From fb882e80f6e06d512885ebf7b8231c1be50e0eb2 Mon Sep 17 00:00:00 2001 From: Arm Patinyasakdikul Date: Tue, 19 Aug 2025 20:24:31 -0500 Subject: [PATCH] Remove noinline attribute from reduceCopyPacks and (#1864) reduceCopyPacksWithBias. --- src/device/common_kernel.h | 8 -------- 1 file changed, 8 deletions(-) diff --git a/src/device/common_kernel.h b/src/device/common_kernel.h index 98113702fc..31dd2b05d3 100644 --- a/src/device/common_kernel.h +++ b/src/device/common_kernel.h @@ -32,11 +32,7 @@ template -#if defined(__gfx942__) || defined(__gfx950__) __device__ __forceinline__ void reduceCopyPacks( -#else -__device__ __attribute__((noinline)) void reduceCopyPacks( -#endif int nThreads, int &thread, uint64_t redArg, uint64_t *preOpArgs, bool postOp, int nSrcs, SrcPtrFn const &srcPtrFn, int nDsts, DstPtrFn const &dstPtrFn, @@ -428,11 +424,7 @@ template -#if defined(__gfx942__) || defined(__gfx950__) __device__ __forceinline__ void reduceCopyPacksWithBias( -#else -__device__ __attribute__((noinline)) void reduceCopyPacksWithBias( -#endif int nThreads, int &thread, uint64_t redArg, uint64_t *preOpArgs, bool postOp, int nSrcs, SrcPtrFn const &srcPtrFn, int nDsts, DstPtrFn const &dstPtrFn,