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
8FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
9
10void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
11 const uint y_idx = i * QUANT_K + 16 * itid;
12 const uint nibble_shift = 4 * (itid & 1);
13 const uint ib32 = itid / 2; // 0..7
14 uint ibi = a_offset + first_row * num_blocks_per_row + i;
15 // Precompute db multiplication factors
16 float db_vals[NUM_ROWS];
17 [[unroll]] for (uint n = 0; n < num_rows; ++n) {
18 const float d = float(data_a[ibi].d);
19 const uint scale_raw = data_a[ibi].scales[ib32];
20 const uint scale = (scale_raw >> nibble_shift) & 0xF;
21 // Merge constant calculations d * (0.5 + scale) * 0.25 = d*0.125 + d*scale*0.25
22 db_vals[n] = d * (0.125f + float(scale) * 0.25f);
23 ibi += num_blocks_per_row;
24 }
25 ibi = a_offset + first_row * num_blocks_per_row + i;
26 [[unroll]] for (uint n = 0; n < num_rows; ++n) {
27 // Preload grid and sign data for all l values
28 vec4 grid0_vals[2], grid1_vals[2];
29 uint sign_vals[2], sign7_vals[2];
30 [[unroll]] for (uint l = 0; l < 2; ++l) {
31 const uint qs = data_a[ibi].qs[2 * itid + l];
32 sign_vals[l] = qs >> 9;
33 sign7_vals[l] = bitCount(sign_vals[l]);
34 const uvec2 grid_data = iq2xs_grid[qs & 511];
35 grid0_vals[l] = vec4(unpack8(grid_data.x));
36 grid1_vals[l] = vec4(unpack8(grid_data.y));
37 }
38 // Preload B data for all j columns (reduce repeated index calculations)
39 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
40 FLOAT_TYPE sum = FLOAT_TYPE(0.0);
41 [[unroll]] for (uint l = 0; l < 2; ++l) {
42 const uint sign = sign_vals[l];
43 const uint sign7 = sign7_vals[l];
44 const vec4 grid0 = grid0_vals[l];
45 const vec4 grid1 = grid1_vals[l];
46 // Precompute indices
47 const uint b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4 + 2 * l;
48 const vec4 b0 = vec4(data_b_v4[b_idx + 0]);
49 const vec4 b4 = vec4(data_b_v4[b_idx + 1]);
50 sum +=
51 fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
52 fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
53 fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
54 fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
55 fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
56 fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
57 fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
58 fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w),
59 FLOAT_TYPE(0.0)))))))));
60 }
61 temp[j][n] = fma(FLOAT_TYPE(db_vals[n]), sum, temp[j][n]);
62 }
63 ibi += num_blocks_per_row;
64 }
65}
66
67void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
68 uint a_offset, b_offset, d_offset;
69 get_offsets(a_offset, b_offset, d_offset);
70
71 const uint num_blocks_per_row = p.ncols / QUANT_K;
72
73 // 16 threads are used to process each block
74 const uint blocks_per_wg = gl_WorkGroupSize.x/16;
75 const uint tid = gl_LocalInvocationID.x;
76 const uint itid = tid % 16; // 0...15
77 const uint ix = tid / 16;
78
79 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
80 [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
81 temp[j][i] = FLOAT_TYPE(0);
82 }
83 }
84
85 [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
86 calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
87
88 reduce_result(temp, d_offset, first_row, num_rows, tid);
89}
90
91void main() {
92 const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
93
94 init_iq_shmem(gl_WorkGroupSize);
95
96 // do NUM_ROWS at a time, unless there aren't enough remaining rows
97 if (first_row + NUM_ROWS <= p.stride_d) {
98 compute_outputs(first_row, NUM_ROWS);
99 } else {
100 if (first_row >= p.stride_d) {
101 return;
102 }
103 compute_outputs(first_row, p.stride_d - first_row);
104 }
105}