LLM inference in C/C++
C++ ggml
ggml-cpu: use UE4M3 LUT in ARM NVFP4 dot product (#25331)
ragz412520a04b220630
diff --git a/ggml/src/ggml-cpu/arch/arm/quants.c b/ggml/src/ggml-cpu/arch/arm/quants.cindex fe62133297..4453667976 100644--- a/ggml/src/ggml-cpu/arch/arm/quants.c+++ b/ggml/src/ggml-cpu/arch/arm/quants.c@@ -812,10 +812,10 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voconst float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);const float32x4_t nvsc = {- ggml_ue4m3_to_fp32(x[ib].d[0]),- ggml_ue4m3_to_fp32(x[ib].d[1]),- ggml_ue4m3_to_fp32(x[ib].d[2]),- ggml_ue4m3_to_fp32(x[ib].d[3])+ GGML_CPU_UE4M3_TO_FP32(x[ib].d[0]),+ GGML_CPU_UE4M3_TO_FP32(x[ib].d[1]),+ GGML_CPU_UE4M3_TO_FP32(x[ib].d[2]),+ GGML_CPU_UE4M3_TO_FP32(x[ib].d[3])};const float32x4_t scales = vmulq_f32(nvsc, (float32x4_t){dy0, dy0, dy1, dy1});diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.hindex be50c25c0a..fca5119e1a 100644--- a/ggml/src/ggml-cpu/simd-mappings.h+++ b/ggml/src/ggml-cpu/simd-mappings.h@@ -131,8 +131,8 @@ extern float ggml_table_f32_ue4m3[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__)+// Use lookup table for UE4M3 on x86 and ARM (faster than bit manipulation)+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__ARM_NEON)#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)