You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

2359 lines
110 KiB
C++

/*
* SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: Apache-2.0
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include <algorithm>
#include <cctype>
#include <cstring>
#include <functional>
#include <iostream>
#include <stdexcept>
#include <string>
#include <vector>
#include "NvInfer.h"
#include "logger.h"
#include "sampleOptions.h"
#include "sampleUtils.h"
using namespace nvinfer1;
namespace sample
{
namespace
{
template <typename T>
T stringToValue(const std::string& option)
{
return T{option};
}
template <>
int32_t stringToValue<int32_t>(const std::string& option)
{
return std::stoi(option);
}
template <>
float stringToValue<float>(const std::string& option)
{
return std::stof(option);
}
template <>
double stringToValue<double>(const std::string& option)
{
return std::stod(option);
}
template <>
bool stringToValue<bool>(const std::string& option)
{
return true;
}
template <>
std::vector<int32_t> stringToValue<std::vector<int32_t>>(const std::string& option)
{
std::vector<int32_t> shape;
std::vector<std::string> dimsStrings = splitToStringVec(option, 'x');
for (const auto& d : dimsStrings)
{
shape.push_back(stringToValue<int32_t>(d));
}
return shape;
}
template <>
nvinfer1::DataType stringToValue<nvinfer1::DataType>(const std::string& option)
{
const std::unordered_map<std::string, nvinfer1::DataType> strToDT{{"fp32", nvinfer1::DataType::kFLOAT},
{"fp16", nvinfer1::DataType::kHALF}, {"int8", nvinfer1::DataType::kINT8}, {"fp8", nvinfer1::DataType::kFP8},
{"int32", nvinfer1::DataType::kINT32}};
const auto& dt = strToDT.find(option);
if (dt == strToDT.end())
{
throw std::invalid_argument("Invalid DataType " + option);
}
return dt->second;
}
template <>
nvinfer1::DeviceType stringToValue<nvinfer1::DeviceType>(std::string const& option)
{
std::unordered_map<std::string, nvinfer1::DeviceType> const strToDevice = {
{"GPU", nvinfer1::DeviceType::kGPU},
{"DLA", nvinfer1::DeviceType::kDLA},
};
auto const& device = strToDevice.find(option);
if (device == strToDevice.end())
{
throw std::invalid_argument("Invalid Device Type " + option);
}
return device->second;
}
template <>
nvinfer1::TensorFormats stringToValue<nvinfer1::TensorFormats>(const std::string& option)
{
std::vector<std::string> optionStrings = splitToStringVec(option, '+');
const std::unordered_map<std::string, nvinfer1::TensorFormat> strToFmt{{"chw", nvinfer1::TensorFormat::kLINEAR},
{"chw2", nvinfer1::TensorFormat::kCHW2}, {"chw4", nvinfer1::TensorFormat::kCHW4},
{"hwc8", nvinfer1::TensorFormat::kHWC8}, {"chw16", nvinfer1::TensorFormat::kCHW16},
{"chw32", nvinfer1::TensorFormat::kCHW32}, {"dhwc8", nvinfer1::TensorFormat::kDHWC8},
{"cdhw32", nvinfer1::TensorFormat::kCDHW32}, {"hwc", nvinfer1::TensorFormat::kHWC},
{"dhwc", nvinfer1::TensorFormat::kDHWC}, {"dla_linear", nvinfer1::TensorFormat::kDLA_LINEAR},
{"dla_hwc4", nvinfer1::TensorFormat::kDLA_HWC4}};
nvinfer1::TensorFormats formats{};
for (auto f : optionStrings)
{
const auto& tf = strToFmt.find(f);
if (tf == strToFmt.end())
{
throw std::invalid_argument(std::string("Invalid TensorFormat ") + f);
}
formats |= 1U << static_cast<int32_t>(tf->second);
}
return formats;
}
template <>
IOFormat stringToValue<IOFormat>(const std::string& option)
{
IOFormat ioFormat{};
const size_t colon = option.find(':');
if (colon == std::string::npos)
{
throw std::invalid_argument(std::string("Invalid IOFormat ") + option);
}
ioFormat.first = stringToValue<nvinfer1::DataType>(option.substr(0, colon));
ioFormat.second = stringToValue<nvinfer1::TensorFormats>(option.substr(colon + 1));
return ioFormat;
}
template <>
SparsityFlag stringToValue<SparsityFlag>(std::string const& option)
{
std::unordered_map<std::string, SparsityFlag> const table{
{"disable", SparsityFlag::kDISABLE}, {"enable", SparsityFlag::kENABLE}, {"force", SparsityFlag::kFORCE}};
auto search = table.find(option);
if (search == table.end())
{
throw std::invalid_argument(std::string("Unknown sparsity mode: ") + option);
}
return search->second;
}
template <typename T>
std::pair<std::string, T> splitNameAndValue(const std::string& s)
{
std::string tensorName;
std::string valueString;
// Support 'inputName':Path format for --loadInputs flag when dealing with Windows paths.
// i.e. 'inputName':c:\inputData
std::vector<std::string> quoteNameRange{ splitToStringVec(s, '\'') };
// splitToStringVec returns the entire string when delimiter is not found, so it's size is always at least 1
if (quoteNameRange.size() != 1)
{
if (quoteNameRange.size() != 3)
{
throw std::invalid_argument(std::string("Found invalid number of \'s when parsing ") + s +
std::string(". Expected: 2, received: ") + std::to_string(quoteNameRange.size() -1));
}
// Everything before the second "'" is the name.
tensorName = quoteNameRange[0] + quoteNameRange[1];
// Path is the last string - ignoring leading ":" so slice it with [1:]
valueString = quoteNameRange[2].substr(1);
return std::pair<std::string, T>(tensorName, stringToValue<T>(valueString));
}
// Split on the last :
std::vector<std::string> nameRange{splitToStringVec(s, ':')};
// Everything before the last : is the name
tensorName = nameRange[0];
for (size_t i = 1; i < nameRange.size() - 1; i++)
{
tensorName += ":" + nameRange[i];
}
// Value is the string element after the last :
valueString = nameRange[nameRange.size() - 1];
return std::pair<std::string, T>(tensorName, stringToValue<T>(valueString));
}
template <typename T>
void splitInsertKeyValue(const std::vector<std::string>& kvList, T& map)
{
for (const auto& kv : kvList)
{
map.insert(splitNameAndValue<typename T::mapped_type>(kv));
}
}
const char* boolToEnabled(bool enable)
{
return enable ? "Enabled" : "Disabled";
}
//! A helper function similar to sep.join(list) in Python.
template <typename T>
std::string joinValuesToString(std::vector<T> const& list, std::string const& sep)
{
std::ostringstream os;
for (int32_t i = 0, n = list.size(); i < n; ++i)
{
os << list[i];
if (i != n - 1)
{
os << sep;
}
}
return os.str();
}
template <typename T, size_t N>
std::string joinValuesToString(std::array<T, N> const& list, std::string const& sep)
{
return joinValuesToString(std::vector<T>(list.begin(), list.end()), sep);
}
//! Check if input option exists in input arguments.
//! If it does: return its value, erase the argument and return true.
//! If it does not: return false.
template <typename T>
bool getAndDelOption(Arguments& arguments, const std::string& option, T& value)
{
const auto match = arguments.find(option);
if (match != arguments.end())
{
value = stringToValue<T>(match->second);
arguments.erase(match);
return true;
}
return false;
}
//! Check if input option exists in input arguments.
//! If it does: return false in value, erase the argument and return true.
//! If it does not: return false.
bool getAndDelNegOption(Arguments& arguments, const std::string& option, bool& value)
{
bool dummy;
if (getAndDelOption(arguments, option, dummy))
{
value = false;
return true;
}
return false;
}
//! Check if input option exists in input arguments.
//! If it does: add all the matched arg values to values vector, erase the argument and return true.
//! If it does not: return false.
template <typename T>
bool getAndDelRepeatedOption(Arguments& arguments, const std::string& option, std::vector<T>& values)
{
const auto match = arguments.equal_range(option);
if (match.first == match.second)
{
return false;
}
auto addToValues
= [&values](Arguments::value_type& argValue) { values.emplace_back(stringToValue<T>(argValue.second)); };
std::for_each(match.first, match.second, addToValues);
arguments.erase(match.first, match.second);
return true;
}
void insertShapesBuild(BuildOptions::ShapeProfile& shapes, nvinfer1::OptProfileSelector selector,
const std::string& name, const std::vector<int32_t>& dims)
{
shapes[name][static_cast<size_t>(selector)] = dims;
}
void insertShapesInference(
InferenceOptions::ShapeProfile& shapes, std::string const& name, std::vector<int32_t> const& dims)
{
shapes[name] = dims;
}
std::string removeSingleQuotationMarks(std::string& str)
{
std::vector<std::string> strList{splitToStringVec(str, '\'')};
// Remove all the escaped single quotation marks
std::string retVal;
// Do not really care about unterminated sequences
for (size_t i = 0; i < strList.size(); i++)
{
retVal += strList[i];
}
return retVal;
}
void getLayerPrecisions(Arguments& arguments, char const* argument, LayerPrecisions& layerPrecisions)
{
std::string list;
if (!getAndDelOption(arguments, argument, list))
{
return;
}
// The layerPrecisions flag contains comma-separated layerName:precision pairs.
std::vector<std::string> precisionList{splitToStringVec(list, ',')};
for (auto const& s : precisionList)
{
auto namePrecisionPair = splitNameAndValue<nvinfer1::DataType>(s);
auto const layerName = removeSingleQuotationMarks(namePrecisionPair.first);
layerPrecisions[layerName] = namePrecisionPair.second;
}
}
void getLayerOutputTypes(Arguments& arguments, char const* argument, LayerOutputTypes& layerOutputTypes)
{
std::string list;
if (!getAndDelOption(arguments, argument, list))
{
return;
}
// The layerOutputTypes flag contains comma-separated layerName:types pairs.
std::vector<std::string> precisionList{splitToStringVec(list, ',')};
for (auto const& s : precisionList)
{
auto namePrecisionPair = splitNameAndValue<std::string>(s);
auto const layerName = removeSingleQuotationMarks(namePrecisionPair.first);
auto const typeStrings = splitToStringVec(namePrecisionPair.second, '+');
std::vector<nvinfer1::DataType> typeVec(typeStrings.size(), nvinfer1::DataType::kFLOAT);
std::transform(typeStrings.begin(), typeStrings.end(), typeVec.begin(), stringToValue<nvinfer1::DataType>);
layerOutputTypes[layerName] = typeVec;
}
}
void getLayerDeviceTypes(Arguments& arguments, char const* argument, LayerDeviceTypes& layerDeviceTypes)
{
std::string list;
if (!getAndDelOption(arguments, argument, list))
{
return;
}
// The layerDeviceTypes flag contains comma-separated layerName:deviceType pairs.
std::vector<std::string> deviceList{splitToStringVec(list, ',')};
for (auto const& s : deviceList)
{
auto nameDevicePair = splitNameAndValue<std::string>(s);
auto const layerName = removeSingleQuotationMarks(nameDevicePair.first);
layerDeviceTypes[layerName] = stringToValue<nvinfer1::DeviceType>(nameDevicePair.second);
}
}
bool getShapesBuild(Arguments& arguments, BuildOptions::ShapeProfile& shapes, char const* argument,
nvinfer1::OptProfileSelector selector)
{
std::string list;
bool retVal = getAndDelOption(arguments, argument, list);
std::vector<std::string> shapeList{splitToStringVec(list, ',')};
for (const auto& s : shapeList)
{
auto nameDimsPair = splitNameAndValue<std::vector<int32_t>>(s);
auto tensorName = removeSingleQuotationMarks(nameDimsPair.first);
auto dims = nameDimsPair.second;
insertShapesBuild(shapes, selector, tensorName, dims);
}
return retVal;
}
bool getShapesInference(Arguments& arguments, InferenceOptions::ShapeProfile& shapes, const char* argument)
{
std::string list;
bool retVal = getAndDelOption(arguments, argument, list);
std::vector<std::string> shapeList{splitToStringVec(list, ',')};
for (const auto& s : shapeList)
{
auto nameDimsPair = splitNameAndValue<std::vector<int32_t>>(s);
auto tensorName = removeSingleQuotationMarks(nameDimsPair.first);
auto dims = nameDimsPair.second;
insertShapesInference(shapes, tensorName, dims);
}
return retVal;
}
void fillShapes(BuildOptions::ShapeProfile& shapes, std::string const& name, ShapeRange const& sourceShapeRange,
nvinfer1::OptProfileSelector minDimsSource, nvinfer1::OptProfileSelector optDimsSource,
nvinfer1::OptProfileSelector maxDimsSource)
{
insertShapesBuild(
shapes, nvinfer1::OptProfileSelector::kMIN, name, sourceShapeRange[static_cast<size_t>(minDimsSource)]);
insertShapesBuild(
shapes, nvinfer1::OptProfileSelector::kOPT, name, sourceShapeRange[static_cast<size_t>(optDimsSource)]);
insertShapesBuild(
shapes, nvinfer1::OptProfileSelector::kMAX, name, sourceShapeRange[static_cast<size_t>(maxDimsSource)]);
}
void processShapes(BuildOptions::ShapeProfile& shapes, bool minShapes, bool optShapes, bool maxShapes, bool calib)
{
// Only accept optShapes only or all three of minShapes, optShapes, maxShapes when calib is set
if (((minShapes || maxShapes) && !optShapes) // minShapes only, maxShapes only, both minShapes and maxShapes
|| (minShapes && !maxShapes && optShapes) // both minShapes and optShapes
|| (!minShapes && maxShapes && optShapes)) // both maxShapes and optShapes
{
if (calib)
{
throw std::invalid_argument(
"Must specify only --optShapesCalib or all of --minShapesCalib, --optShapesCalib, --maxShapesCalib");
}
}
if (!minShapes && !optShapes && !maxShapes)
{
return;
}
BuildOptions::ShapeProfile newShapes;
for (auto& s : shapes)
{
nvinfer1::OptProfileSelector minDimsSource, optDimsSource, maxDimsSource;
minDimsSource = nvinfer1::OptProfileSelector::kMIN;
optDimsSource = nvinfer1::OptProfileSelector::kOPT;
maxDimsSource = nvinfer1::OptProfileSelector::kMAX;
// Populate missing minShapes
if (!minShapes)
{
if (optShapes)
{
minDimsSource = optDimsSource;
sample::gLogWarning << "optShapes is being broadcasted to minShapes for tensor " << s.first
<< std::endl;
}
else
{
minDimsSource = maxDimsSource;
sample::gLogWarning << "maxShapes is being broadcasted to minShapes for tensor " << s.first
<< std::endl;
}
}
// Populate missing optShapes
if (!optShapes)
{
if (maxShapes)
{
optDimsSource = maxDimsSource;
sample::gLogWarning << "maxShapes is being broadcasted to optShapes for tensor " << s.first
<< std::endl;
}
else
{
optDimsSource = minDimsSource;
sample::gLogWarning << "minShapes is being broadcasted to optShapes for tensor " << s.first
<< std::endl;
}
}
// Populate missing maxShapes
if (!maxShapes)
{
if (optShapes)
{
maxDimsSource = optDimsSource;
sample::gLogWarning << "optShapes is being broadcasted to maxShapes for tensor " << s.first
<< std::endl;
}
else
{
maxDimsSource = minDimsSource;
sample::gLogWarning << "minShapes is being broadcasted to maxShapes for tensor " << s.first
<< std::endl;
}
}
fillShapes(newShapes, s.first, s.second, minDimsSource, optDimsSource, maxDimsSource);
}
shapes = newShapes;
}
template <typename T>
void printShapes(std::ostream& os, const char* phase, const T& shapes)
{
if (shapes.empty())
{
os << "Input " << phase << " shapes: model" << std::endl;
}
else
{
for (const auto& s : shapes)
{
os << "Input " << phase << " shape: " << s.first << "=" << s.second << std::endl;
}
}
}
std::ostream& printBatch(std::ostream& os, int32_t maxBatch)
{
if (maxBatch != maxBatchNotProvided)
{
os << maxBatch;
}
else
{
os << "explicit batch";
}
return os;
}
std::ostream& printTacticSources(
std::ostream& os, nvinfer1::TacticSources enabledSources, nvinfer1::TacticSources disabledSources)
{
if (!enabledSources && !disabledSources)
{
os << "Using default tactic sources";
}
else
{
auto const addSource = [&](uint32_t source, std::string const& name) {
if (enabledSources & source)
{
os << name << " [ON], ";
}
else if (disabledSources & source)
{
os << name << " [OFF], ";
}
};
addSource(1U << static_cast<uint32_t>(nvinfer1::TacticSource::kCUBLAS), "cublas");
addSource(1U << static_cast<uint32_t>(nvinfer1::TacticSource::kCUBLAS_LT), "cublasLt");
addSource(1U << static_cast<uint32_t>(nvinfer1::TacticSource::kCUDNN), "cudnn");
addSource(1U << static_cast<uint32_t>(nvinfer1::TacticSource::kEDGE_MASK_CONVOLUTIONS), "edge mask convolutions");
addSource(1U << static_cast<uint32_t>(nvinfer1::TacticSource::kJIT_CONVOLUTIONS), "JIT convolutions");
}
return os;
}
std::ostream& printPrecision(std::ostream& os, BuildOptions const& options)
{
os << "FP32";
if (options.fp16)
{
os << "+FP16";
}
if (options.int8)
{
os << "+INT8";
}
if (options.fp8)
{
os << "+FP8";
}
if (options.precisionConstraints == PrecisionConstraints::kOBEY)
{
os << " (obey precision constraints)";
}
if (options.precisionConstraints == PrecisionConstraints::kPREFER)
{
os << " (prefer precision constraints)";
}
return os;
}
std::ostream& printTempfileControls(std::ostream& os, TempfileControlFlags const tempfileControls)
{
auto getFlag = [&](TempfileControlFlag f) -> char const* {
bool allowed = !!(tempfileControls & (1U << static_cast<int64_t>(f)));
return allowed ? "allow" : "deny";
};
auto const inMemory = getFlag(TempfileControlFlag::kALLOW_IN_MEMORY_FILES);
auto const temporary = getFlag(TempfileControlFlag::kALLOW_TEMPORARY_FILES);
os << "{ in_memory: " << inMemory << ", temporary: " << temporary << " }";
return os;
}
std::ostream& printTimingCache(std::ostream& os, TimingCacheMode const& timingCacheMode)
{
switch (timingCacheMode)
{
case TimingCacheMode::kGLOBAL: os << "global"; break;
case TimingCacheMode::kLOCAL: os << "local"; break;
case TimingCacheMode::kDISABLE: os << "disable"; break;
}
return os;
}
std::ostream& printSparsity(std::ostream& os, BuildOptions const& options)
{
switch (options.sparsity)
{
case SparsityFlag::kDISABLE: os << "Disabled"; break;
case SparsityFlag::kENABLE: os << "Enabled"; break;
case SparsityFlag::kFORCE: os << "Forced"; break;
}
return os;
}
std::ostream& printMemoryPools(std::ostream& os, BuildOptions const& options)
{
auto const printValueOrDefault = [&os](double const val) {
if (val >= 0)
{
os << val << " MiB";
}
else
{
os << "default";
}
};
os << "workspace: ";
printValueOrDefault(options.workspace);
os << ", ";
os << "dlaSRAM: ";
printValueOrDefault(options.dlaSRAM);
os << ", ";
os << "dlaLocalDRAM: ";
printValueOrDefault(options.dlaLocalDRAM);
os << ", ";
os << "dlaGlobalDRAM: ";
printValueOrDefault(options.dlaGlobalDRAM);
return os;
}
std::string previewFeatureToString(PreviewFeature feature)
{
// clang-format off
switch (feature)
{
case PreviewFeature::kFASTER_DYNAMIC_SHAPES_0805: return "kFASTER_DYNAMIC_SHAPES_0805";
case PreviewFeature::kDISABLE_EXTERNAL_TACTIC_SOURCES_FOR_CORE_0805: return "kDISABLE_EXTERNAL_TACTIC_SOURCES_FOR_CORE_0805";
case PreviewFeature::kPROFILE_SHARING_0806: return "kPROFILE_SHARING_0806";
}
return "Invalid Preview Feature";
// clang-format on
}
std::ostream& printPreviewFlags(std::ostream& os, BuildOptions const& options)
{
if (options.previewFeatures.empty())
{
os << "Use default preview flags.";
return os;
}
auto const addFlag = [&](PreviewFeature feat) {
int32_t featVal = static_cast<int32_t>(feat);
if (options.previewFeatures.find(featVal) != options.previewFeatures.end())
{
os << previewFeatureToString(feat) << (options.previewFeatures.at(featVal) ? " [ON], " : " [OFF], ");
}
};
addFlag(PreviewFeature::kFASTER_DYNAMIC_SHAPES_0805);
addFlag(PreviewFeature::kDISABLE_EXTERNAL_TACTIC_SOURCES_FOR_CORE_0805);
addFlag(PreviewFeature::kPROFILE_SHARING_0806);
return os;
}
} // namespace
Arguments argsToArgumentsMap(int32_t argc, char* argv[])
{
Arguments arguments;
for (int32_t i = 1; i < argc; ++i)
{
auto valuePtr = strchr(argv[i], '=');
if (valuePtr)
{
std::string value{valuePtr + 1};
arguments.emplace(std::string(argv[i], valuePtr - argv[i]), value);
}
else
{
arguments.emplace(argv[i], "");
}
}
return arguments;
}
void BaseModelOptions::parse(Arguments& arguments)
{
if (getAndDelOption(arguments, "--onnx", model))
{
format = ModelFormat::kONNX;
}
else if (getAndDelOption(arguments, "--uff", model))
{
format = ModelFormat::kUFF;
}
else if (getAndDelOption(arguments, "--model", model))
{
format = ModelFormat::kCAFFE;
}
}
void UffInput::parse(Arguments& arguments)
{
getAndDelOption(arguments, "--uffNHWC", NHWC);
std::vector<std::string> args;
if (getAndDelRepeatedOption(arguments, "--uffInput", args))
{
for (const auto& i : args)
{
std::vector<std::string> values{splitToStringVec(i, ',')};
if (values.size() == 4)
{
nvinfer1::Dims3 dims{std::stoi(values[1]), std::stoi(values[2]), std::stoi(values[3])};
inputs.emplace_back(values[0], dims);
}
else
{
throw std::invalid_argument(std::string("Invalid uffInput ") + i);
}
}
}
}
void ModelOptions::parse(Arguments& arguments)
{
baseModel.parse(arguments);
switch (baseModel.format)
{
case ModelFormat::kCAFFE:
{
getAndDelOption(arguments, "--deploy", prototxt);
break;
}
case ModelFormat::kUFF:
{
uffInputs.parse(arguments);
if (uffInputs.inputs.empty())
{
throw std::invalid_argument("Uff models require at least one input");
}
break;
}
case ModelFormat::kONNX: break;
case ModelFormat::kANY:
{
if (getAndDelOption(arguments, "--deploy", prototxt))
{
baseModel.format = ModelFormat::kCAFFE;
}
break;
}
}
// The --output flag should only be used with Caffe and UFF. It has no effect on ONNX.
std::vector<std::string> outArgs;
if (getAndDelRepeatedOption(arguments, "--output", outArgs))
{
for (const auto& o : outArgs)
{
for (auto& v : splitToStringVec(o, ','))
{
outputs.emplace_back(std::move(v));
}
}
}
if (baseModel.format == ModelFormat::kCAFFE || baseModel.format == ModelFormat::kUFF)
{
if (outputs.empty())
{
throw std::invalid_argument("Caffe and Uff models require at least one output");
}
}
else if (baseModel.format == ModelFormat::kONNX)
{
if (!outputs.empty())
{
throw std::invalid_argument("The --output flag should not be used with ONNX models.");
}
}
}
void getTempfileControls(Arguments& arguments, char const* argument, TempfileControlFlags& tempfileControls)
{
std::string list;
if (!getAndDelOption(arguments, argument, list))
{
return;
}
std::vector<std::string> controlList{splitToStringVec(list, ',')};
for (auto const& s : controlList)
{
auto controlAllowPair = splitNameAndValue<std::string>(s);
bool allowed{false};
int32_t offset{-1};
if (controlAllowPair.second.compare("allow") == 0)
{
allowed = true;
}
else if (controlAllowPair.second.compare("deny") != 0)
{
throw std::invalid_argument("--tempfileControls value should be `deny` or `allow`");
}
if (controlAllowPair.first.compare("in_memory") == 0)
{
offset = static_cast<int32_t>(TempfileControlFlag::kALLOW_IN_MEMORY_FILES);
}
else if (controlAllowPair.first.compare("temporary") == 0)
{
offset = static_cast<int32_t>(TempfileControlFlag::kALLOW_TEMPORARY_FILES);
}
else
{
throw std::invalid_argument(std::string{"Unknown --tempfileControls key "} + controlAllowPair.first);
}
if (allowed)
{
tempfileControls |= (1U << offset);
}
else
{
tempfileControls &= ~(1U << offset);
}
}
}
void BuildOptions::parse(Arguments& arguments)
{
auto getFormats = [&arguments](std::vector<IOFormat>& formatsVector, const char* argument) {
std::string list;
getAndDelOption(arguments, argument, list);
std::vector<std::string> formats{splitToStringVec(list, ',')};
for (const auto& f : formats)
{
formatsVector.push_back(stringToValue<IOFormat>(f));
}
};
getFormats(inputFormats, "--inputIOFormats");
getFormats(outputFormats, "--outputIOFormats");
bool addedExplicitBatchFlag{false};
getAndDelOption(arguments, "--explicitBatch", addedExplicitBatchFlag);
if (addedExplicitBatchFlag)
{
sample::gLogWarning << "--explicitBatch flag has been deprecated and has no effect!" << std::endl;
sample::gLogWarning << "Explicit batch dim is automatically enabled if input model is ONNX or if dynamic "
<< "shapes are provided when the engine is built." << std::endl;
}
bool minShapes = getShapesBuild(arguments, shapes, "--minShapes", nvinfer1::OptProfileSelector::kMIN);
bool optShapes = getShapesBuild(arguments, shapes, "--optShapes", nvinfer1::OptProfileSelector::kOPT);
bool maxShapes = getShapesBuild(arguments, shapes, "--maxShapes", nvinfer1::OptProfileSelector::kMAX);
processShapes(shapes, minShapes, optShapes, maxShapes, false);
bool minShapesCalib
= getShapesBuild(arguments, shapesCalib, "--minShapesCalib", nvinfer1::OptProfileSelector::kMIN);
bool optShapesCalib
= getShapesBuild(arguments, shapesCalib, "--optShapesCalib", nvinfer1::OptProfileSelector::kOPT);
bool maxShapesCalib
= getShapesBuild(arguments, shapesCalib, "--maxShapesCalib", nvinfer1::OptProfileSelector::kMAX);
processShapes(shapesCalib, minShapesCalib, optShapesCalib, maxShapesCalib, true);
bool addedExplicitPrecisionFlag{false};
getAndDelOption(arguments, "--explicitPrecision", addedExplicitPrecisionFlag);
if (addedExplicitPrecisionFlag)
{
sample::gLogWarning << "--explicitPrecision flag has been deprecated and has no effect!" << std::endl;
}
if (getAndDelOption(arguments, "--workspace", workspace))
{
sample::gLogWarning << "--workspace flag has been deprecated by --memPoolSize flag." << std::endl;
}
std::string memPoolSizes;
getAndDelOption(arguments, "--memPoolSize", memPoolSizes);
std::vector<std::string> memPoolSpecs{splitToStringVec(memPoolSizes, ',')};
for (auto const& memPoolSpec : memPoolSpecs)
{
std::string memPoolName;
double memPoolSize;
std::tie(memPoolName, memPoolSize) = splitNameAndValue<double>(memPoolSpec);
if (memPoolSize < 0)
{
throw std::invalid_argument(std::string("Negative memory pool size: ") + std::to_string(memPoolSize));
}
if (memPoolName == "workspace")
{
workspace = memPoolSize;
}
else if (memPoolName == "dlaSRAM")
{
dlaSRAM = memPoolSize;
}
else if (memPoolName == "dlaLocalDRAM")
{
dlaLocalDRAM = memPoolSize;
}
else if (memPoolName == "dlaGlobalDRAM")
{
dlaGlobalDRAM = memPoolSize;
}
else if (!memPoolName.empty())
{
throw std::invalid_argument(std::string("Unknown memory pool: ") + memPoolName);
}
}
getAndDelOption(arguments, "--maxBatch", maxBatch);
getAndDelOption(arguments, "--minTiming", minTiming);
getAndDelOption(arguments, "--avgTiming", avgTiming);
bool best{false};
getAndDelOption(arguments, "--best", best);
if (best)
{
int8 = true;
fp16 = true;
}
getAndDelOption(arguments, "--refit", refittable);
// --vc and --versionCompatible are synonyms
getAndDelOption(arguments, "--vc", versionCompatible);
if (!versionCompatible)
{
getAndDelOption(arguments, "--versionCompatible", versionCompatible);
}
getAndDelOption(arguments, "--excludeLeanRuntime", excludeLeanRuntime);
getAndDelNegOption(arguments, "--noTF32", tf32);
getAndDelOption(arguments, "--fp16", fp16);
getAndDelOption(arguments, "--int8", int8);
getAndDelOption(arguments, "--fp8", fp8);
if (fp8 && int8)
{
throw std::invalid_argument("Invalid usage, fp8 and int8 aren't allowed to be enabled together.");
}
getAndDelOption(arguments, "--safe", safe);
getAndDelOption(arguments, "--consistency", consistency);
getAndDelOption(arguments, "--restricted", restricted);
if (getAndDelOption(arguments, "--buildOnly", skipInference))
{
sample::gLogWarning << "--buildOnly flag has been deprecated by --skipInference flag." << std::endl;
}
getAndDelOption(arguments, "--skipInference", skipInference);
getAndDelOption(arguments, "--directIO", directIO);
std::string precisionConstraintsString;
getAndDelOption(arguments, "--precisionConstraints", precisionConstraintsString);
if (!precisionConstraintsString.empty())
{
const std::unordered_map<std::string, PrecisionConstraints> precisionConstraintsMap
= {{"obey", PrecisionConstraints::kOBEY}, {"prefer", PrecisionConstraints::kPREFER},
{"none", PrecisionConstraints::kNONE}};
auto it = precisionConstraintsMap.find(precisionConstraintsString);
if (it == precisionConstraintsMap.end())
{
throw std::invalid_argument(std::string("Unknown precision constraints: ") + precisionConstraintsString);
}
precisionConstraints = it->second;
}
else
{
precisionConstraints = PrecisionConstraints::kNONE;
}
getLayerPrecisions(arguments, "--layerPrecisions", layerPrecisions);
getLayerOutputTypes(arguments, "--layerOutputTypes", layerOutputTypes);
getLayerDeviceTypes(arguments, "--layerDeviceTypes", layerDeviceTypes);
if (layerPrecisions.empty() && layerOutputTypes.empty() && precisionConstraints != PrecisionConstraints::kNONE)
{
sample::gLogWarning << R"(When --precisionConstraints flag is set to "obey" or "prefer", please add )"
<< "--layerPrecision/--layerOutputTypes flags to set layer-wise precisions and output "
<< "types." << std::endl;
}
else if ((!layerPrecisions.empty() || !layerOutputTypes.empty())
&& precisionConstraints == PrecisionConstraints::kNONE)
{
sample::gLogWarning << "--layerPrecision/--layerOutputTypes flags have no effect when --precisionConstraints "
<< R"(flag is set to "none".)" << std::endl;
}
getAndDelOption(arguments, "--sparsity", sparsity);
bool calibCheck = getAndDelOption(arguments, "--calib", calibration);
if (int8 && calibCheck && !shapes.empty() && shapesCalib.empty())
{
shapesCalib = shapes;
}
std::string profilingVerbosityString;
if (getAndDelOption(arguments, "--nvtxMode", profilingVerbosityString))
{
sample::gLogWarning << "--nvtxMode flag has been deprecated by --profilingVerbosity flag." << std::endl;
}
getAndDelOption(arguments, "--profilingVerbosity", profilingVerbosityString);
if (profilingVerbosityString == "layer_names_only")
{
profilingVerbosity = nvinfer1::ProfilingVerbosity::kLAYER_NAMES_ONLY;
}
else if (profilingVerbosityString == "none")
{
profilingVerbosity = nvinfer1::ProfilingVerbosity::kNONE;
}
else if (profilingVerbosityString == "detailed")
{
profilingVerbosity = nvinfer1::ProfilingVerbosity::kDETAILED;
}
else if (profilingVerbosityString == "default")
{
sample::gLogWarning << "--profilingVerbosity=default has been deprecated by "
"--profilingVerbosity=layer_names_only."
<< std::endl;
profilingVerbosity = nvinfer1::ProfilingVerbosity::kLAYER_NAMES_ONLY;
}
else if (profilingVerbosityString == "verbose")
{
sample::gLogWarning << "--profilingVerbosity=verbose has been deprecated by --profilingVerbosity=detailed."
<< std::endl;
profilingVerbosity = nvinfer1::ProfilingVerbosity::kDETAILED;
}
else if (!profilingVerbosityString.empty())
{
throw std::invalid_argument(std::string("Unknown profilingVerbosity: ") + profilingVerbosityString);
}
if (getAndDelOption(arguments, "--loadEngine", engine))
{
load = true;
}
if (getAndDelOption(arguments, "--saveEngine", engine))
{
save = true;
}
if (load && save)
{
throw std::invalid_argument("Incompatible load and save engine options selected");
}
std::string tacticSourceArgs;
if (getAndDelOption(arguments, "--tacticSources", tacticSourceArgs))
{
std::vector<std::string> tacticList = splitToStringVec(tacticSourceArgs, ',');
for (auto& t : tacticList)
{
bool enable{false};
if (t.front() == '+')
{
enable = true;
}
else if (t.front() != '-')
{
throw std::invalid_argument(
"Tactic source must be prefixed with + or -, indicating whether it should be enabled or disabled "
"respectively.");
}
t.erase(0, 1);
const auto toUpper = [](std::string& sourceName) {
std::transform(
sourceName.begin(), sourceName.end(), sourceName.begin(), [](char c) { return std::toupper(c); });
return sourceName;
};
nvinfer1::TacticSource source{};
t = toUpper(t);
if (t == "CUBLAS")
{
source = nvinfer1::TacticSource::kCUBLAS;
}
else if (t == "CUBLASLT" || t == "CUBLAS_LT")
{
source = nvinfer1::TacticSource::kCUBLAS_LT;
}
else if (t == "CUDNN")
{
source = nvinfer1::TacticSource::kCUDNN;
}
else if (t == "EDGE_MASK_CONVOLUTIONS")
{
source = nvinfer1::TacticSource::kEDGE_MASK_CONVOLUTIONS;
}
else if (t == "JIT_CONVOLUTIONS")
{
source = nvinfer1::TacticSource::kJIT_CONVOLUTIONS;
}
else
{
throw std::invalid_argument(std::string("Unknown tactic source: ") + t);
}
uint32_t sourceBit = 1U << static_cast<uint32_t>(source);
if (enable)
{
enabledTactics |= sourceBit;
}
else
{
disabledTactics |= sourceBit;
}
if (enabledTactics & disabledTactics)
{
throw std::invalid_argument(std::string("Cannot enable and disable ") + t);
}
}
}
bool noBuilderCache{false};
getAndDelOption(arguments, "--noBuilderCache", noBuilderCache);
getAndDelOption(arguments, "--timingCacheFile", timingCacheFile);
if (noBuilderCache)
{
timingCacheMode = TimingCacheMode::kDISABLE;
}
else if (!timingCacheFile.empty())
{
timingCacheMode = TimingCacheMode::kGLOBAL;
}
else
{
timingCacheMode = TimingCacheMode::kLOCAL;
}
getAndDelOption(arguments, "--heuristic", heuristic);
getAndDelOption(arguments, "--builderOptimizationLevel", builderOptimizationLevel);
std::string hardwareCompatibleArgs;
getAndDelOption(arguments, "--hardwareCompatibilityLevel", hardwareCompatibleArgs);
if (hardwareCompatibleArgs == "none" || hardwareCompatibleArgs.empty())
{
hardwareCompatibilityLevel = HardwareCompatibilityLevel::kNONE;
}
else if (samplesCommon::toLower(hardwareCompatibleArgs) == "ampere+")
{
hardwareCompatibilityLevel = HardwareCompatibilityLevel::kAMPERE_PLUS;
}
else
{
throw std::invalid_argument(std::string("Unknown hardwareCompatibilityLevel: ") + hardwareCompatibleArgs
+ ". Valid options: none, ampere+.");
}
getAndDelOption(arguments, "--maxAuxStreams", maxAuxStreams);
std::string previewFeaturesBuf;
getAndDelOption(arguments, "--preview", previewFeaturesBuf);
std::vector<std::string> previewFeaturesVec{splitToStringVec(previewFeaturesBuf, ',')};
for (auto featureName : previewFeaturesVec)
{
bool enable{false};
if (featureName.front() == '+')
{
enable = true;
}
else if (featureName.front() != '-')
{
throw std::invalid_argument(
"Preview features must be prefixed with + or -, indicating whether it should be enabled or disabled "
"respectively.");
}
featureName.erase(0, 1);
PreviewFeature feat{};
if (featureName == "profileSharing0806")
{
feat = PreviewFeature::kPROFILE_SHARING_0806;
}
else if (featureName == "fasterDynamicShapes0805")
{
feat = PreviewFeature::kFASTER_DYNAMIC_SHAPES_0805;
}
else if (featureName == "disableExternalTacticSourcesForCore0805")
{
feat = PreviewFeature::kDISABLE_EXTERNAL_TACTIC_SOURCES_FOR_CORE_0805;
}
else
{
throw std::invalid_argument(std::string("Unknown preview feature: ") + featureName);
}
previewFeatures[static_cast<int32_t>(feat)] = enable;
}
getAndDelOption(arguments, "--tempdir", tempdir);
getTempfileControls(arguments, "--tempfileControls", tempfileControls);
std::string runtimeMode;
getAndDelOption(arguments, "--useRuntime", runtimeMode);
if (runtimeMode == "full")
{
useRuntime = RuntimeMode::kFULL;
}
else if (runtimeMode == "dispatch")
{
useRuntime = RuntimeMode::kDISPATCH;
}
else if (runtimeMode == "lean")
{
useRuntime = RuntimeMode::kLEAN;
}
else if (!runtimeMode.empty())
{
throw std::invalid_argument(std::string("Unknown useRuntime: ") + runtimeMode);
}
if ((useRuntime == RuntimeMode::kDISPATCH || useRuntime == RuntimeMode::kLEAN) && !versionCompatible)
{
versionCompatible = true;
sample::gLogWarning << "Implicitly enabling --versionCompatible since --useRuntime=" << runtimeMode
<< " is set." << std::endl;
}
if (useRuntime != RuntimeMode::kFULL && !load)
{
throw std::invalid_argument(std::string("Building a TensorRT engine requires --useRuntime=full."));
}
getAndDelOption(arguments, "--leanDLLPath", leanDLLPath);
}
void SystemOptions::parse(Arguments& arguments)
{
getAndDelOption(arguments, "--device", device);
getAndDelOption(arguments, "--useDLACore", DLACore);
getAndDelOption(arguments, "--allowGPUFallback", fallback);
std::string pluginName;
while (getAndDelOption(arguments, "--plugins", pluginName))
{
sample::gLogWarning << "--plugins flag has been deprecated, use --staticPlugins flag instead." << std::endl;
plugins.emplace_back(pluginName);
}
while (getAndDelOption(arguments, "--staticPlugins", pluginName))
{
plugins.emplace_back(pluginName);
}
while (getAndDelOption(arguments, "--setPluginsToSerialize", pluginName))
{
setPluginsToSerialize.emplace_back(pluginName);
}
while (getAndDelOption(arguments, "--dynamicPlugins", pluginName))
{
dynamicPlugins.emplace_back(pluginName);
}
getAndDelOption(arguments, "--ignoreParsedPluginLibs", ignoreParsedPluginLibs);
}
void InferenceOptions::parse(Arguments& arguments)
{
if (getAndDelOption(arguments, "--streams", infStreams))
{
sample::gLogWarning << "--streams flag has been deprecated, use --infStreams flag instead." << std::endl;
}
getAndDelOption(arguments, "--infStreams", infStreams);
getAndDelOption(arguments, "--iterations", iterations);
getAndDelOption(arguments, "--duration", duration);
getAndDelOption(arguments, "--warmUp", warmup);
getAndDelOption(arguments, "--sleepTime", sleep);
getAndDelOption(arguments, "--idleTime", idle);
bool exposeDMA{false};
if (getAndDelOption(arguments, "--exposeDMA", exposeDMA))
{
overlap = !exposeDMA;
}
getAndDelOption(arguments, "--noDataTransfers", skipTransfers);
getAndDelOption(arguments, "--useManagedMemory", useManaged);
getAndDelOption(arguments, "--useSpinWait", spin);
getAndDelOption(arguments, "--threads", threads);
getAndDelOption(arguments, "--useCudaGraph", graph);
getAndDelOption(arguments, "--separateProfileRun", rerun);
getAndDelOption(arguments, "--timeDeserialize", timeDeserialize);
getAndDelOption(arguments, "--timeRefit", timeRefit);
getAndDelOption(arguments, "--persistentCacheRatio", persistentCacheRatio);
std::string list;
getAndDelOption(arguments, "--loadInputs", list);
std::vector<std::string> inputsList{splitToStringVec(list, ',')};
splitInsertKeyValue(inputsList, inputs);
getShapesInference(arguments, shapes, "--shapes");
getAndDelOption(arguments, "--batch", batch);
}
void ReportingOptions::parse(Arguments& arguments)
{
getAndDelOption(arguments, "--avgRuns", avgs);
getAndDelOption(arguments, "--verbose", verbose);
getAndDelOption(arguments, "--dumpRefit", refit);
getAndDelOption(arguments, "--dumpOutput", output);
getAndDelOption(arguments, "--dumpRawBindingsToFile", dumpRawBindings);
getAndDelOption(arguments, "--dumpProfile", profile);
getAndDelOption(arguments, "--dumpLayerInfo", layerInfo);
getAndDelOption(arguments, "--exportTimes", exportTimes);
getAndDelOption(arguments, "--exportOutput", exportOutput);
getAndDelOption(arguments, "--exportProfile", exportProfile);
getAndDelOption(arguments, "--exportLayerInfo", exportLayerInfo);
std::string percentileString;
getAndDelOption(arguments, "--percentile", percentileString);
std::vector<std::string> percentileStrings = splitToStringVec(percentileString, ',');
if (!percentileStrings.empty())
{
percentiles.clear();
}
for (const auto& p : percentileStrings)
{
percentiles.push_back(stringToValue<float>(p));
}
for (auto percentile : percentiles)
{
if (percentile < 0.F || percentile > 100.F)
{
throw std::invalid_argument(std::string("Percentile ") + std::to_string(percentile) + "is not in [0,100]");
}
}
}
bool parseHelp(Arguments& arguments)
{
bool helpLong{false};
bool helpShort{false};
getAndDelOption(arguments, "--help", helpLong);
getAndDelOption(arguments, "-h", helpShort);
return helpLong || helpShort;
}
void AllOptions::parse(Arguments& arguments)
{
model.parse(arguments);
build.parse(arguments);
system.parse(arguments);
inference.parse(arguments);
// Use explicitBatch when input model is ONNX or when dynamic shapes are used.
const bool isOnnx{model.baseModel.format == ModelFormat::kONNX};
const bool hasDynamicShapes{!build.shapes.empty() || !inference.shapes.empty()};
const bool detectedExplicitBatch = isOnnx || hasDynamicShapes;
// Throw an error if user tries to use --batch or --maxBatch when the engine has explicit batch dim.
const bool maxBatchWasSet{build.maxBatch != maxBatchNotProvided};
const bool batchWasSet{inference.batch != batchNotProvided};
if (detectedExplicitBatch && (maxBatchWasSet || batchWasSet))
{
throw std::invalid_argument(
"The --batch and --maxBatch flags should not be used when the input model is ONNX or when dynamic shapes "
"are provided. Please use --optShapes and --shapes to set input shapes instead.");
}
if (build.useRuntime != RuntimeMode::kFULL && inference.timeRefit)
{
throw std::invalid_argument("--timeRefit requires --useRuntime=full.");
}
// If batch and/or maxBatch is not set and the engine has implicit batch dim, set them to default values.
if (!detectedExplicitBatch)
{
// If batch is not set, set it to default value.
if (!batchWasSet)
{
inference.batch = defaultBatch;
}
// If maxBatch is not set, set it to be equal to batch.
if (!maxBatchWasSet)
{
build.maxBatch = inference.batch;
}
// MaxBatch should not be less than batch.
if (build.maxBatch < inference.batch)
{
throw std::invalid_argument("Build max batch " + std::to_string(build.maxBatch)
+ " is less than inference batch " + std::to_string(inference.batch));
}
}
// Propagate shape profile between builder and inference
for (auto const& s : build.shapes)
{
if (inference.shapes.find(s.first) == inference.shapes.end())
{
insertShapesInference(
inference.shapes, s.first, s.second[static_cast<size_t>(nvinfer1::OptProfileSelector::kOPT)]);
}
}
for (auto const& s : inference.shapes)
{
if (build.shapes.find(s.first) == build.shapes.end())
{
// assume min/opt/max all the same
insertShapesBuild(build.shapes, nvinfer1::OptProfileSelector::kMIN, s.first, s.second);
insertShapesBuild(build.shapes, nvinfer1::OptProfileSelector::kOPT, s.first, s.second);
insertShapesBuild(build.shapes, nvinfer1::OptProfileSelector::kMAX, s.first, s.second);
}
}
// Set nvtxVerbosity to be the same as build-time profilingVerbosity.
inference.nvtxVerbosity = build.profilingVerbosity;
reporting.parse(arguments);
helps = parseHelp(arguments);
if (!helps)
{
if (!build.load && model.baseModel.format == ModelFormat::kANY)
{
throw std::invalid_argument("Model missing or format not recognized");
}
if (build.safe && system.DLACore >= 0)
{
auto checkSafeDLAFormats = [](std::vector<IOFormat> const& fmt) {
return fmt.empty() ? false : std::all_of(fmt.begin(), fmt.end(), [](IOFormat const& pair) {
bool supported{false};
bool const isDLA_LINEAR{
pair.second == 1U << static_cast<int32_t>(nvinfer1::TensorFormat::kDLA_LINEAR)};
bool const isCHW4{pair.second == 1U << static_cast<int32_t>(nvinfer1::TensorFormat::kCHW4)};
bool const isCHW32{pair.second == 1U << static_cast<int32_t>(nvinfer1::TensorFormat::kCHW32)};
bool const isCHW16{pair.second == 1U << static_cast<int32_t>(nvinfer1::TensorFormat::kCHW16)};
supported |= pair.first == nvinfer1::DataType::kINT8 && (isDLA_LINEAR || isCHW4 || isCHW32);
supported |= pair.first == nvinfer1::DataType::kHALF && (isDLA_LINEAR || isCHW4 || isCHW16);
return supported;
});
};
if (!checkSafeDLAFormats(build.inputFormats) || !checkSafeDLAFormats(build.outputFormats))
{
throw std::invalid_argument(
"I/O formats for safe DLA capability are restricted to fp16/int8:dla_linear, fp16:chw16 or "
"int8:chw32");
}
if (system.fallback)
{
throw std::invalid_argument("GPU fallback (--allowGPUFallback) not allowed for safe DLA capability");
}
}
}
}
void TaskInferenceOptions::parse(Arguments& arguments)
{
getAndDelOption(arguments, "engine", engine);
getAndDelOption(arguments, "device", device);
getAndDelOption(arguments, "batch", batch);
getAndDelOption(arguments, "DLACore", DLACore);
getAndDelOption(arguments, "graph", graph);
getAndDelOption(arguments, "persistentCacheRatio", persistentCacheRatio);
}
void SafeBuilderOptions::parse(Arguments& arguments)
{
auto getFormats = [&arguments](std::vector<IOFormat>& formatsVector, const char* argument) {
std::string list;
getAndDelOption(arguments, argument, list);
std::vector<std::string> formats{splitToStringVec(list, ',')};
for (const auto& f : formats)
{
formatsVector.push_back(stringToValue<IOFormat>(f));
}
};
getAndDelOption(arguments, "--serialized", serialized);
getAndDelOption(arguments, "--onnx", onnxModelFile);
getAndDelOption(arguments, "--help", help);
getAndDelOption(arguments, "-h", help);
getAndDelOption(arguments, "--verbose", verbose);
getAndDelOption(arguments, "-v", verbose);
getFormats(inputFormats, "--inputIOFormats");
getFormats(outputFormats, "--outputIOFormats");
getAndDelOption(arguments, "--int8", int8);
getAndDelOption(arguments, "--calib", calibFile);
getAndDelOption(arguments, "--consistency", consistency);
getAndDelOption(arguments, "--std", standard);
std::string pluginName;
while (getAndDelOption(arguments, "--plugins", pluginName))
{
sample::gLogWarning << "--plugins flag has been deprecated, use --staticPlugins flag instead." << std::endl;
plugins.emplace_back(pluginName);
}
while (getAndDelOption(arguments, "--staticPlugins", pluginName))
{
plugins.emplace_back(pluginName);
}
bool noBuilderCache{false};
getAndDelOption(arguments, "--noBuilderCache", noBuilderCache);
getAndDelOption(arguments, "--timingCacheFile", timingCacheFile);
getAndDelOption(arguments, "--minTiming", minTiming);
getAndDelOption(arguments, "--avgTiming", avgTiming);
if (noBuilderCache)
{
timingCacheMode = TimingCacheMode::kDISABLE;
}
else if (!timingCacheFile.empty())
{
timingCacheMode = TimingCacheMode::kGLOBAL;
}
else
{
timingCacheMode = TimingCacheMode::kLOCAL;
}
getAndDelOption(arguments, "--sparsity", sparsity);
}
std::ostream& operator<<(std::ostream& os, const BaseModelOptions& options)
{
os << "=== Model Options ===" << std::endl;
os << "Format: ";
switch (options.format)
{
case ModelFormat::kCAFFE:
{
os << "Caffe";
break;
}
case ModelFormat::kONNX:
{
os << "ONNX";
break;
}
case ModelFormat::kUFF:
{
os << "UFF";
break;
}
case ModelFormat::kANY: os << "*"; break;
}
os << std::endl << "Model: " << options.model << std::endl;
return os;
}
std::ostream& operator<<(std::ostream& os, const UffInput& input)
{
os << "Uff Inputs Layout: " << (input.NHWC ? "NHWC" : "NCHW") << std::endl;
for (const auto& i : input.inputs)
{
os << "Input: " << i.first << "," << i.second.d[0] << "," << i.second.d[1] << "," << i.second.d[2] << std::endl;
}
return os;
}
std::ostream& operator<<(std::ostream& os, const ModelOptions& options)
{
os << options.baseModel;
switch (options.baseModel.format)
{
case ModelFormat::kCAFFE:
{
os << "Prototxt: " << options.prototxt << std::endl;
break;
}
case ModelFormat::kUFF:
{
os << options.uffInputs;
break;
}
case ModelFormat::kONNX: // Fallthrough: No options to report for ONNX or the generic case
case ModelFormat::kANY: break;
}
os << "Output:";
for (const auto& o : options.outputs)
{
os << " " << o;
}
os << std::endl;
return os;
}
std::ostream& operator<<(std::ostream& os, nvinfer1::DataType dtype)
{
switch (dtype)
{
case nvinfer1::DataType::kFLOAT:
{
os << "fp32";
break;
}
case nvinfer1::DataType::kHALF:
{
os << "fp16";
break;
}
case nvinfer1::DataType::kINT8:
{
os << "int8";
break;
}
case nvinfer1::DataType::kINT32:
{
os << "int32";
break;
}
case nvinfer1::DataType::kBOOL:
{
os << "bool";
break;
}
case nvinfer1::DataType::kUINT8:
{
os << "uint8";
break;
}
case nvinfer1::DataType::kFP8:
{
os << "fp8";
break;
}
}
return os;
}
std::ostream& operator<<(std::ostream& os, IOFormat const& format)
{
os << format.first << ":";
for (int32_t f = 0; f < nvinfer1::EnumMax<nvinfer1::TensorFormat>(); ++f)
{
if ((1U << f) & format.second)
{
if (f)
{
os << "+";
}
switch (nvinfer1::TensorFormat(f))
{
case nvinfer1::TensorFormat::kLINEAR:
{
os << "chw";
break;
}
case nvinfer1::TensorFormat::kCHW2:
{
os << "chw2";
break;
}
case nvinfer1::TensorFormat::kHWC8:
{
os << "hwc8";
break;
}
case nvinfer1::TensorFormat::kHWC16:
{
os << "hwc16";
break;
}
case nvinfer1::TensorFormat::kCHW4:
{
os << "chw4";
break;
}
case nvinfer1::TensorFormat::kCHW16:
{
os << "chw16";
break;
}
case nvinfer1::TensorFormat::kCHW32:
{
os << "chw32";
break;
}
case nvinfer1::TensorFormat::kDHWC8:
{
os << "dhwc8";
break;
}
case nvinfer1::TensorFormat::kCDHW32:
{
os << "cdhw32";
break;
}
case nvinfer1::TensorFormat::kHWC:
{
os << "hwc";
break;
}
case nvinfer1::TensorFormat::kDHWC:
{
os << "dhwc";
break;
}
case nvinfer1::TensorFormat::kDLA_LINEAR:
{
os << "dla_linear";
break;
}
case nvinfer1::TensorFormat::kDLA_HWC4:
{
os << "dla_hwc4";
break;
}
}
}
}
return os;
}
std::ostream& operator<<(std::ostream& os, nvinfer1::DeviceType devType)
{
switch (devType)
{
case nvinfer1::DeviceType::kGPU:
{
os << "GPU";
break;
}
case nvinfer1::DeviceType::kDLA:
{
os << "DLA";
break;
}
}
return os;
}
std::ostream& operator<<(std::ostream& os, const ShapeRange& dims)
{
int32_t i = 0;
for (const auto& d : dims)
{
if (!d.size())
{
break;
}
os << (i ? "+" : "") << d;
++i;
}
return os;
}
std::ostream& operator<<(std::ostream& os, LayerPrecisions const& layerPrecisions)
{
int32_t i = 0;
for (auto const& layerPrecision : layerPrecisions)
{
os << (i ? "," : "") << layerPrecision.first << ":" << layerPrecision.second;
++i;
}
return os;
}
std::ostream& operator<<(std::ostream& os, LayerDeviceTypes const& layerDeviceTypes)
{
int32_t i = 0;
for (auto const& layerDevicePair : layerDeviceTypes)
{
os << (i++ ? ", " : "") << layerDevicePair.first << ":" << layerDevicePair.second;
}
return os;
}
std::ostream& operator<<(std::ostream& os, const BuildOptions& options)
{
// clang-format off
os << "=== Build Options ===" << std::endl <<
"Max batch: "; printBatch(os, options.maxBatch) << std::endl <<
"Memory Pools: "; printMemoryPools(os, options) << std::endl <<
"minTiming: " << options.minTiming << std::endl <<
"avgTiming: " << options.avgTiming << std::endl <<
"Precision: "; printPrecision(os, options) << std::endl <<
"LayerPrecisions: " << options.layerPrecisions << std::endl <<
"Layer Device Types: " << options.layerDeviceTypes << std::endl <<
"Calibration: " << (options.int8 && options.calibration.empty() ? "Dynamic" : options.calibration.c_str()) << std::endl <<
"Refit: " << boolToEnabled(options.refittable) << std::endl <<
"Version Compatible: " << boolToEnabled(options.versionCompatible) << std::endl <<
"TensorRT runtime: " << options.useRuntime << std::endl <<
"Lean DLL Path: " << options.leanDLLPath << std::endl <<
"Tempfile Controls: "; printTempfileControls(os, options.tempfileControls) << std::endl <<
"Exclude Lean Runtime: " << boolToEnabled(options.excludeLeanRuntime) << std::endl <<
"Sparsity: "; printSparsity(os, options) << std::endl <<
"Safe mode: " << boolToEnabled(options.safe) << std::endl <<
"DirectIO mode: " << boolToEnabled(options.directIO) << std::endl <<
"Restricted mode: " << boolToEnabled(options.restricted) << std::endl <<
"Skip inference: " << boolToEnabled(options.skipInference) << std::endl <<
"Save engine: " << (options.save ? options.engine : "") << std::endl <<
"Load engine: " << (options.load ? options.engine : "") << std::endl <<
"Profiling verbosity: " << static_cast<int32_t>(options.profilingVerbosity) << std::endl <<
"Tactic sources: "; printTacticSources(os, options.enabledTactics, options.disabledTactics) << std::endl <<
"timingCacheMode: "; printTimingCache(os, options.timingCacheMode) << std::endl <<
"timingCacheFile: " << options.timingCacheFile << std::endl <<
"Heuristic: " << boolToEnabled(options.heuristic) << std::endl <<
"Preview Features: "; printPreviewFlags(os, options) << std::endl <<
"MaxAuxStreams: " << options.maxAuxStreams << std::endl <<
"BuilderOptimizationLevel: " << options.builderOptimizationLevel << std::endl;
// clang-format on
auto printIOFormats = [](std::ostream& os, const char* direction, const std::vector<IOFormat> formats) {
if (formats.empty())
{
os << direction << "s format: fp32:CHW" << std::endl;
}
else
{
for (const auto& f : formats)
{
os << direction << ": " << f << std::endl;
}
}
};
printIOFormats(os, "Input(s)", options.inputFormats);
printIOFormats(os, "Output(s)", options.outputFormats);
printShapes(os, "build", options.shapes);
printShapes(os, "calibration", options.shapesCalib);
return os;
}
std::ostream& operator<<(std::ostream& os, const SystemOptions& options)
{
// clang-format off
os << "=== System Options ===" << std::endl <<
"Device: " << options.device << std::endl <<
"DLACore: " << (options.DLACore != -1 ? std::to_string(options.DLACore) : "") <<
(options.DLACore != -1 && options.fallback ? "(With GPU fallback)" : "") << std::endl;
os << "Plugins:";
for (const auto& p : options.plugins)
{
os << " " << p;
}
os << std::endl;
os << "setPluginsToSerialize:";
for (const auto& p : options.setPluginsToSerialize)
{
os << " " << p;
}
os << std::endl;
os << "dynamicPlugins:";
for (const auto& p : options.dynamicPlugins)
{
os << " " << p;
}
os << std::endl;
os << "ignoreParsedPluginLibs: " << options.ignoreParsedPluginLibs << std::endl;
os << std::endl;
return os;
// clang-format on
}
std::ostream& operator<<(std::ostream& os, const InferenceOptions& options)
{
// clang-format off
os << "=== Inference Options ===" << std::endl <<
"Batch: ";
if (options.batch && options.shapes.empty())
{
os << options.batch << std::endl;
}
else
{
os << "Explicit" << std::endl;
}
printShapes(os, "inference", options.shapes);
os << "Iterations: " << options.iterations << std::endl <<
"Duration: " << options.duration << "s (+ "
<< options.warmup << "ms warm up)" << std::endl <<
"Sleep time: " << options.sleep << "ms" << std::endl <<
"Idle time: " << options.idle << "ms" << std::endl <<
"Inference Streams: " << options.infStreams << std::endl <<
"ExposeDMA: " << boolToEnabled(!options.overlap) << std::endl <<
"Data transfers: " << boolToEnabled(!options.skipTransfers) << std::endl <<
"Spin-wait: " << boolToEnabled(options.spin) << std::endl <<
"Multithreading: " << boolToEnabled(options.threads) << std::endl <<
"CUDA Graph: " << boolToEnabled(options.graph) << std::endl <<
"Separate profiling: " << boolToEnabled(options.rerun) << std::endl <<
"Time Deserialize: " << boolToEnabled(options.timeDeserialize) << std::endl <<
"Time Refit: " << boolToEnabled(options.timeRefit) << std::endl <<
"NVTX verbosity: " << static_cast<int32_t>(options.nvtxVerbosity) << std::endl <<
"Persistent Cache Ratio: " << static_cast<float>(options.persistentCacheRatio) << std::endl;
// clang-format on
os << "Inputs:" << std::endl;
for (const auto& input : options.inputs)
{
os << input.first << "<-" << input.second << std::endl;
}
return os;
}
std::ostream& operator<<(std::ostream& os, const ReportingOptions& options)
{
// clang-format off
os << "=== Reporting Options ===" << std::endl <<
"Verbose: " << boolToEnabled(options.verbose) << std::endl <<
"Averages: " << options.avgs << " inferences" << std::endl <<
"Percentiles: " << joinValuesToString(options.percentiles, ",") << std::endl <<
"Dump refittable layers:" << boolToEnabled(options.refit) << std::endl <<
"Dump output: " << boolToEnabled(options.output) << std::endl <<
"Profile: " << boolToEnabled(options.profile) << std::endl <<
"Export timing to JSON file: " << options.exportTimes << std::endl <<
"Export output to JSON file: " << options.exportOutput << std::endl <<
"Export profile to JSON file: " << options.exportProfile << std::endl;
// clang-format on
return os;
}
std::ostream& operator<<(std::ostream& os, const AllOptions& options)
{
os << options.model << options.build << options.system << options.inference << options.reporting << std::endl;
return os;
}
std::ostream& operator<<(std::ostream& os, const SafeBuilderOptions& options)
{
auto printIOFormats = [](std::ostream& os, const char* direction, const std::vector<IOFormat> formats) {
if (formats.empty())
{
os << direction << "s format: fp32:CHW" << std::endl;
}
else
{
for (const auto& f : formats)
{
os << direction << ": " << f << std::endl;
}
}
};
os << "=== Build Options ===" << std::endl;
os << "Model ONNX: " << options.onnxModelFile << std::endl;
os << "Precision: FP16";
if (options.int8)
{
os << " + INT8";
}
if (options.fp8)
{
os << " + FP8";
}
os << std::endl;
os << "Calibration file: " << options.calibFile << std::endl;
os << "Serialized Network: " << options.serialized << std::endl;
printIOFormats(os, "Input(s)", options.inputFormats);
printIOFormats(os, "Output(s)", options.outputFormats);
os << "Plugins:";
for (const auto& p : options.plugins)
{
os << " " << p;
}
os << "timingCacheMode: ";
printTimingCache(os, options.timingCacheMode) << std::endl;
os << "timingCacheFile: " << options.timingCacheFile << std::endl;
os << std::endl;
return os;
}
void BaseModelOptions::help(std::ostream& os)
{
// clang-format off
os << " --uff=<file> UFF model" << std::endl <<
" --onnx=<file> ONNX model" << std::endl <<
" --model=<file> Caffe model (default = no model, random weights used)" << std::endl;
// clang-format on
}
void UffInput::help(std::ostream& os)
{
// clang-format off
os << " --uffInput=<name>,X,Y,Z Input blob name and its dimensions (X,Y,Z=C,H,W), it can be specified "
"multiple times; at least one is required for UFF models" << std::endl <<
" --uffNHWC Set if inputs are in the NHWC layout instead of NCHW (use " <<
"X,Y,Z=H,W,C order in --uffInput)" << std::endl;
// clang-format on
}
void ModelOptions::help(std::ostream& os)
{
// clang-format off
os << "=== Model Options ===" << std::endl;
BaseModelOptions::help(os);
os << " --deploy=<file> Caffe prototxt file" << std::endl <<
" --output=<name>[,<name>]* Output names (it can be specified multiple times); at least one output "
"is required for UFF and Caffe" << std::endl;
UffInput::help(os);
// clang-format on
}
void BuildOptions::help(std::ostream& os)
{
// clang-format off
os << "=== Build Options ===" "\n"
" --maxBatch Set max batch size and build an implicit batch engine (default = same size as --batch)" "\n"
" This option should not be used when the input model is ONNX or when dynamic shapes are provided." "\n"
" --minShapes=spec Build with dynamic shapes using a profile with the min shapes provided" "\n"
" --optShapes=spec Build with dynamic shapes using a profile with the opt shapes provided" "\n"
" --maxShapes=spec Build with dynamic shapes using a profile with the max shapes provided" "\n"
" --minShapesCalib=spec Calibrate with dynamic shapes using a profile with the min shapes provided" "\n"
" --optShapesCalib=spec Calibrate with dynamic shapes using a profile with the opt shapes provided" "\n"
" --maxShapesCalib=spec Calibrate with dynamic shapes using a profile with the max shapes provided" "\n"
" Note: All three of min, opt and max shapes must be supplied." "\n"
" However, if only opt shapes is supplied then it will be expanded so" "\n"
" that min shapes and max shapes are set to the same values as opt shapes." "\n"
" Input names can be wrapped with escaped single quotes (ex: 'Input:0')." "\n"
" Example input shapes spec: input0:1x3x256x256,input1:1x3x128x128" "\n"
" Each input shape is supplied as a key-value pair where key is the input name and" "\n"
" value is the dimensions (including the batch dimension) to be used for that input." "\n"
" Each key-value pair has the key and value separated using a colon (:)." "\n"
" Multiple input shapes can be provided via comma-separated key-value pairs." "\n"
" --inputIOFormats=spec Type and format of each of the input tensors (default = all inputs in fp32:chw)" "\n"
" See --outputIOFormats help for the grammar of type and format list." "\n"
" Note: If this option is specified, please set comma-separated types and formats for all" "\n"
" inputs following the same order as network inputs ID (even if only one input" "\n"
" needs specifying IO format) or set the type and format once for broadcasting." "\n"
" --outputIOFormats=spec Type and format of each of the output tensors (default = all outputs in fp32:chw)" "\n"
" Note: If this option is specified, please set comma-separated types and formats for all" "\n"
" outputs following the same order as network outputs ID (even if only one output" "\n"
" needs specifying IO format) or set the type and format once for broadcasting." "\n"
R"( IO Formats: spec ::= IOfmt[","spec])" "\n"
" IOfmt ::= type:fmt" "\n"
R"( type ::= "fp32"|"fp16"|"int32"|"int8")" "\n"
R"( fmt ::= ("chw"|"chw2"|"chw4"|"hwc8"|"chw16"|"chw32"|"dhwc8"|)" "\n"
R"( "cdhw32"|"hwc"|"dla_linear"|"dla_hwc4")["+"fmt])" "\n"
" --workspace=N Set workspace size in MiB." "\n"
" --memPoolSize=poolspec Specify the size constraints of the designated memory pool(s) in MiB." "\n"
" Note: Also accepts decimal sizes, e.g. 0.25MiB. Will be rounded down to the nearest integer bytes." "\n"
R"( Pool constraint: poolspec ::= poolfmt[","poolspec])" "\n"
" poolfmt ::= pool:sizeInMiB" "\n"
R"( pool ::= "workspace"|"dlaSRAM"|"dlaLocalDRAM"|"dlaGlobalDRAM")" "\n"
" --profilingVerbosity=mode Specify profiling verbosity. mode ::= layer_names_only|detailed|none (default = layer_names_only)" "\n"
" --minTiming=M Set the minimum number of iterations used in kernel selection (default = "
<< defaultMinTiming << ")" "\n"
" --avgTiming=M Set the number of times averaged in each iteration for kernel selection (default = "
<< defaultAvgTiming << ")" "\n"
" --refit Mark the engine as refittable. This will allow the inspection of refittable layers " "\n"
" and weights within the engine." "\n"
" --versionCompatible, --vc Mark the engine as version compatible. This allows the engine to be used with newer versions" "\n"
" of TensorRT on the same host OS, as well as TensorRT's dispatch and lean runtimes." "\n"
" Only supported with explicit batch." "\n"
R"( --useRuntime=runtime TensorRT runtime to execute engine. "lean" and "dispatch" require loading VC engine and do)" "\n"
" not support building an engine." "\n"
R"( runtime::= "full"|"lean"|"dispatch")" "\n"
" --leanDLLPath=<file> External lean runtime DLL to use in version compatiable mode." "\n"
" --excludeLeanRuntime When --versionCompatible is enabled, this flag indicates that the generated engine should" "\n"
" not include an embedded lean runtime. If this is set, the user must explicitly specify a" "\n"
" valid lean runtime to use when loading the engine. Only supported with explicit batch" "\n"
" and weights within the engine." "\n"
" --sparsity=spec Control sparsity (default = disabled). " "\n"
R"( Sparsity: spec ::= "disable", "enable", "force")" "\n"
" Note: Description about each of these options is as below" "\n"
" disable = do not enable sparse tactics in the builder (this is the default)" "\n"
" enable = enable sparse tactics in the builder (but these tactics will only be" "\n"
" considered if the weights have the right sparsity pattern)" "\n"
" force = enable sparse tactics in the builder and force-overwrite the weights to have" "\n"
" a sparsity pattern (even if you loaded a model yourself)" "\n"
" --noTF32 Disable tf32 precision (default is to enable tf32, in addition to fp32)" "\n"
" --fp16 Enable fp16 precision, in addition to fp32 (default = disabled)" "\n"
" --int8 Enable int8 precision, in addition to fp32 (default = disabled)" "\n"
" --fp8 Enable fp8 precision, in addition to fp32 (default = disabled)" "\n"
" --best Enable all precisions to achieve the best performance (default = disabled)" "\n"
" --directIO Avoid reformatting at network boundaries. (default = disabled)" "\n"
" --precisionConstraints=spec Control precision constraint setting. (default = none)" "\n"
R"( Precision Constraints: spec ::= "none" | "obey" | "prefer")" "\n"
" none = no constraints" "\n"
" prefer = meet precision constraints set by --layerPrecisions/--layerOutputTypes if possible" "\n"
" obey = meet precision constraints set by --layerPrecisions/--layerOutputTypes or fail" "\n"
" otherwise" "\n"
" --layerPrecisions=spec Control per-layer precision constraints. Effective only when precisionConstraints is set to" "\n"
R"( "obey" or "prefer". (default = none))" "\n"
R"( The specs are read left-to-right, and later ones override earlier ones. "*" can be used as a)" "\n"
" layerName to specify the default precision for all the unspecified layers." "\n"
R"( Per-layer precision spec ::= layerPrecision[","spec])" "\n"
R"( layerPrecision ::= layerName":"precision)" "\n"
R"( precision ::= "fp32"|"fp16"|"int32"|"int8")" "\n"
" --layerOutputTypes=spec Control per-layer output type constraints. Effective only when precisionConstraints is set to" "\n"
R"( "obey" or "prefer". (default = none)" "\n"
R"( The specs are read left-to-right, and later ones override earlier ones. "*" can be used as a)" "\n"
" layerName to specify the default precision for all the unspecified layers. If a layer has more than""\n"
R"( one output, then multiple types separated by "+" can be provided for this layer.)" "\n"
R"( Per-layer output type spec ::= layerOutputTypes[","spec])" "\n"
R"( layerOutputTypes ::= layerName":"type)" "\n"
R"( type ::= "fp32"|"fp16"|"int32"|"int8"["+"type])" "\n"
" --layerDeviceTypes=spec Specify layer-specific device type." "\n"
" The specs are read left-to-right, and later ones override earlier ones. If a layer does not have" "\n"
" a device type specified, the layer will opt for the default device type." "\n"
R"( Per-layer device type spec ::= layerDeviceTypePair[","spec])" "\n"
R"( layerDeviceTypePair ::= layerName":"deviceType)" "\n"
R"( deviceType ::= "GPU"|"DLA")" "\n"
" --calib=<file> Read INT8 calibration cache file" "\n"
" --safe Enable build safety certified engine" "\n"
" --consistency Perform consistency checking on safety certified engine" "\n"
" --restricted Enable safety scope checking with kSAFETY_SCOPE build flag" "\n"
" --saveEngine=<file> Save the serialized engine" "\n"
" --loadEngine=<file> Load a serialized engine" "\n"
" --tacticSources=tactics Specify the tactics to be used by adding (+) or removing (-) tactics from the default " "\n"
" tactic sources (default = all available tactics)." "\n"
" Note: Currently only cuDNN, cuBLAS, cuBLAS-LT, and edge mask convolutions are listed as optional" "\n"
" tactics." "\n"
R"( Tactic Sources: tactics ::= [","tactic])" "\n"
" tactic ::= (+|-)lib" "\n"
R"( lib ::= "CUBLAS"|"CUBLAS_LT"|"CUDNN"|"EDGE_MASK_CONVOLUTIONS")" "\n"
R"( |"JIT_CONVOLUTIONS")" "\n"
" For example, to disable cudnn and enable cublas: --tacticSources=-CUDNN,+CUBLAS" "\n"
" --noBuilderCache Disable timing cache in builder (default is to enable timing cache)" "\n"
" --heuristic Enable tactic selection heuristic in builder (default is to disable the heuristic)" "\n"
" --timingCacheFile=<file> Save/load the serialized global timing cache" "\n"
" --preview=features Specify preview feature to be used by adding (+) or removing (-) preview features from the default" "\n"
R"( Preview Features: features ::= [","feature])" "\n"
" feature ::= (+|-)flag" "\n"
R"( flag ::= "fasterDynamicShapes0805")" "\n"
R"( |"disableExternalTacticSourcesForCore0805")" "\n"
R"( |"profileSharing0806")" "\n"
" --builderOptimizationLevel Set the builder optimization level. (default is 3" "\n"
" Higher level allows TensorRT to spend more building time for more optimization options." "\n"
" The default level is 3. Valid values include integers from 0 to the maximum optimization level," "\n"
" which is currently 5." "\n"
" --hardwareCompatibilityLevel=mode Make the engine file compatible with other GPU architectures. (default = none)" "\n"
R"( Hardware Compatibility Level: mode ::= "none" | "ampere+")" "\n"
" none = no compatibility" "\n"
" ampere+ = compatible with Ampere and newer GPUs" "\n"
" --tempdir=<dir> Overrides the default temporary directory TensorRT will use when creating temporary files." "\n"
" See IRuntime::setTemporaryDirectory API documentation for more information." "\n"
" --tempfileControls=controls Controls what TensorRT is allowed to use when creating temporary executable files." "\n"
" Should be a comma-separated list with entries in the format (in_memory|temporary):(allow|deny)." "\n"
" in_memory: Controls whether TensorRT is allowed to create temporary in-memory executable files." "\n"
" temporary: Controls whether TensorRT is allowed to create temporary executable files in the" "\n"
" filesystem (in the directory given by --tempdir)." "\n"
" For example, to allow in-memory files and disallow temporary files:" "\n"
" --tempfileControls=in_memory:allow,temporary:deny" "\n"
R"( If a flag is unspecified, the default behavior is "allow".)" "\n"
" --maxAuxStreams=N Set maximum number of auxiliary streams per inference stream that TRT is allowed to use to run " "\n"
" kernels in parallel if the network contains ops that can run in parallel, with the cost of more " "\n"
" memory usage. Set this to 0 for optimal memory usage. (default = using heuristics)" "\n"
;
// clang-format on
os << std::flush;
}
void SystemOptions::help(std::ostream& os)
{
// clang-format off
os << "=== System Options ===" << std::endl <<
" --device=N Select cuda device N (default = " << defaultDevice << ")" << std::endl <<
" --useDLACore=N Select DLA core N for layers that support DLA (default = none)" << std::endl <<
" --allowGPUFallback When DLA is enabled, allow GPU fallback for unsupported layers "
"(default = disabled)" << std::endl <<
" --staticPlugins Plugin library (.so) to load statically (can be specified multiple times)" << std::endl <<
" --dynamicPlugins Plugin library (.so) to load dynamically and may be serialized with the engine if they are included in --setPluginsToSerialize (can be specified multiple times)" << std::endl <<
" --setPluginsToSerialize Plugin library (.so) to be serialized with the engine (can be specified multiple times)" << std::endl <<
" --ignoreParsedPluginLibs By default, when building a version-compatible engine, plugin libraries specified by the ONNX parser " << std::endl <<
" are implicitly serialized with the engine (unless --excludeLeanRuntime is specified) and loaded dynamically. " << std::endl <<
" Enable this flag to ignore these plugin libraries instead." << std::endl;
// clang-format on
}
void InferenceOptions::help(std::ostream& os)
{
// clang-format off
os << "=== Inference Options ===" << std::endl <<
" --batch=N Set batch size for implicit batch engines (default = " << defaultBatch << ")" << std::endl <<
" This option should not be used when the engine is built from an ONNX model or when dynamic" << std::endl <<
" shapes are provided when the engine is built." << std::endl <<
" --shapes=spec Set input shapes for dynamic shapes inference inputs." << std::endl <<
R"( Note: Input names can be wrapped with escaped single quotes (ex: 'Input:0').)" << std::endl <<
" Example input shapes spec: input0:1x3x256x256, input1:1x3x128x128" << std::endl <<
" Each input shape is supplied as a key-value pair where key is the input name and" << std::endl <<
" value is the dimensions (including the batch dimension) to be used for that input." << std::endl <<
" Each key-value pair has the key and value separated using a colon (:)." << std::endl <<
" Multiple input shapes can be provided via comma-separated key-value pairs." << std::endl <<
" --loadInputs=spec Load input values from files (default = generate random inputs). Input names can be "
"wrapped with single quotes (ex: 'Input:0')" << std::endl <<
R"( Input values spec ::= Ival[","spec])" << std::endl <<
R"( Ival ::= name":"file)" << std::endl <<
" --iterations=N Run at least N inference iterations (default = " << defaultIterations << ")" << std::endl <<
" --warmUp=N Run for N milliseconds to warmup before measuring performance (default = "
<< defaultWarmUp << ")" << std::endl <<
" --duration=N Run performance measurements for at least N seconds wallclock time (default = "
<< defaultDuration << ")" << std::endl <<
" --sleepTime=N Delay inference start with a gap of N milliseconds between launch and compute "
"(default = " << defaultSleep << ")" << std::endl <<
" --idleTime=N Sleep N milliseconds between two continuous iterations"
"(default = " << defaultIdle << ")" << std::endl <<
" --infStreams=N Instantiate N engines to run inference concurrently (default = " << defaultStreams << ")" << std::endl <<
" --exposeDMA Serialize DMA transfers to and from device (default = disabled)." << std::endl <<
" --noDataTransfers Disable DMA transfers to and from device (default = enabled)." << std::endl <<
" --useManagedMemory Use managed memory instead of separate host and device allocations (default = disabled)." << std::endl <<
" --useSpinWait Actively synchronize on GPU events. This option may decrease synchronization time but "
"increase CPU usage and power (default = disabled)" << std::endl <<
" --threads Enable multithreading to drive engines with independent threads"
" or speed up refitting (default = disabled) " << std::endl <<
" --useCudaGraph Use CUDA graph to capture engine execution and then launch inference (default = disabled)." << std::endl <<
" This flag may be ignored if the graph capture fails." << std::endl <<
" --timeDeserialize Time the amount of time it takes to deserialize the network and exit." << std::endl <<
" --timeRefit Time the amount of time it takes to refit the engine before inference." << std::endl <<
" --separateProfileRun Do not attach the profiler in the benchmark run; if profiling is enabled, a second "
"profile run will be executed (default = disabled)" << std::endl <<
" --skipInference Exit after the engine has been built and skip inference perf measurement "
"(default = disabled)" << std::endl <<
" --persistentCacheRatio Set the persistentCacheLimit in ratio, 0.5 represent half of max persistent L2 size "
"(default = 0)" << std::endl;
// clang-format on
}
void ReportingOptions::help(std::ostream& os)
{
// clang-format off
os << "=== Reporting Options ===" << std::endl <<
" --verbose Use verbose logging (default = false)" << std::endl <<
" --avgRuns=N Report performance measurements averaged over N consecutive "
"iterations (default = " << defaultAvgRuns << ")" << std::endl <<
" --percentile=P1,P2,P3,... Report performance for the P1,P2,P3,... percentages (0<=P_i<=100, 0 "
"representing max perf, and 100 representing min perf; (default"
" = " << joinValuesToString(defaultPercentiles, ",") << "%)" << std::endl <<
" --dumpRefit Print the refittable layers and weights from a refittable "
"engine" << std::endl <<
" --dumpOutput Print the output tensor(s) of the last inference iteration "
"(default = disabled)" << std::endl <<
" --dumpRawBindingsToFile Print the input/output tensor(s) of the last inference iteration to file"
"(default = disabled)" << std::endl <<
" --dumpProfile Print profile information per layer (default = disabled)" << std::endl <<
" --dumpLayerInfo Print layer information of the engine to console "
"(default = disabled)" << std::endl <<
" --exportTimes=<file> Write the timing results in a json file (default = disabled)" << std::endl <<
" --exportOutput=<file> Write the output tensors to a json file (default = disabled)" << std::endl <<
" --exportProfile=<file> Write the profile information per layer in a json file "
"(default = disabled)" << std::endl <<
" --exportLayerInfo=<file> Write the layer information of the engine in a json file "
"(default = disabled)" << std::endl;
// clang-format on
}
void TaskInferenceOptions::help(std::ostream& os)
{
// clang-format off
os << "=== Task Inference Options ===" << std::endl <<
" engine=<file> Specify a serialized engine for this task" << std::endl <<
" device=N Specify a GPU device for this task" << std::endl <<
" DLACore=N Specify a DLACore for this task" << std::endl <<
" batch=N Set batch size for implicit batch engines (default = " << defaultBatch << ")" << std::endl <<
" This option should not be used for explicit batch engines" << std::endl <<
" graph=1 Use cuda graph for this task" << std::endl <<
" persistentCacheRatio=[0-1] Set the persistentCacheLimit ratio for this task (default = 0)" << std::endl;
// clang-format on
}
void helpHelp(std::ostream& os)
{
// clang-format off
os << "=== Help ===" << std::endl <<
" --help, -h Print this message" << std::endl;
// clang-format on
}
void AllOptions::help(std::ostream& os)
{
ModelOptions::help(os);
os << std::endl;
BuildOptions::help(os);
os << std::endl;
InferenceOptions::help(os);
os << std::endl;
// clang-format off
os << "=== Build and Inference Batch Options ===" << std::endl <<
" When using implicit batch, the max batch size of the engine, if not given, " << std::endl <<
" is set to the inference batch size;" << std::endl <<
" when using explicit batch, if shapes are specified only for inference, they " << std::endl <<
" will be used also as min/opt/max in the build profile; if shapes are " << std::endl <<
" specified only for the build, the opt shapes will be used also for inference;" << std::endl <<
" if both are specified, they must be compatible; and if explicit batch is " << std::endl <<
" enabled but neither is specified, the model must provide complete static" << std::endl <<
" dimensions, including batch size, for all inputs" << std::endl <<
" Using ONNX models automatically forces explicit batch." << std::endl <<
std::endl;
// clang-format on
ReportingOptions::help(os);
os << std::endl;
SystemOptions::help(os);
os << std::endl;
helpHelp(os);
}
void SafeBuilderOptions::printHelp(std::ostream& os)
{
// clang-format off
os << "=== Mandatory ===" << std::endl <<
" --onnx=<file> ONNX model" << std::endl <<
" " << std::endl <<
"=== Optional ===" << std::endl <<
" --inputIOFormats=spec Type and format of each of the input tensors (default = all inputs in fp32:chw)" << std::endl <<
" See --outputIOFormats help for the grammar of type and format list." << std::endl <<
" Note: If this option is specified, please set comma-separated types and formats for all" << std::endl <<
" inputs following the same order as network inputs ID (even if only one input" << std::endl <<
" needs specifying IO format) or set the type and format once for broadcasting." << std::endl <<
" --outputIOFormats=spec Type and format of each of the output tensors (default = all outputs in fp32:chw)" << std::endl <<
" Note: If this option is specified, please set comma-separated types and formats for all" << std::endl <<
" outputs following the same order as network outputs ID (even if only one output" << std::endl <<
" needs specifying IO format) or set the type and format once for broadcasting." << std::endl <<
R"( IO Formats: spec ::= IOfmt[","spec])" << std::endl <<
" IOfmt ::= type:fmt" << std::endl <<
R"( type ::= "fp32"|"fp16"|"int32"|"int8")" << std::endl <<
R"( fmt ::= ("chw"|"chw2"|"chw4"|"hwc8"|"chw16"|"chw32"|"dhwc8"|)" << std::endl <<
R"( "cdhw32"|"hwc"|"dla_linear"|"dla_hwc4")["+"fmt])" << std::endl <<
" --int8 Enable int8 precision, in addition to fp16 (default = disabled)" << std::endl <<
" --consistency Enable consistency check for serialized engine, (default = disabled)" << std::endl <<
" --std Build standard serialized engine, (default = disabled)" << std::endl <<
" --calib=<file> Read INT8 calibration cache file" << std::endl <<
" --serialized=<file> Save the serialized network" << std::endl <<
" --staticPlugins Plugin library (.so) to load statically (can be specified multiple times)" << std::endl <<
" --verbose or -v Use verbose logging (default = false)" << std::endl <<
" --help or -h Print this message" << std::endl <<
" --noBuilderCache Disable timing cache in builder (default is to enable timing cache)" << std::endl <<
" --timingCacheFile=<file> Save/load the serialized global timing cache" << std::endl <<
" --sparsity=spec Control sparsity (default = disabled). " << std::endl <<
R"( Sparsity: spec ::= "disable", "enable", "force")" << std::endl <<
" Note: Description about each of these options is as below" << std::endl <<
" disable = do not enable sparse tactics in the builder (this is the default)" << std::endl <<
" enable = enable sparse tactics in the builder (but these tactics will only be" << std::endl <<
" considered if the weights have the right sparsity pattern)" << std::endl <<
" force = enable sparse tactics in the builder and force-overwrite the weights to have" << std::endl <<
" a sparsity pattern" << std::endl <<
" --minTiming=M Set the minimum number of iterations used in kernel selection (default = " << std::endl <<
"" << defaultMinTiming << ")" << std::endl <<
" --avgTiming=M Set the number of times averaged in each iteration for kernel selection (default = " << std::endl <<
"" << defaultAvgTiming << ")" << std::endl <<
"" << std::endl;
// clang-format on
}
} // namespace sample