Skip to content

Commit 7b820e7

Browse files
remyoudomphengremyoudompheng
committed
vulkan: add MMV kernels for IQ3 quants
1 parent cc99d98 commit 7b820e7

File tree

3 files changed

+179
-1
lines changed

3 files changed

+179
-1
lines changed

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

Lines changed: 90 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,90 @@
1+
#version 450
2+
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
3+
4+
#include "mul_mat_vec_base.comp"
5+
6+
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
7+
8+
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
9+
10+
void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, 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 + 32 * ib32;
12+
13+
uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
14+
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
15+
const float d = float(data_a[ibi].d);
16+
const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
17+
const float dscale = d * (1 + 2 * scale);
18+
const uint qh = data_a[ibi].qh[ib32];
19+
FLOAT_TYPE sum[NUM_COLS];
20+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
21+
sum[j] = 0.0;
22+
}
23+
[[unroll]] for (uint l = 0; l < 4; ++l) {
24+
const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
25+
const uint sign = data_a[ibi].signs[4 * ib32 + l];
26+
const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
27+
const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));
28+
29+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
30+
const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
31+
const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
32+
33+
sum[j] =
34+
fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
35+
fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
36+
fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
37+
fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
38+
fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
39+
fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
40+
fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
41+
fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
42+
sum[j]))))))));
43+
}
44+
}
45+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
46+
temp[j][n] = fma(dscale, sum[j], temp[j][n]);
47+
}
48+
ibi += num_blocks_per_row;
49+
}
50+
}
51+
52+
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
53+
uint a_offset, b_offset, d_offset;
54+
get_offsets(a_offset, b_offset, d_offset);
55+
56+
const uint num_blocks_per_row = p.ncols / QUANT_K;
57+
58+
// 8 threads are used to process each block
59+
const uint blocks_per_wg = gl_WorkGroupSize.x/8;
60+
const uint tid = gl_LocalInvocationID.x;
61+
const uint itid = tid % 8; // 0...7
62+
const uint ix = tid / 8;
63+
64+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
65+
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
66+
temp[j][i] = FLOAT_TYPE(0);
67+
}
68+
}
69+
70+
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
71+
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
72+
73+
reduce_result(temp, d_offset, first_row, num_rows, tid);
74+
}
75+
76+
void main() {
77+
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
78+
79+
init_iq_shmem(gl_WorkGroupSize);
80+
81+
// do NUM_ROWS at a time, unless there aren't enough remaining rows
82+
if (first_row + NUM_ROWS <= p.stride_d) {
83+
compute_outputs(first_row, NUM_ROWS);
84+
} else {
85+
if (first_row >= p.stride_d) {
86+
return;
87+
}
88+
compute_outputs(first_row, p.stride_d - first_row);
89+
}
90+
}

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp

Lines changed: 88 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,88 @@
1+
#version 450
2+
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
3+
4+
#include "mul_mat_vec_base.comp"
5+
6+
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
7+
8+
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
9+
10+
void 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 ib32 = itid / 2; // 0..7
13+
14+
uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
15+
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
16+
const float d = float(data_a[ibi].d);
17+
const uint signscale = pack32(u16vec2(
18+
data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32],
19+
data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32 + 1]));
20+
const float db = d * 0.5 * (0.5 + (signscale >> 28));
21+
[[unroll]] for (uint l = 0; l < 2; ++l) {
22+
const uint qs0 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l];
23+
const uint qs1 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l + 1];
24+
const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7);
25+
const uint sign7 = bitCount(sign);
26+
const vec4 grid0 = vec4(unpack8(iq3xxs_grid[qs0]));
27+
const vec4 grid1 = vec4(unpack8(iq3xxs_grid[qs1]));
28+
29+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
30+
const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
31+
const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
32+
33+
FLOAT_TYPE sum =
34+
fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
35+
fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
36+
fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
37+
fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
38+
fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
39+
fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
40+
fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
41+
fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w),
42+
FLOAT_TYPE(0.0)))))))));
43+
temp[j][n] = fma(db, sum, temp[j][n]);
44+
}
45+
}
46+
ibi += num_blocks_per_row;
47+
}
48+
}
49+
50+
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
51+
uint a_offset, b_offset, d_offset;
52+
get_offsets(a_offset, b_offset, d_offset);
53+
54+
const uint num_blocks_per_row = p.ncols / QUANT_K;
55+
56+
// 16 threads are used to process each block
57+
const uint blocks_per_wg = gl_WorkGroupSize.x/16;
58+
const uint tid = gl_LocalInvocationID.x;
59+
const uint itid = tid % 16; // 0...15
60+
const uint ix = tid / 16;
61+
62+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
63+
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
64+
temp[j][i] = FLOAT_TYPE(0);
65+
}
66+
}
67+
68+
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
69+
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
70+
71+
reduce_result(temp, d_offset, first_row, num_rows, tid);
72+
}
73+
74+
void main() {
75+
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
76+
77+
init_iq_shmem(gl_WorkGroupSize);
78+
79+
// do NUM_ROWS at a time, unless there aren't enough remaining rows
80+
if (first_row + NUM_ROWS <= p.stride_d) {
81+
compute_outputs(first_row, NUM_ROWS);
82+
} else {
83+
if (first_row >= p.stride_d) {
84+
return;
85+
}
86+
compute_outputs(first_row, p.stride_d - first_row);
87+
}
88+
}

‎ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -396,7 +396,7 @@ void process_shaders() {
396396
for (const auto& tname : type_names) {
397397
// mul mat vec
398398
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
399-
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
399+
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
400400

401401
string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
402402
string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));

0 commit comments

Comments
 (0)