GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-21 12:07:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Prevent nvcc from miscompiling the cuda metakernel. Unfortunately this reintroduces some compulation warnings but it's much better than having to deal with random assertion failures.

Browse Source
This commit is contained in:
Benoit Steiner 2016-01-08 13:53:40 -08:00
parent 6639b7d6e8
commit 53749ff415
3 changed files with 7 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
View File

@ -238,14 +238,10 @@ struct GpuDevice {
};
#ifndef __CUDA_ARCH__
#define LAUNCH_CUDA_KERNEL(kernel, gridsize, blocksize, sharedmem, device, ...) \
(kernel) <<< (gridsize), (blocksize), (sharedmem), (device).stream() >>> (__VA_ARGS__); \
assert(cudaGetLastError() == cudaSuccess);
#else
#define LAUNCH_CUDA_KERNEL(...) \
eigen_assert(false && "Cannot launch a kernel from another kernel");
#endif
// FIXME: Should be device and kernel specific.
#ifdef __CUDACC__

16
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@ -156,14 +156,14 @@ template <typename Expression>
class TensorExecutor<Expression, GpuDevice, false> {
public:
typedef typename Expression::Index Index;
EIGEN_DEVICE_FUNC static void run(const Expression& expr, const GpuDevice& device);
static void run(const Expression& expr, const GpuDevice& device);
};
template <typename Expression>
class TensorExecutor<Expression, GpuDevice, true> {
public:
typedef typename Expression::Index Index;
EIGEN_DEVICE_FUNC static void run(const Expression& expr, const GpuDevice& device);
static void run(const Expression& expr, const GpuDevice& device);
};
#if defined(__CUDACC__)
@ -213,9 +213,8 @@ EigenMetaKernel_Vectorizable(Evaluator memcopied_eval, Index size) {
/*static*/
template <typename Expression>
EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, false>::run(const Expression& expr, const GpuDevice& device)
inline void TensorExecutor<Expression, GpuDevice, false>::run(const Expression& expr, const GpuDevice& device)
{
#ifndef __CUDA_ARCH__
TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
@ -228,17 +227,13 @@ EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, false>::run(
LAUNCH_CUDA_KERNEL((EigenMetaKernel_NonVectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size);
}
evaluator.cleanup();
#else
eigen_assert(false && "Cannot launch a kernel from another kernel");
#endif
}
/*static*/
template<typename Expression>
EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, true>::run(const Expression& expr, const GpuDevice& device)
inline void TensorExecutor<Expression, GpuDevice, true>::run(const Expression& expr, const GpuDevice& device)
{
#ifndef __CUDA_ARCH__
TensorEvaluator<Expression, GpuDevice> evaluator(expr, device);
const bool needs_assign = evaluator.evalSubExprsIfNeeded(NULL);
if (needs_assign)
@ -251,9 +246,6 @@ EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, true>::run(c
LAUNCH_CUDA_KERNEL((EigenMetaKernel_Vectorizable<TensorEvaluator<Expression, GpuDevice>, Index>), num_blocks, block_size, 0, device, evaluator, size);
}
evaluator.cleanup();
#else
eigen_assert(false && "Cannot launch a kernel from another kernel");
#endif
}
#endif // __CUDACC__

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

@ -115,11 +115,11 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
internal::is_same<typename Self::CoeffReturnType, float>::value;
template <typename OutputType>
EIGEN_DEVICE_FUNC static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
static void run(const Self& self, Op& reducer, const GpuDevice& device, OutputType* output) {
assert(false && "Should only be called on floats");
}
EIGEN_DEVICE_FUNC static void run(const Self& self, Op& reducer, const GpuDevice& device, float* output) {
static void run(const Self& self, Op& reducer, const GpuDevice& device, float* output) {
typedef typename Self::Index Index;
const Index num_coeffs = array_prod(self.m_impl.dimensions());