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

ggml: move fp16/bf16 conversion optimizations to CPU backend + export conversion APIs (#13107)

Browse source 
* ggml: dynamic x86_64 feature detection for FP32 <-> FP16/BF16 conversion

* move fp converter to ggml-cpu

* Switch ggml_compute_forward_get_rows_f16/bf16 to new ggml_cpu_fp16/bf16_to_fp32
This commit is contained in:
SXX 2025-04-26 22:05:31 +08:00 • committed by GitHub
parent d5fe4e81bd
commit 77d5e9a76a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 101 additions and 50 deletions
Download patch file Download diff file Expand all files Collapse all files

51
ggml/src/ggml.c
View file

@ -4,6 +4,7 @@
#include "ggml-backend.h"
#include "ggml-impl.h"
#include "ggml-threading.h"
#include "ggml-cpu.h"
#include "ggml.h"
// FIXME: required here for quantization functions
@ -382,58 +383,16 @@ void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
}
}
// FIXME: these functions must detect the instruction set at runtime, since they are part of the core ggml library
// currently, the ggml_cpu_has_* functions are entirely compile-time
void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
int64_t i = 0;
#if defined(__F16C__)
//if (ggml_cpu_has_f16c()) {
for (; i + 7 < n; i += 8) {
__m256 x_vec = _mm256_loadu_ps(x + i);
__m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
_mm_storeu_si128((__m128i *)(y + i), y_vec);
}
for(; i + 3 < n; i += 4) {
__m128 x_vec = _mm_loadu_ps(x + i);
__m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
_mm_storel_epi64((__m128i *)(y + i), y_vec);
}
//}
#endif
for (; i < n; i++) {
int i = 0;
for (; i < n; ++i) {
y[i] = GGML_FP32_TO_FP16(x[i]);
}
}
void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
int64_t i = 0;
#if defined(__AVX512F__)
//if (ggml_cpu_has_avx512()) {
for (; i + 16 <= n; i += 16) {
_mm512_storeu_ps(y + i,
_mm512_castsi512_ps(
_mm512_slli_epi32(
_mm512_cvtepu16_epi32(
_mm256_loadu_si256(
(const __m256i *)(x + i))),
16)));
}
//}
#endif
#if defined(__AVX2__)
//if (ggml_cpu_has_avx2()) {
for (; i + 8 <= n; i += 8) {
_mm256_storeu_ps(y + i,
_mm256_castsi256_ps(
_mm256_slli_epi32(
_mm256_cvtepu16_epi32(
_mm_loadu_si128(
(const __m128i *)(x + i))),
16)));
}
//}
#endif
for (; i < n; i++) {
int i = 0;
for (; i < n; ++i) {
y[i] = GGML_BF16_TO_FP32(x[i]);
}
}