Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: add cpy bf16 -> f32 pipelines (#22677)

Browse Source
This commit is contained in:
Pascal
2026-05-17 11:31:20 +02:00
committed by GitHub
parent 7ba22c6a09
commit fcae601e44
4 changed files with 23 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp
+6 -2
View File
@@ -19,7 +19,9 @@ void main() {
if (idx + (num_iter-1)*num_threads < p.ne) {
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
#if defined(DATA_D_BF16)
#if defined(DATA_A_BF16)
data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx])));
#elif defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + idx]);
data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
@@ -35,7 +37,9 @@ void main() {
continue;
}
#if defined(DATA_D_BF16)
#if defined(DATA_A_BF16)
data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx])));
#elif defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + idx]);
data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
ggml/src/ggml-vulkan/vulkan-shaders/copy.comp
+3 -1
View File
@@ -12,7 +12,9 @@ void main() {
return;
}
#if defined(DATA_D_BF16)
#if defined(DATA_A_BF16)
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + src0_idx(idx)])));
#elif defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + src0_idx(idx)]);
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+2
View File
@@ -731,6 +731,7 @@ void process_shaders() {
string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("cpy_f16_f32", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
string_to_spv("cpy_bf16_f32","copy.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("contig_cpy_f32_i32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
string_to_spv("contig_cpy_i32_f32", "contig_copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
@@ -738,6 +739,7 @@ void process_shaders() {
string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("contig_cpy_f16_f32", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
string_to_spv("contig_cpy_bf16_f32","contig_copy.comp",{{"A_TYPE", "uint16_t"}, {"D_TYPE", "float"}, {"DATA_A_BF16", "1"}});
string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});