/external/tensorflow/tensorflow/core/kernels/rnn/ |
D | blas_gemm.cc | 29 se::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory) { in AsDeviceMemory() function 48 auto a_ptr = AsDeviceMemory(a); in operator ()() 49 auto b_ptr = AsDeviceMemory(b); in operator ()() 50 auto c_ptr = AsDeviceMemory(c); in operator ()()
|
/external/tensorflow/tensorflow/core/kernels/ |
D | cudnn_pooling_gpu.cc | 92 auto input_data = AsDeviceMemory(transformed_input.template flat<T>().data(), in Compute() 95 AsDeviceMemory(transformed_output.template flat<T>().data(), in Compute() 213 AsDeviceMemory(transformed_output.template flat<T>().data(), in Compute() 216 AsDeviceMemory(transformed_input.template flat<T>().data(), in Compute() 219 AsDeviceMemory(transformed_output_backprop.template flat<T>().data(), in Compute() 222 AsDeviceMemory(transformed_input_backprop.template flat<T>().data(), in Compute()
|
D | fused_batch_norm_op.cc | 768 y_ptr = StreamExecutorUtil::AsDeviceMemory<T>(*y); in operator ()() 786 y_ptr = StreamExecutorUtil::AsDeviceMemory<T>(y_transformed); in operator ()() 812 auto x_ptr = StreamExecutorUtil::AsDeviceMemory<T>(x_maybe_transformed); in operator ()() 813 auto scale_ptr = StreamExecutorUtil::AsDeviceMemory<U>(scale); in operator ()() 814 auto offset_ptr = StreamExecutorUtil::AsDeviceMemory<U>(offset); in operator ()() 816 StreamExecutorUtil::AsDeviceMemory<U>(estimated_mean); in operator ()() 818 StreamExecutorUtil::AsDeviceMemory<U>(estimated_variance); in operator ()() 821 ? StreamExecutorUtil::AsDeviceMemory<U>(*side_input) in operator ()() 823 auto batch_mean_ptr = StreamExecutorUtil::AsDeviceMemory<U>(*batch_mean); in operator ()() 825 auto batch_var_ptr = StreamExecutorUtil::AsDeviceMemory<U>(*batch_var); in operator ()() [all …]
|
D | lrn_op.cc | 216 auto input_data = StreamExecutorUtil::AsDeviceMemory<T>(in); in launch() 217 auto output_data = StreamExecutorUtil::AsDeviceMemory<T>(*output); in launch() 270 AsDeviceMemory(transformed_input.template flat<T>().data(), in launch() 273 AsDeviceMemory(transformed_output.template flat<T>().data(), in launch() 508 auto input_grads_data = StreamExecutorUtil::AsDeviceMemory<T>(in_grads); in launch() 509 auto input_image_data = StreamExecutorUtil::AsDeviceMemory<T>(in_image); in launch() 510 auto output_image_data = StreamExecutorUtil::AsDeviceMemory<T>(out_image); in launch() 511 auto output_grads_data = StreamExecutorUtil::AsDeviceMemory<T>(*output); in launch() 584 AsDeviceMemory(transformed_in_grads.template flat<T>().data(), in launch() 587 AsDeviceMemory(transformed_in_image.template flat<T>().data(), in launch() [all …]
|
D | cudnn_rnn_ops.cc | 293 const DeviceMemory<T> AsDeviceMemory(const Tensor* tensor) { in AsDeviceMemory() function 300 DeviceMemory<T> AsDeviceMemory(Tensor* tensor) { in AsDeviceMemory() function 466 return AsDeviceMemory<uint8>(handle_.AccessTensor(context_)); in AllocateBytes() 809 auto input_data = AsDeviceMemory<T>(input); in DoForward() 810 auto input_h_data = AsDeviceMemory<T>(input_h); in DoForward() 813 input_c_data = AsDeviceMemory<T>(input_c); in DoForward() 816 auto params_data = AsDeviceMemory<T>(params); in DoForward() 817 auto output_data = AsDeviceMemory<T>(output); in DoForward() 818 auto output_h_data = AsDeviceMemory<T>(output_h); in DoForward() 821 output_c_data = AsDeviceMemory<T>(output_c); in DoForward() [all …]
|
D | batch_matmul_op_impl.h | 257 se::DeviceMemory<T> AsDeviceMemory(const T* gpu_memory) { 336 a_device_memory.push_back(AsDeviceMemory(a_base_ptr + i * m * k)); 337 b_device_memory.push_back(AsDeviceMemory(b_base_ptr + i * k * n)); 338 c_device_memory.push_back(AsDeviceMemory(c_base_ptr + i * m * n)); 347 a_device_memory.push_back(AsDeviceMemory(a_base_ptr + i * m * k)); 350 b_device_memory.push_back(AsDeviceMemory(b_base_ptr + i * k * n)); 353 c_device_memory.push_back(AsDeviceMemory(c_base_ptr + i * m * n)); 472 a_device_memory.push_back(AsDeviceMemory(a_base_ptr + i * m * k)); 473 b_device_memory.push_back(AsDeviceMemory(b_base_ptr + i * k * n)); 474 c_device_memory.push_back(AsDeviceMemory(c_base_ptr + i * m * n)); [all …]
|
D | pooling_ops_common.cc | 237 AsDeviceMemory(reinterpret_cast<const typename RawType<T>::type*>( in Compute() 242 AsDeviceMemory(reinterpret_cast<const typename RawType<T>::type*>( in Compute() 434 AsDeviceMemory(transformed_output.template flat<T>().data(), in Compute() 437 AsDeviceMemory(transformed_input.template flat<T>().data(), in Compute() 440 AsDeviceMemory(transformed_output_backprop.template flat<T>().data(), in Compute() 443 AsDeviceMemory(transformed_input_backprop.template flat<T>().data(), in Compute()
|
D | conv_ops_3d.cc | 258 auto a_ptr = AsDeviceMemory(input.template flat<T>().data(), in launch() 260 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in launch() 262 auto c_ptr = AsDeviceMemory(output->template flat<T>().data(), in launch() 285 auto a_ptr = AsDeviceMemory(input.template flat<T>().data(), in launch() 287 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in launch() 289 auto c_ptr = AsDeviceMemory(output->template flat<T>().data(), in launch() 414 auto input_ptr = AsDeviceMemory(input.template flat<T>().data(), in launch() 417 AsDeviceMemory(transformed_filter.template flat<T>().data(), in launch() 420 AsDeviceMemory(transformed_output.template flat<T>().data(), in launch()
|
D | matrix_triangular_solve_op_impl.h | 53 se::DeviceMemory<Scalar> AsDeviceMemory(const Scalar* gpu_memory) { in AsDeviceMemory() function 267 auto src_device_mem = AsDeviceMemory(in_y.template flat<Scalar>().data()); 268 auto dst_device_mem = AsDeviceMemory(out->template flat<Scalar>().data()); 286 auto src_device_mem = AsDeviceMemory(b_tmp_ptrs[b_batch_indices[i]]); 288 AsDeviceMemory(out->template flat<Scalar>().data() + i * m * n);
|
D | conv_grad_ops_3d.cc | 1189 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in Compute() 1191 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in Compute() 1193 auto c_ptr = AsDeviceMemory(in_backprop->template flat<T>().data(), in Compute() 1219 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in Compute() 1221 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in Compute() 1223 auto c_ptr = AsDeviceMemory(in_backprop->template flat<T>().data(), in Compute() 1340 AsDeviceMemory(transformed_out_backprop.template flat<T>().data(), in Compute() 1343 AsDeviceMemory(transformed_filter.template flat<T>().data(), in Compute() 1346 AsDeviceMemory(pre_transformed_in_backprop.template flat<T>().data(), in Compute() 1661 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in Compute() [all …]
|
D | conv_grad_filter_ops.cc | 728 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in operator ()() 734 auto b_ptr = AsDeviceMemory(input.template flat<T>().data(), in operator ()() 740 auto c_ptr = AsDeviceMemory(filter_backprop->template flat<T>().data(), in operator ()() 767 auto a_ptr = AsDeviceMemory(input.template flat<T>().data(), in operator ()() 769 auto b_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in operator ()() 771 auto c_ptr = AsDeviceMemory(filter_backprop->template flat<T>().data(), in operator ()() 945 AsDeviceMemory(transformed_out_backprop.template flat<T>().data(), in operator ()() 948 AsDeviceMemory(pre_transformed_filter_backprop.template flat<T>().data(), in operator ()() 950 auto input_ptr = AsDeviceMemory(transformed_input.template flat<T>().data(), in operator ()()
|
D | fft_ops.cc | 344 se::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory) { in AsDeviceMemory() function 351 se::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory, uint64 size) { in AsDeviceMemory() function 387 AsDeviceMemory(temporary_memory.flat<uint8>().data(), in AllocateBytes() 532 auto src = AsDeviceMemory<InT>(in.flat<InT>().data()); in DoFFTInternal() 533 auto dst = AsDeviceMemory<OutT>(out->flat<OutT>().data()); in DoFFTInternal()
|
D | conv_ops.cc | 687 auto a_ptr = AsDeviceMemory(input.template flat<T>().data(), in operator ()() 689 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in operator ()() 691 auto c_ptr = AsDeviceMemory(output->template flat<T>().data(), in operator ()() 715 auto a_ptr = AsDeviceMemory(input.template flat<T>().data(), in operator ()() 717 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in operator ()() 719 auto c_ptr = AsDeviceMemory(output->template flat<T>().data(), in operator ()() 951 auto input_ptr = AsDeviceMemory(input.template flat<T>().data(), in operator ()() 954 AsDeviceMemory(transformed_filter.template flat<T>().data(), in operator ()() 957 AsDeviceMemory(transformed_output.template flat<T>().data(), in operator ()()
|
D | ctc_loss_op.cc | 339 auto probs_data = StreamExecutorUtil::AsDeviceMemory<float>(*inputs); in Compute() 340 auto costs_data = StreamExecutorUtil::AsDeviceMemory<float>(*loss); in Compute() 341 auto grads_data = StreamExecutorUtil::AsDeviceMemory<float>(*gradient); in Compute()
|
D | conv_grad_input_ops.cc | 905 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in operator ()() 907 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in operator ()() 909 auto c_ptr = AsDeviceMemory(in_backprop->template flat<T>().data(), in operator ()() 938 auto a_ptr = AsDeviceMemory(out_backprop.template flat<T>().data(), in operator ()() 940 auto b_ptr = AsDeviceMemory(filter.template flat<T>().data(), in operator ()() 942 auto c_ptr = AsDeviceMemory(in_backprop->template flat<T>().data(), in operator ()() 1111 AsDeviceMemory(transformed_out_backprop.template flat<T>().data(), in operator ()() 1114 AsDeviceMemory(transformed_filter.template flat<T>().data(), in operator ()() 1117 AsDeviceMemory(pre_transformed_in_backprop.template flat<T>().data(), in operator ()()
|
D | tridiagonal_solve_op_gpu.cu.cc | 71 se::DeviceMemory<Scalar> AsDeviceMemory(const Scalar* cuda_memory) { in AsDeviceMemory() function 80 auto src_device_mem = AsDeviceMemory(src); in CopyDeviceToDevice() 81 auto dst_device_mem = AsDeviceMemory(dst); in CopyDeviceToDevice()
|
D | matmul_op.cc | 280 auto a_ptr = AsDeviceMemory(a.template flat<T>().data(), in launch() 282 auto b_ptr = AsDeviceMemory(b.template flat<T>().data(), in launch() 284 auto c_ptr = AsDeviceMemory(out->template flat<T>().data(), in launch()
|
D | conv_ops_gpu.h | 86 AsDeviceMemory(temporary_memory.flat<uint8>().data(), in AllocateBytes()
|
D | gpu_utils.h | 63 inline se::DeviceMemory<T> AsDeviceMemory(const T* cuda_memory, uint64 size) { in AsDeviceMemory() function
|
D | conv_ops_fused_impl.h | 573 return AsDeviceMemory(t.template flat<T>().data(), 584 AsDeviceMemory(static_cast<T*>(nullptr), 0);
|
/external/tensorflow/tensorflow/core/util/ |
D | stream_executor_util.h | 33 static se::DeviceMemory<T> AsDeviceMemory(const Tensor& t) { in AsDeviceMemory() function
|
/external/tensorflow/tensorflow/core/common_runtime/gpu/ |
D | gpu_util.h | 73 static se::DeviceMemory<T> AsDeviceMemory(const Tensor& t) { in AsDeviceMemory() function
|
/external/tensorflow/tensorflow/core/nccl/ |
D | nccl_manager_test.cc | 155 auto in_gpu_mem = AsDeviceMemory(in_gpu.flat<Scalar>().data()); in MakeReductionTestCase() 196 auto in_gpu_mem = AsDeviceMemory(in_gpu.flat<Scalar>().data()); in MakeGatherTestCase() 232 auto in_gpu_mem = AsDeviceMemory(in_gpu.flat<Scalar>().data()); in MakeBroadcastTestCase() 268 auto out_gpu_mem = AsDeviceMemory(out_gpu.flat<Scalar>().data()); in VerifyResults() 432 static se::DeviceMemory<Scalar> AsDeviceMemory(const Scalar* cuda_memory) { in AsDeviceMemory() function in tensorflow::NcclManagerTest
|