ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

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

Browse source 
* Vulkan: Remove float16 use in shaders

* Fix validation error about subgroup_size_control extension
This commit is contained in:
0cc4m 2025-01-10 06:39:33 +01:00 • committed by GitHub
parent ee7136c6d1
commit c3f9d25706
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 50 additions and 51 deletions
Download patch file Download diff file Expand all files Collapse all files

12
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
View file

@ -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"
@ -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);
@ -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)));