diff --git a/projects/hip/include/hip/hcc_detail/device_functions.h b/projects/hip/include/hip/hcc_detail/device_functions.h index a24df22315..b88df2e168 100644 --- a/projects/hip/include/hip/hcc_detail/device_functions.h +++ b/projects/hip/include/hip/hcc_detail/device_functions.h @@ -254,14 +254,22 @@ __device__ static inline float __hip_ds_permutef(int index, float src) { template __device__ static inline unsigned __hip_ds_swizzle_N(unsigned int src) { union { int i; unsigned u; float f; } tmp; tmp.u = src; +#if defined(__HCC__) + tmp.i = __llvm_amdgcn_ds_swizzle(tmp.i, pattern); +#else tmp.i = __builtin_amdgcn_ds_swizzle(tmp.i, pattern); +#endif return tmp.u; } template __device__ static inline float __hip_ds_swizzlef_N(float src) { union { int i; unsigned u; float f; } tmp; tmp.f = src; +#if defined(__HCC__) + tmp.i = __llvm_amdgcn_ds_swizzle(tmp.i, pattern); +#else tmp.i = __builtin_amdgcn_ds_swizzle(tmp.i, pattern); +#endif return tmp.f; } diff --git a/projects/hip/include/hip/hcc_detail/hip_runtime.h b/projects/hip/include/hip/hcc_detail/hip_runtime.h index 9e7bb618fe..62be5794fb 100644 --- a/projects/hip/include/hip/hcc_detail/hip_runtime.h +++ b/projects/hip/include/hip/hcc_detail/hip_runtime.h @@ -336,7 +336,7 @@ typedef int hipLaunchParm; #define hipLaunchKernel(kernelName, numblocks, numthreads, memperblock, streamId, ...) \ do { \ - kernelName<<<(numblocks), (numthreads), (memperblock), (streamId)>>>(hipLaunchParam{}, ##__VA_ARGS__); \ + kernelName<<<(numblocks), (numthreads), (memperblock), (streamId)>>>(hipLaunchParm{}, ##__VA_ARGS__); \ } while (0) #define hipLaunchKernelGGL(kernelName, numblocks, numthreads, memperblock, streamId, ...) \ diff --git a/projects/hip/src/hip_memory.cpp b/projects/hip/src/hip_memory.cpp index d540ab782b..e3823c504e 100644 --- a/projects/hip/src/hip_memory.cpp +++ b/projects/hip/src/hip_memory.cpp @@ -339,21 +339,25 @@ hipError_t hipHostMalloc(void** ptr, size_t sizeBytes, unsigned int flags) { hip_status = hipErrorInvalidValue; } else { auto device = ctx->getWriteableDevice(); - +#if (__hcc_workweek__ >= 19115) + //Avoid mapping host pinned memory to all devices by HCC + unsigned amFlags = amHostUnmapped; +#else unsigned amFlags = 0; +#endif if (flags & hipHostMallocCoherent) { - amFlags = amHostCoherent; + amFlags |= amHostCoherent; } else if (flags & hipHostMallocNonCoherent) { - amFlags = amHostNonCoherent; + amFlags |= amHostNonCoherent; } else { // depends on env variables: - amFlags = HIP_HOST_COHERENT ? amHostCoherent : amHostNonCoherent; + amFlags |= HIP_HOST_COHERENT ? amHostCoherent : amHostNonCoherent; } *ptr = hip_internal::allocAndSharePtr( (amFlags & amHostCoherent) ? "finegrained_host" : "pinned_host", sizeBytes, ctx, - (trueFlags & hipHostMallocPortable) /*shareWithAll*/, amFlags, flags, 0); + true /*shareWithAll*/, amFlags, flags, 0); if (sizeBytes && (*ptr == NULL)) { hip_status = hipErrorMemoryAllocation;