Main Content

Analyze Performance of Code Generated for Deep Learning Networks

This example shows you how to analyze and optimize the performance of generated CUDA® code for deep learning networks by using the gpuPerformanceAnalyzer function.

The gpuPerformanceAnalyzer function runs a software-in-the-loop (SIL) execution that collects metrics on CPU and GPU activities in the generated code. The function provides a report containing a chronological timeline plot that you can use to visualize, identify, and mitigate performance bottlenecks in the generated CUDA code.

This example generates the performance analysis report for the Generate Digit Images on NVIDIA GPU Using Variational Autoencoder example from GPU Coder. For more information, see Generate Digit Images on NVIDIA GPU Using Variational Autoencoder.

Third-Party Prerequisites

  • CUDA-enabled NVIDIA® GPU.

  • NVIDIA CUDA toolkit and driver. For information on the supported versions of the compilers and libraries, see Third-Party Hardware.

  • Environment variables for the compilers and libraries. For setting up the environment variables, see Setting Up the Prerequisite Products.

  • Permissions to access GPU performance counters. From CUDA toolkit v10.1, NVIDIA restricts access to performance counters to only admin users. To enable GPU performance counters to for all users, see the instructions provided in Permission issue with Performance Counters (NVIDIA).

Verify GPU Environment

To verify that the compilers and libraries for this example are set up correctly, use the coder.checkGpuInstall function.

envCfg = coder.gpuEnvConfig("host");
envCfg.DeepLibTarget = "cudnn";
envCfg.Profiling = 1;
envCfg.DeepCodegen = 1;
envCfg.Quiet = 1;

When the Quiet property of the coder.gpuEnvConfig object is set to true, the coder.checkGpuInstall function returns only warning or error messages.

Pretrained Variational Autoencoder Network (VAE)

Autoencoders have two parts: the encoder and the decoder. The encoder takes an image input and outputs a compressed representation that is a vector of size latent_dim. In this example, the value of latent_dim is equal to 20. The decoder takes this compressed representation, decodes it, and recreates the original image.

VAEs differ from regular autoencoders in that they do not use the encoding-decoding process to reconstruct an input. Instead, they impose a probability distribution on the latent space, and learn the distribution so that the distribution of outputs from the decoder matches that of the observed data. Then, they sample from this distribution to generate new data.

This example uses the decoder network trained in the Train Variational Autoencoder (VAE) to Generate Images example. To train the network yourself, see Train Variational Autoencoder (VAE) to Generate Images (Deep Learning Toolbox).


The generateVAE Entry-Point Function

The generateVAE entry-point function loads the dlnetwork object from the trainedDecoderVAENet MAT-file into a persistent variable and reuses the persistent object during subsequent prediction calls. It initializes a dlarray object containing 25 randomly generated encodings, passes them through the decoder network, and extracts the numeric data of the generated image from the deep learning array object.

function generatedImage =  generateVAE(decoderNetFileName,latentDim,Environment) %#codegen
% Copyright 2020-2021 The MathWorks, Inc.

persistent decoderNet;
if isempty(decoderNet)
    decoderNet = coder.loadDeepLearningNetwork(decoderNetFileName);

% Generate random noise
randomNoise = dlarray(randn(1,1,latentDim,25,'single'),'SSCB');

if'MATLAB') && strcmp(Environment,'gpu')
    randomNoise = gpuArray(randomNoise);

% Generate new image from noise
generatedImage = sigmoid(predict(decoderNet,randomNoise));

% Extract numeric data from dlarray
generatedImage = extractdata(generatedImage);


Generate GPU Performance Analyzer Report

To analyze the performance of the generated code by using the gpuPerformanceAnalyzer function, create a code configuration object with a dynamic library build type by using the dll input argument. Because the gpuPerformanceAnalyzer function accepts only an Embedded Coder® configuration object, enable the option to create a coder.EmbeddedCodeConfig configuration object.

cfg = coder.gpuConfig("dll","ecoder",true);

Use the coder.DeepLearningConfig function to create a CuDNN deep learning configuration object and assign it to the DeepLearningConfig property of the GPU code configuration object.

cfg.TargetLang = "C++";
cfg.GpuConfig.EnableMemoryManager = true;
cfg.DeepLearningConfig = coder.DeepLearningConfig("cudnn");

Run gpuPerformanceAnalyzer with the default iteration count of 2.

latentDim = 20;
Env = "gpu";
matfile = "trainedDecoderVAENet.mat";
inputs  = {coder.Constant(matfile), coder.Constant(latentDim), coder.Constant(Env)};
designFileName = "generateVAE";

gpuPerformanceAnalyzer(designFileName, inputs, ...
    "Config", cfg, "NumIterations", 2);
### Starting GPU code generation
Code generation successful: View report

### GPU code generation finished
### Starting application profiling
### Starting SIL execution for 'generateVAE'
    To terminate execution: clear generateVAE_sil
### Application stopped
### Stopping SIL execution for 'generateVAE'
### Application profiling finished
### Starting profiling data processing
### Profiling data processing finished
### Showing profiling data

Generate Performance Analyzer Report Using codegen

You can use the -gpuprofile option of the codegen command to enable GPU profiling and create the GPU performance analyzer report. For example,

codegen -config cfg -gpuprofile generateVAE.m -args inputs

When code generation completes, the software generates a generateVAE_sil executable. Run the SIL executable.


Click the clear generateVAE_sil link in the MATLAB® Command Window. The GPU performance analyzer report is available as a link in the Command Window after terminating the SIL executable. For example,

### Application stopped
### Stopping SIL execution for 'fog_rectification'
### Starting profiling data processing
### Profiling data processing finished
    Open GPU Performance Analyzer report: open('/home/test/gpucoder-ex87489778/codegen/dll/fog_rectification/html/gpuProfiler.mldatx')

GPU Performance Analyzer

The GPU performance analyzer report lists GPU and CPU activities, events, and performance metrics in a chronological timeline plot that you can use to visualize, identify, and address performance bottlenecks in the generated CUDA code.

These numbers are representative. The actual values depend on your hardware setup. The profiling in this example was performed using MATLAB® R2023b on a machine with an 6 core, 3.5GHz Intel® Xeon® CPU, and an NVIDIA TITAN XP GPU.

Profiling Timeline

The profiling timeline shows the complete trace of all events that have a runtime higher than the threshold value. This image shows a snippet of the profiling trace when the threshold value is set to 0.0 ms.

You can use the mouse wheel or the equivalent touchpad option to control the zoom level of the timeline. Alternatively, you can use the timeline summary at the top of the panel to control the zoom level and navigate the timeline plot.

The tooltips on each event indicate the start time, end time, and duration of the selected event on the CPU and the GPU. The tooltips also indicate the time elapsed between the kernel launch on the CPU and the actual execution of the kernel on the GPU.

Use the right-click context menu on each event to add a trace between the CPU and corresponding GPU events. You can also use the right-click menu to view the generated CUDA code that corresponds to an event on the code pane.

Event Statistics

The event statistics pane shows additional information for the selected event. For example, the Crop2dImpl kernel shows the following statistics:


The insights pane includes pie charts that provide an overview of the GPU and CPU activities. The pie chart changes according to the zoom level of the profiling timeline. This image shows a snippet of the insights. Within the region selected on the timeline, it shows that the GPU utilization is 42%.

Trace Code

You can use the code pane to trace from the MATLAB® code to the CUDA code or from the CUDA code to the MATLAB code. Traceable code is marked with blue on the side that you are tracing from and with orange on the side that you are tracing to. As you point to the traceable code, the pane highlights the code in purple and traces the corresponding code on the other side. When you select a code section, the pane highlights the code in yellow. The code remains selected until you press Esc or select different code. To change the side that you are tracing from, select code on the other side.

Call Tree

This section lists the GPU events called from the CPU. Each event in the call tree lists the execution times as percentages of the caller function. This metric can help you to identify performance bottlenecks in the generated code. You can also navigate to specific events on the profiling timeline by clicking on the corresponding events in the call tree.


This section provides filtering options for the report.

  • View Mode — View profiling results for the entire application, including initialization and terminate, or the design function (without initialization and terminate).

  • Event Threshold — Skip events shorter than the given threshold.

  • Memory Allocation/Free — Show GPU device memory allocation and deallocation related events on the CPU activities bar.

  • Memory Transfers — Show host-to-device and device-to-host memory transfers.

  • Kernels — Show CPU kernel launches and GPU kernel activities.

  • Others — Show other GPU related events such as synchronization and waiting for GPU.

See Also



Related Topics