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

ggml-cpu: use LUT for converting e8->f32 scales on x86 (#19288)

Browse Source 
* ggml-cpu: use LUT for converting e8->f32 scales on x86

* add dispatch based on macro
This commit is contained in:
Aman Gupta
2026-02-04 09:43:29 +08:00
committed by GitHub
parent 44008ce8f9
commit 2ceda3f662
3 changed files with 28 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/simd-mappings.h
+11
View File
@@ -116,6 +116,17 @@ extern "C" {
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
extern float ggml_table_f32_f16[1 << 16];
// precomputed f32 table for e8m0 half (1 KB)
// defined in ggml-cpu.c, initialized in ggml_cpu_init()
extern float ggml_table_f32_e8m0_half[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)]
#else
#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(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.