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

vulkan: request round-to-even for fp16 in im2col/rope_head (#10767)

Browse source 
Vulkan doesn't mandate a specific rounding mode, but the shader_float_controls
feature allows rounding mode to be requested if the implementation supports it.
This commit is contained in:
Jeff Bolz 2024-12-10 14:23:17 -06:00 • committed by GitHub
parent dafae66cc2
commit b685daf386
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 31 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
View file

@ -1,6 +1,11 @@
#version 450
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_spirv_intrinsics: enable
#if RTE16
spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
#endif
layout (push_constant) uniform parameter
{

5
ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp
View file

@ -1,6 +1,11 @@
#include "types.comp"
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_spirv_intrinsics: enable
#if RTE16
spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
#endif
layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;

3
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View file

@ -461,9 +461,11 @@ void process_shaders() {
string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
@ -471,6 +473,7 @@ void process_shaders() {
string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
string_to_spv("im2col_f32_f16_rte", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}));
string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));