devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix cuda device warnings

Browse Source
This commit is contained in:
Antonio Sanchez
2025-02-28 15:21:23 -08:00
parent c53002f5fb
commit 23b1682723
53 changed files with 655 additions and 600 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@@ -678,15 +678,15 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType, MakePointer_>,
template <typename S, typename O, bool V> friend struct internal::FullReducerShard;
#endif
#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
template <int B, int N, typename S, typename R, typename I> friend __global__ void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*, unsigned int*);
#ifdef EIGEN_HAS_CUDA_FP16
template <typename S, typename R, typename I> friend void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
template <int NPT, typename S, typename R, typename I> friend void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
template <typename S, typename R, typename I> friend __global__ void internal::ReductionInitFullReduxKernelHalfFloat(R, const S, I, half2*);
template <int B, int N, typename S, typename R, typename I> friend __global__ void internal::FullReductionKernelHalfFloat(R, const S, I, half*, half2*);
template <int NPT, typename S, typename R, typename I> friend __global__ void internal::InnerReductionKernelHalfFloat(R, const S, I, I, half*);
#endif
template <int NPT, typename S, typename R, typename I> friend void internal::InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I> friend __global__ void internal::InnerReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I> friend void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
template <int NPT, typename S, typename R, typename I> friend __global__ void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
#endif
template <typename S, typename O, typename D> friend struct internal::InnerReducer;

20
unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
View File

@@ -168,7 +168,12 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
#pragma unroll
for (int offset = warpSize/2; offset > 0; offset /= 2) {
#if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
reducer.reduce(__shfl_down(accum, offset, warpSize), &accum);
#else
reducer.reduce(__shfl_down_sync(0xFFFFFFFF, accum, offset, warpSize), &accum);
#endif
}
if ((threadIdx.x & (warpSize - 1)) == 0) {
@@ -244,7 +249,11 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
#pragma unroll
for (int offset = warpSize/2; offset > 0; offset /= 2) {
#if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
reducer.reducePacket(__shfl_down(accum, offset, warpSize), &accum);
#else
reducer.reducePacket(__shfl_down_sync(0xFFFFFFFF, accum, offset, warpSize), &accum);
#endif
}
if ((threadIdx.x & (warpSize - 1)) == 0) {
@@ -426,7 +435,11 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
#pragma unroll
for (int offset = warpSize/2; offset > 0; offset /= 2) {
#if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
reducer.reduce(__shfl_down(reduced_val, offset), &reduced_val);
#else
reducer.reduce(__shfl_down_sync(0xFFFFFFFF, reduced_val, offset), &reduced_val);
#endif
}
if ((threadIdx.x & (warpSize - 1)) == 0) {
@@ -516,8 +529,15 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
#pragma unroll
for (int offset = warpSize/2; offset > 0; offset /= 2) {
#if defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
reducer.reducePacket(__shfl_down(reduced_val1, offset, warpSize), &reduced_val1);
reducer.reducePacket(__shfl_down(reduced_val2, offset, warpSize), &reduced_val2);
#else
reducer.reducePacket(__shfl_down_sync(0xFFFFFFFF, reduced_val1, offset, warpSize), &reduced_val1);
reducer.reducePacket(__shfl_down_sync(0xFFFFFFFF, reduced_val2, offset, warpSize), &reduced_val2);
#endif
}
half val1 = __low2half(reduced_val1);

12
unsupported/Eigen/src/SpecialFunctions/SpecialFunctionsPacketMath.h
View File

@@ -15,27 +15,27 @@ namespace Eigen {
namespace internal {
/** \internal \returns the ln(|gamma(\a a)|) (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet plgamma(const Packet& a) { using numext::lgamma; return lgamma(a); }
/** \internal \returns the derivative of lgamma, psi(\a a) (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet pdigamma(const Packet& a) { using numext::digamma; return digamma(a); }
/** \internal \returns the zeta function of two arguments (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet pzeta(const Packet& x, const Packet& q) { using numext::zeta; return zeta(x, q); }
/** \internal \returns the polygamma function (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet ppolygamma(const Packet& n, const Packet& x) { using numext::polygamma; return polygamma(n, x); }
/** \internal \returns the erf(\a a) (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet perf(const Packet& a) { using numext::erf; return erf(a); }
/** \internal \returns the erfc(\a a) (coeff-wise) */
template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet perfc(const Packet& a) { using numext::erfc; return erfc(a); }
/** \internal \returns the incomplete gamma function igamma(\a a, \a x) */

9
unsupported/test/CMakeLists.txt
View File

@@ -216,17 +216,14 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
message(STATUS "Flags used to compile cuda code: " ${CMAKE_CXX_FLAGS})
if( (NOT EIGEN_TEST_CXX11) OR (CMAKE_VERSION VERSION_LESS 3.3))
set(EIGEN_CUDA_CXX11_FLAG "-std=c++11")
else()
# otherwise the flag has already been added because of the above set(CMAKE_CXX_STANDARD 11)
set(EIGEN_CUDA_CXX11_FLAG "")
string(APPEND EIGEN_CUDA_CXX11_FLAGS " -std=c++11")
endif()
if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
set(CUDA_NVCC_FLAGS "-ccbin ${CMAKE_C_COMPILER}" CACHE STRING "nvcc flags" FORCE)
endif()
if(EIGEN_TEST_CUDA_CLANG)
string(APPEND CMAKE_CXX_FLAGS " --cuda-path=${CUDA_TOOLKIT_ROOT_DIR} ${EIGEN_CUDA_CXX11_FLAG}")
string(APPEND CMAKE_CXX_FLAGS " --cuda-path=${CUDA_TOOLKIT_ROOT_DIR} ${EIGEN_CUDA_CXX11_FLAGS}")
foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
string(APPEND CMAKE_CXX_FLAGS " --cuda-gpu-arch=sm_${ARCH}")
endforeach()
@@ -246,7 +243,7 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
foreach(ARCH IN LISTS EIGEN_CUDA_COMPUTE_ARCH)
string(APPEND NVCC_ARCH_FLAGS " -gencode arch=compute_${ARCH},code=sm_${ARCH}")
endforeach()
set(CUDA_NVCC_FLAGS "--expt-relaxed-constexpr -Xcudafe \"--display_error_number\" ${NVCC_ARCH_FLAGS} ${CUDA_NVCC_FLAGS} ${EIGEN_CUDA_CXX11_FLAG}")
set(CUDA_NVCC_FLAGS "--expt-relaxed-constexpr -Xcudafe \"--display_error_number\" ${NVCC_ARCH_FLAGS} ${CUDA_NVCC_FLAGS} ${EIGEN_CUDA_CXX11_FLAGS}")
cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include")
endif()