experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add helper macro to check TRT version. Use to disable call to setType for 5.1.3 since the bug is fixed

Browse Source
This commit is contained in:
Trevor Morris 2019-03-12 17:20:08 -07:00
parent e7f508844e
commit dc72ab94e0
3 changed files with 19 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
tensorflow/compiler/tf2tensorrt/convert/convert_nodes.cc
View File

@ -367,11 +367,13 @@ nvinfer1::ITensor* Converter::CreateConstantLayer(
if (!layer) return nullptr; if (!layer) return nullptr;
const nvinfer1::DataType trt_dtype = trt_weights.type; const nvinfer1::DataType trt_dtype = trt_weights.type;
nvinfer1::ITensor* trt_tensor = layer->getOutput(0); nvinfer1::ITensor* trt_tensor = layer->getOutput(0);
#if !IS_TRT_VERSION_GE(5, 1, 3)
// TODO(laigd): there is a bug in TensorRT 5.0 library that, if we don't set // TODO(laigd): there is a bug in TensorRT 5.0 library that, if we don't set
// the data type below, it will always be kFLOAT regardless what the data type // the data type below, it will always be kFLOAT regardless what the data type
// of the weights is. Once NVIDIA fixes this bug, we should remove the data // of the weights is. Once NVIDIA fixes this bug, we should remove the data
// type setting logic below and test should still pass. // type setting logic below and test should still pass.
trt_tensor->setType(trt_dtype); trt_tensor->setType(trt_dtype);
#endif
return trt_tensor; return trt_tensor;
} }
@ -574,13 +576,13 @@ class TRT_TensorOrWeights::SimpleITensor : public nvinfer1::ITensor {
void setLocation(nvinfer1::TensorLocation location) override {} void setLocation(nvinfer1::TensorLocation location) override {}
#if NV_TENSORRT_MAJOR >= 5 #if IS_TRT_VERSION_GE(5, 0, 0)
bool setDynamicRange(float min, float max) override { return true; } bool setDynamicRange(float min, float max) override { return true; }
float getDynamicRange() const override { return 0; } float getDynamicRange() const override { return 0; }
#endif #endif
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
bool dynamicRangeIsSet() const override { return true; } bool dynamicRangeIsSet() const override { return true; }
void resetDynamicRange() override {} void resetDynamicRange() override {}
@ -1281,7 +1283,7 @@ void Converter::MaybeApplyQuantizationRanges() {
// Infer ranges across marked ops. // Infer ranges across marked ops.
PropagateQuantizationRanges(); PropagateQuantizationRanges();
// Apply ranges. // Apply ranges.
#if NV_TENSORRT_MAJOR >= 5 #if IS_TRT_VERSION_GE(5, 0, 0)
for (auto pair : quantization_ranges_) { for (auto pair : quantization_ranges_) {
nvinfer1::ITensor* tensor = pair.first; nvinfer1::ITensor* tensor = pair.first;
const float range = pair.second; const float range = pair.second;
@ -2297,9 +2299,7 @@ Status ConvertStridedSliceHelper(OpConverterParams* params,
} }
// TRT 5.1 adds a slice layer. For older versions, we attempt to use the // TRT 5.1 adds a slice layer. For older versions, we attempt to use the
// padding layer with negative padding. // padding layer with negative padding.
#if (NV_TENSORRT_MAJOR > 5 || \ #if IS_TRT_VERSION_GE(5, 1, 0) && 0
(NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1)) && \
0
// TODO(laigd): TRT 5.1 RC has a bug when ISliceLayer is used along with // TODO(laigd): TRT 5.1 RC has a bug when ISliceLayer is used along with
// IConcatenationLayer, so disable ISliceLayer for now until it's fixed. // IConcatenationLayer, so disable ISliceLayer for now until it's fixed.
// Use ISliceLayer. // Use ISliceLayer.
@ -3220,7 +3220,7 @@ UnaryOperationMap() {
{"Sqrt", nvinfer1::UnaryOperation::kSQRT}, {"Sqrt", nvinfer1::UnaryOperation::kSQRT},
{"Abs", nvinfer1::UnaryOperation::kABS}, {"Abs", nvinfer1::UnaryOperation::kABS},
{"Reciprocal", nvinfer1::UnaryOperation::kRECIP}, {"Reciprocal", nvinfer1::UnaryOperation::kRECIP},
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
{"Sin", nvinfer1::UnaryOperation::kSIN}, {"Sin", nvinfer1::UnaryOperation::kSIN},
{"Cos", nvinfer1::UnaryOperation::kCOS}, {"Cos", nvinfer1::UnaryOperation::kCOS},
{"Tan", nvinfer1::UnaryOperation::kTAN}, {"Tan", nvinfer1::UnaryOperation::kTAN},

6
tensorflow/compiler/tf2tensorrt/convert/convert_nodes.h
View File

@ -43,6 +43,12 @@ extern const char* const kOutputPHName;
namespace convert { namespace convert {
#define IS_TRT_VERSION_GE(major, minor, patch) \
((NV_TENSORRT_MAJOR > major) || \
(NV_TENSORRT_MAJOR == major && NV_TENSORRT_MINOR > minor) || \
(NV_TENSORRT_MAJOR == major && NV_TENSORRT_MINOR == minor && \
NV_TENSORRT_PATCH >= patch))
struct EngineConnection { struct EngineConnection {
// Constructs a non-control edge. // Constructs a non-control edge.
EngineConnection(const string& outside, int out_id, int out_port, EngineConnection(const string& outside, int out_id, int out_port,

12
tensorflow/compiler/tf2tensorrt/convert/convert_nodes_test.cc
View File

@ -233,7 +233,7 @@ class FakeITensor : public nvinfer1::ITensor {
location_ = location; location_ = location;
} }
#if NV_TENSORRT_MAJOR >= 5 #if IS_TRT_VERSION_GE(5, 0, 0)
bool setDynamicRange(float min, float max) override { bool setDynamicRange(float min, float max) override {
dynamic_range_ = std::max(std::abs(min), std::abs(max)); dynamic_range_ = std::max(std::abs(min), std::abs(max));
return true; return true;
@ -242,7 +242,7 @@ class FakeITensor : public nvinfer1::ITensor {
float getDynamicRange() const override { return dynamic_range_; } float getDynamicRange() const override { return dynamic_range_; }
#endif #endif
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
bool dynamicRangeIsSet() const override { return true; } bool dynamicRangeIsSet() const override { return true; }
void resetDynamicRange() override {} void resetDynamicRange() override {}
@ -845,7 +845,7 @@ TEST_F(ConverterTest, MaybeApplyQuantizationRanges) {
// Input range should be inferred along the chain and applied to tensors. // Input range should be inferred along the chain and applied to tensors.
int8_converter.MaybeApplyQuantizationRanges(); int8_converter.MaybeApplyQuantizationRanges();
#if NV_TENSORRT_MAJOR >= 5 #if IS_TRT_VERSION_GE(5, 0, 0)
EXPECT_EQ(input.getDynamicRange(), 5.0f); EXPECT_EQ(input.getDynamicRange(), 5.0f);
EXPECT_EQ(infer_1.getDynamicRange(), 5.0f); EXPECT_EQ(infer_1.getDynamicRange(), 5.0f);
EXPECT_EQ(infer_2.getDynamicRange(), 5.0f); EXPECT_EQ(infer_2.getDynamicRange(), 5.0f);
@ -2672,7 +2672,7 @@ TEST_F(OpConverterTest, ConvertStridedSlice) {
RunValidationAndConversion(node_def); RunValidationAndConversion(node_def);
} }
// TRT 5.1+ supports strides // TRT 5.1+ supports strides
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
{ {
// Negative strides, should fail. // Negative strides, should fail.
Reset(); Reset();
@ -2735,7 +2735,7 @@ TEST_F(OpConverterTest, ConvertStridedSlice) {
// Same input is used for all tests. // Same input is used for all tests.
const std::vector<float> ok_input = {1, 2, 3, 4, 5, 6}; const std::vector<float> ok_input = {1, 2, 3, 4, 5, 6};
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
const int kStridedSliceOKCases = 23; const int kStridedSliceOKCases = 23;
#else #else
const int kStridedSliceOKCases = 19; const int kStridedSliceOKCases = 19;
@ -2862,7 +2862,7 @@ TEST_F(OpConverterTest, ConvertStridedSlice) {
/*end_mask=*/get_mask({1, 0, 0, 0}), /*end_mask=*/get_mask({1, 0, 0, 0}),
/*expected_output_dims=*/{1, 2, 3}, /*expected_output_dims=*/{1, 2, 3},
/*expected_output=*/{1, 2, 3, 4, 5, 6}}, /*expected_output=*/{1, 2, 3, 4, 5, 6}},
#if NV_TENSORRT_MAJOR > 5 || (NV_TENSORRT_MAJOR == 5 && NV_TENSORRT_MINOR >= 1) #if IS_TRT_VERSION_GE(5, 1, 0)
// Strides // Strides
TestParams{/*input_dims=*/{6}, TestParams{/*input_dims=*/{6},
/*begin=*/{0, 0}, /*end=*/{0, 5}, /*strides=*/{1, 2}, /*begin=*/{0, 0}, /*end=*/{0, 5}, /*strides=*/{1, 2},