From b30ef05f7bf17e33591b573d62640fa3a21a2519 Mon Sep 17 00:00:00 2001
From: bri-prism <288398250+bri-prism@users.noreply.github.com>
Date: Tue, 29 Sep 2026 00:29:51 -0700
Subject: [PATCH] metal: FWHT perf optimizations (llama/29602)

Assisted-by: Claude Code
---
 ggml/src/ggml-metal/ggml-metal-impl.h  | 2 +-
 ggml/src/ggml-metal/kernels/misc.metal | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
index 8bb70d077..eb1868f62 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 d29786e1c..fb27983db 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>;