Skip to content

LLM inference in C/C++

C++ ggml

metal: FWHT perf optimizations (#29602)

bri-prism18bbc46b4697
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
index 8bb70d0777..eb1868f62e 100644
--- a/ggml/src/ggml-metal/ggml-metal-impl.h
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -1242,7 +1242,7 @@ typedef struct {
// widths at or above this use the threadgroup FWHT kernel, one row per threadgroup
// with GGML_METAL_FWHT_TG_NT threads, instead of one row per simdgroup
-#define GGML_METAL_FWHT_TG_MIN_N 1024
+#define GGML_METAL_FWHT_TG_MIN_N 512
#define GGML_METAL_FWHT_TG_NT 256
typedef struct {
diff --git a/ggml/src/ggml-metal/kernels/misc.metal b/ggml/src/ggml-metal/kernels/misc.metal
index d29786e1c6..fb27983db3 100644
--- a/ggml/src/ggml-metal/kernels/misc.metal
+++ b/ggml/src/ggml-metal/kernels/misc.metal
@@ -541,18 +541,18 @@ typedef decltype(kernel_fwht<64, half>) kernel_fwht_f16_t;
template [[host_name("kernel_fwht_f32_64")]] kernel kernel_fwht_f32_t kernel_fwht<64, float>;
template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_f32_t kernel_fwht<128, float>;
template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_f32_t kernel_fwht<256, float>;
-template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_f32_t kernel_fwht<512, float>;
template [[host_name("kernel_fwht_f16_64")]] kernel kernel_fwht_f16_t kernel_fwht<64, half>;
template [[host_name("kernel_fwht_f16_128")]] kernel kernel_fwht_f16_t kernel_fwht<128, half>;
template [[host_name("kernel_fwht_f16_256")]] kernel kernel_fwht_f16_t kernel_fwht<256, half>;
-template [[host_name("kernel_fwht_f16_512")]] kernel kernel_fwht_f16_t kernel_fwht<512, half>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_f32_t kernel_fwht_tg<512, GGML_METAL_FWHT_TG_NT, float>;
template [[host_name("kernel_fwht_f32_1024")]] kernel kernel_fwht_f32_t kernel_fwht_tg<1024, GGML_METAL_FWHT_TG_NT, float>;
template [[host_name("kernel_fwht_f32_2048")]] kernel kernel_fwht_f32_t kernel_fwht_tg<2048, GGML_METAL_FWHT_TG_NT, float>;
template [[host_name("kernel_fwht_f32_4096")]] kernel kernel_fwht_f32_t kernel_fwht_tg<4096, GGML_METAL_FWHT_TG_NT, float>;
template [[host_name("kernel_fwht_f32_8192")]] kernel kernel_fwht_f32_t kernel_fwht_tg<8192, GGML_METAL_FWHT_TG_NT, float>;
+template [[host_name("kernel_fwht_f16_512")]] kernel kernel_fwht_f16_t kernel_fwht_tg<512, GGML_METAL_FWHT_TG_NT, half>;
template [[host_name("kernel_fwht_f16_1024")]] kernel kernel_fwht_f16_t kernel_fwht_tg<1024, GGML_METAL_FWHT_TG_NT, half>;
template [[host_name("kernel_fwht_f16_2048")]] kernel kernel_fwht_f16_t kernel_fwht_tg<2048, GGML_METAL_FWHT_TG_NT, half>;
template [[host_name("kernel_fwht_f16_4096")]] kernel kernel_fwht_f16_t kernel_fwht_tg<4096, GGML_METAL_FWHT_TG_NT, half>;