Skip to content

Commit ab2f56e

Browse files
remyoudomphengremyoudompheng
committed
vulkan: define MMV kernels for IQ1 quantizations
1 parent fa92caa commit ab2f56e

File tree

3 files changed

+169
-1
lines changed

3 files changed

+169
-1
lines changed

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

Lines changed: 82 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,82 @@
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 uint16_t[4] scales = data_a[ibi].scales;
16+
const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
17+
const float d = float(uint16BitsToHalf(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)));
18+
19+
const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
20+
[[unroll]] for (uint l = 0; l < 4; ++l) {
21+
const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
22+
const uint qs = data_a[ibi].qs[4 * ib32 + l];
23+
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
24+
const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
25+
26+
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
27+
28+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
29+
vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
30+
vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
31+
32+
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
33+
[[unroll]] for (int k = 0; k < 4; ++k) {
34+
sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
35+
fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
36+
}
37+
temp[j][n] = fma(dl, sum, temp[j][n]);
38+
}
39+
}
40+
ibi += num_blocks_per_row;
41+
}
42+
}
43+
44+
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
45+
uint a_offset, b_offset, d_offset;
46+
get_offsets(a_offset, b_offset, d_offset);
47+
48+
const uint num_blocks_per_row = p.ncols / QUANT_K;
49+
50+
// 8 threads are used to process each block
51+
const uint blocks_per_wg = gl_WorkGroupSize.x/8;
52+
const uint tid = gl_LocalInvocationID.x;
53+
const uint itid = tid % 8; // 0...7
54+
const uint ix = tid / 8;
55+
56+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
57+
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
58+
temp[j][i] = FLOAT_TYPE(0);
59+
}
60+
}
61+
62+
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
63+
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
64+
65+
reduce_result(temp, d_offset, first_row, num_rows, tid);
66+
}
67+
68+
void main() {
69+
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
70+
71+
init_iq_shmem(gl_WorkGroupSize);
72+
73+
// do NUM_ROWS at a time, unless there aren't enough remaining rows
74+
if (first_row + NUM_ROWS <= p.stride_d) {
75+
compute_outputs(first_row, NUM_ROWS);
76+
} else {
77+
if (first_row >= p.stride_d) {
78+
return;
79+
}
80+
compute_outputs(first_row, p.stride_d - first_row);
81+
}
82+
}

‎ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

Lines changed: 79 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,79 @@
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 qh = data_a[ibi].qh[ib32];
17+
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
18+
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
19+
20+
[[unroll]] for (uint l = 0; l < 4; ++l) {
21+
const uint qs = data_a[ibi].qs[4 * ib32 + l];
22+
const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
23+
const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
24+
25+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
26+
vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
27+
vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
28+
29+
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
30+
[[unroll]] for (int k = 0; k < 4; ++k) {
31+
sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
32+
fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
33+
}
34+
temp[j][n] = fma(dl, sum, temp[j][n]);
35+
}
36+
}
37+
ibi += num_blocks_per_row;
38+
}
39+
}
40+
41+
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
42+
uint a_offset, b_offset, d_offset;
43+
get_offsets(a_offset, b_offset, d_offset);
44+
45+
const uint num_blocks_per_row = p.ncols / QUANT_K;
46+
47+
// 8 threads are used to process each block
48+
const uint blocks_per_wg = gl_WorkGroupSize.x/8;
49+
const uint tid = gl_LocalInvocationID.x;
50+
const uint itid = tid % 8; // 0...7
51+
const uint ix = tid / 8;
52+
53+
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
54+
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
55+
temp[j][i] = FLOAT_TYPE(0);
56+
}
57+
}
58+
59+
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
60+
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
61+
62+
reduce_result(temp, d_offset, first_row, num_rows, tid);
63+
}
64+
65+
void main() {
66+
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
67+
68+
init_iq_shmem(gl_WorkGroupSize);
69+
70+
// do NUM_ROWS at a time, unless there aren't enough remaining rows
71+
if (first_row + NUM_ROWS <= p.stride_d) {
72+
compute_outputs(first_row, NUM_ROWS);
73+
} else {
74+
if (first_row >= p.stride_d) {
75+
return;
76+
}
77+
compute_outputs(first_row, p.stride_d - first_row);
78+
}
79+
}

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

Lines changed: 8 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -184,6 +184,13 @@ std::string to_uppercase(const std::string& input) {
184184
return result;
185185
}
186186

187+
bool string_starts_with(const std::string& str, const std::string& prefix) {
188+
if (prefix.size() > str.size()) {
189+
return false;
190+
}
191+
return std::equal(prefix.begin(), prefix.end(), str.begin());
192+
}
193+
187194
bool string_ends_with(const std::string& str, const std::string& suffix) {
188195
if (suffix.size() > str.size()) {
189196
return false;
@@ -389,7 +396,7 @@ void process_shaders() {
389396
for (const auto& tname : type_names) {
390397
// mul mat vec
391398
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
392-
std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
399+
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
393400

394401
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"}}));
395402
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)