1#version 450
2#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
3
4#include "mul_mat_vec_base.glsl"
5
6layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
7
8shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][2][8];
9
10FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
11uint csel = 0;
12
13void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) {
14 const uint y_idx = i * QUANT_K + y_offset;
15
16 [[unroll]] for (uint n = 0; n < num_rows; ++n) {
17 const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
18 csel ^= 1;
19
20 if (!all_threads) { // when we don't have enough blocks to use all threads
21 if (i < num_blocks_per_row)
22 sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
23 barrier();
24
25 if (i >= num_blocks_per_row)
26 continue;
27 }
28
29 const uint32_t hmk = ~(uint32_t(data_a_packed16[ib0 + i].hmask[v_in]) | (uint32_t(data_a_packed16[ib0 + i].hmask[v_in + 8]) << 16));
30 const vec4 hmk_0 = vec4(unpack8(((hmk & hm_m[0]) >> ( v_im4)) << 2));
31 const vec4 hmk_1 = vec4(unpack8(((hmk & hm_m[1]) >> (1 + v_im4)) << 2));
32 const vec4 hmk_2 = vec4(unpack8(((hmk & hm_m[2]) >> (2 + v_im4)) << 2));
33 const vec4 hmk_3 = vec4(unpack8(((hmk & hm_m[3]) >> (3 + v_im4)) << 2));
34
35 // 0, 1, 16, 17
36 uint32_t qs_u32 = uint32_t(data_a[ib0 + i].qs[q_offset]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 1]) << 8);
37 qs_u32 |= (uint32_t(data_a[ib0 + i].qs[q_offset + 16]) | (uint32_t(data_a[ib0 + i].qs[q_offset + 17]) << 8)) << 16;
38 const vec4 qs_u32_0 = vec4(unpack8(qs_u32 & 0x03030303));
39 const vec4 qs_u32_2 = vec4(unpack8((qs_u32 >> 2) & 0x03030303));
40 const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
41 const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
42
43 if (all_threads) {
44 sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32);
45 barrier();
46 }
47
48 const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
49
50 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
51 vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]);
52 vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]);
53 vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
54 vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
55 vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
56 vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
57 vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
58 vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);
59
60 FLOAT_TYPE sum = FLOAT_TYPE(0.0);
61 [[unroll]] for (int l = 0; l < 2; ++l) {
62 sum = fma(FLOAT_TYPE( b0[l]) * sccache[csel][ix][v_im][0], qs_u32_0[l ] - hmk_0[l ],
63 fma(FLOAT_TYPE( b16[l]) * sccache[csel][ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2],
64 fma(FLOAT_TYPE( b32[l]) * sccache[csel][ix][v_im][2], qs_u32_2[l ] - hmk_1[l ],
65 fma(FLOAT_TYPE( b48[l]) * sccache[csel][ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2],
66 fma(FLOAT_TYPE( b64[l]) * sccache[csel][ix][v_im][4], qs_u32_4[l ] - hmk_2[l ],
67 fma(FLOAT_TYPE( b80[l]) * sccache[csel][ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2],
68 fma(FLOAT_TYPE( b96[l]) * sccache[csel][ix][v_im][6], qs_u32_6[l ] - hmk_3[l ],
69 fma(FLOAT_TYPE(b112[l]) * sccache[csel][ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum))))))));
70 }
71 temp[j][n] = fma(d, sum, temp[j][n]);
72 }
73 }
74}
75
76void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
77 uint a_offset, b_offset, d_offset;
78 get_offsets(a_offset, b_offset, d_offset);
79
80 const uint num_blocks_per_row = p.ncols / QUANT_K;
81
82 // 16 threads are used to process each block
83 const uint it_size = gl_WorkGroupSize.x/16;
84 const uint tid = gl_LocalInvocationID.x;
85 const uint itid = tid%16; // 0...15
86 const uint ix = tid/16;
87 const uint itid8 = itid%8;
88
89 const uint v_im = itid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
90 const uint v_im4 = v_im*4;
91 const uint v_in = itid - 8*v_im; // 0...7
92
93 const uint32_t m = 0x01010101 << (4 * v_im);
94 uint32_t hm_m[4];
95 [[unroll]] for (uint j = 0; j < 4; ++j)
96 hm_m[j] = m << j;
97
98 const uint l0 = 2*v_in; // 0...15
99 const uint q_offset = 32*v_im + l0;
100 const uint y_offset = 128*v_im + l0;
101
102 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
103 [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
104 temp[j][i] = FLOAT_TYPE(0);
105 }
106 }
107
108 const uint s_shift = v_im4 + 2*(itid8/4);
109
110 const uint nbr_par_th = num_blocks_per_row%it_size;
111 const uint nbr_all_th = num_blocks_per_row - nbr_par_th;
112 uint i0 = 0;
113 [[unroll]] for (; i0 < nbr_all_th; i0 += it_size)
114 calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, true);
115 calc_superblock(a_offset, b_offset, ix, itid8, v_im, v_im4, v_in, hm_m, q_offset, y_offset, s_shift, i0 + ix, num_blocks_per_row, first_row, num_rows, false);
116
117 reduce_result(temp, d_offset, first_row, num_rows, tid);
118}
119
120void main() {
121 const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
122
123 // do NUM_ROWS at a time, unless there aren't enough remaining rows
124 if (first_row + NUM_ROWS <= p.stride_d) {
125 compute_outputs(first_row, NUM_ROWS);
126 } else {
127 if (first_row >= p.stride_d) {
128 return;
129 }
130 compute_outputs(first_row, p.stride_d - first_row);
131 }
132}