vulkan: implement specialized MMV kernels for IQ2 quantizations
This commit is contained in:
Rémy O
parent
07820effd9
commit
cc99d98e87
6 changed files with 275 additions and 2 deletions
|
|
@ -29,7 +29,7 @@ void main() {
|
|||
uint qs = data_a[ib].qs[4 * ib32 + l];
|
||||
const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
|
||||
qs |= (qh << (8 - 2 * l)) & 0x300;
|
||||
const uvec2 grid = iq2s_grid[qs & 511];
|
||||
const uvec2 grid = iq2s_grid[qs];
|
||||
const u8vec4 grid0 = unpack8(grid.x);
|
||||
const u8vec4 grid1 = unpack8(grid.y);
|
||||
data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));
|
||||
|
|
|
|||
90
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp
Normal file
90
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp
Normal file
|
|
@ -0,0 +1,90 @@
|
|||
#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
|
||||
const uint y_idx = i * QUANT_K + 16 * itid;
|
||||
const uint nibble_shift = 4 * (itid & 1);
|
||||
const uint ib32 = itid / 2; // 0..7
|
||||
|
||||
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 scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
|
||||
const float db = d * (0.5 + scale) * 0.25;
|
||||
|
||||
const uint qh = data_a[ibi].qh[ib32];
|
||||
const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
|
||||
const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
|
||||
[[unroll]] for (uint l = 0; l < 2; ++l) {
|
||||
const uint8_t sign = sign16[l];
|
||||
const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);
|
||||
const uvec2 grid = iq2s_grid[qs];
|
||||
const vec4 grid0 = vec4(unpack8(grid.x));
|
||||
const vec4 grid1 = vec4(unpack8(grid.y));
|
||||
|
||||
[[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 =
|
||||
fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
|
||||
fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
|
||||
fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
|
||||
fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
|
||||
fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
|
||||
fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
|
||||
fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
|
||||
fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
|
||||
FLOAT_TYPE(0.0)))))))));
|
||||
temp[j][n] = fma(db, 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;
|
||||
|
||||
// 16 threads are used to process each block
|
||||
const uint blocks_per_wg = gl_WorkGroupSize.x/16;
|
||||
const uint tid = gl_LocalInvocationID.x;
|
||||
const uint itid = tid % 16; // 0...15
|
||||
const uint ix = tid / 16;
|
||||
|
||||
[[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);
|
||||
}
|
||||
}
|
||||
87
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp
Normal file
87
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp
Normal file
|
|
@ -0,0 +1,87 @@
|
|||
#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
|
||||
const uint y_idx = i * QUANT_K + 16 * itid;
|
||||
const uint nibble_shift = 4 * (itid & 1);
|
||||
const uint ib32 = itid / 2; // 0..7
|
||||
|
||||
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 scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
|
||||
const float db = d * (0.5 + scale) * 0.25;
|
||||
|
||||
[[unroll]] for (uint l = 0; l < 2; ++l) {
|
||||
const uint qs = data_a[ibi].qs[2 * itid + l];
|
||||
const uint sign = qs >> 9;
|
||||
const uint sign7 = bitCount(sign);
|
||||
const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
|
||||
const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
|
||||
|
||||
[[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 =
|
||||
fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
|
||||
fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
|
||||
fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
|
||||
fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
|
||||
fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
|
||||
fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
|
||||
fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
|
||||
fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w),
|
||||
FLOAT_TYPE(0.0)))))))));
|
||||
temp[j][n] = fma(db, 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;
|
||||
|
||||
// 16 threads are used to process each block
|
||||
const uint blocks_per_wg = gl_WorkGroupSize.x/16;
|
||||
const uint tid = gl_LocalInvocationID.x;
|
||||
const uint itid = tid % 16; // 0...15
|
||||
const uint ix = tid / 16;
|
||||
|
||||
[[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);
|
||||
}
|
||||
}
|
||||
87
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp
Normal file
87
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp
Normal file
|
|
@ -0,0 +1,87 @@
|
|||
#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
|
||||
const uint y_idx = i * QUANT_K + 16 * itid;
|
||||
const uint ib32 = itid / 2; // 0..7
|
||||
|
||||
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 signscale = pack32(u16vec2(
|
||||
data_a_packed16[ibi].qs[4 * ib32 + 2],
|
||||
data_a_packed16[ibi].qs[4 * ib32 + 3]));
|
||||
const float db = d * 0.25 * (0.5 + (signscale >> 28));
|
||||
[[unroll]] for (uint l = 0; l < 2; ++l) {
|
||||
const uint qs = data_a[ibi].qs[8 * ib32 + 2 * (itid & 1) + l];
|
||||
const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7);
|
||||
const uint sign7 = bitCount(sign);
|
||||
const vec4 grid0 = vec4(unpack8(iq2xxs_grid[qs].x));
|
||||
const vec4 grid1 = vec4(unpack8(iq2xxs_grid[qs].y));
|
||||
|
||||
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
|
||||
const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
|
||||
const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
|
||||
|
||||
FLOAT_TYPE sum =
|
||||
fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x),
|
||||
fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y),
|
||||
fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z),
|
||||
fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w),
|
||||
fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x),
|
||||
fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y),
|
||||
fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z),
|
||||
fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w),
|
||||
FLOAT_TYPE(0.0)))))))));
|
||||
temp[j][n] = fma(db, 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;
|
||||
|
||||
// 16 threads are used to process each block
|
||||
const uint blocks_per_wg = gl_WorkGroupSize.x/16;
|
||||
const uint tid = gl_LocalInvocationID.x;
|
||||
const uint itid = tid % 16; // 0...15
|
||||
const uint ix = tid / 16;
|
||||
|
||||
[[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);
|
||||
}
|
||||
}
|
||||
|
|
@ -753,6 +753,14 @@ struct block_iq2_s
|
|||
uint8_t scales[QUANT_K_IQ2_S/32];
|
||||
};
|
||||
|
||||
struct block_iq2_s_packed16
|
||||
{
|
||||
float16_t d;
|
||||
uint16_t qs[QUANT_K_IQ2_S/8];
|
||||
uint16_t qh[QUANT_K_IQ2_S/64];
|
||||
uint16_t scales[QUANT_K_IQ2_S/64];
|
||||
};
|
||||
|
||||
#if defined(DATA_A_IQ2_S)
|
||||
|
||||
const uvec2 iq2s_grid_const[1024] = {
|
||||
|
|
@ -1028,6 +1036,7 @@ void init_iq_shmem(uvec3 wgsize)
|
|||
#define QUANT_K QUANT_K_IQ2_S
|
||||
#define QUANT_R QUANT_R_IQ2_S
|
||||
#define A_TYPE block_iq2_s
|
||||
#define A_TYPE_PACKED16 block_iq2_s_packed16
|
||||
#endif
|
||||
|
||||
#define QUANT_K_IQ3_XXS 256
|
||||
|
|
|
|||
|
|
@ -396,7 +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") || string_starts_with(tname, "iq1_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
|
||||
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";
|
||||
|
||||
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"}}));
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue