1#version 450
2
3#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
4
5#include "mul_mat_vec_base.glsl"
6
7layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
8
9FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
10
11void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im, const uint l0, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
12 const uint y1_idx = i * QUANT_K + y_offset;
13 const uint y2_idx = y1_idx + 128;
14
15 [[unroll]] for (uint n = 0; n < num_rows; ++n) {
16 const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
17 const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
18
19 const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im ];
20 const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
21 const uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
22
23 const uint32_t scale_0_4_l = (scale4_u32 << 16) | scale0_u32;
24 const uint32_t scale_0_4_h = (scale_0_4_l & 0xC0C0C0C0) >> 2;
25 const vec4 scale_0_4_l_f = vec4(unpack8(scale_0_4_l & 0x3F3F3F3F));
26 const vec4 scale8_f = vec4(unpack8((((scale8_u32 << 12) | scale8_u32) & 0x0F0F0F0F) | scale_0_4_h));
27
28 const FLOAT_TYPE sc0 = scale_0_4_l_f.x;
29 const FLOAT_TYPE sc1 = scale_0_4_l_f.y;
30 const FLOAT_TYPE sc2 = scale_0_4_l_f.z;
31 const FLOAT_TYPE sc3 = scale_0_4_l_f.w;
32 const FLOAT_TYPE sc4 = scale8_f.x;
33 const FLOAT_TYPE sc5 = scale8_f.y;
34 const FLOAT_TYPE sc6 = scale8_f.z;
35 const FLOAT_TYPE sc7 = scale8_f.w;
36
37 const uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
38 const uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
39
40 uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
41 uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
42 uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
43 uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
44
45 const uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
46
47 const uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
48 const uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
49 const uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010);
50 const uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
51
52 qs0_16_u32_lo4 += qs0_16_lo4_offset16;
53 qs0_16_u32_hi4 += qs0_16_hi4_offset16;
54 qs64_80_u32_lo4 += qs64_80_lo4_offset16;
55 qs64_80_u32_hi4 += qs64_80_hi4_offset16;
56
57 const vec4 qs0_16_lo4 = vec4(unpack8(qs0_16_u32_lo4));
58 const vec4 qs64_80_lo4 = vec4(unpack8(qs64_80_u32_lo4));
59 const vec4 qs0_16_hi4 = vec4(unpack8(qs0_16_u32_hi4));
60 const vec4 qs64_80_hi4 = vec4(unpack8(qs64_80_u32_hi4));
61
62 const FLOAT_TYPE q4_0 = qs0_16_lo4.x;
63 const FLOAT_TYPE q4_1 = qs0_16_lo4.y;
64 const FLOAT_TYPE q4_2 = qs0_16_lo4.z;
65 const FLOAT_TYPE q4_3 = qs0_16_lo4.w;
66 const FLOAT_TYPE q4_4 = qs0_16_hi4.x;
67 const FLOAT_TYPE q4_5 = qs0_16_hi4.y;
68 const FLOAT_TYPE q4_6 = qs0_16_hi4.z;
69 const FLOAT_TYPE q4_7 = qs0_16_hi4.w;
70 const FLOAT_TYPE q4_8 = qs64_80_lo4.x;
71 const FLOAT_TYPE q4_9 = qs64_80_lo4.y;
72 const FLOAT_TYPE q4_10 = qs64_80_lo4.z;
73 const FLOAT_TYPE q4_11 = qs64_80_lo4.w;
74 const FLOAT_TYPE q4_12 = qs64_80_hi4.x;
75 const FLOAT_TYPE q4_13 = qs64_80_hi4.y;
76 const FLOAT_TYPE q4_14 = qs64_80_hi4.z;
77 const FLOAT_TYPE q4_15 = qs64_80_hi4.w;
78
79 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
80 vec2 by10 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 ]);
81 vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8]);
82 vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]);
83 vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]);
84 vec2 by20 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 ]);
85 vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8]);
86 vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]);
87 vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]);
88
89 const FLOAT_TYPE sx =
90 fma(FLOAT_TYPE(by10.x), q4_0,
91 fma(FLOAT_TYPE(by10.y), q4_1,
92 fma(FLOAT_TYPE(by116.x), q4_2,
93 FLOAT_TYPE(by116.y) * q4_3)));
94 const FLOAT_TYPE sy =
95 fma(FLOAT_TYPE(by132.x), q4_4,
96 fma(FLOAT_TYPE(by132.y), q4_5,
97 fma(FLOAT_TYPE(by148.x), q4_6,
98 FLOAT_TYPE(by148.y) * q4_7)));
99 const FLOAT_TYPE sz =
100 fma(FLOAT_TYPE(by20.x), q4_8,
101 fma(FLOAT_TYPE(by20.y), q4_9,
102 fma(FLOAT_TYPE(by216.x), q4_10,
103 FLOAT_TYPE(by216.y) * q4_11)));
104 const FLOAT_TYPE sw =
105 fma(FLOAT_TYPE(by232.x), q4_12,
106 fma(FLOAT_TYPE(by232.y), q4_13,
107 fma(FLOAT_TYPE(by248.x), q4_14,
108 FLOAT_TYPE(by248.y) * q4_15)));
109 const FLOAT_TYPE smin =
110 fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
111 fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
112 fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
113 (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
114 temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
115 }
116 }
117}
118
119void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
120 uint a_offset, b_offset, d_offset;
121 get_offsets(a_offset, b_offset, d_offset);
122
123 const uint num_blocks_per_row = p.ncols / QUANT_K;
124
125 // 16 threads are used to process each block
126 const uint it_size = gl_WorkGroupSize.x/16;
127 const uint tid = gl_LocalInvocationID.x;
128 const uint itid = tid%16; // 0...15
129 const uint ix = tid/16;
130
131 const uint il = itid/4; // 0...3
132 const uint ir = itid - 4*il; // 0...3
133
134 const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
135 const uint v_in = il % 2;
136
137 const uint l0 = 4*ir + 2*v_in; // 0...15
138 const uint q_offset = 32*v_im + l0;
139 const uint y_offset = 64*v_im + l0;
140
141 [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
142 [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
143 temp[j][i] = FLOAT_TYPE(0);
144 }
145 }
146
147 [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size)
148 calc_superblock(a_offset, b_offset, v_im, l0, q_offset, y_offset, i, num_blocks_per_row, first_row, num_rows);
149
150 reduce_result(temp, d_offset, first_row, num_rows, tid);
151}
152
153void main() {
154 const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
155
156 // do NUM_ROWS at a time, unless there aren't enough remaining rows
157 if (first_row + NUM_ROWS <= p.stride_d) {
158 compute_outputs(first_row, NUM_ROWS);
159 } else {
160 if (first_row >= p.stride_d) {
161 return;
162 }
163 compute_outputs(first_row, p.stride_d - first_row);
164 }
165}