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 "tensorflow/core/kernels/sparse_tensor_dense_matmul_op.h" 21 22 #include "tensorflow/core/framework/bounds_check.h" 23 #include "tensorflow/core/framework/register_types.h" 24 #include "tensorflow/core/util/cuda_kernel_helper.h" 25 26 namespace tensorflow { 27 28 typedef Eigen::GpuDevice GPUDevice; 29 30 template <typename T, typename Tindices, bool ADJ_A, bool ADJ_B> 31 __global__ void SparseTensorDenseMatMulKernel(int nnz, int m, int b_rows, 32 int b_cols, int p, 33 const Tindices* a_indices, 34 const T* a_values, const T* b, 35 T* out) { 36 // out_{ij} = sum_k {a_ik b_kj} 37 // out = A * B', out_{ij} = sum_k {a_ik (b')_kj}; b'_{kj} = b_{jk} 38 const int n = (ADJ_B) ? b_cols : b_rows; 39 CUDA_1D_KERNEL_LOOP(index, nnz * p) { 40 const int a_ix = index / p; 41 const int j = index % p; 42 const int i = ldg(a_indices + 2 * a_ix + ((ADJ_A) ? 1 : 0)); 43 const int k = ldg(a_indices + 2 * a_ix + ((ADJ_A) ? 0 : 1)); 44 if (!FastBoundsCheck(i, m)) { 45 continue; // Nowhere to signal an error :( 46 } 47 // out[i, j] 48 T* out_location = out + i * p + j; 49 if (!FastBoundsCheck(k, n)) { 50 CudaAtomicAdd(out_location, std::numeric_limits<T>::quiet_NaN()); 51 continue; 52 } 53 54 // a_value == (ADJ_A) ? a[k, i] : a[i, k] 55 const T a_value = ldg(a_values + a_ix); 56 57 // b_value == (ADJ_B) ? b[j, k] : b[k, j] 58 const T b_value = ldg(b + ((ADJ_B) ? j * b_cols + k : k * b_cols + j)); 59 CudaAtomicAdd(out_location, a_value * b_value); 60 } 61 } 62 63 namespace functor { 64 65 template <typename T, typename Tindices, bool ADJ_A, bool ADJ_B> 66 struct SparseTensorDenseMatMulFunctor<GPUDevice, T, Tindices, ADJ_A, ADJ_B> { 67 static EIGEN_ALWAYS_INLINE Status 68 Compute(const GPUDevice& d, typename TTypes<T>::Matrix out, 69 typename TTypes<Tindices>::ConstMatrix a_indices, 70 typename TTypes<T>::ConstVec a_values, 71 typename TTypes<T>::ConstMatrix b) { 72 out.device(d) = out.constant(T(0)); 73 int nnz = a_values.size(); 74 // out = A * B, A is [m x n] and B is [n x p], out is [m x p] 75 int m = out.dimension(0); 76 int p = out.dimension(1); 77 int b_rows = b.dimension(0); 78 int b_cols = b.dimension(1); 79 80 // TODO(ebrevdo): Should this be alpha * nnz instead of 81 // out.size()? Perhaps p * nnz ? 82 CudaLaunchConfig config = GetCudaLaunchConfig(p * nnz, d); 83 84 TF_CHECK_OK(CudaLaunchKernel( 85 SparseTensorDenseMatMulKernel<T, Tindices, ADJ_A, ADJ_B>, 86 config.block_count, config.thread_per_block, 0, d.stream(), nnz, m, 87 b_rows, b_cols, p, a_indices.data(), a_values.data(), b.data(), 88 out.data())); 89 90 return Status::OK(); 91 } 92 }; 93 94 } // namespace functor 95 96 #define DEFINE(T, Tindices) \ 97 template struct functor::SparseTensorDenseMatMulFunctor< \ 98 GPUDevice, T, Tindices, false, false>; \ 99 template struct functor::SparseTensorDenseMatMulFunctor< \ 100 GPUDevice, T, Tindices, false, true>; \ 101 template struct functor::SparseTensorDenseMatMulFunctor< \ 102 GPUDevice, T, Tindices, true, false>; \ 103 template struct functor::SparseTensorDenseMatMulFunctor< \ 104 GPUDevice, T, Tindices, true, true>; 105 106 DEFINE(float, int32); 107 DEFINE(float, int64); 108 #undef DEFINE 109 110 } // end namespace tensorflow 111 112 #endif // GOOGLE_CUDA 113