Skip to content

Commit

Permalink
add scatter, and update results
Browse files Browse the repository at this point in the history
  • Loading branch information
@hughperkins
hughperkins committed May 28, 2017
1 parent 7b9bb63 commit a1813a1
Show file tree
Hide file tree
Showing 9 changed files with 51 additions and 31 deletions.
6 changes: 3 additions & 3 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -60,16 +60,16 @@ Please see the main repository for full Tensorflow documentation. This readme w

| test | Mac Sierra, using Radeon Pro 450 GPU |
|----- |-------|
| unit tests (`py.test -v`) | pass | pass |
| unit tests (`py.test -v`) | pass |
| [linear_regression.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/2_BasicModels/linear_regression.py) | slow, but works |
| [logistic_regression.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/2_BasicModels/logistic_regression.py) | ok |
| [nearest_neighbor.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/2_BasicModels/nearest_neighbor.py) | ok (accuracy 0.92) |
| [autoencoder.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/autoencoder.py)| runs ok, random numbers working now :-) |
| [multilayer_perceptron.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/multilayer_perceptron.py) | runs ok, loss decreases, random numbers working now :-) |
| [recurrent_network.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/recurrent_network.py)| runs ok, loss goes down |
| [bidirectional_rnn.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/birectional_rnn.py)| runs ok, loss goes down |
| [dynamic_rnn.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/dynamic_rnn.py) | Missing Reshape gradients |
| [convolutional_network.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/convolutional_network.py) | Missing conv operation |
| [dynamic_rnn.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/dynamic_rnn.py) | Runs, but loss is nan currently |
| [convolutional_network.py](https://github.com/hughperkins/TensorFlow-Examples/blob/enforce-gpu/examples/3_NeuralNetworks/convolutional_network.py) | Missing conv |

## Installation

Expand Down
10 changes: 5 additions & 5 deletions tensorflow/core/kernels/scatter_functor.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#if GOOGLE_CUDA
// #if GOOGLE_CUDA

#include "tensorflow/core/kernels/scatter_functor.h"
#include "tensorflow/core/framework/register_types.h"
Expand Down Expand Up @@ -43,8 +43,8 @@ namespace functor {
DECLARE_GPU_SPECS_OP(T, Index, scatter_op::UpdateOp::DIV);

#define DECLARE_GPU_SPECS(T) \
DECLARE_GPU_SPECS_INDEX(T, int32); \
DECLARE_GPU_SPECS_INDEX(T, int64);
DECLARE_GPU_SPECS_INDEX(T, int32);
// DECLARE_GPU_SPECS_INDEX(T, int64);

TF_CALL_GPU_NUMBER_TYPES_NO_HALF(DECLARE_GPU_SPECS);

Expand All @@ -55,8 +55,8 @@ TF_CALL_GPU_NUMBER_TYPES_NO_HALF(DECLARE_GPU_SPECS);
} // namespace functor
} // namespace tensorflow

#else
// #else

#include "tensorflow/core/kernels/scatter_functor.h"

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA
10 changes: 5 additions & 5 deletions tensorflow/core/kernels/scatter_functor_gpu.cu.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#if GOOGLE_CUDA
// #if GOOGLE_CUDA

#define EIGEN_USE_GPU

Expand All @@ -34,11 +34,11 @@ typedef Eigen::GpuDevice GPUDevice;
DEFINE_GPU_SPECS_OP(T, Index, scatter_op::UpdateOp::DIV);

#define DEFINE_GPU_SPECS(T) \
DEFINE_GPU_SPECS_INDEX(T, int32); \
DEFINE_GPU_SPECS_INDEX(T, int64);
DEFINE_GPU_SPECS_INDEX(T, int32);
// DEFINE_GPU_SPECS_INDEX(T, int64);

DEFINE_GPU_SPECS(float);
DEFINE_GPU_SPECS(double);
// DEFINE_GPU_SPECS(double);
// TODO(b/27222123): The following fails to compile due to lack of support for
// fp16.
// TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPECS);
Expand All @@ -49,4 +49,4 @@ DEFINE_GPU_SPECS(double);

} // namespace tensorflow

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA
4 changes: 2 additions & 2 deletions tensorflow/core/kernels/scatter_functor_gpu.cu.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ limitations under the License.
#ifndef TENSORFLOW_KERNELS_SCATTER_FUNCTOR_GPU_CU_H_
#define TENSORFLOW_KERNELS_SCATTER_FUNCTOR_GPU_CU_H_

#if GOOGLE_CUDA
// #if GOOGLE_CUDA

#define EIGEN_USE_GPU

Expand Down Expand Up @@ -94,6 +94,6 @@ struct ScatterFunctor<GPUDevice, T, Index, op> {
} // namespace functor
} // namespace tensorflow

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA

#endif // TENSORFLOW_KERNELS_SCATTER_FUNCTOR_GPU_CU_H_
24 changes: 22 additions & 2 deletions tensorflow/core/kernels/scatter_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -159,7 +159,27 @@ TF_CALL_NUMBER_TYPES(REGISTER_SCATTER_ARITHEMTIC_CPU);
TF_CALL_ALL_TYPES(REGISTER_SCATTER_UPDATE_CPU);

// Registers GPU kernels.
#if GOOGLE_CUDA
// #if GOOGLE_CUDA

// added by Hugh:
#undef REGISTER_SCATTER_KERNEL
#undef REGISTER_SCATTER_ARITHEMTIC
#undef REGISTER_SCATTER_UPDATE

#define REGISTER_SCATTER_KERNEL(type, dev, name, op) \
REGISTER_SCATTER_KERNEL_INDEX(type, int32, dev, name, op);
// REGISTER_SCATTER_KERNEL_INDEX(type, int64, dev, name, op);

#define REGISTER_SCATTER_ARITHEMTIC(type, dev) \
REGISTER_SCATTER_KERNEL(type, dev, "ScatterAdd", scatter_op::UpdateOp::ADD); \
REGISTER_SCATTER_KERNEL(type, dev, "ScatterDiv", scatter_op::UpdateOp::DIV); \
REGISTER_SCATTER_KERNEL(type, dev, "ScatterMul", scatter_op::UpdateOp::MUL); \
REGISTER_SCATTER_KERNEL(type, dev, "ScatterSub", scatter_op::UpdateOp::SUB);

#define REGISTER_SCATTER_UPDATE(type, dev) \
REGISTER_SCATTER_KERNEL(type, dev, "ScatterUpdate", \
scatter_op::UpdateOp::ASSIGN);

#define REGISTER_SCATTER_ARITHEMTIC_GPU(type) \
REGISTER_SCATTER_ARITHEMTIC(type, GPU);

Expand All @@ -168,7 +188,7 @@ TF_CALL_ALL_TYPES(REGISTER_SCATTER_UPDATE_CPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_ARITHEMTIC_GPU);
TF_CALL_GPU_NUMBER_TYPES_NO_HALF(REGISTER_SCATTER_UPDATE_GPU);

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA

#undef REGISTER_SCATTER_ADD
#undef REGISTER_SCATTER_ARITHEMTIC
Expand Down
10 changes: 5 additions & 5 deletions tensorflow/core/kernels/scatter_op_gpu.cu.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#if GOOGLE_CUDA
// #if GOOGLE_CUDA

#define EIGEN_USE_GPU

Expand All @@ -35,11 +35,11 @@ typedef Eigen::GpuDevice GPUDevice;
DEFINE_GPU_SPECS_OP(T, Index, scatter_op::UpdateOp::DIV);

#define DEFINE_GPU_SPECS(T) \
DEFINE_GPU_SPECS_INDEX(T, int32); \
DEFINE_GPU_SPECS_INDEX(T, int64);
DEFINE_GPU_SPECS_INDEX(T, int32);
// DEFINE_GPU_SPECS_INDEX(T, int64);

DEFINE_GPU_SPECS(float);
DEFINE_GPU_SPECS(double);
// DEFINE_GPU_SPECS(double);
// TODO: The following fails to compile.
// TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPECS);

Expand All @@ -49,4 +49,4 @@ DEFINE_GPU_SPECS(double);

} // namespace tensorflow

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA
8 changes: 4 additions & 4 deletions tensorflow/core/kernels/segment_reduction_ops.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -310,7 +310,7 @@ TF_CALL_NUMBER_TYPES(REGISTER_CPU_UNSORTED_KERNELS_ALL);
#undef REGISTER_CPU_UNSORTED_KERNELS
#undef REGISTER_CPU_UNSORTED_KERNELS_ALL

#if GOOGLE_CUDA
// #if GOOGLE_CUDA
#define REGISTER_GPU_UNSORTED_KERNELS(type, index_type) \
REGISTER_KERNEL_BUILDER(Name("UnsortedSegmentSum") \
.Device(DEVICE_GPU) \
Expand All @@ -320,13 +320,13 @@ TF_CALL_NUMBER_TYPES(REGISTER_CPU_UNSORTED_KERNELS_ALL);
UnsortedSegmentSumOp<GPUDevice, type, index_type>);

#define REGISTER_GPU_UNSORTED_KERNELS_ALL(type) \
REGISTER_GPU_UNSORTED_KERNELS(type, int32); \
REGISTER_GPU_UNSORTED_KERNELS(type, int64);
REGISTER_GPU_UNSORTED_KERNELS(type, int32);
// REGISTER_GPU_UNSORTED_KERNELS(type, int64);

TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU_UNSORTED_KERNELS_ALL);
#undef REGISTER_GPU_UNSORTED_KERNELS
#undef REGISTER_GPU_UNSORTED_KERNELS_ALL
#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA

// Same as SegmentReductionOp but takes as input a "sparse" tensor, represented
// by two dense tensors, one containing the data, and the other containing
Expand Down
8 changes: 4 additions & 4 deletions tensorflow/core/kernels/segment_reduction_ops_gpu.cu.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#if GOOGLE_CUDA
// #if GOOGLE_CUDA

#define EIGEN_USE_GPU

Expand Down Expand Up @@ -95,8 +95,8 @@ struct UnsortedSegmentSumFunctor<GPUDevice, T, Index> {
template struct UnsortedSegmentSumFunctor<GPUDevice, T, Index>

#define DEFINE_GPU_SPECS(T) \
DEFINE_GPU_SPECS_INDEX(T, int32); \
DEFINE_GPU_SPECS_INDEX(T, int64);
DEFINE_GPU_SPECS_INDEX(T, int32);
// DEFINE_GPU_SPECS_INDEX(T, int64);

TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPECS);

Expand All @@ -106,4 +106,4 @@ TF_CALL_GPU_NUMBER_TYPES(DEFINE_GPU_SPECS);
} // namespace functor
} // namespace tensorflow

#endif // GOOGLE_CUDA
// #endif // GOOGLE_CUDA
2 changes: 1 addition & 1 deletion third_party/coriander
Submodule coriander updated 6 files
+1 −1 src/function_names_map.cpp
+1 −1 src/kernel_dumper.cpp
+29 −51 src/new_instruction_dumper.cpp
+1 −0 src/new_instruction_dumper.h
+19 −0 src/shims.cpp
+23 −0 test/test_maths.py

0 comments on commit a1813a1

Please sign in to comment.