From 9937cc22dac63f95a60924bf0690288760056a1a Mon Sep 17 00:00:00 2001 From: Vincent Rabaud Date: Mon, 1 Jun 2026 13:16:58 +0200 Subject: [PATCH] Make sure pagedAttnAVGemmKernel uses no more than the stack Divide FAST_GEMM_MAX_STACKBUF by 2 because there are two AutoBuffers. --- .../dnn/src/layers/cpu_kernels/fast_gemm_kernels.default.hpp | 2 +- modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.simd.hpp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.default.hpp b/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.default.hpp index 38c2f66594..179086c440 100644 --- a/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.default.hpp +++ b/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.default.hpp @@ -13,7 +13,7 @@ #include // parallel_for_ #define FAST_GEMM_STORAGE (1<<20) // 2^20 -#define FAST_GEMM_MAX_STACKBUF (1 << 14) +#define FAST_GEMM_MAX_STACKBUF (1 << 13) #define FAST_GEMM_F32_MC 64 #define FAST_GEMM_F32_NC 240 diff --git a/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.simd.hpp b/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.simd.hpp index 47f91f2071..09a05068f9 100644 --- a/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.simd.hpp +++ b/modules/dnn/src/layers/cpu_kernels/fast_gemm_kernels.simd.hpp @@ -13,7 +13,7 @@ #include // parallel_for_ #define FAST_GEMM_STORAGE (1<<20) // 2^20 -#define FAST_GEMM_MAX_STACKBUF (1 << 14) +#define FAST_GEMM_MAX_STACKBUF (1 << 13) #if CV_AVX #define FAST_GEMM_F32_MC 60