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

Clean up CUDA/NVCC version macros and their use in Eigen, and a few other CUDA build failures.

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2019-05-31 15:26:06 -07:00
parent 5614400581
commit b08527b0c1
11 changed files with 77 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
unsupported/Eigen/CXX11/src/Tensor/TensorContractionGpu.h
View File

@@ -388,7 +388,7 @@ EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
// the sum across all big k blocks of the product of little k block of index (x, y)
// with block of index (y, z). To compute the final output, we need to reduce
// the 8 threads over y by summation.
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
#else
#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor_sync(0xFFFFFFFF, res(i, j), mask)
@@ -621,7 +621,7 @@ EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rh
x1 = rhs_pf0.x;
x2 = rhs_pf0.z;
}
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
x1 = __shfl_xor(x1, 4);
x2 = __shfl_xor(x2, 4);
#else

4
unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h
View File

@@ -81,8 +81,8 @@
// gpu_assert can be overridden
#ifndef gpu_assert
#if defined(EIGEN_HIP_DEVICE_COMPILE) || (defined(EIGEN_CUDACC) && (EIGEN_CUDACC_VER==0))
// clang-cuda and HIPCC do not support the use of assert on the GPU side.
#if defined(EIGEN_HIP_DEVICE_COMPILE)
// HIPCC do not support the use of assert on the GPU side.
#define gpu_assert(COND)
#else
#define gpu_assert(COND) assert(COND)

8
unsupported/Eigen/CXX11/src/Tensor/TensorReductionGpu.h
View File

@@ -177,7 +177,7 @@ __global__ void FullReductionKernel(Reducer reducer, const Self input, Index num
} else {
reducer.reduce(__shfl_down(static_cast<int>(accum), offset, warpSize), &accum);
}
#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
#elif 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);
@@ -269,7 +269,7 @@ __global__ void FullReductionKernelHalfFloat(Reducer reducer, const Self input,
wka_in.h = accum;
wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
reducer.reducePacket(wka_out.h, &accum);
#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
#elif defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000
reducer.reducePacket(__shfl_down(accum, offset, warpSize), &accum);
#else
int temp = __shfl_down_sync(0xFFFFFFFF, *(int*)(&accum), (unsigned)offset, warpSize);
@@ -466,7 +466,7 @@ __global__ void InnerReductionKernel(Reducer reducer, const Self input, Index nu
} else {
reducer.reduce(__shfl_down(static_cast<int>(reduced_val), offset), &reduced_val);
}
#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
#elif 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);
@@ -571,7 +571,7 @@ __global__ void InnerReductionKernelHalfFloat(Reducer reducer, const Self input,
wka_in.h = reduced_val2;
wka_out.i = __shfl_down(wka_in.i, offset, warpSize);
reducer.reducePacket(wka_out.h, &reduced_val2);
#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
#elif 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