); + const int i02 = get_group_id(1); + const int i01 = get_group_id(0); + + global float4 * x = (global float4 *)(src0_ptr + src0_offset + i01*nb01 + i02*nb02 + i03*nb03); + global float4 * w = (global float4 *)(src1_ptr + src1_offset + (i01%ne11)*nb11 + (i02%ne12)*nb12 + (i03%ne13)*nb13); + global float4 * b = (global float4 *)(src2_ptr + src2_offset + (i01%ne21)*nb21 + (i02%ne22)*nb22 + (i03%ne23)*nb23); + global float4 * y = (global float4 *)(dst_ptr + dst_offset + i01*nbd1 + i02*nbd2 + i03*nbd3); + + float p_sum = 0.0f; + float p_sum_sq = 0.0f; + + const int n_chunks = ne00 / 4; + for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) { + float4 val = x[i00]; + p_sum += val.x + val.y + val.z + val.w; + p_sum_sq += dot(val, val); + } + + p_sum = sub_group_reduce_add(p_sum); + p_sum_sq = sub_group_reduce_add(p_sum_sq); + + if (get_sub_group_local_id() == 0) { + sums[get_sub_group_id()] = (float2)(p_sum, p_sum_sq); + } + barrier(CLK_LOCAL_MEM_FENCE); + + if (get_local_id(0) == 0) { + float sum = 0.0f; + float sum_sq = 0.0f; + for (uint i = 0; i < get_num_sub_groups(); ++i) { + float2 s = sums[i]; + sum += s.x; + sum_sq += s.y; + } + + const float inv_ne00 = 1.0f / (float)ne00; + const float mean = sum * inv_ne00; + const float variance = mad(-mean, mean, sum_sq * inv_ne00); + + sums[0] = (float2)(mean, rsqrt(variance + eps)); + } + barrier(CLK_LOCAL_MEM_FENCE); + + const float2 mean_scale = sums[0]; + const float mean = mean_scale.x; + const float scale = mean_scale.y; + const float neg_mean_scale = -mean * scale; + + for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) { + const int w_idx = ne10 > 1 ? i00 : 0; + const int b_idx = ne20 > 1 ? i00 : 0; + const float4 norm_x = mad(x[i00], (float4)scale, (float4)neg_mean_scale); + y[i00] = mad(norm_x, w[w_idx], b[b_idx]); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl new file mode 100644 index 0000000..31fb7cc --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl @@ -0,0 +1,39 @@ +kernel void kernel_pad( + global void * src0, + ulong offset0, + global void * dst, + ulong offsetd, + int ne00, int ne01, int ne02, int ne03, + ulong nb00, ulong nb01, ulong nb02, ulong nb03, + int ne0, int ne1, int ne2, int ne3, + ulong nb0, ulong nb1, ulong nb2, ulong nb3, + int lp0, int rp0, + int lp1, int rp1, + int lp2, int rp2, + int lp3, int rp3 +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int i0 = get_global_id(0); + int i1 = get_group_id(1); + int i2 = get_group_id(2) % ne2; + int i3 = get_group_id(2) / ne2; + + if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { + return; + } + + uint src0_idx = (i3 - lp3)*nb03 + (i2 - lp2)*nb02 + (i1 - lp1)*nb01 + (i0 - lp0)*nb00; + uint dst_idx = i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0; + + global float * src0_ptr = (global float *)((global char *)src0 + src0_idx); + global float * dst_ptr = (global float *)((global char *)dst + dst_idx); + + bool in_src_bounds = (i0 >= lp0 && i0 < ne0 - rp0) && + (i1 >= lp1 && i1 < ne1 - rp1) && + (i2 >= lp2 && i2 < ne2 - rp2) && + (i3 >= lp3 && i3 < ne3 - rp3); + + *dst_ptr = in_src_bounds ? *src0_ptr : 0.0f; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl new file mode 100644 index 0000000..60ff28a --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl @@ -0,0 +1,16 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// relu +//------------------------------------------------------------------------------ +kernel void kernel_relu( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = fmax(0.0f, src0[get_global_id(0)]); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl new file mode 100644 index 0000000..53951a5 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl @@ -0,0 +1,38 @@ +kernel void kernel_repeat_f32( + global const char * src0, + ulong offset0, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + dst = dst + offsetd; + + const int i3 = get_group_id(2); + const int i2 = get_group_id(1); + const int i1 = get_group_id(0); + + const int i03 = i3%ne03; + const int i02 = i2%ne02; + const int i01 = i1%ne01; + + global const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * dst_ptr = dst + i3*nb3 + i2*nb2 + i1*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i00 = i0%ne00; + *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i00*nb00)); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl new file mode 100644 index 0000000..4b18d17 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl @@ -0,0 +1,190 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +//------------------------------------------------------------------------------ +// rms_norm +//------------------------------------------------------------------------------ +// This kernel depends on subgroup size. +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_32 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_rms_norm( + global void * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb01, + ulong nb02, + ulong nb03, + float eps, + local float * sum // Note, the size depends on number of subgroups +) { + src0 = (global void*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); + global float * x_scalar = (global float *) x; + float4 sumf = 0; + float all_sum = 0; + + // parallel sum + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + sumf += x[i00] * x[i00]; + } + all_sum = sumf.s0 + sumf.s1 + sumf.s2 + sumf.s3; + all_sum = sub_group_reduce_add(all_sum); + if (get_sub_group_local_id() == 0) { + sum[get_sub_group_id()] = all_sum; + } + + barrier(CLK_LOCAL_MEM_FENCE); + // broadcast + for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) { + if (get_local_id(0) < i) { + sum[get_local_id(0)] += sum[get_local_id(0) + i]; + } + } + if (get_local_id(0) == 0) { + for (int i = 4 * (ne00 / 4); i < ne00; i++) { + sum[0] += x_scalar[i]; + } + sum[0] /= ne00; + } + + barrier(CLK_LOCAL_MEM_FENCE); + + const float mean = sum[0]; + const float scale = 1.0f/sqrt(mean + eps); + + global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + global float * y_scalar = (global float *) y; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + y[i00] = x[i00] * scale; + } + if (get_local_id(0) == 0) { + for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) { + y_scalar[i00] = x_scalar[i00] * scale; + } + } +} + +//------------------------------------------------------------------------------ +// rms_norm_mul +//------------------------------------------------------------------------------ +#ifdef INTEL_GPU +REQD_SUBGROUP_SIZE_32 +#elif defined (ADRENO_GPU) +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_rms_norm_mul( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb1, + ulong nb2, + ulong nb3, + float eps, + local float * sum +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + // The size of sum is sizeof(float)*subgroup_size. + // Each subgroup writes its partial sum to this array. + // So the number of subgroups per workgroup for this kernel cannot exceed the subgroup size. + // This is generally true - + // for subgroup size 64, workgroup size should be less than 4096 (the max is usually 1024). + if (get_sub_group_id() == 0) { + sum[get_sub_group_local_id()] = 0.0f; + } + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float4 * x = (global float4 *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); + global float4 * f = (global float4 *) (src1 + (i03%ne13)*nb13 + (i02%ne12)*nb12 + (i01%ne11)*nb11); + + float sumf = 0; + + // parallel sum + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + sumf += dot(x[i00], x[i00]); + } + sumf = sub_group_reduce_add(sumf); + + barrier(CLK_LOCAL_MEM_FENCE); + + if (get_sub_group_local_id() == 0) { + sum[get_sub_group_id()] = sumf; + } + + barrier(CLK_LOCAL_MEM_FENCE); + + //for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) { + // if (get_local_id(0) < i) { + // sum[get_local_id(0)] += sum[get_local_id(0) + i]; + // } + //} + //if (get_local_id(0) == 0) { + // sum[0] /= ne00; + //} + + //barrier(CLK_LOCAL_MEM_FENCE); + + sumf = sum[get_sub_group_local_id()]; + sumf = sub_group_reduce_add(sumf); + + float mean = sumf / ne00; + float scale = 1.0f/sqrt(mean + eps); + + global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1); + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + y[i00] = (x[i00] * scale) * f[i00%(ne10/4)]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl new file mode 100644 index 0000000..82f4cd8 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl @@ -0,0 +1,747 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// kernel_rope +//------------------------------------------------------------------------------ +float rope_yarn_ramp(float low, float high, int i0) { + const float y = (i0 / 2 - low) / max(0.001f, high - low); + return 1.0f - min(1.0f, max(0.0f, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +float2 rope_yarn( + float theta_extrap, float freq_scale, float2 corr_dims, int i0, float ext_factor, float mscale +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.s0, corr_dims.s1, i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * log(1.0f / freq_scale); + } + return (float2)(cos(theta) * mscale, sin(theta) * mscale); +} + +// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get +// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))` +float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) { + return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base)); +} + +float2 rope_yarn_corr_dims( + int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow +) { + // start and end correction dims + return (float2)( + max(0.0f, floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base))), + min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base))) + ); +} + +kernel void kernel_rope_norm_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + float theta_base = (float) pos[i2]; + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + float theta = theta_base * pow(freq_base, inv_ndims*i0); + + float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + float x0 = src[0]; + float x1 = src[1]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_norm_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + float theta_base = (float) pos[i2]; + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + float theta = theta_base * pow(freq_base, inv_ndims*i0); + + float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + float x0 = src[0]; + float x1 = src[1]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_neox_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + float theta_base = (float) pos[i2]; + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_neox_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + float theta_base = (float) pos[i2]; + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_multi_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections, + int is_imrope +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const int sector = (i0 / 2) % sect_dims; + float theta_base = 0.0f; + + if (is_imrope) { + if (sector % 3 == 1 && sector < 3 * sections.s1) { // h + theta_base = (float) pos[i2 + ne02 * 1]; + } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w + theta_base = (float) pos[i2 + ne02 * 2]; + } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t + theta_base = (float) pos[i2 + ne02 * 0]; + } else { // e + theta_base = (float) pos[i2 + ne02 * 3]; + } + } else { + if (sector < sections.s0) { + theta_base = pos[i2]; + } + else if (sector >= sections.s0 && sector < sec_w) { + theta_base = pos[i2 + ne2 * 1]; + } + else if (sector >= sec_w && sector < sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 2]; + } + else if (sector >= sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 3]; + } + } + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_multi_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections, + int is_imrope +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const int sector = (i0 / 2) % sect_dims; + float theta_base = 0.0f; + + if (is_imrope) { + if (sector % 3 == 1 && sector < 3 * sections.s1) { // h + theta_base = (float) pos[i2 + ne02 * 1]; + } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w + theta_base = (float) pos[i2 + ne02 * 2]; + } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t + theta_base = (float) pos[i2 + ne02 * 0]; + } else { // e + theta_base = (float) pos[i2 + ne02 * 3]; + } + } else { + if (sector < sections.s0) { + theta_base = pos[i2]; + } + else if (sector >= sections.s0 && sector < sec_w) { + theta_base = pos[i2 + ne2 * 1]; + } + else if (sector >= sec_w && sector < sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 2]; + } + else if (sector >= sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 3]; + } + } + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_vision_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + int ic = i0/2; + + const int sector = (i0/2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + const int p = sector; + theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p); + } else if (sector >= sections.s0 && sector < sec_w) { + const int p = sector - sections.s0; + theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p); + } + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } +} + +kernel void kernel_rope_vision_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + int ic = i0/2; + + const int sector = (i0/2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + const int p = sector; + theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p); + } else if (sector >= sections.s0 && sector < sec_w) { + const int p = sector - sections.s0; + theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p); + } + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl new file mode 100644 index 0000000..17ed97f --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl @@ -0,0 +1,27 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_scale_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + float scale, + float bias +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias; +} + +kernel void kernel_scale_f32_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd, + float scale, + float bias +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl new file mode 100644 index 0000000..fc3ff7a --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl @@ -0,0 +1,208 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +// v = { mp, L, d } +inline uint fastdiv(uint n, uint4 v) { + uint msbs; + msbs = mul_hi(n, v.s0); + return (msbs + n) >> v.s1; +} +inline uint fastmod(uint n, uint4 v) { + uint q = fastdiv(n, v); + return n - q * v.s2; +} + +kernel void kernel_set_rows_f32_i64( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + uint4 ne11, + uint4 ne12, + ulong nb10, + ulong nb11, + ulong nb12, + int nblk0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1); + + if (i01 >= ne01) { + return; + } + + //int i12 = i03%ne12; + //int i11 = i02%ne11; + int i12 = fastmod(i03, ne12); + int i11 = fastmod(i02, ne11); + + int i10 = i01; + long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0]; + + global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3); + global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03); + + for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) { + dst_row[ind] = (float)src_row[ind]; + } +} + +kernel void kernel_set_rows_f16_i64( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + uint4 ne11, + uint4 ne12, + ulong nb10, + ulong nb11, + ulong nb12, + int nblk0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1); + + if (i01 >= ne01) { + return; + } + + //int i12 = i03%ne12; + //int i11 = i02%ne11; + int i12 = fastmod(i03, ne12); + int i11 = fastmod(i02, ne11); + + int i10 = i01; + long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0]; + + global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3); + global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03); + + for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) { + dst_row[ind] = src_row[ind]; + } +} + +kernel void kernel_set_rows_f32_i32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + uint4 ne11, + uint4 ne12, + ulong nb10, + ulong nb11, + ulong nb12, + int nblk0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1); + + if (i01 >= ne01) { + return; + } + + //int i12 = i03%ne12; + //int i11 = i02%ne11; + int i12 = fastmod(i03, ne12); + int i11 = fastmod(i02, ne11); + + int i10 = i01; + int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0]; + + global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3); + global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03); + + for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) { + dst_row[ind] = (float)src_row[ind]; + } +} + +kernel void kernel_set_rows_f16_i32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + int ne01, + ulong nb01, + ulong nb02, + ulong nb03, + uint4 ne11, + uint4 ne12, + ulong nb10, + ulong nb11, + ulong nb12, + int nblk0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1); + + if (i01 >= ne01) { + return; + } + + //int i12 = i03%ne12; + //int i11 = i02%ne11; + int i12 = fastmod(i03, ne12); + int i11 = fastmod(i02, ne11); + + int i10 = i01; + int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0]; + + global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3); + global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03); + + for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) { + dst_row[ind] = src_row[ind]; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl new file mode 100644 index 0000000..e3f669d --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl @@ -0,0 +1,29 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// sigmoid +//------------------------------------------------------------------------------ + +kernel void kernel_sigmoid_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)])); +} + +kernel void kernel_sigmoid_f16( + global half * src0, + ulong offset0, + global half * dst, + ulong offsetd +) { + src0 = (global half*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)])); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl new file mode 100644 index 0000000..1d95e1b --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl @@ -0,0 +1,30 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// silu +//------------------------------------------------------------------------------ +kernel void kernel_silu( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + float x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x / (1.0f + exp(-x)); +} + +kernel void kernel_silu_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + float4 x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x / (1.0f + exp(-x)); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl new file mode 100644 index 0000000..571d165 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl @@ -0,0 +1,108 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max_4_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + int ne13, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb1, + ulong nb2, + ulong nb3, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + src2 = src2 + offset2; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03%ne13; + int i12 = i02%ne12; + int i11 = i01; + + global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03); + global half4 * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0; + global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0; + global float4 * pdst4 = (global float4 *)(dst + i01*nb1 + i02*nb2 + i03*nb3); + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)); + } + float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3)); + + const float max = sub_group_reduce_max(lmax); + + // parallel sum + float4 lsum4 = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max); + lsum4 += exp_psrc4; + pdst4[i00] = exp_psrc4; + } + float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3; + + float sum = sub_group_reduce_add(lsum); + + if (psrc2) { + sum += exp(psrc2[i02] - max); + } + + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + pdst4[i00] /= sum; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl new file mode 100644 index 0000000..1f944b2 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl @@ -0,0 +1,108 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max_4( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + int ne13, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb1, + ulong nb2, + ulong nb3, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + src2 = src2 + offset2; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03%ne13; + int i12 = i02%ne12; + int i11 = i01; + + global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03); + global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0; + global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0; + global float4 * pdst4 = (global float4 *)(dst + i01*nb1 + i02*nb2 + i03*nb3); + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + } + float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3)); + + const float max = sub_group_reduce_max(lmax); + + // parallel sum + float4 lsum4 = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max); + lsum4 += exp_psrc4; + pdst4[i00] = exp_psrc4; + } + float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3; + + float sum = sub_group_reduce_add(lsum); + + if (psrc2) { + sum += exp(psrc2[i02] - max); + } + + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + pdst4[i00] /= sum; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl new file mode 100644 index 0000000..4baa6c2 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl @@ -0,0 +1,107 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + int ne13, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb1, + ulong nb2, + ulong nb3, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + src2 = src2 + offset2; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03%ne13; + int i12 = i02%ne12; + int i11 = i01; + + global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03); + global half * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0; + global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0; + global float * pdst = (global float *)(dst + i01*nb1 + i02*nb2 + i03*nb3); + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float lmax = psrc2 ? psrc2[i02] : -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + } + float max = sub_group_reduce_max(lmax); + + // parallel sum + float lsum = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max); + lsum += exp_psrc0; + // Remember the result of exp here. exp is expensive, so we really do not + // wish to compute it twice. + pdst[i00] = exp_psrc0; + } + + float sum = sub_group_reduce_add(lsum); + + if (psrc2) { + sum += exp(psrc2[i02] - max); + } + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + pdst[i00] /= sum; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl new file mode 100644 index 0000000..d503190 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl @@ -0,0 +1,107 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +#ifdef cl_intel_subgroups +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#endif + +#ifdef cl_intel_required_subgroup_size +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * src2, + ulong offset2, + global char * dst, + ulong offsetd, + int ne00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne12, + int ne13, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb1, + ulong nb2, + ulong nb3, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + src2 = src2 + offset2; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03%ne13; + int i12 = i02%ne12; + int i11 = i01; + + global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03); + global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0; + global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0; + global float * pdst = (global float *)(dst + i01*nb1 + i02*nb2 + i03*nb3); + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float lmax = psrc2 ? psrc2[i02] : -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + } + float max = sub_group_reduce_max(lmax); + + // parallel sum + float lsum = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max); + lsum += exp_psrc0; + // Remember the result of exp here. exp is expensive, so we really do not + // wish to compute it twice. + pdst[i00] = exp_psrc0; + } + + float sum = sub_group_reduce_add(lsum); + + if (psrc2) { + sum += exp(psrc2[i02] - max); + } + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + pdst[i00] /= sum; + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl new file mode 100644 index 0000000..033766e --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl @@ -0,0 +1,88 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// softplus +//------------------------------------------------------------------------------ +inline float softplus_f32(float x){ + float ax = fabs(x); + float m = fmax(x, 0.0f); + return log1p(exp(-ax)) + m; +} + +kernel void kernel_softplus_f32_nd( + global void * p_src0_base, + ulong off_src0_abs, + global void * p_dst_base, + ulong off_dst_abs, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13 +) { + int i0 = get_global_id(0); + int i1 = get_global_id(1); + int i2 = get_global_id(2); + + if (i0 < ne10 && i1 < ne11 && i2 < ne12) { + for (int i3 = 0; i3 < ne13; ++i3) { + ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03; + global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor); + + ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13; + global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor); + + *dst_val_ptr = softplus_f32(*src_val_ptr); + } + } +} + +kernel void kernel_softplus_f16_nd( + global void * p_src0_base, + ulong off_src0_abs, + global void * p_dst_base, + ulong off_dst_abs, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13 +) { + int i0 = get_global_id(0); + int i1 = get_global_id(1); + int i2 = get_global_id(2); + + if (i0 < ne10 && i1 < ne11 && i2 < ne12) { + for (int i3 = 0; i3 < ne13; ++i3) { + ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03; + global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor); + + ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13; + global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor); + + *dst_val_ptr = (half)(softplus_f32((float)(*src_val_ptr))); + } + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl new file mode 100644 index 0000000..80745fc --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl @@ -0,0 +1,51 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// solve_tri +//------------------------------------------------------------------------------ +kernel void kernel_solve_tri_f32( + global uchar * src0, + ulong offset0, + global uchar * src1, + ulong offset1, + global uchar * dst, + ulong offsetd, + int n, + int k, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + int col = get_global_id(0); + int i2 = get_global_id(1); + int i3 = get_global_id(2); + + global const uchar * Lb = src0 + offset0 + i2 * nb02 + i3 * nb03; + global const uchar * Bb = src1 + offset1 + i2 * nb12 + i3 * nb13; + global uchar * Xb = dst + offsetd + i2 * nb2 + i3 * nb3; + + for(int row = 0; row < n; ++row){ + global const float *pB = (global const float *)(Bb + row * nb11 + col * nb10); + + float sum = 0.0f; + for(int j = 0; j < row; ++j){ + global const float *pL = (global const float *)(Lb + row * nb01 + j * nb00); + global const float *pX = (global const float *)(Xb + j * nb1 + col * nb0); + sum += (*pL) * (*pX); + } + + global const float * pDiag = (global const float *)(Lb + row * nb01 + row *nb00); + global float * pOut = (global float *)(Xb + row * nb1 + col *nb0); + + *pOut = ((* pB) - sum) / (*pDiag); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl new file mode 100644 index 0000000..4310906 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl @@ -0,0 +1,53 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_sqr_cont_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = src0[gid] * src0[gid]; +} + +kernel void kernel_sqr_cont_f32_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = src0[gid] * src0[gid]; +} + +kernel void kernel_sqr_cont_f16( + global half * src0, + ulong offset0, + global half * dst, + ulong offsetd +) { + src0 = (global half*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = src0[gid] * src0[gid]; +} + +kernel void kernel_sqr_cont_f16_4( + global half4 * src0, + ulong offset0, + global half4 * dst, + ulong offsetd +) { + src0 = (global half4*)((global char*)src0 + offset0); + dst = (global half4*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = src0[gid] * src0[gid]; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl new file mode 100644 index 0000000..c59fbe0 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl @@ -0,0 +1,53 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_sqrt_cont_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = sqrt(src0[gid]); +} + +kernel void kernel_sqrt_cont_f32_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = sqrt(src0[gid]); +} + +kernel void kernel_sqrt_cont_f16( + global half * src0, + ulong offset0, + global half * dst, + ulong offsetd +) { + src0 = (global half*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = convert_half(sqrt(convert_float(src0[gid]))); +} + +kernel void kernel_sqrt_cont_f16_4( + global half4 * src0, + ulong offset0, + global half4 * dst, + ulong offsetd +) { + src0 = (global half4*)((global char*)src0 + offset0); + dst = (global half4*)((global char*)dst + offsetd); + + uint gid = get_global_id(0); + dst[gid] = convert_half4(sqrt(convert_float4(src0[gid]))); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl new file mode 100644 index 0000000..7ae21ac --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl @@ -0,0 +1,77 @@ +kernel void kernel_ssm_conv_f32_f32( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + int ne10, + ulong nb11, + ulong nb0, + ulong nb1, + ulong nb2 +){ + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int ir = get_global_id(0); + int i2 = get_global_id(1); + int i3 = get_global_id(2); + + int nc = ne10; + + global float * s = (global float *) (src0 + ir*nb01 + i2*nb00 + i3*nb02); + global float * c = (global float *) (src1 + ir*nb11); + global float * d = (global float *) (dst + ir*nb0 + i2*nb1 + i3*nb2); + + float sumf = 0.0f; + + for (int i0 = 0; i0 < nc; ++i0) { + sumf += s[i0] * c[i0]; + } + + d[0] = sumf; +} + +kernel void kernel_ssm_conv_f32_f32_4( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + int ne10, + ulong nb11, + ulong nb0, + ulong nb1, + ulong nb2 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int ir = get_global_id(0); + int i2 = get_global_id(1); + int i3 = get_global_id(2); + + int nc = ne10; + + global float4 * s = (global float4 *) (src0 + ir*nb01 + i2*nb00 + i3*nb02); + global float4 * c = (global float4 *) (src1 + ir*nb11); + global float * d = (global float *) (dst + ir*nb0 + i2*nb1 + i3*nb2); + + float sumf = 0.0f; + + for (int i0 = 0; i0 < nc/4; ++i0) { + sumf += dot(s[i0], c[i0]); + } + + d[0] = sumf; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl new file mode 100644 index 0000000..423ed59 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl @@ -0,0 +1,138 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// div +//------------------------------------------------------------------------------ +kernel void kernel_sub( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) - *((global float *)(src1_ptr + i10*nb10)); + } +} + +// assumption: src1 is a row +// broadcast src1 into src0 +kernel void kernel_sub_row( + global float4 * src0, + ulong offset0, + global float4 * src1, + ulong offset1, + global float4 * dst, + ulong offsetd, + int ne +) { + src0 = (global float4*)((global char*)src0 + offset0); + src1 = (global float4*)((global char*)src1 + offset1); + dst = (global float4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] - src1[idx1]; +} + +kernel void kernel_sub_f16( + global char * src0, + ulong offset0, + global char * src1, + ulong offset1, + global char * dst, + ulong offsetd, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + ulong nb10, + ulong nb11, + ulong nb12, + ulong nb13, + int ne0, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + src1 = src1 + offset1; + dst = dst + offsetd; + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + int i13 = i03 % ne13; + int i12 = i02 % ne12; + int i11 = i01 % ne11; + + global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01; + global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1; + + for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) { + const int i10 = i0 % ne10; + *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) - *((global half *)(src1_ptr + i10*nb10)); + } +} + +kernel void kernel_sub_row_f16( + global half4 * src0, + ulong offset0, + global half4 * src1, + ulong offset1, + global half4 * dst, + ulong offsetd, + int ne +) { + src0 = (global half4*)((global char*)src0 + offset0); + src1 = (global half4*)((global char*)src1 + offset1); + dst = (global half4*)((global char*)dst + offsetd); + + // This performs better than using %. + uint gid = get_global_id(0); + uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne + dst[gid] = src0[gid] - src1[idx1]; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl new file mode 100644 index 0000000..c5f7c57 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl @@ -0,0 +1,39 @@ + +kernel void kernel_sum_rows_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = (global float *)((global char *)src0 + offset0); + dst = (global float *)((global char *)dst + offsetd); + + int i3 = get_global_id(2); + int i2 = get_global_id(1); + int i1 = get_global_id(0); + + if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) { + return; + } + + global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03); + global float * dst_row = (global float *) ((global char *) dst + i1*nb1 + i2*nb2 + i3*nb3); + + float row_sum = 0; + + for (int i0 = 0; i0 < ne00; i0++) { + row_sum += src_row[i0]; + } + + dst_row[0] = row_sum; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl new file mode 100644 index 0000000..2c4887a --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl @@ -0,0 +1,109 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +kernel void kernel_tanh_f32( + global const float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = tanh(src0[get_global_id(0)]); +} + +kernel void kernel_tanh_f32_4( + global const float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = tanh(src0[get_global_id(0)]); +} + +kernel void kernel_tanh_f16( + global const half * src0, + ulong offset0, + global half * dst, + ulong offsetd +) { + src0 = (global half*)((global char*)src0 + offset0); + dst = (global half*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = tanh(src0[get_global_id(0)]); +} + +kernel void kernel_tanh_f16_4( + global const half4 * src0, + ulong offset0, + global half4 * dst, + ulong offsetd +) { + src0 = (global half4*)((global char*)src0 + offset0); + dst = (global half4*)((global char*)dst + offsetd); + + dst[get_global_id(0)] = tanh(src0[get_global_id(0)]); +} + +kernel void kernel_tanh_f32_nc( + global const char * src0, + ulong offset0, + global char * dst, + ulong offsetd, + int ne00, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + dst = dst + offsetd; + + const int i3 = get_group_id(2); + const int i2 = get_group_id(1); + const int i1 = get_group_id(0); + + for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) { + global const float * x = (global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + *y = tanh(*x); + } +} + +kernel void kernel_tanh_f16_nc( + global const char * src0, + ulong offset0, + global char * dst, + ulong offsetd, + int ne00, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3 +) { + src0 = src0 + offset0; + dst = dst + offsetd; + + const int i3 = get_group_id(2); + const int i2 = get_group_id(1); + const int i1 = get_group_id(0); + + for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) { + global const half * x = (global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * y = (global half *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + *y = tanh(*x); + } +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl new file mode 100644 index 0000000..1279b65 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl @@ -0,0 +1,117 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +// 16-bit transpose, loading/storing a 4x4 tile of elements +kernel void kernel_transpose_16( + __read_only image1d_buffer_t input, + __write_only image1d_buffer_t output, + const uint rows, + const uint cols +) { + + const int i = get_global_id(0); + const int j = get_global_id(1); + const int i_2 = i<<2; + const int j_2 = j<<2; + + half4 temp0 = read_imageh(input, (j_2+0)*cols+i); + half4 temp1 = read_imageh(input, (j_2+1)*cols+i); + half4 temp2 = read_imageh(input, (j_2+2)*cols+i); + half4 temp3 = read_imageh(input, (j_2+3)*cols+i); + + write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); + write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1)); + write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2)); + write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3)); +} + +// Padded kernel for irregular shape +kernel void kernel_transpose_16_4x1( + __read_only image1d_buffer_t input, + __write_only image1d_buffer_t output, + const uint rows, + const uint cols +) { + + const int i = get_global_id(0); + const int j = get_global_id(1); + const int j_2 = j << 2; + + half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x; + half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x; + half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x; + half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x; + + write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3)); +} + +// Transpose treating each element as 16-bit using buffer +kernel void kernel_transpose_16_buf( + global const ushort * input, + global ushort * output, + const int ldi, + const int ldo +) { + const int x = get_global_id(0); + const int y = get_global_id(1); + + output[x*ldo + y] = input[y*ldi + x]; +} + +// 32-bit transpose, loading/storing a 4x4 tile of elements +kernel void kernel_transpose_32( + __read_only image1d_buffer_t input, + __write_only image1d_buffer_t output, + const uint rows, + const uint cols +) { + + const int i = get_global_id(0); + const int j = get_global_id(1); + const int i_2 = i<<2; + const int j_2 = j<<2; + + float4 temp0 = read_imagef(input, (j_2+0)*cols+i); + float4 temp1 = read_imagef(input, (j_2+1)*cols+i); + float4 temp2 = read_imagef(input, (j_2+2)*cols+i); + float4 temp3 = read_imagef(input, (j_2+3)*cols+i); + + write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); + write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1)); + write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2)); + write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3)); + +} + +// 32-bit transpose, loading/storing a 4x4 tile of elements +// Only used for activations +// converts to FP16 +// also adds zero padding for non multiple of 8 prompt lengths +kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) { + + const int i = get_global_id(0); + const int j = get_global_id(1); + const int i_2 = i<<2; + const int j_2 = j<<2; + half4 temp0 = {0,0,0,0}; // initialize outputs to 0 + half4 temp1 = {0,0,0,0}; + half4 temp2 = {0,0,0,0}; + half4 temp3 = {0,0,0,0}; + + if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0 + temp0 = read_imageh(input, (j_2+0)*cols+i); + } + if((j_2+1)*cols+i*4+3 < rows*cols*16){ + temp1 = read_imageh(input, (j_2+1)*cols+i); + } + if((j_2+2)*cols+i*4+3 < rows*cols*16){ + temp2 = read_imageh(input, (j_2+2)*cols+i); + } + if((j_2+3)*cols+i*4+3 < rows*cols*16){ + temp3 = read_imageh(input, (j_2+3)*cols+i); + } + + write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding + write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1)); + write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2)); + write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3)); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl new file mode 100644 index 0000000..35cdd54 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl @@ -0,0 +1,32 @@ +#pragma OPENCL EXTENSION cl_khr_fp16 : enable + +//------------------------------------------------------------------------------ +// tri +//------------------------------------------------------------------------------ +__kernel void kernel_tri_f32( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd, + int n, + int ne0, + int ne1, + int tri_type +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + int idx = get_global_id(0); + if (idx >= n) return; + + int i0 = idx % ne0; + int i1 = (idx / ne0) % ne1; + + int keep = 0; + if (tri_type == 0) keep = (i0 >= i1); + else if (tri_type == 1) keep = (i0 > i1); + else if (tri_type == 2) keep = (i0 <= i1); + else keep = (i0 < i1); + + dst[idx] = keep ? src0[idx] : 0.0f; +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl new file mode 100644 index 0000000..21444bd --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl @@ -0,0 +1,48 @@ +kernel void kernel_timestep_embedding( + global const void * p_timesteps, + ulong off_timesteps, + global void * p_dst, + ulong off_dst, + int dst_nb1_bytes, + int logical_dim, + int max_period +) { + int local_i; + int local_j; + int local_half_dim; + float local_timestep_val; + float local_freq; + float local_arg; + global float * local_embed_data_ptr; + global const float * local_timesteps_input_ptr; + global float * local_dst_output_base_ptr; + + local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps); + local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst); + + local_i = get_global_id(1); + local_j = get_global_id(0); + + local_half_dim = logical_dim / 2; + local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes); + + if (logical_dim % 2 != 0 && local_j == local_half_dim) { + local_embed_data_ptr[2 * local_half_dim] = 0.0f; + } + + if (local_j >= local_half_dim) { + return; + } + + local_timestep_val = local_timesteps_input_ptr[local_i]; + + if (local_half_dim == 0) { + local_freq = 1.0f; + } else { + local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim); + } + + local_arg = local_timestep_val * local_freq; + local_embed_data_ptr[local_j] = cos(local_arg); + local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg); +} diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl new file mode 100644 index 0000000..25c6835 --- /dev/null +++ b/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl @@ -0,0 +1,120 @@ +kernel void kernel_upscale( + global const void * p_src0, + ulong off_src0, + global void * p_dst, + ulong off_dst, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne10, + int ne11, + int ne12, + int ne13, + float sf0, + float sf1, + float sf2, + float sf3 +) { + global const char * src_base = (global const char *)p_src0 + off_src0; + global float * dst_base = (global float *)((global char *)p_dst + off_dst); + + int index = get_global_id(0); + int dst_total_elements = ne10 * ne11 * ne12 * ne13; + + if (index >= dst_total_elements) { + return; + } + + int i10 = index % ne10; + int i11 = (index / ne10) % ne11; + int i12 = (index / (ne10 * ne11)) % ne12; + int i13 = index / (ne10 * ne11 * ne12); + + int i00 = (int)(i10 / sf0); + int i01 = (int)(i11 / sf1); + int i02 = (int)(i12 / sf2); + int i03 = (int)(i13 / sf3); + + ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00; + global const float * src_element_ptr = (global const float *)(src_base + offset_src_element); + + dst_base[index] = *src_element_ptr; +} + +kernel void kernel_upscale_bilinear( + global const void * p_src0, + ulong off_src0, + global void * p_dst, + ulong off_dst, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne00_src, + int ne01_src, + int ne10_dst, + int ne11_dst, + int ne12_dst, + int ne13_dst, + float sf0, + float sf1, + float sf2, + float sf3, + float pixel_offset +) { + global const char * src_base = (global const char *)p_src0 + off_src0; + global float * dst_base = (global float *)((global char *)p_dst + off_dst); + + int index = get_global_id(0); + int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst; + + if (index >= dst_total_elements) { + return; + } + + int i10_dst = index % ne10_dst; + int i11_dst = (index / ne10_dst) % ne11_dst; + int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst; + int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst); + + int i02_src = (int)(i12_dst / sf2); + int i03_src = (int)(i13_dst / sf3); + + float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset; + long y0_src = (long)floor(y_src_f); + long y1_src = y0_src + 1; + + y0_src = max(0L, min(y0_src, (long)ne01_src - 1)); + y1_src = max(0L, min(y1_src, (long)ne01_src - 1)); + + float dy = y_src_f - (float)y0_src; + dy = max(0.0f, min(dy, 1.0f)); + + float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset; + long x0_src = (long)floor(x_src_f); + long x1_src = x0_src + 1; + + x0_src = max(0L, min(x0_src, (long)ne00_src - 1)); + x1_src = max(0L, min(x1_src, (long)ne00_src - 1)); + + float dx = x_src_f - (float)x0_src; + dx = max(0.0f, min(dx, 1.0f)); + + global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03); + global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03); + global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03); + global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03); + + const float val_a = *p_a; + const float val_b = *p_b; + const float val_c = *p_c; + const float val_d = *p_d; + + float result = val_a * (1.0f - dx) * (1.0f - dy) + + val_b * dx * (1.0f - dy) + + val_c * (1.0f - dx) * dy + + val_d * dx * dy; + + dst_base[index] = result; +} |