1#include "conv-transpose-1d.cuh"
 2
 3static  __global__ void conv_transpose_1d_kernel(
 4        const int s0, const int p0, const int d0, const int output_size,
 5        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
 6        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
 7        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
 8        const float * src0, const float * src1,  float * dst) {
 9    int global_index = threadIdx.x + blockIdx.x * blockDim.x;
10    if (global_index >= output_size) {
11        return;
12    }
13
14    int out_index = global_index / dst_ne0;
15
16    float accumulator = 0;
17
18    for (int c = 0; c < src0_ne2; c++) {
19        int idx = global_index % dst_ne0;
20
21        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
22        int input_offset = src1_ne0 * c;
23
24        for (int i = 0; i < src1_ne0; i++) {
25            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
26                continue;
27            }
28            int weight_idx = idx - i*s0;
29
30            float kernel_weight = src0[kernel_offset + weight_idx];
31            float input_value =  src1[input_offset+i];
32
33            accumulator += kernel_weight * input_value;
34        }
35    }
36    dst[global_index] = accumulator;
37    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
38}
39
40static void conv_transpose_1d_f32_f32_cuda(
41        const int s0, const int p0, const int d0, const int output_size,
42        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
43        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
44        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
45        const float * src0, const float * src1,  float * dst,
46        cudaStream_t stream) {
47
48    const int num_blocks = (output_size + CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE;
49    conv_transpose_1d_kernel<<<num_blocks,CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE, 0, stream>>>(
50        s0,p0,d0,output_size,
51        src0_ne0, src0_ne1,  src0_ne2, src0_ne3,
52        src1_ne0, src1_ne1,  src1_ne2, src1_ne3,
53        dst_ne0,  dst_ne1,   dst_ne2,  dst_ne3,
54        src0,src1, dst);
55}
56
57void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
58    const ggml_tensor * src0 = dst->src[0];
59    const float * src0_d = (const float *)src0->data;
60
61    const ggml_tensor * src1 = dst->src[1];
62    const float * src1_d = (const float *)src1->data;
63
64    float * dst_d = (float *)dst->data;
65    cudaStream_t stream = ctx.stream();
66
67    GGML_ASSERT(src0->type == GGML_TYPE_F32);
68    GGML_ASSERT( dst->type == GGML_TYPE_F32);
69
70    GGML_ASSERT(ggml_is_contiguous(src0));
71    GGML_ASSERT(ggml_is_contiguous(src1));
72
73    const int32_t * opts = (const int32_t *)dst->op_params;
74
75    const int s0 = opts[0];
76    const int p0 = 0;//opts[3];
77    const int d0 = 1;//opts[4];
78
79    const int64_t output_size = ggml_nelements(dst);
80
81    conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size,
82        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
83        src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
84        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
85        src0_d, src1_d, dst_d, stream);
86}