从零开始 TensorRT（4）命令行工具篇：trtexec 基本功能

前言

学习资料：
TensorRT 源码示例
B站视频：TensorRT 教程 | 基于 8.6.1 版本
视频配套代码 cookbook

参考源码：cookbook → 07-Tool → trtexec
官方文档：trtexec

在 TensorRT 的安装目录 xxx/TensorRT-8.6.1.6/bin 下有命令行工具 trtexec，主要功能：
（1）由 ONNX 文件生成 TensorRT 引擎并序列化为 plan 文件
（2）查看 ONNX 或 plan 文件的网络逐层信息
（3）模型性能测试，即测试 TensorRT 引擎在随机输入或给定输入下的性能

示例一：解析 ONNX 生成引擎并推理测速

  目录结构

├── resnet18.onnx
└── test_trtexec.sh

  在终端中临时添加环境变量后运行脚本

export PATH=xxx/TensorRT-8.6.1.6/bin${PATH:+:${PATH}}
bash test_trtexec.sh

  脚本内容如下，使用之前 Python 篇生成的 ResNet18 ONNX 文件生成 TensorRT 引擎，并且会做推理测试，输出日志信息存储在 result-fp32.txt 中。

trtexec \
    --onnx=resnet18.onnx \
    --saveEngine=resnet18-fp32.plan \
    --minShapes=x:1x3x224x224 \
    --optShapes=x:4x3x224x224 \
    --maxShapes=x:16x3x224x224 \
    --memPoolSize=workspace:1024MiB \
    --verbose \
    > result-fp32.txt 2>&1

  > 为输出重定向操作符，将输出重定向到指定位置。
  2>&1 中 2 为标准错误流 stderr，1 为标准输出流 stdout。表示把报错信息和正常输出一起保存到文件中

命令行常用选项

（1）构建阶段

--onnx=resnet18.onnx				指定模型文件
--saveEngine=resnet18-fp32.plan		输出引擎文件名

--minShapes=x:1x3x224x224
--optShapes=x:4x3x224x224
--maxShapes=x:16x3x224x224			输入形状的最小值、常见值、最大值
--memPoolSize=workspace:1024MiB		优化过程可使用的显存最大值
--verbose							打印详细日志, 但无法设置日志等级
--skipInference						只创建引擎不进行推理

--fp16
--int8
--noTF32
--best								指定引擎精度
--sparsity=[disable|enable|force]	指定稀疏性
--builderOptimizationLevel=5		设置优化等级(默认2)
--timingCacheFile=timing.cache		指定输出优化计时缓存文件
--profilingVerbosity=detailed		构建期保留更多的逐层信息

（2）运行阶段

--loadEngine=resnet18-fp32.plan		读取引擎文件
--shapes=x:4x3x224x224				指定输入形状
--warmUp=1000						预热阶段最短运行时间(单位:ms)
--duration=10						测试阶段最短运行时间(单位:s)
--iterations=100					测试阶段运行最小迭代次数
--useCudaGraph						使用 CudaGraph 捕获和执行推理过程
--noDataTransfers					关闭 Host 和 Device 之间数据传输
--streams=2							使用多个 stream 运行推理
--dumpProfile						显示逐层性能数据
--exportProfile=layerProfile.txt	保存逐层性能数据

解析输出日志信息

  输出的日志信息非常多，分块进行解读

（1）各种参数配置

=== Model Options ===
Format: ONNX
Model: resnet18.onnx
Output:

=== Build Options ===
Max batch: explicit batch
Memory Pools: workspace: 1024 MiB, dlaSRAM: default, dlaLocalDRAM: default, dlaGlobalDRAM: default
minTiming: 1
avgTiming: 8
Precision: FP32
LayerPrecisions: 
Layer Device Types: 
Calibration: 
Refit: Disabled
Version Compatible: Disabled
TensorRT runtime: full
Lean DLL Path: 
Tempfile Controls: { in_memory: allow, temporary: allow }
Exclude Lean Runtime: Disabled
Sparsity: Disabled
Safe mode: Disabled
Build DLA standalone loadable: Disabled
Allow GPU fallback for DLA: Disabled
DirectIO mode: Disabled
Restricted mode: Disabled
Skip inference: Disabled
Save engine: resnet18-fp32.plan
Load engine: 
Profiling verbosity: 0
Tactic sources: Using default tactic sources
timingCacheMode: local
timingCacheFile: 
Heuristic: Disabled
Preview Features: Use default preview flags.
MaxAuxStreams: -1
BuilderOptimizationLevel: -1
Input(s)s format: fp32:CHW
Output(s)s format: fp32:CHW
Input build shape: x=1x3x224x224+4x3x224x224+16x3x224x224
Input calibration shapes: model

=== System Options ===
Device: 0
DLACore: 
Plugins:
setPluginsToSerialize:
dynamicPlugins:
ignoreParsedPluginLibs: 0

=== Inference Options ===
Batch: Explicit
Input inference shape: x=4x3x224x224
Iterations: 10
Duration: 3s (+ 200ms warm up)
Sleep time: 0ms
Idle time: 0ms
Inference Streams: 1
ExposeDMA: Disabled
Data transfers: Enabled
Spin-wait: Disabled
Multithreading: Disabled
CUDA Graph: Disabled
Separate profiling: Disabled
Time Deserialize: Disabled
Time Refit: Disabled
NVTX verbosity: 0
Persistent Cache Ratio: 0
Inputs:

=== Reporting Options ===
Verbose: Enabled
Averages: 10 inferences
Percentiles: 90,95,99
Dump refittable layers:Disabled
Dump output: Disabled
Profile: Disabled
Export timing to JSON file: 
Export output to JSON file: 
Export profile to JSON file: 

=== Device Information ===
Selected Device: NVIDIA GeForce RTX 3090
Compute Capability: 8.6
SMs: 82
Device Global Memory: 24258 MiB
Shared Memory per SM: 100 KiB
Memory Bus Width: 384 bits (ECC disabled)
Application Compute Clock Rate: 1.695 GHz
Application Memory Clock Rate: 9.751 GHz

Note: The application clock rates do not reflect the actual clock rates that the GPU is currently running at.

TensorRT version: 8.6.1
"较长的一段内容在加载标准插件"
Loading standard plugins
Registered plugin creator - ::BatchedNMSDynamic_TRT version 1
...
Registered plugin creator - ::VoxelGeneratorPlugin version 1
[MemUsageChange] Init CUDA: CPU +13, GPU +0, now: CPU 19, GPU 774 (MiB)
Trying to load shared library libnvinfer_builder_resource.so.8.6.1
Loaded shared library libnvinfer_builder_resource.so.8.6.1
[MemUsageChange] Init builder kernel library: CPU +1450, GPU +266, now: CPU 1545, GPU 1034 (MiB)
CUDA lazy loading is enabled.

（2）Parse 过程

Start parsing network model.
----------------------------------------------------------------
"onnx 模型相关信息"
Input filename:   resnet18.onnx
ONNX IR version:  0.0.7
Opset version:    12
Producer name:    pytorch
Producer version: 2.0.1
Domain:           
Model version:    0
Doc string:       
----------------------------------------------------------------
"已加载插件"
Plugin creator already registered - ::xxx

"添加输入、构建网络"
Adding network input: x with dtype: float32, dimensions: (-1, 3, 224, 224)
Registering tensor: x for ONNX tensor: x
Importing initializer: fc.weight
Importing initializer: fc.bias
Importing initializer: onnx::Conv_193
...
Importing initializer: onnx::Conv_251

"每个 Parsing node 作为一段进行查看, 主要关于输入输出和层"
Parsing node: /conv1/Conv [Conv]
Searching for input: x
Searching for input: onnx::Conv_193
Searching for input: onnx::Conv_194
/conv1/Conv [Conv] inputs: [x -> (-1, 3, 224, 224)[FLOAT]], [onnx::Conv_193 -> (64, 3, 7, 7)[FLOAT]], [onnx::Conv_194 -> (64)[FLOAT]], 
Convolution input dimensions: (-1, 3, 224, 224)
Registering layer: /conv1/Conv for ONNX node: /conv1/Conv
Using kernel: (7, 7), strides: (2, 2), prepadding: (3, 3), postpadding: (3, 3), dilations: (1, 1), numOutputs: 64
Convolution output dimensions: (-1, 64, 112, 112)
Registering tensor: /conv1/Conv_output_0 for ONNX tensor: /conv1/Conv_output_0
/conv1/Conv [Conv] outputs: [/conv1/Conv_output_0 -> (-1, 64, 112, 112)[FLOAT]], 

Parsing node: /relu/Relu [Relu]
Searching for input: /conv1/Conv_output_0
/relu/Relu [Relu] inputs: [/conv1/Conv_output_0 -> (-1, 64, 112, 112)[FLOAT]], 
Registering layer: /relu/Relu for ONNX node: /relu/Relu
Registering tensor: /relu/Relu_output_0 for ONNX tensor: /relu/Relu_output_0
/relu/Relu [Relu] outputs: [/relu/Relu_output_0 -> (-1, 64, 112, 112)[FLOAT]], 
...
Finished parsing network model. Parse time: 0.144613

"对模型进行的一系列优化和融合操作, 优化后的层数, 耗时等信息"
Original: 53 layers
After dead-layer removal: 53 layers
Graph construction completed in 0.00150762 seconds.
Running: ConstShuffleFusion on fc.bias
ConstShuffleFusion: Fusing fc.bias with (Unnamed Layer* 56) [Shuffle]
After Myelin optimization: 52 layers
...
After scale fusion: 49 layers
...
After dupe layer removal: 24 layers
After final dead-layer removal: 24 layers
After tensor merging: 24 layers
After vertical fusions: 24 layers
After dupe layer removal: 24 layers
After final dead-layer removal: 24 layers
After tensor merging: 24 layers
After slice removal: 24 layers
After concat removal: 24 layers
Trying to split Reshape and strided tensor
Graph optimization time: 0.0239815 seconds.
Building graph using backend strategy 2
Local timing cache in use. Profiling results in this builder pass will not be stored.
Constructing optimization profile number 0 [1/1].
Applying generic optimizations to the graph for inference.
Reserving memory for host IO tensors. Host: 0 bytes

TODO：
（1）哪些层可以优化？如何优化？
（2）优化前后结构有何区别？

（3）自动优化

=============== Computing costs for /conv1/Conv + /relu/Relu
*************** Autotuning format combination: Float(150528,50176,224,1) -> Float(802816,12544,112,1) ***************
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskConvolution[0x80000009])
"Tactic Name: 策略名称, Tactic: 策略编号, Time: 耗时"
Tactic Name: ampere_scudnn_128x128_relu_medium_nn_v1 Tactic: 0xf067e6205da31c2e Time: 0.11264
...
Tactic Name: ampere_scudnn_128x64_relu_medium_nn_v1 Tactic: 0xf64396b97c889179 Time: 0.0694857
"分析耗时, 最快策略编号及其耗时"
/conv1/Conv + /relu/Relu (CaskConvolution[0x80000009]) profiling completed in 0.124037 seconds. Fastest Tactic: 0xf64396b97c889179 Time: 0.0694857
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CudnnConvolution[0x80000000])
"无有效策略, 跳过"
CudnnConvolution has no valid tactics for this config, skipping
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskFlattenConvolution[0x80000036])
CaskFlattenConvolution has no valid tactics for this config, skipping
>>>>>>>>>>>>>>> Chose Runner Type: CaskConvolution Tactic: 0xf64396b97c889179

*************** Autotuning format combination: Float(150528,1,672,3) -> Float(802816,1,7168,64) ***************
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskConvolution[0x80000009])
Tactic Name: sm80_xmma_fprop_implicit_gemm_indexed_f32f32_f32f32_f32_nhwckrsc_nhwc_tilesize64x64x8_stage3_warpsize1x4x1_g1_ffma_aligna4_alignc4 Tactic: 0x19b688348f983aa0 Time: 0.155648
...
Tactic Name: sm80_xmma_fprop_implicit_gemm_indexed_f32f32_f32f32_f32_nhwckrsc_nhwc_tilesize128x32x8_stage3_warpsize2x2x1_g1_ffma_aligna4_alignc4 Tactic: 0xa6448a1e79f1ca6f Time: 0.174373
/conv1/Conv + /relu/Relu (CaskConvolution[0x80000009]) profiling completed in 0.0186587 seconds. Fastest Tactic: 0xf231cca3335919a4 Time: 0.063488
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskFlattenConvolution[0x80000036])
CaskFlattenConvolution has no valid tactics for this config, skipping
>>>>>>>>>>>>>>> Chose Runner Type: CaskConvolution Tactic: 0xf231cca3335919a4

*************** Autotuning format combination: Float(50176,1:4,224,1) -> Float(802816,12544,112,1) ***************
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskConvolution[0x80000009])
Tactic Name: sm80_xmma_fprop_implicit_gemm_indexed_f32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x128x16_stage4_warpsize2x2x1_g1_tensor16x8x8_alignc4 Tactic: 0xe8f7b6a5bab325f8 Time: 0.35957
Tactic Name: sm80_xmma_fprop_implicit_gemm_indexed_f32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x128x16_stage4_warpsize2x2x1_g1_tensor16x8x8 Tactic: 0xe0a307ffe0ffb6a5 Time: 0.344503
/conv1/Conv + /relu/Relu (CaskConvolution[0x80000009]) profiling completed in 0.0105646 seconds. Fastest Tactic: 0xe0a307ffe0ffb6a5 Time: 0.344503
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskFlattenConvolution[0x80000036])
CaskFlattenConvolution has no valid tactics for this config, skipping
>>>>>>>>>>>>>>> Chose Runner Type: CaskConvolution Tactic: 0xe0a307ffe0ffb6a5

*************** Autotuning format combination: Float(50176,1:4,224,1) -> Float(200704,1:4,1792,16) ***************
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskConvolution[0x80000009])
Tactic Name: ampere_scudnn_128x64_sliced1x2_ldg4_relu_exp_large_nhwc_tn_v1 Tactic: 0xbdfdef6b84f7ccc9 Time: 0.279845
...
Tactic Name: sm80_xmma_fprop_implicit_gemm_indexed_wo_smem_f32f32_tf32f32_f32_nhwckrsc_nhwc_tilesize128x16x32_stage1_warpsize4x1x1_g1_tensor16x8x8 Tactic: 0xae48d3ccfe1edfcd Time: 0.0678278
/conv1/Conv + /relu/Relu (CaskConvolution[0x80000009]) profiling completed in 0.0911707 seconds. Fastest Tactic: 0x9cb304e2edbc1221 Time: 0.059392
--------------- Timing Runner: /conv1/Conv + /relu/Relu (CaskFlattenConvolution[0x80000036])
CaskFlattenConvolution has no valid tactics for this config, skipping
>>>>>>>>>>>>>>> Chose Runner Type: CaskConvolution Tactic: 0x9cb304e2edbc1221
...

"计算重新格式化成本"
=============== Computing reformatting costs
=============== Computing reformatting costs: 
*************** Autotuning Reformat: Float(150528,50176,224,1) -> Float(150528,1,672,3) ***************
--------------- Timing Runner: Optimizer Reformat(x -> <out>) (Reformat[0x80000006])
Tactic: 0x00000000000003e8 Time: 0.00663667
Tactic: 0x00000000000003ea Time: 0.016319
Tactic: 0x0000000000000000 Time: 0.0147323
Optimizer Reformat(x -> <out>) (Reformat[0x80000006]) profiling completed in 0.00680021 seconds. Fastest Tactic: 0x00000000000003e8 Time: 0.00663667
*************** Autotuning Reformat: Float(150528,50176,224,1) -> Float(50176,1:4,224,1) ***************
--------------- Timing Runner: Optimizer Reformat(x -> <out>) (Reformat[0x80000006])
Tactic: 0x00000000000003e8 Time: 0.0116016
Tactic: 0x00000000000003ea Time: 0.0174405
Tactic: 0x0000000000000000 Time: 0.0147739
Optimizer Reformat(x -> <out>) (Reformat[0x80000006]) profiling completed in 0.00589612 seconds. Fastest Tactic: 0x00000000000003e8 Time: 0.0116016
...
=============== Computing reformatting costs

"添加重新格式化层"
Adding reformat layer: Reformatted Input Tensor 0 to /fc/Gemm (/avgpool/GlobalAveragePool_output_0) from Float(512,1,1,1) to Float(128,1:4,128,128)
Adding reformat layer: Reformatted Input Tensor 0 to reshape_after_/fc/Gemm (/fc/Gemm_out_tensor) from Float(250,1:4,250,250) to Float(1000,1,1,1)
Formats and tactics selection completed in 6.81485 seconds.
After reformat layers: 26 layers
Total number of blocks in pre-optimized block assignment: 26
Detected 1 inputs and 1 output network tensors.
"内存、显存、临时内存"
Layer: /conv1/Conv + /relu/Relu Host Persistent: 4016 Device Persistent: 75776 Scratch Memory: 0
...
Layer: /fc/Gemm Host Persistent: 7200 Device Persistent: 0 Scratch Memory: 0
Skipped printing memory information for 3 layers with 0 memory size i.e. Host Persistent + Device Persistent + Scratch Memory == 0.
Total Host Persistent Memory: 86272
Total Device Persistent Memory: 75776
Total Scratch Memory: 4608
"峰值"
[MemUsageStats] Peak memory usage of TRT CPU/GPU memory allocators: CPU 25 MiB, GPU 98 MiB

"块偏移, 用于管理和分配不同层的内存块, 以便在GPU上高效执行计算"
[BlockAssignment] Started assigning block shifts. This will take 27 steps to complete.
STILL ALIVE: Started step 26 of 27
[BlockAssignment] Algorithm ShiftNTopDown took 0.430364ms to assign 4 blocks to 27 nodes requiring 77074944 bytes.
Total number of blocks in optimized block assignment: 4
Total Activation Memory: 77074944
Finalize: /conv1/Conv + /relu/Relu Set kernel index: 0
...
Finalize: /fc/Gemm Set kernel index: 10
Total number of generated kernels selected for the engine: 11
Kernel: 0 CASK_STATIC
...
Kernel: 10 CASK_STATIC
"禁用未使用的策略源, 提高引擎生成的效率"
Disabling unused tactic source: JIT_CONVOLUTIONS
"引擎生成总耗时"
Engine generation completed in 6.96604 seconds.
"删除计时缓存"
Deleting timing cache: 144 entries, served 172 hits since creation.
Engine Layer Information:
Layer(CaskConvolution): /conv1/Conv + /relu/Relu, Tactic: 0xf64396b97c889179, x (Float[-1,3,224,224]) -> /relu/Relu_output_0 (Float[-1,64,112,112])
...
Layer(NoOp): reshape_after_/fc/Gemm, Tactic: 0x0000000000000000, Reformatted Input Tensor 0 to reshape_after_/fc/Gemm (Float[-1,1000,1,1]) -> y (Float[-1,1000])
[MemUsageChange] TensorRT-managed a
Adding 1 engine(s) to plan file.
Engine built in 18.0591 sec.
Loaded engine size: 73 MiB
"反序列化需20415微秒"
Deserialization required 20415 microseconds.
"反序列化中内存变化"
[MemUsageChange] TensorRT-managed allocation in engine deserialization: CPU +0, GPU +72, now: CPU 0, GPU 72 (MiB)
Engine deserialized in 0.0260023 sec.
Total per-runner device persistent memory is 75776
Total per-runner host persistent memory is 86272
Allocated activation device memory of size 77074944
"执行上下文创建过程中内存变化"
[MemUsageChange] TensorRT-managed allocation in IExecutionContext creation: CPU +0, GPU +74, now: CPU 0, GPU 146 (MiB)

TODO：
（1）策略名称包含了关于硬件架构、操作类型、数据类型、优化级别等信息，具体每一项代表了什么含义？
（2）Autotuning format combination 自动优化时使用不同的数据格式，这些输入输出形状是如何确定的？

（4）推理测速

CUDA lazy loading is enabled.
Setting persistentCacheLimit to 0 bytes.
Using enqueueV3.
Using random values for input x
Input binding for x with dimensions 4x3x224x224 is created.
Output binding for y with dimensions 4x1000 is created.
Starting inference
Warmup completed 185 queries over 200 ms
Timing trace has 2523 queries over 3.00267 s

=== Trace details ===
Trace averages of 10 runs:
Average on 10 runs - GPU latency: 1.11339 ms - Host latency: 1.52892 ms (enqueue 0.235403 ms)
...
Average on 10 runs - GPU latency: 1.11299 ms - Host latency: 1.55708 ms (enqueue 0.262207 ms)

=== Performance summary ===
Throughput: 840.251 qps
Latency: min = 1.4353 ms, max = 5.20972 ms, mean = 1.62134 ms, median = 1.55011 ms, percentile(90%) = 1.91345 ms, percentile(95%) = 2.07056 ms, percentile(99%) = 2.51172 ms
Enqueue Time: min = 0.174561 ms, max = 4.94312 ms, mean = 0.261935 ms, median = 0.261475 ms, percentile(90%) = 0.29071 ms, percentile(95%) = 0.315125 ms, percentile(99%) = 0.348938 ms
H2D Latency: min = 0.362305 ms, max = 0.622559 ms, mean = 0.430081 ms, median = 0.433594 ms, percentile(90%) = 0.449829 ms, percentile(95%) = 0.45874 ms, percentile(99%) = 0.493317 ms
GPU Compute Time: min = 1.05573 ms, max = 4.69727 ms, mean = 1.18524 ms, median = 1.11011 ms, percentile(90%) = 1.48682 ms, percentile(95%) = 1.6311 ms, percentile(99%) = 2.10425 ms
D2H Latency: min = 0.00427246 ms, max = 0.0371094 ms, mean = 0.00601927 ms, median = 0.00561523 ms, percentile(90%) = 0.00732422 ms, percentile(95%) = 0.00775146 ms, percentile(99%) = 0.0085144 ms
Total Host Walltime: 3.00267 s
Total GPU Compute Time: 2.99037 s
"GPU计算时间不稳定, 方差系数=18.1606%, 锁定GPU时钟频率或添加--useSpinWait可能提高稳定性"
* GPU compute time is unstable, with coefficient of variance = 18.1606%.
  If not already in use, locking GPU clock frequency or adding --useSpinWait may improve the stability.
Explanations of the performance metrics are printed in the verbose logs.
"性能指标说明, 此处注释根据官方文档中的解释进行补充"
=== Explanations of the performance metrics ===
"预热过后第一个 query 加入队列到最后一个 query 完成的时间"
Total Host Walltime: the host walltime from when the first query (after warmups) is enqueued to when the last query is completed.

"GPU执行一个 query 的延迟"
GPU Compute Time: the GPU latency to execute the kernels for a query.

"GPU执行所有 query 的延迟"
"如果明显比 Total Host Walltime 短, 可能由于 host 端开销或数据传输导致 GPU 利用低效"
Total GPU Compute Time: the summation of the GPU Compute Time of all the queries. If this is significantly shorter than Total Host Walltime, the GPU may be under-utilized because of host-side overheads or data transfers.

"吞吐量: 每秒完成 query 数量"
"如果明显小于 GPU Compute Time 的倒数, 可能由于 host 端开销或数据传输导致 GPU 利用低效"
"throughput = the number of inferences / Total Host Walltime"
"使用 CUDA graphs(--useCudaGraph) 或禁用 H2D/D2H 传输(--noDataTransfer) 可能会提高 GPU 利用率"
"检测到 GPU 未充分利用时， 输出日志会提供相关指导"
Throughput: the observed throughput computed by dividing the number of queries by the Total Host Walltime. If this is significantly lower than the reciprocal of GPU Compute Time, the GPU may be under-utilized because of host-side overheads or data transfers.

"query 排队的 host 延迟"
"如果比 GPU Compute Time 长, GPU 利用低效, 吞吐量可能由 host 端开销主导"
"包括调用 H2D/D2H CUDA APIs、运行 host-side heuristics(host端启发式算法)、启动 CUDA 内核"
"使用 CUDA graphs(--useCudaGraph) 可以减少 Enqueue Time"
Enqueue Time: the host latency to enqueue a query. If this is longer than GPU Compute Time, the GPU may be under-utilized.

H2D Latency: the latency for host-to-device data transfers for input tensors of a single query.

D2H Latency: the latency for device-to-host data transfers for output tensors of a single query.

"官方文档中为 Host Latency, 单个推理的延迟"
"Host Latency = H2D Latency + GPU Compute Time + D2H Latency"
Latency: the summation of H2D Latency, GPU Compute Time, and D2H Latency. This is the latency to infer a single query.

  query 在官方文档中是 inference，理解为一次推理过程；host-side heuristics 翻译为主机端启发式算法，暂不理解具体是什么。

  下图源于官方文档，辅助理解推理过程。

示例二：7项功能

# 当命令返回值不等于0时, 立刻退出脚本, 不会执行后续命令
set -e
# 执行每个命令前, 打印命令及其参数
set -x

clear
rm -rf ./*log ./*.plan ./*.cache ./*.lock ./*.json ./*.raw

# 01 运行onnx
trtexec \
    --onnx=resnet18.onnx \
    > 01-run_onnx.log 2>&1

# 02 parse onnx生成engine
trtexec \
    --onnx=resnet18.onnx \
    --saveEngine=resnet18.plan \
    --timingCacheFile=resnet18.cache \
    --minShapes=x:1x3x224x224 \
    --optShapes=x:4x3x224x224 \
    --maxShapes=x:16x3x224x224 \
    --fp16 \
    --noTF32 \
    --memPoolSize=workspace:1024MiB \
    --builderOptimizationLevel=5 \
    --maxAuxStreams=4 \
    --skipInference \
    --verbose \
    > 02-generate_engine.log 2>&1

# 03 运行engine
trtexec \
    --loadEngine=resnet18.plan \
    --shapes=x:4x3x224x224 \
    --noDataTransfers \
    --useSpinWait \
    --useCudaGraph \
    --verbose \
    > 03-run_engine.log 2>&1

# 04 导出engine信息
trtexec \
    --onnx=resnet18.onnx \
    --skipInference \
    --profilingVerbosity=detailed \
    --dumpLayerInfo \
    --exportLayerInfo="./04-exportLayerInfo.log" \
    > 04-export_layer_info.log 2>&1

# 05 导出profiling信息
trtexec \
    --loadEngine=resnet18.plan \
    --dumpProfile \
    --exportTimes="./05-exportTimes.json" \
    --exportProfile="./05-exportProfile.json" \
    > 05-export_profile.log 2>&1

# 06 保存输入输出数据
trtexec \
    --loadEngine=resnet18.plan \
    --dumpOutput \
    --dumpRawBindingsToFile \
    > 06-save_data.log 2>&1

# 07 读取数据进行推理
trtexec \
    --loadEngine=resnet18.plan \
    --loadInputs=x:x.input.1.3.224.224.Float.raw \
    --dumpOutput \
    > 07-load_data.log 2>&1

  在 cookbook 中，第三个命令无法正常运行就做了下修改；第八个命令需要 plugin 便去除了，后续到 plugin 部分再专门研究。

# 03-Load TensorRT engine built above and do inference
trtexec model-02.plan \
    --trt \
    --shapes=tensorX:4x1x28x28 \
    --noDataTransfers \
    --useSpinWait \
    --useCudaGraph \
    --verbose \
    > result-03.log 2>&1