EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases 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
9 changed files with 50 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_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"
@@ -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);
@@ -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,