Home | History | Annotate | Download | only in kernels 
      1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
      2 
      3 Licensed under the Apache License, Version 2.0 (the "License");
      4 you may not use this file except in compliance with the License.
      5 You may obtain a copy of the License at
      6 
      7     http://www.apache.org/licenses/LICENSE-2.0
      8 
      9 Unless required by applicable law or agreed to in writing, software
     10 distributed under the License is distributed on an "AS IS" BASIS,
     11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     12 See the License for the specific language governing permissions and
     13 limitations under the License.
     14 ==============================================================================*/
     15 
     16 #if GOOGLE_CUDA
     17 
     18 #define EIGEN_USE_GPU
     19 
     20 #include <complex>
     21 #include "tensorflow/core/framework/register_types.h"
     22 #include "tensorflow/core/kernels/diag_op.h"
     23 #include "tensorflow/core/util/cuda_kernel_helper.h"
     24 
     25 namespace tensorflow {
     26 namespace functor {
     27 
     28 typedef Eigen::GpuDevice GPUDevice;
     29 
     30 template <typename T>
     31 __global__ void DiagCudaKernel(const int num_threads, const int64 size,
     32                                const T* in, T* out) {
     33   CUDA_1D_KERNEL_LOOP(index, num_threads) {
     34     // Fill the diagonal elements or set to zero in other place.
     35     if (index % (1 + size) == 0) {
     36       out[index] = in[index / (1 + size)];
     37     } else {
     38       out[index] = T(0);
     39     }
     40   }
     41 }
     42 
     43 template <typename T>
     44 struct DiagFunctor<GPUDevice, T> {
     45   EIGEN_ALWAYS_INLINE Status operator()(OpKernelContext* context,
     46                                         const int64 size, const T* in, T* out) {
     47     // Empty tensor couldn't launch the kernel.
     48     if (size == 0) {
     49       return Status::OK();
     50     }
     51 
     52     // CudaLaunchConfig uses an int for virtual_thread_count,
     53     // so this may overflow for `size*size` in extreme cases,
     54     // here is checking the multiplication overflow for integer.
     55     if (size && (int(size * size) / size) != size) {
     56       return errors::Internal("DiagOp got input size too large.");
     57     }
     58     int virtual_thread_count = int(size * size);
     59 
     60     // Launch the GPU kernel.
     61     const GPUDevice& device = context->eigen_device<GPUDevice>();
     62     CudaLaunchConfig diag_config =
     63         GetCudaLaunchConfig(virtual_thread_count, device);
     64     DiagCudaKernel<<<diag_config.block_count, diag_config.thread_per_block, 0,
     65                      device.stream()>>>(diag_config.virtual_thread_count, size,
     66                                         in, out);
     67 
     68     auto err = cudaGetLastError();
     69     if (err != cudaSuccess) {
     70       return errors::Internal(
     71           "Could not launch DiagOp kernel: ", cudaGetErrorString(err), ".");
     72     }
     73     return Status::OK();
     74   }
     75 };
     76 
     77 template struct DiagFunctor<GPUDevice, double>;
     78 template struct DiagFunctor<GPUDevice, float>;
     79 template struct DiagFunctor<GPUDevice, int32>;
     80 template struct DiagFunctor<GPUDevice, int64>;
     81 template struct DiagFunctor<GPUDevice, complex64>;
     82 template struct DiagFunctor<GPUDevice, complex128>;
     83 
     84 template <typename T>
     85 __global__ void DiagPartCudaKernel(const int num_threads, const int64 size,
     86                                    const T* in, T* out) {
     87   CUDA_1D_KERNEL_LOOP(index, num_threads) {
     88     out[index] = in[(1 + size) * index];
     89   }
     90 }
     91 
     92 template <typename T>
     93 struct DiagPartFunctor<GPUDevice, T> {
     94   EIGEN_ALWAYS_INLINE Status operator()(OpKernelContext* context,
     95                                         const int64 size, const T* in, T* out) {
     96     // Empty tensor couldn't launch the kernel.
     97     if (size == 0) {
     98       return Status::OK();
     99     }
    100     const GPUDevice& device = context->eigen_device<GPUDevice>();
    101 
    102     // Extract the diagonal elements.
    103     CudaLaunchConfig diag_config = GetCudaLaunchConfig(size, device);
    104     DiagPartCudaKernel<<<diag_config.block_count, diag_config.thread_per_block,
    105                          0, device.stream()>>>(diag_config.virtual_thread_count,
    106                                                size, in, out);
    107 
    108     auto err = cudaGetLastError();
    109     if (err != cudaSuccess) {
    110       return errors::Internal(
    111           "Could not launch DiagPartOp kernel: ", cudaGetErrorString(err), ".");
    112     }
    113     return Status::OK();
    114   }
    115 };
    116 
    117 template struct DiagPartFunctor<GPUDevice, double>;
    118 template struct DiagPartFunctor<GPUDevice, float>;
    119 template struct DiagPartFunctor<GPUDevice, int32>;
    120 template struct DiagPartFunctor<GPUDevice, int64>;
    121 template struct DiagPartFunctor<GPUDevice, complex64>;
    122 template struct DiagPartFunctor<GPUDevice, complex128>;
    123 
    124 }  // end namespace functor
    125 }  // end namespace tensorflow
    126 
    127 #endif  // GOOGLE_CUDA
    128