Generate C/C++/CUDA Code from an AI Model

Generate deployable C/C++ or CUDA code from an AI model using MATLAB Coder or GPU Coder. The workflow follows a common pattern regardless of model framework: load, inspect, write entry-point, generate MEX, verify, then generate production code.

When to Use

• User wants to generate C/C++/CUDA code from an AI model (PyTorch, LiteRT)
• User has a model file (.pt2, .tflite) and wants to load it into MATLAB
• User wants MEX acceleration for an AI model
• User wants to generate CUDA code or GPU-accelerated MEX from an AI model
• User wants to deploy an AI model to hardware
• User wants to use a PyTorch or LiteRT model in Simulink (simulation or code generation)
• User wants to verify AI model numerics between the source framework and MATLAB

When NOT to Use

• General MATLAB Coder usage (codegen syntax, config tuning, writing codegen-ready code)
• Editable dlnetwork for Deep Learning Toolbox workflows (quantization, compression, transfer learning) — use importNetworkFromPyTorch (PyTorch), importNetworkFromTensorFlow (SavedModel), or importNetworkFromKeras (.keras/.h5) which return a dlnetwork. For deployment of an editable dlnetwork with model compression (INT8 quantization via dlquantizer, pruning, projection) or exportNetworkToSimulink workflows — use matlab-deploy-embedded-ai (Pattern 1).
• Training or fine-tuning — this skill is for inference code generation only

Supported Frameworks

Framework	Model format	Load function	Status
PyTorch	.pt2	loadPyTorchExportedProgram	Supported (R2026a+)
LiteRT / TFLite	.tflite	loadLiteRTModel	Supported (R2026a+)

For PyTorch-specific details (API routing, entry-point pattern, export workflow, data layout, common mistakes): see references/pytorch-workflow.md.

For LiteRT-specific details (API routing, entry-point pattern, variable-size inputs, Simulink integration, conventions): see references/litert-workflow.md. For converting TensorFlow/Keras/.h5 to .tflite: see references/tensorflow-to-litert-conversion.md.

Generic Workflow

The code generation workflow follows the same steps for any framework:

1. Load and Inspect

Load the model and check its input/output specifications to determine expected shapes and types.

2. Write Entry-Point Function

Create a codegen-compatible entry-point function that:

• Loads the model from a file path
• Runs inference on an input
• Returns the output

The model file path must be wrapped with coder.Constant so it's known at compile time.

3. Verify Numerics

Compare MATLAB inference output against the source framework to confirm correct loading. Use the same input data in both environments and compare with tolerance.

4. Generate MEX (First!)

Always generate MEX before lib/exe to verify on the host machine:

CPU MEX:

cfg = coder.config("mex");
codegen -config cfg -args {coder.Constant("model_file"), input} entryPoint

CUDA MEX (GPU acceleration):

cfg = coder.gpuConfig("mex");
codegen -config cfg -args {coder.Constant("model_file"), input} entryPoint

For CPU MEX SIMD acceleration (SIMDAcceleration = 'Full' for AVX2 on Intel/AMD), see the matlab-generate-code skill. For the DNN- inference-specific MEX AVX2 ceiling, see references/dnn-codegen-options.md.

5. Verify MEX Output

Compare MEX output against MATLAB reference using matlab.unittest with tolerance:

refOut = entryPoint("model_file", input);
mexOut = entryPoint_mex("model_file", input);
testCase = matlab.unittest.TestCase.forInteractiveUse;
testCase.verifyThat(mexOut, matlab.unittest.constraints.IsEqualTo(refOut, ...
    'Within', matlab.unittest.constraints.AbsoluteTolerance(single(1e-5))));

6. Generate Library/Executable

Once MEX is verified, generate production code:

cfgLib = coder.config("lib");
cfgLib.TargetLang = "C++";  % set to "C++" for C++ output; default is "C"
codegen -config cfgLib -args {coder.Constant("model_file"), input} entryPoint

For DLL: coder.config("dll"). For executable: coder.config("exe").

CUDA variants: Replace coder.config with coder.gpuConfig.

Performance tuning:

• Generic knobs (SIMD instruction sets, reduction-loop vectorization, multithreaded loops): see the matlab-generate-code skill.
• MATLAB Coder ↔ Simulink Coder property naming duality and slbuild set_param patterns: see the matlab-deploy-embedded-code skill.
• DNN-inference-specific knobs (DLTargetLibrary / DeepLearningConfig to disable third-party DL libraries, LargeConstantGeneration to serialize weights to data files): see references/dnn-codegen-options.md.

7. Use in Simulink

For Simulink integration, use the dedicated PyTorch ExportedProgram block from dlosslib — set ModelFilePath to the .pt2 file and it auto-detects input/output shapes. No entry-point function or coder.Constant needed.

Pre/post-processing can be done with Simulink blocks around the dedicated block. If you need everything in a single block, use a MATLAB Function block with loadPyTorchExportedProgram + invoke (same pattern as the entry-point, but the model path is a string literal — no coder.Constant).

Both paths support slbuild code generation (requires fixed-step solver + ERT or GRT target). See references/simulink-workflow.md for full details.

8. Deploy to Hardware (Optional — requires Embedded Coder)

For embedded deployment, use the same entry-point function with an Embedded Coder configuration. See the matlab-deploy-embedded-code skill for ERT config, hardware settings, PIL/SIL verification, and target-specific options. Ask the user to install the skill if it is not installed

Key Functions

Function	Purpose	Package	Since
coder.Constant	Make argument a compile-time constant	MATLAB Coder	R2011a
coder.gpuConfig	Create GPU (CUDA) code generation config	GPU Coder	R2017b
codegen	Generate code	MATLAB Coder	R2011a
loadPyTorchExportedProgram	Load .pt2 into MATLAB	MATLAB Coder Support Package for PyTorch and LiteRT Models	R2026a
loadLiteRTModel	Load .tflite into MATLAB	MATLAB Coder Support Package for PyTorch and LiteRT Models	R2026a

Conventions

• Always check the model's input specifications for correct input shape and type
• Always generate MEX first, verify, then proceed to lib/exe
• Always use coder.Constant for the model file path argument
• Input data is typically single-precision (check model input specs to confirm)
• Do NOT use importNetworkFromPyTorch, importNetworkFromTensorFlow, or importNetworkFromKeras in this skill — they return a dlnetwork on a different path. If the user needs a dlnetwork for quantization, projection, pruning, or exportNetworkToSimulink before code generation, route to matlab-deploy-embedded-ai instead

References

• references/pytorch-workflow.md — Full PyTorch-specific workflow: API routing, entry-point pattern, export guidance, common mistakes, and conventions. Consult for any PyTorch/.pt2 model code generation task. Links to deeper PyTorch references (API signatures, data layout, numeric verification, supported models).
• references/export-pytorch-models.md — Exporting an eager-mode PyTorch model to .pt2 with torch.export (upstream of loading). Consult when the user has a PyTorch model but no .pt2 file yet, or hits torch.export SerializeError / kwarg-mismatch errors. Links to pytorch-export-patterns.md (per-source templates) and pytorch-export-gotchas.md (torch 2.11 serialization fixes).
• references/simulink-workflow.md — Simulink integration: dedicated PyTorch ExportedProgram block (Path A) vs MATLAB Function block (Path B), block mask parameters, code generation config, and key differences from command-line codegen.
• references/litert-workflow.md — Full LiteRT-specific workflow: API routing (loadLiteRTModel → inputSpecifications → invoke), entry-point pattern, variable-size input handling, Simulink integration, and conventions. Consult for any LiteRT/.tflite model code generation task.
• references/tensorflow-to-litert-conversion.md — Converting TensorFlow SavedModel/Keras/.h5 to .tflite via the Python tf.lite.TFLiteConverter API. Consult when the user has a TensorFlow model but no .tflite file yet.
• references/litert-numeric-verification.md — Verifying MEX numerics against MATLAB reference for LiteRT models (tolerance guidance, common mismatches).
• references/codegen-workflow.md — Shared code generation steps for both PyTorch and LiteRT: MEX generation, MEX verification, library/executable targets, coder.Constant usage, and coder.DeepLearningConfig notes.
• references/dnn-codegen-options.md — DNN-INFERENCE-SPECIFIC codegen options: DLTargetLibrary / DeepLearningConfig('none') for the plain-C DL path, LargeConstantGeneration for serializing large DNN weights to data files, and MEX SIMD ceiling in a DNN-inference context. Read this when the generic file's knobs need DNN-specific framing (e.g., "the MEX SIMD cap matters because inference is the target").

See Also

• matlab-generate-code — Generic MATLAB Coder tuning (SIMD instruction sets, OpenMP multi-threading, reduction-loop vectorization). Consult for performance options not specific to DNN inference.
• matlab-deploy-embedded-ai — dlnetwork-based codegen with model compression (quantization, pruning, projection) and exportNetworkToSimulink workflows (Pattern 1). Use it when the source is an editable dlnetwork in MATLAB rather than a .pt2 / .tflite file.
• matlab-optimize-gpu-codegen — CUDA-target codegen tuning (coder.gpuConfig, kernel fusion, memory-hierarchy options) when the deployment target is an NVIDIA GPU rather than CPU or embedded hardware.

---

Copyright 2026 The MathWorks, Inc.

---