From 46f8a18567731925e06a7389a6c611e1dc420ea8 Mon Sep 17 00:00:00 2001 From: Deven Desai Date: Wed, 5 Aug 2020 01:46:34 +0000 Subject: Adding an explicit launch_bounds(1024) attribute for GPU kernels. Starting with ROCm 3.5, the HIP compiler will change from HCC to hip-clang. This compiler change introduce a change in the default value of the `__launch_bounds__` attribute associated with a GPU kernel. (default value means the value assumed by the compiler as the `__launch_bounds attribute__` value, when it is not explicitly specified by the user) Currently (i.e. for HIP with ROCm 3.3 and older), the default value is 1024. That changes to 256 with ROCm 3.5 (i.e. hip-clang compiler). As a consequence of this change, if a GPU kernel with a `__luanch_bounds__` attribute of 256 is launched at runtime with a threads_per_block value > 256, it leads to a runtime error. This is leading to a couple of Eigen unit test failures with ROCm 3.5. This commit adds an explicit `__launch_bounds(1024)__` attribute to every GPU kernel that currently does not have it explicitly specified (and hence will end up getting the default value of 256 with the change to hip-clang) --- .../Eigen/CXX11/src/Tensor/TensorReductionGpu.h | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h') diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h index 36df03d62..cfc49166a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h @@ -121,7 +121,7 @@ __device__ inline void atomicReduce(float* output, float accum, SumReducer -__global__ void ReductionInitKernel(const CoeffType val, Index num_preserved_coeffs, CoeffType* output) { +__global__ __launch_bounds__(1024) void ReductionInitKernel(const CoeffType val, Index num_preserved_coeffs, CoeffType* output) { const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x; const Index num_threads = blockDim.x * gridDim.x; for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) { @@ -132,7 +132,7 @@ __global__ void ReductionInitKernel(const CoeffType val, Index num_preserved_coe template -__global__ void FullReductionKernel(Reducer reducer, const Self input, Index num_coeffs, +__global__ __launch_bounds__(1024) void FullReductionKernel(Reducer reducer, const Self input, Index num_coeffs, typename Self::CoeffReturnType* output, unsigned int* semaphore) { #if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300) // Initialize the output value @@ -214,7 +214,7 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num #ifdef EIGEN_HAS_GPU_FP16 template -__global__ void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, +__global__ __launch_bounds__(1024) void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, packet_traits::type* scratch) { eigen_assert(blockDim.x == 1); eigen_assert(gridDim.x == 1); @@ -239,7 +239,7 @@ __global__ void ReductionInitFullReduxKernelHalfFloat(Reducer reducer, const Sel template -__global__ void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output) { +__global__ __launch_bounds__(1024) void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output) { const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x; const Index num_threads = blockDim.x * gridDim.x; typedef typename packet_traits::type PacketType; @@ -259,7 +259,7 @@ __global__ void ReductionInitKernelHalfFloat(Reducer reducer, const Self input, template -__global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, +__global__ __launch_bounds__(1024) void FullReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs, half* output, packet_traits::type* scratch) { typedef typename packet_traits::type PacketType; const int packet_width = unpacket_traits::size; @@ -358,7 +358,7 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input, } template -__global__ void ReductionCleanupKernelHalfFloat(Op reducer, half* output, packet_traits::type* scratch) { +__global__ __launch_bounds__(1024) void ReductionCleanupKernelHalfFloat(Op reducer, half* output, packet_traits::type* scratch) { eigen_assert(threadIdx.x == 1); half2* pscratch = reinterpret_cast(scratch); half tmp = __float2half(0.f); @@ -476,7 +476,7 @@ struct FullReducer { template -__global__ void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, +__global__ __launch_bounds__(1024) void InnerReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, typename Self::CoeffReturnType* output) { #if (defined(EIGEN_HIP_DEVICE_COMPILE) && defined(__HIP_ARCH_HAS_WARP_SHUFFLE__)) || (EIGEN_CUDA_ARCH >= 300) typedef typename Self::CoeffReturnType Type; @@ -561,7 +561,7 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu template -__global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, +__global__ __launch_bounds__(1024) void InnerReductionKernelHalfFloat(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, half* output) { eigen_assert(blockDim.y == 1); eigen_assert(blockDim.z == 1); @@ -868,7 +868,7 @@ struct InnerReducer { template -__global__ void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, +__global__ __launch_bounds__(1024) void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs, typename Self::CoeffReturnType* output) { const Index num_threads = blockDim.x * gridDim.x; const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x; -- cgit v1.2.3