From f40cf089f0ecd0316c0150918b9bbfc930fba955 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Fri, 26 Jun 2020 22:05:12 +0000
Subject: [PATCH 01/10] completed allocate_temp
---
tensorflow/c/kernels.cc | 24 +++++
tensorflow/c/kernels.h | 9 ++
tensorflow/c/kernels_test.cc | 166 +++++++++++++++++++++++++++--------
3 files changed, 163 insertions(+), 36 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index a0ed0d9f245..80b5234b52d 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -26,6 +26,9 @@ limitations under the License.
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/types.h"
+#include "tensorflow/core/framework/tensor_shape.h"
+#include "tensorflow/core/lib/gtl/array_slice.h"
+
// This file forms the basis of a stable ABI for third-party kernel
// implementations. It is crucial that changes to this file are made cautiously
// and with a focus on maintaining both source and binary compatibility.
@@ -260,3 +263,24 @@ TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context, int index,
}
return result;
}
+
+TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
+ int64_t* dims, int num_dims, TF_Status* status){
+ auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelContext*>(context);
+ TF_SetStatus(status, TF_OK, "");
+ tensorflow::TensorShape shape;
+ for(int i = 0; i < num_dims; ++i){
+ shape.AddDim(dims[i]);
+ }
+ tensorflow::Status s;
+ tensorflow::Tensor tensor_temp;
+ TF_Tensor* tf_tensor_temp;
+ s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor_temp);
+ if (s.ok()){
+ tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ }
+ if (s.ok()){
+ ::tensorflow::Set_TF_Status_from_Status(status, s);
+ return tf_tensor_temp;
+ }
+}
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index 084717c1d9e..e450511da3a 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -190,6 +190,15 @@ TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,
int64_t* dims, int num_dims,
size_t len, TF_Status* status);
+// Allocates a temporary Tensor of the specified type and shape. The
+// Tensor must not be used after kernel construction is
+// complete.
+
+// num_dims must equal the size of array dims
+TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
+ int64_t* dims, int num_dims, TF_Status* status);
+
+
#ifdef __cplusplus
} /* end extern "C" */
#endif
diff --git a/tensorflow/c/kernels_test.cc b/tensorflow/c/kernels_test.cc
index 423302741de..738c1e12c80 100644
--- a/tensorflow/c/kernels_test.cc
+++ b/tensorflow/c/kernels_test.cc
@@ -360,6 +360,17 @@ class DeviceKernelOpTest : public OpsTestBase {
#endif
};
+// Helper function for tests that validates that the tensor has
+// shape and type corresponding to dims and dtype.
+void validate_tensor(TF_Tensor* tensor, int64_t* dims, int64_t num_dims,
+ TF_DataType dtype);
+
+// Helper function for tests that copies data of length
+// tensor_size_bytes from values to tensor
+template <typename T>
+void set_tensor_data(TF_Tensor* tensor, T* values, size_t tensor_size_bytes,
+ TF_OpKernelContext* ctx);
+
REGISTER_OP("AllocateOutputOp1").Output("output1: float");
TEST_F(DeviceKernelOpTest, TestAllocateOutputSizeOne) {
@@ -371,22 +382,11 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSizeOne) {
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
/*num_dims=*/1, /*len=*/tensor_size_bytes, s);
- EXPECT_EQ(TF_OK, TF_GetCode(s));
- EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
- EXPECT_EQ(1, TF_NumDims(output));
- EXPECT_EQ(1, TF_Dim(output, 0));
-
+ validate_tensor(output, &dim, 1, TF_FLOAT);
+
// Set output to 3
- float* data = reinterpret_cast<float*>(TF_TensorData(output));
- float value = 3.0f;
-#if GOOGLE_CUDA
- OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
- cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, &value,
- tensor_size_bytes);
-#else
- *data = value;
-#endif
-
+ float values[1] = {3.0f};
+ set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -409,12 +409,8 @@ TEST_F(DeviceKernelOpTest, TestAllocateEmptyOutput) {
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
/*num_dims=*/1, /*len=*/0, s);
-
EXPECT_EQ(TF_OK, TF_GetCode(s));
- EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
- EXPECT_EQ(1, TF_NumDims(output));
- EXPECT_EQ(0, TF_Dim(output, 0));
-
+ validate_tensor(output, &dim, 1, TF_FLOAT);
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -434,27 +430,16 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSize2x3) {
TF_Status* s = TF_NewStatus();
// Allocate 2x3 output
int64_t dim[2] = {2, 3};
- size_t tensor_size_bytes = 6 * TF_DataTypeSize(TF_FLOAT);
+ size_t tensor_size_bytes = TF_DataTypeSize(TF_FLOAT) * 6;
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/dim,
/*num_dims=*/2, /*len=*/tensor_size_bytes, s);
EXPECT_EQ(TF_OK, TF_GetCode(s));
- EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
- EXPECT_EQ(2, TF_NumDims(output));
- EXPECT_EQ(2, TF_Dim(output, 0));
- EXPECT_EQ(3, TF_Dim(output, 1));
+ validate_tensor(output, dim, 2, TF_FLOAT);
// Set output to [1 2 3 4 5 6]
- void* data = TF_TensorData(output);
- float value[6] = {1, 2, 3, 4, 5, 6};
-#if GOOGLE_CUDA
- OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
- cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, value,
- tensor_size_bytes);
-#else
- memcpy(data, value, tensor_size_bytes);
-#endif
-
+ float values[6] = {1, 2, 3, 4, 5, 6};
+ set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -466,4 +451,113 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSize2x3) {
EXPECT_EQ("Tensor<type: float shape: [2,3] values: [1 2 3][4 5 6]>",
output->DebugString(100));
}
-} // namespace tensorflow
+
+REGISTER_OP("AllocateTempOp1").Output("output1: float");
+
+TEST_F(DeviceKernelOpTest, TestAllocateTempSizeOne) {
+ auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
+ // Allocate output
+ TF_Status* s = TF_NewStatus();
+ int64_t dim = 1;
+ TF_Tensor* output = TF_AllocateTemp(
+ /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
+ /*num_dims=*/1, s);
+ size_t tensor_size_bytes = TF_DataTypeSize(TF_FLOAT);
+ EXPECT_EQ(TF_OK, TF_GetCode(s));
+ validate_tensor(output, &dim, 1, TF_FLOAT);
+
+ // Set output to 3
+ float values[1] = {3.0f};
+ set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
+ TF_SetOutput(ctx, 0, output, s);
+ TF_DeleteStatus(s);
+ TF_DeleteTensor(output);
+ };
+
+ SetupOp("AllocateTempOp1", "AllocateTemp1", my_compute_func);
+
+ TF_ASSERT_OK(RunOpKernel());
+ Tensor* output = GetOutput(0);
+ EXPECT_EQ("Tensor<type: float shape: [1] values: 3>",
+ output->DebugString(100));
+}
+
+REGISTER_OP("AllocateTempOp0").Output("output1: float");
+
+TEST_F(DeviceKernelOpTest, TestAllocateTempEmpty) {
+ auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
+ TF_Status* s = TF_NewStatus();
+ // Allocate empty output
+ int64_t dim = 0;
+ TF_Tensor* output = TF_AllocateTemp(
+ /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
+ /*num_dims=*/1, s);
+ EXPECT_EQ(TF_OK, TF_GetCode(s));
+ validate_tensor(output, &dim, 1, TF_FLOAT);
+ TF_SetOutput(ctx, 0, output, s);
+ TF_DeleteStatus(s);
+ TF_DeleteTensor(output);
+ };
+
+ SetupOp("AllocateTempOp0", "AllocateTemp0", my_compute_func);
+
+ TF_ASSERT_OK(RunOpKernel());
+ Tensor* output = GetOutput(0);
+ EXPECT_EQ("Tensor<type: float shape: [0] values: >",
+ output->DebugString(100));
+}
+
+REGISTER_OP("AllocateTempOp2x3").Output("output1: float");
+
+TEST_F(DeviceKernelOpTest, TestAllocateTempSize2x3) {
+ auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
+ TF_Status* s = TF_NewStatus();
+ size_t tensor_size_bytes = 6 * TF_DataTypeSize(TF_FLOAT);
+ // Allocate 2x3 output
+ int64_t dim[2] = {2, 3};
+ TF_Tensor* output = TF_AllocateTemp(
+ /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/dim,
+ /*num_dims=*/2, s);
+ EXPECT_EQ(TF_OK, TF_GetCode(s));
+ validate_tensor(output, dim, 2, TF_FLOAT);
+
+ // Set output to [1 2 3 4 5 6]
+ void* data = TF_TensorData(output);
+ float values[6] = {1, 2, 3, 4, 5, 6};
+ set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
+ TF_SetOutput(ctx, 0, output, s);
+ TF_DeleteStatus(s);
+ TF_DeleteTensor(output);
+ };
+
+ SetupOp("AllocateTempOp2x3", "AllocateTempOp2x3", my_compute_func);
+
+ TF_ASSERT_OK(RunOpKernel());
+ Tensor* output = GetOutput(0);
+ EXPECT_EQ("Tensor<type: float shape: [2,3] values: [1 2 3][4 5 6]>",
+ output->DebugString(100));
+}
+
+void validate_tensor(TF_Tensor* tensor, int64_t* dims, int64_t num_dims,
+ TF_DataType dtype){
+ EXPECT_EQ(TF_FLOAT, TF_TensorType(tensor));
+ EXPECT_EQ(num_dims, TF_NumDims(tensor));
+ for(int i = 0; i < num_dims; ++i){
+ EXPECT_EQ(dims[i], TF_Dim(tensor, i));
+ }
+}
+
+template <typename T>
+void set_tensor_data(TF_Tensor* tensor, T* values, size_t tensor_size_bytes,
+ TF_OpKernelContext* ctx){
+ T* data = reinterpret_cast<T*>(TF_TensorData(tensor));
+#if GOOGLE_CUDA
+ OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
+ cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, values,
+ tensor_size_bytes);
+#else
+ memcpy(data, values, tensor_size_bytes);
+#endif
+}
+
+} // namespace tensorflow
\ No newline at end of file
From 8b07609b2ee977a8a97120dbaad4c3fabc151b0f Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Fri, 26 Jun 2020 22:33:19 +0000
Subject: [PATCH 02/10] added priority function to TF_KernelDefBuilder
---
tensorflow/c/kernels.cc | 5 +++++
tensorflow/c/kernels.h | 4 ++++
2 files changed, 9 insertions(+)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 80b5234b52d..905219c6e16 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -100,6 +100,11 @@ void TF_KernelBuilder_HostMemory(TF_KernelBuilder* kernel_builder,
kernel_builder->cc_builder->HostMemory(arg_name);
}
+void TF_KernelBuilder_Priority(TF_KernelBuilder* kernel_builder,
+ int32_t priority_number){
+ kernel_builder->cc_builder->Priority(priority_number);
+}
+
namespace tensorflow {
namespace {
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index e450511da3a..b245dd8a7fc 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -107,6 +107,10 @@ TF_CAPI_EXPORT extern void TF_KernelBuilder_TypeConstraint(
TF_CAPI_EXPORT extern void TF_KernelBuilder_HostMemory(
TF_KernelBuilder* kernel_builder, const char* arg_name);
+// Specify a priority number for this kernel.
+TF_CAPI_EXPORT extern void TF_KernelBuilder_Priority(
+ TF_KernelBuilder* kernel_builder, int32_t priority_number);
+
// Register the given kernel builder with the TensorFlow runtime. If
// registration fails, the given status will be populated.
//
From 70feb59205f75ed0642eb091bd63960f7358039e Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Sun, 28 Jun 2020 20:55:41 +0000
Subject: [PATCH 03/10] fixed indentation and comments for allocate_temp
---
tensorflow/c/kernels.cc | 11 +++++++----
tensorflow/c/kernels.h | 10 +++++-----
2 files changed, 12 insertions(+), 9 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 905219c6e16..02703d97bbe 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -281,11 +281,14 @@ TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
tensorflow::Tensor tensor_temp;
TF_Tensor* tf_tensor_temp;
s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor_temp);
- if (s.ok()){
- tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ if (!s.ok()){
+ ::tensorflow::Set_TF_Status_from_Status(status, s);
+ return nullptr;
}
- if (s.ok()){
+ tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ if (!s.ok()){
::tensorflow::Set_TF_Status_from_Status(status, s);
- return tf_tensor_temp;
+ return nullptr;
}
+ return tf_tensor_temp;
}
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index b245dd8a7fc..8ed3488988d 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -194,13 +194,13 @@ TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,
int64_t* dims, int num_dims,
size_t len, TF_Status* status);
-// Allocates a temporary Tensor of the specified type and shape. The
-// Tensor must not be used after kernel construction is
-// complete.
+// Allocates a temporary Tensor of the specified type and shape. Devices
+// such as GPUs that enqueue Ops for lazy execution may retain references
+// to the temporary tensors after the Op's Compute method has run.
// num_dims must equal the size of array dims
-TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
- int64_t* dims, int num_dims, TF_Status* status);
+TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context,
+ TF_DataType dtype, int64_t* dims, int num_dims, TF_Status* status);
#ifdef __cplusplus
From b58635dca87103807acd14b67545122d4d112ee0 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Sun, 28 Jun 2020 20:55:41 +0000
Subject: [PATCH 04/10] took out array_slice.h
---
tensorflow/c/kernels.cc | 13 +++++++------
tensorflow/c/kernels.h | 10 +++++-----
2 files changed, 12 insertions(+), 11 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 905219c6e16..3a8170575ad 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -25,9 +25,7 @@ limitations under the License.
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/types.h"
-
#include "tensorflow/core/framework/tensor_shape.h"
-#include "tensorflow/core/lib/gtl/array_slice.h"
// This file forms the basis of a stable ABI for third-party kernel
// implementations. It is crucial that changes to this file are made cautiously
@@ -281,11 +279,14 @@ TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
tensorflow::Tensor tensor_temp;
TF_Tensor* tf_tensor_temp;
s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor_temp);
- if (s.ok()){
- tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ if (!s.ok()){
+ ::tensorflow::Set_TF_Status_from_Status(status, s);
+ return nullptr;
}
- if (s.ok()){
+ tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ if (!s.ok()){
::tensorflow::Set_TF_Status_from_Status(status, s);
- return tf_tensor_temp;
+ return nullptr;
}
+ return tf_tensor_temp;
}
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index b245dd8a7fc..8ed3488988d 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -194,13 +194,13 @@ TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,
int64_t* dims, int num_dims,
size_t len, TF_Status* status);
-// Allocates a temporary Tensor of the specified type and shape. The
-// Tensor must not be used after kernel construction is
-// complete.
+// Allocates a temporary Tensor of the specified type and shape. Devices
+// such as GPUs that enqueue Ops for lazy execution may retain references
+// to the temporary tensors after the Op's Compute method has run.
// num_dims must equal the size of array dims
-TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
- int64_t* dims, int num_dims, TF_Status* status);
+TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context,
+ TF_DataType dtype, int64_t* dims, int num_dims, TF_Status* status);
#ifdef __cplusplus
From fb70daadc449ffc50d6d0cdce17375bd72b6ab07 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Wed, 1 Jul 2020 00:15:09 +0000
Subject: [PATCH 05/10] fixed allocate_temp indent errors
---
tensorflow/c/kernels.cc | 2 +-
tensorflow/c/kernels.h | 4 +++-
2 files changed, 4 insertions(+), 2 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 3a8170575ad..749e6e89b8b 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -268,7 +268,7 @@ TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context, int index,
}
TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
- int64_t* dims, int num_dims, TF_Status* status){
+ int64_t* dims, int num_dims, TF_Status* status){
auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelContext*>(context);
TF_SetStatus(status, TF_OK, "");
tensorflow::TensorShape shape;
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index 8ed3488988d..1891ce31a23 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -200,7 +200,9 @@ TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,
// num_dims must equal the size of array dims
TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context,
- TF_DataType dtype, int64_t* dims, int num_dims, TF_Status* status);
+ TF_DataType dtype,
+ int64_t* dims, int num_dims,
+ TF_Status* status);
#ifdef __cplusplus
From e52df6b5f34a76f3974132a5b3bf2653bed1e660 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Mon, 6 Jul 2020 23:27:50 +0000
Subject: [PATCH 06/10] fixed naming for variables
---
tensorflow/c/kernels.cc | 10 +++++-----
1 file changed, 5 insertions(+), 5 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 749e6e89b8b..5f5bd9779b1 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -276,17 +276,17 @@ TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
shape.AddDim(dims[i]);
}
tensorflow::Status s;
- tensorflow::Tensor tensor_temp;
- TF_Tensor* tf_tensor_temp;
- s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor_temp);
+ tensorflow::Tensor tensor;
+ TF_Tensor* tf_tensor;
+ s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor);
if (!s.ok()){
::tensorflow::Set_TF_Status_from_Status(status, s);
return nullptr;
}
- tf_tensor_temp = TF_TensorFromTensor(tensor_temp, &s);
+ tf_tensor = TF_TensorFromTensor(tensor, &s);
if (!s.ok()){
::tensorflow::Set_TF_Status_from_Status(status, s);
return nullptr;
}
- return tf_tensor_temp;
+ return tf_tensor;
}
From 7b0a2473467e922f7a3b00076d7c3878e7a132ae Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Wed, 15 Jul 2020 20:57:34 +0000
Subject: [PATCH 07/10] added Scalar Shape to TF_ShapeInfererence functions
---
tensorflow/c/ops.cc | 6 ++++++
tensorflow/c/ops.h | 5 +++++
2 files changed, 11 insertions(+)
diff --git a/tensorflow/c/ops.cc b/tensorflow/c/ops.cc
index 118385ed72c..b1a69922e5d 100644
--- a/tensorflow/c/ops.cc
+++ b/tensorflow/c/ops.cc
@@ -104,6 +104,12 @@ TF_ShapeHandle* TF_NewShapeHandle() {
return reinterpret_cast<TF_ShapeHandle*>(new ShapeHandle);
}
+TF_ShapeHandle* TF_ShapeInferenceContextScalar(TF_ShapeInferenceContext* ctx) {
+ auto* handle = new ShapeHandle;
+ *handle = reinterpret_cast<InferenceContext*>(ctx)->Scalar();
+ return reinterpret_cast<TF_ShapeHandle*>(handle);
+}
+
TF_ShapeHandle* TF_ShapeInferenceContextVectorFromSize(
TF_ShapeInferenceContext* ctx, size_t size) {
auto* handle = new ShapeHandle;
diff --git a/tensorflow/c/ops.h b/tensorflow/c/ops.h
index 14868e40260..91cad73101f 100644
--- a/tensorflow/c/ops.h
+++ b/tensorflow/c/ops.h
@@ -280,6 +280,11 @@ extern void TF_ShapeInferenceContextSetOutput(TF_ShapeInferenceContext* ctx,
int i, TF_ShapeHandle* handle,
TF_Status* status);
+// Returns a newly-allocated scalar shape handle. The returned handle should
+// be freed with TF_DeleteShapeHandle.
+TF_CAPI_EXPORT extern TF_ShapeHandle* TF_ShapeInferenceContextScalar(
+ TF_ShapeInferenceContext* ctx);
+
// Returns a newly-allocate shape handle representing a vector of the given
// size. The returned handle should be freed with TF_DeleteShapeHandle.
TF_CAPI_EXPORT extern TF_ShapeHandle* TF_ShapeInferenceContextVectorFromSize(
From 23ebf04dbf7d11f7989bd85085844a9ffecd3c2f Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Wed, 15 Jul 2020 21:26:21 +0000
Subject: [PATCH 08/10] updated paths to upstream master for cleaner merge
---
tensorflow/c/kernels.cc | 25 ------
tensorflow/c/kernels.h | 11 ---
tensorflow/c/kernels_test.cc | 166 ++++++++---------------------------
3 files changed, 36 insertions(+), 166 deletions(-)
diff --git a/tensorflow/c/kernels.cc b/tensorflow/c/kernels.cc
index 7aef824cc69..8fa50711a8d 100644
--- a/tensorflow/c/kernels.cc
+++ b/tensorflow/c/kernels.cc
@@ -25,7 +25,6 @@ limitations under the License.
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/types.h"
-#include "tensorflow/core/framework/tensor_shape.h"
// This file forms the basis of a stable ABI for third-party kernel
// implementations. It is crucial that changes to this file are made cautiously
@@ -273,27 +272,3 @@ TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context, int index,
}
return tf_tensor;
}
-
-TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context, TF_DataType dtype,
- int64_t* dims, int num_dims, TF_Status* status){
- auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelContext*>(context);
- TF_SetStatus(status, TF_OK, "");
- tensorflow::TensorShape shape;
- for(int i = 0; i < num_dims; ++i){
- shape.AddDim(dims[i]);
- }
- tensorflow::Status s;
- tensorflow::Tensor tensor;
- TF_Tensor* tf_tensor;
- s = cc_ctx->allocate_temp(static_cast<tensorflow::DataType>(dtype), shape, &tensor);
- if (!s.ok()){
- ::tensorflow::Set_TF_Status_from_Status(status, s);
- return nullptr;
- }
- tf_tensor = TF_TensorFromTensor(tensor, &s);
- if (!s.ok()){
- ::tensorflow::Set_TF_Status_from_Status(status, s);
- return nullptr;
- }
- return tf_tensor;
-}
diff --git a/tensorflow/c/kernels.h b/tensorflow/c/kernels.h
index e1397402dae..1428f7ab928 100644
--- a/tensorflow/c/kernels.h
+++ b/tensorflow/c/kernels.h
@@ -194,17 +194,6 @@ TF_CAPI_EXPORT TF_Tensor* TF_AllocateOutput(TF_OpKernelContext* context,
int64_t* dims, int num_dims,
size_t len, TF_Status* status);
-// Allocates a temporary Tensor of the specified type and shape. Devices
-// such as GPUs that enqueue Ops for lazy execution may retain references
-// to the temporary tensors after the Op's Compute method has run.
-
-// num_dims must equal the size of array dims
-TF_CAPI_EXPORT extern TF_Tensor* TF_AllocateTemp(TF_OpKernelContext* context,
- TF_DataType dtype,
- int64_t* dims, int num_dims,
- TF_Status* status);
-
-
#ifdef __cplusplus
} /* end extern "C" */
#endif
diff --git a/tensorflow/c/kernels_test.cc b/tensorflow/c/kernels_test.cc
index 738c1e12c80..423302741de 100644
--- a/tensorflow/c/kernels_test.cc
+++ b/tensorflow/c/kernels_test.cc
@@ -360,17 +360,6 @@ class DeviceKernelOpTest : public OpsTestBase {
#endif
};
-// Helper function for tests that validates that the tensor has
-// shape and type corresponding to dims and dtype.
-void validate_tensor(TF_Tensor* tensor, int64_t* dims, int64_t num_dims,
- TF_DataType dtype);
-
-// Helper function for tests that copies data of length
-// tensor_size_bytes from values to tensor
-template <typename T>
-void set_tensor_data(TF_Tensor* tensor, T* values, size_t tensor_size_bytes,
- TF_OpKernelContext* ctx);
-
REGISTER_OP("AllocateOutputOp1").Output("output1: float");
TEST_F(DeviceKernelOpTest, TestAllocateOutputSizeOne) {
@@ -382,11 +371,22 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSizeOne) {
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
/*num_dims=*/1, /*len=*/tensor_size_bytes, s);
- validate_tensor(output, &dim, 1, TF_FLOAT);
-
+ EXPECT_EQ(TF_OK, TF_GetCode(s));
+ EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
+ EXPECT_EQ(1, TF_NumDims(output));
+ EXPECT_EQ(1, TF_Dim(output, 0));
+
// Set output to 3
- float values[1] = {3.0f};
- set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
+ float* data = reinterpret_cast<float*>(TF_TensorData(output));
+ float value = 3.0f;
+#if GOOGLE_CUDA
+ OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
+ cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, &value,
+ tensor_size_bytes);
+#else
+ *data = value;
+#endif
+
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -409,8 +409,12 @@ TEST_F(DeviceKernelOpTest, TestAllocateEmptyOutput) {
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
/*num_dims=*/1, /*len=*/0, s);
+
EXPECT_EQ(TF_OK, TF_GetCode(s));
- validate_tensor(output, &dim, 1, TF_FLOAT);
+ EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
+ EXPECT_EQ(1, TF_NumDims(output));
+ EXPECT_EQ(0, TF_Dim(output, 0));
+
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -430,16 +434,27 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSize2x3) {
TF_Status* s = TF_NewStatus();
// Allocate 2x3 output
int64_t dim[2] = {2, 3};
- size_t tensor_size_bytes = TF_DataTypeSize(TF_FLOAT) * 6;
+ size_t tensor_size_bytes = 6 * TF_DataTypeSize(TF_FLOAT);
TF_Tensor* output = TF_AllocateOutput(
/*context=*/ctx, /*index=*/0, /*dtype=*/TF_FLOAT, /*dims=*/dim,
/*num_dims=*/2, /*len=*/tensor_size_bytes, s);
EXPECT_EQ(TF_OK, TF_GetCode(s));
- validate_tensor(output, dim, 2, TF_FLOAT);
+ EXPECT_EQ(TF_FLOAT, TF_TensorType(output));
+ EXPECT_EQ(2, TF_NumDims(output));
+ EXPECT_EQ(2, TF_Dim(output, 0));
+ EXPECT_EQ(3, TF_Dim(output, 1));
// Set output to [1 2 3 4 5 6]
- float values[6] = {1, 2, 3, 4, 5, 6};
- set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
+ void* data = TF_TensorData(output);
+ float value[6] = {1, 2, 3, 4, 5, 6};
+#if GOOGLE_CUDA
+ OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
+ cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, value,
+ tensor_size_bytes);
+#else
+ memcpy(data, value, tensor_size_bytes);
+#endif
+
TF_DeleteStatus(s);
TF_DeleteTensor(output);
};
@@ -451,113 +466,4 @@ TEST_F(DeviceKernelOpTest, TestAllocateOutputSize2x3) {
EXPECT_EQ("Tensor<type: float shape: [2,3] values: [1 2 3][4 5 6]>",
output->DebugString(100));
}
-
-REGISTER_OP("AllocateTempOp1").Output("output1: float");
-
-TEST_F(DeviceKernelOpTest, TestAllocateTempSizeOne) {
- auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
- // Allocate output
- TF_Status* s = TF_NewStatus();
- int64_t dim = 1;
- TF_Tensor* output = TF_AllocateTemp(
- /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
- /*num_dims=*/1, s);
- size_t tensor_size_bytes = TF_DataTypeSize(TF_FLOAT);
- EXPECT_EQ(TF_OK, TF_GetCode(s));
- validate_tensor(output, &dim, 1, TF_FLOAT);
-
- // Set output to 3
- float values[1] = {3.0f};
- set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
- TF_SetOutput(ctx, 0, output, s);
- TF_DeleteStatus(s);
- TF_DeleteTensor(output);
- };
-
- SetupOp("AllocateTempOp1", "AllocateTemp1", my_compute_func);
-
- TF_ASSERT_OK(RunOpKernel());
- Tensor* output = GetOutput(0);
- EXPECT_EQ("Tensor<type: float shape: [1] values: 3>",
- output->DebugString(100));
-}
-
-REGISTER_OP("AllocateTempOp0").Output("output1: float");
-
-TEST_F(DeviceKernelOpTest, TestAllocateTempEmpty) {
- auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
- TF_Status* s = TF_NewStatus();
- // Allocate empty output
- int64_t dim = 0;
- TF_Tensor* output = TF_AllocateTemp(
- /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/&dim,
- /*num_dims=*/1, s);
- EXPECT_EQ(TF_OK, TF_GetCode(s));
- validate_tensor(output, &dim, 1, TF_FLOAT);
- TF_SetOutput(ctx, 0, output, s);
- TF_DeleteStatus(s);
- TF_DeleteTensor(output);
- };
-
- SetupOp("AllocateTempOp0", "AllocateTemp0", my_compute_func);
-
- TF_ASSERT_OK(RunOpKernel());
- Tensor* output = GetOutput(0);
- EXPECT_EQ("Tensor<type: float shape: [0] values: >",
- output->DebugString(100));
-}
-
-REGISTER_OP("AllocateTempOp2x3").Output("output1: float");
-
-TEST_F(DeviceKernelOpTest, TestAllocateTempSize2x3) {
- auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
- TF_Status* s = TF_NewStatus();
- size_t tensor_size_bytes = 6 * TF_DataTypeSize(TF_FLOAT);
- // Allocate 2x3 output
- int64_t dim[2] = {2, 3};
- TF_Tensor* output = TF_AllocateTemp(
- /*context=*/ctx, /*dtype=*/TF_FLOAT, /*dims=*/dim,
- /*num_dims=*/2, s);
- EXPECT_EQ(TF_OK, TF_GetCode(s));
- validate_tensor(output, dim, 2, TF_FLOAT);
-
- // Set output to [1 2 3 4 5 6]
- void* data = TF_TensorData(output);
- float values[6] = {1, 2, 3, 4, 5, 6};
- set_tensor_data<float>(output, values, tensor_size_bytes, ctx);
- TF_SetOutput(ctx, 0, output, s);
- TF_DeleteStatus(s);
- TF_DeleteTensor(output);
- };
-
- SetupOp("AllocateTempOp2x3", "AllocateTempOp2x3", my_compute_func);
-
- TF_ASSERT_OK(RunOpKernel());
- Tensor* output = GetOutput(0);
- EXPECT_EQ("Tensor<type: float shape: [2,3] values: [1 2 3][4 5 6]>",
- output->DebugString(100));
-}
-
-void validate_tensor(TF_Tensor* tensor, int64_t* dims, int64_t num_dims,
- TF_DataType dtype){
- EXPECT_EQ(TF_FLOAT, TF_TensorType(tensor));
- EXPECT_EQ(num_dims, TF_NumDims(tensor));
- for(int i = 0; i < num_dims; ++i){
- EXPECT_EQ(dims[i], TF_Dim(tensor, i));
- }
-}
-
-template <typename T>
-void set_tensor_data(TF_Tensor* tensor, T* values, size_t tensor_size_bytes,
- TF_OpKernelContext* ctx){
- T* data = reinterpret_cast<T*>(TF_TensorData(tensor));
-#if GOOGLE_CUDA
- OpKernelContext* cc_ctx = reinterpret_cast<OpKernelContext*>(ctx);
- cc_ctx->eigen_gpu_device().memcpyHostToDevice(data, values,
- tensor_size_bytes);
-#else
- memcpy(data, values, tensor_size_bytes);
-#endif
-}
-
-} // namespace tensorflow
\ No newline at end of file
+} // namespace tensorflow
From 843987549843c7113efccf998c65fb004006a116 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Wed, 15 Jul 2020 23:40:50 +0000
Subject: [PATCH 09/10] added test for TF_ShapeInferenceContextScalar
---
tensorflow/c/ops_test.cc | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/tensorflow/c/ops_test.cc b/tensorflow/c/ops_test.cc
index 482413f966c..63105b1ad00 100644
--- a/tensorflow/c/ops_test.cc
+++ b/tensorflow/c/ops_test.cc
@@ -316,5 +316,15 @@ TEST(OpsTest, ShapeInferenceSubshape) {
TF_DeleteShapeHandle(handle);
}
+TEST(OpsTest, ShapeInferenceScalarShape) {
+ NodeDef def;
+ shape_inference::InferenceContext c(0, def, MakeOpDef(0, 0),
+ {S({})}, {}, {}, {});
+ TF_ShapeHandle* TF_scalar_shape = TF_ShapeInferenceContextScalar(C_CTX(&c));
+ shape_inference::ShapeHandle* scalar_shape =
+ reinterpret_cast<shape_inference::ShapeHandle*>(TF_scalar_shape);
+ ASSERT_EQ("[]", c.DebugString(*scalar_shape));
+}
+
} // namespace
} // namespace tensorflow
From 3f324403a59e312ee9a428bdd1b6c4baa529c127 Mon Sep 17 00:00:00 2001
From: Daniel Nguyen <nguyendaniel@google.com>
Date: Wed, 22 Jul 2020 17:49:59 +0000
Subject: [PATCH 10/10] fixed mem leak in ops_test
---
tensorflow/c/ops_test.cc | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/tensorflow/c/ops_test.cc b/tensorflow/c/ops_test.cc
index 63105b1ad00..5d6ec8df1e3 100644
--- a/tensorflow/c/ops_test.cc
+++ b/tensorflow/c/ops_test.cc
@@ -323,7 +323,8 @@ TEST(OpsTest, ShapeInferenceScalarShape) {
TF_ShapeHandle* TF_scalar_shape = TF_ShapeInferenceContextScalar(C_CTX(&c));
shape_inference::ShapeHandle* scalar_shape =
reinterpret_cast<shape_inference::ShapeHandle*>(TF_scalar_shape);
- ASSERT_EQ("[]", c.DebugString(*scalar_shape));
+ ASSERT_EQ("[]", c.DebugString(*scalar_shape));
+ TF_DeleteShapeHandle(TF_scalar_shape);
}
} // namespace