vulkan: define MMV kernels for IQ1 quantizations

This commit is contained in:
Rémy O 2025-02-01 19:27:21 +01:00
parent fa92caae18
commit ab2f56e270
3 changed files with 169 additions and 1 deletions

View file

@ -0,0 +1,82 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#include "mul_mat_vec_base.comp"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
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) {
const uint y_idx = i * QUANT_K + 32 * ib32;
uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint16_t[4] scales = data_a[ibi].scales;
const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
const float d = float(uint16BitsToHalf(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)));
const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
[[unroll]] for (uint l = 0; l < 4; ++l) {
const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
const uint qs = data_a[ibi].qs[4 * ib32 + l];
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int k = 0; k < 4; ++k) {
sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
}
temp[j][n] = fma(dl, sum, temp[j][n]);
}
}
ibi += num_blocks_per_row;
}
}
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
// 8 threads are used to process each block
const uint blocks_per_wg = gl_WorkGroupSize.x/8;
const uint tid = gl_LocalInvocationID.x;
const uint itid = tid % 8; // 0...7
const uint ix = tid / 8;
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
init_iq_shmem(gl_WorkGroupSize);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -0,0 +1,79 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#include "mul_mat_vec_base.comp"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
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) {
const uint y_idx = i * QUANT_K + 32 * ib32;
uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const float d = float(data_a[ibi].d);
const uint qh = data_a[ibi].qh[ib32];
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
[[unroll]] for (uint l = 0; l < 4; ++l) {
const uint qs = data_a[ibi].qs[4 * ib32 + l];
const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int k = 0; k < 4; ++k) {
sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
}
temp[j][n] = fma(dl, sum, temp[j][n]);
}
}
ibi += num_blocks_per_row;
}
}
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
const uint num_blocks_per_row = p.ncols / QUANT_K;
// 8 threads are used to process each block
const uint blocks_per_wg = gl_WorkGroupSize.x/8;
const uint tid = gl_LocalInvocationID.x;
const uint itid = tid % 8; // 0...7
const uint ix = tid / 8;
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
init_iq_shmem(gl_WorkGroupSize);
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}

View file

@ -184,6 +184,13 @@ std::string to_uppercase(const std::string& input) {
return result;
}
bool string_starts_with(const std::string& str, const std::string& prefix) {
if (prefix.size() > str.size()) {
return false;
}
return std::equal(prefix.begin(), prefix.end(), str.begin());
}
bool string_ends_with(const std::string& str, const std::string& suffix) {
if (suffix.size() > str.size()) {
return false;
@ -389,7 +396,7 @@ void process_shaders() {
for (const auto& tname : type_names) {
// mul mat vec
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
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"}}));
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"}}));