diff options
author | 2016-04-20 18:08:20 -0700 | |
---|---|---|
committer | 2016-04-20 18:08:20 -0700 | |
commit | 2dde1b102866e1928e925678951463f2a7051af1 (patch) | |
tree | d6405558903617aa9f5fabe4be8b5601fe6d406a /unsupported/Eigen | |
parent | a792cd357d31f0a4fce62ed1fa4cc0334cf2f143 (diff) |
Don't crash when attempting to reduce empty tensors.
Diffstat (limited to 'unsupported/Eigen')
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 2 | ||||
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h | 12 | ||||
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h | 7 |
3 files changed, 17 insertions, 4 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h index bf6e10a7b..c3edae477 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h @@ -238,7 +238,7 @@ inline void TensorExecutor<Expression, GpuDevice, Vectorizable>::run( device.maxCudaThreadsPerMultiProcessor() / block_size; const Index size = array_prod(evaluator.dimensions()); // Create a least one block to ensure we won't crash when tensorflow calls with tensors of size 0. - const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1); + const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, divup<int>(size, block_size)), 1); LAUNCH_CUDA_KERNEL( (EigenMetaKernel<TensorEvaluator<Expression, GpuDevice>, Index>), diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h index 6af2d45d4..cd04716bd 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h @@ -24,9 +24,17 @@ const T2& choose(Cond<false>, const T1&, const T2& second) { return second; } -template <typename T> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE + +template <typename T, typename X, typename Y> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE +T divup(const X x, const Y y) { + return static_cast<T>((x + y - 1) / y); +} + +template <typename T> +EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE T divup(const T x, const T y) { - return (x + y - 1) / y; + return static_cast<T>((x + y - 1) / y); } template <size_t n> struct max_n_1 { diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h index afa5a257a..fd2587dd5 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h @@ -134,9 +134,14 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> { typedef typename Self::Index Index; const Index num_coeffs = array_prod(self.m_impl.dimensions()); + // Don't crash when we're called with an input tensor of size 0. + if (num_coeffs == 0) { + return; + } + const int block_size = 256; const int num_per_thread = 128; - const int num_blocks = numext::ceil(static_cast<float>(num_coeffs) / (block_size * num_per_thread)); + const int num_blocks = divup<int>(num_coeffs, block_size * num_per_thread); if (num_blocks > 1) { // We initialize the outputs outside the reduction kernel when we can't be sure that there |