Skip to content

Vulkan: Fix float16 use on devices without float16 support + fix subgroup_size_control validation error #11161

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jan 10, 2025
Merged

Vulkan: Fix float16 use on devices without float16 support + fix subgroup_size_control validation error #11161

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5a39219
Vulkan: Remove float16 use in shaders
0cc4m Jan 9, 2025
914a82d
Fix validation error about subgroup_size_control extension
0cc4m Jan 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2277,6 +2277,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
if (device->subgroup_size_control) {
device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
device_extensions.push_back("VK_EXT_subgroup_size_control");
}

device->subgroup_size_control = device->subgroup_size_control &&
Expand All @@ -2285,7 +2286,6 @@ static vk_device ggml_vk_get_device(size_t idx) {

if (device->subgroup_size_control) {
device->subgroup_require_full_support = subgroup_size_control_features.computeFullSubgroups;
device_extensions.push_back("VK_EXT_subgroup_size_control");
}

#if defined(VK_KHR_cooperative_matrix)
Expand Down
9 changes: 3 additions & 6 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,9 +1,6 @@
#version 450

#ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif
#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand All @@ -27,8 +24,8 @@ void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const

#if K_PER_ITER == 8
#if QUANT_R == 2
const B_TYPE_VEC4 bv02 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4];
const B_TYPE_VEC4 bv13 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4];
const vec4 bv02 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
const vec4 bv13 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4]);
const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
#else
Expand Down
24 changes: 12 additions & 12 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand Down Expand Up @@ -40,9 +40,9 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {

[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;
vec2 d = vec2(data_a[ib0 + i].d);
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);

uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
Expand All @@ -63,14 +63,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec2 qs16 = uvec2(unpack8(qs16_u16));

[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]);
vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]);
vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);

FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
Expand Down
18 changes: 9 additions & 9 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand Down Expand Up @@ -60,14 +60,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {

[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {

B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
vec2 b0 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0]);
vec2 b16 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8]);
vec2 b32 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16]);
vec2 b48 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24]);
vec2 b64 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32]);
vec2 b80 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40]);
vec2 b96 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48]);
vec2 b112 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56]);

FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
Expand Down
12 changes: 6 additions & 6 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
#version 450

#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand Down Expand Up @@ -45,7 +45,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {

[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
vec2 d = vec2(data_a[ib0 + i].d);
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);

Expand Down Expand Up @@ -96,10 +96,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint32_t q4_15 = qs64_hi4.w;

[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC4 by10 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4];
B_TYPE_VEC4 by132 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8];
B_TYPE_VEC4 by20 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4];
B_TYPE_VEC4 by232 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8];
vec4 by10 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 ]);
vec4 by132 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8]);
vec4 by20 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 ]);
vec4 by232 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8]);

const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x), q4_0, fma(FLOAT_TYPE(by10.y), q4_1, fma(FLOAT_TYPE(by10.z), q4_2, FLOAT_TYPE(by10.w) * q4_3)));
const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x), q4_4, fma(FLOAT_TYPE(by132.y), q4_5, fma(FLOAT_TYPE(by132.z), q4_6, FLOAT_TYPE(by132.w) * q4_7)));
Expand Down
20 changes: 10 additions & 10 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
#version 450

#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand Down Expand Up @@ -42,7 +42,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {

[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
vec2 d = vec2(data_a[ib0 + i].d);
const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);

Expand Down Expand Up @@ -105,14 +105,14 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint32_t q4_15 = qs64_80_hi4.w;

[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 by10 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2];
B_TYPE_VEC2 by116 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8];
B_TYPE_VEC2 by132 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16];
B_TYPE_VEC2 by148 = data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24];
B_TYPE_VEC2 by20 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2];
B_TYPE_VEC2 by216 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8];
B_TYPE_VEC2 by232 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16];
B_TYPE_VEC2 by248 = data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24];
vec2 by10 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 ]);
vec2 by116 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 8]);
vec2 by132 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 16]);
vec2 by148 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y1_idx) / 2 + 24]);
vec2 by20 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 ]);
vec2 by216 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 8]);
vec2 by232 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 16]);
vec2 by248 = vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y2_idx) / 2 + 24]);

const FLOAT_TYPE sx =
fma(FLOAT_TYPE(by10.x), q4_0,
Expand Down
10 changes: 5 additions & 5 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
#version 450

#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require

#include "mul_mat_vec_base.comp"

Expand Down Expand Up @@ -77,10 +77,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec4 q3 = uvec4(unpack8(q3_u32));

[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC4 by0 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4];
B_TYPE_VEC4 by32 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8];
B_TYPE_VEC4 by64 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16];
B_TYPE_VEC4 by96 = data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24];
vec4 by0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 ]);
vec4 by32 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 8]);
vec4 by64 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 16]);
vec4 by96 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 24]);

FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 4; ++l) {
Expand Down
1 change: 0 additions & 1 deletion ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
#version 450

#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#extension GL_EXT_control_flow_attributes : enable

layout (push_constant) uniform parameter
Expand Down
5 changes: 4 additions & 1 deletion ggml/src/ggml-vulkan/vulkan-shaders/types.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,10 @@
#if !defined(GGML_TYPES_COMP)
#define GGML_TYPES_COMP

#extension GL_EXT_shader_explicit_arithmetic_types : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
#extension GL_EXT_shader_16bit_storage : require

#if defined(DATA_A_F32)
#define QUANT_K 1
Expand Down
Loading