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

9
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View File

@@ -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"
@@ -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