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

ggml-cpu: add AVX2 optimization for nvfp4 dot product and use UE4M3 LUT (#23961)

Browse Source
This commit is contained in:
ragz4125
2026-07-01 15:31:20 +08:00
committed by GitHub
parent 9d88e7cedd
commit 6dbc1174b8
5 changed files with 164 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/simd-mappings.h
+11
View File
@@ -120,6 +120,10 @@ extern float ggml_table_f32_f16[1 << 16];
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
extern float ggml_table_f32_e8m0_half[1 << 8];
// precomputed f32 table for ue4m3 (1 KB)
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
extern float ggml_table_f32_ue4m3[1 << 8];
// Use lookup table for E8M0 on x86 (faster than bit manipulation)
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
@@ -127,6 +131,13 @@ extern float ggml_table_f32_e8m0_half[1 << 8];
#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
#endif
// Use lookup table for UE4M3 on x86 (faster than bit manipulation)
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
#define GGML_CPU_UE4M3_TO_FP32(x) ggml_table_f32_ue4m3[(uint8_t)(x)]
#else
#define GGML_CPU_UE4M3_TO_FP32(x) ggml_ue4m3_to_fp32(x)
#endif
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.