Deploy Embedded Code

Configure MATLAB Coder with Embedded Coder for production-quality code generation targeting embedded hardware, and verify correctness with processor-in-the-loop (PIL) testing.

When to Use

• Generating C/C++ code for a microcontroller or embedded Linux board
• Setting up PIL or SIL verification for generated code
• Configuring code generation with no dynamic memory allocation
• Configuring Embedded Coder for an AI model entry-point function prepared by matlab-deploy-embedded-ai; for the full AI compression and model-loading workflow, trigger matlab-deploy-embedded-ai first
• Selecting a hardware target board (STM32, Raspberry Pi)

When NOT to Use

• Generating MEX or desktop libraries — use standard codegen workflows
• Simulink-based deployment — use Simulink Coder / Embedded Coder workflows directly
• GPU code generation (CUDA) — use GPU Coder

Workflow

1. Create an ERT-Based Configuration

cfg = coder.config("lib", "ecoder", true);

The "ecoder", true flag creates an ERT-based (Embedded Real-Time) configuration that generates production-quality code with no OS dependencies.

2. Select Target Hardware

With coder.hardware:

cfg.Hardware = coder.hardware("STM32F746G-Discovery");

Without the support package, configure hardware manually:

cfg.HardwareImplementation.ProdHWDeviceType = 'ARM Compatible->ARM Cortex-M';
cfg.HardwareImplementation.ProdBitPerFloat = 32;
cfg.HardwareImplementation.ProdBitPerDouble = 64;

See references/supported-hardware.md for the full list of supported boards and their constraints.

3. Configure Memory for Bare-Metal Targets

cfg.EnableDynamicMemoryAllocation = false;
cfg.StackUsageMax = 512;

• EnableDynamicMemoryAllocation = false — disables malloc/free for targets where heap is unavailable or non-deterministic. All arrays must be bounded at compile time.
• StackUsageMax — set based on target SRAM. The code generation report shows actual usage after compilation.

For entry-points that use deep learning inference (invoke, predict):

cfg.DeepLearningConfig = coder.DeepLearningConfig('none');
cfg.LargeConstantGeneration = "KeepInSourceFiles";

• DeepLearningConfig('none') — generates C with no external DL library dependencies (MKL-DNN, cuDNN, TensorRT). Required for bare-metal targets. Without this, codegen may attempt to link an unavailable library and fail.
• LargeConstantGeneration = "KeepInSourceFiles" — keeps weight constants in source files rather than separate data files. Needed for bare-metal targets where external data file linking is unsupported.

4. Configure Performance (SIMD and OpenMP)

SIMD vectorization for embedded ARM targets:

cfg.InstructionSetExtensions = 'Neon v7';  % ARM Cortex-A (128-bit, 4x float32)

Target	Value	Notes
ARM Cortex-A (Raspberry Pi)	'Neon v7'	128-bit SIMD
ARM Cortex-M	Do not set — use CodeReplacementLibrary instead	Different mechanism

For non-embedded targets (Intel x86-64 and the full InstructionSetExtensions ladder), OptimizeReductions, and OpenMP, see the matlab-generate-code skill. For the MATLAB Coder ↔ Simulink Coder property naming duality, see references/simulink-config.md.

Code replacement library (CRL) — routes supported ops to vendor-optimized math library implementations.

On Cortex-A: set BOTH InstructionSetExtensions AND a non-SIMD CodeReplacementLibrary. They target different layers. ISE emits NEON SIMD intrinsics inline for vectorizable loops in generated code. A non-SIMD CRL routes higher-level math ops — trig, filtering, matrix — to library implementations you don't have to generate. ARM Cortex-A CMSIS maps to the ARM CMSIS-DSP library (hand-tuned Cortex-A kernels shipped by ARM); it does not itself emit NEON, so it composes with an explicit InstructionSetExtensions = 'Neon v7'. Setting only one of ISE or CRL leaves performance on the table:

cfg.HardwareImplementation.ProdHWDeviceType = 'ARM Compatible->ARM Cortex-A';
cfg.InstructionSetExtensions = 'Neon v7';           % SIMD intrinsics (also shown in §4 above)
cfg.CodeReplacementLibrary   = 'ARM Cortex-A CMSIS'; % non-SIMD CRL — routes math ops to CMSIS-DSP

For Cortex-M, select the CRL matching your compiler (e.g. 'ARM Cortex-M' for generic; vendor-specific CRLs are shipped with the corresponding support package). Cortex-M does not use InstructionSetExtensions.

OpenMP — enable on multi-core targets (Cortex-A); disable on single-core (Cortex-M — no OS/threading support, will fail to compile):

cfg.EnableOpenMP = true;   % multi-core targets (Cortex-A)
cfg.EnableOpenMP = false;  % single-core targets (Cortex-M)

5. Set Up PIL Verification

PIL compiles the generated code, deploys it to the physical board, sends test vectors, and compares outputs against MATLAB. This catches precision differences, stack overflows, and memory issues that SIL cannot detect.

Cortex-M (serial transport):

cfg.VerificationMode = "PIL";
cfg.Hardware.PILInterface = "Serial";
cfg.Hardware.PILCOMPort = "COM4";  % adjust to your system

Cortex-A / Raspberry Pi (SSH transport):

cfg.VerificationMode = "PIL";
cfg.Hardware = coder.hardware("Raspberry Pi");
cfg.Hardware.DeviceAddress = "192.168.1.10";
cfg.Hardware.Username = "<your-pi-username>";
cfg.Hardware.Password = "<your-pi-password>";
cfg.Hardware.BuildDir = "/home/pi/mymodel";  % optional: defaults to /home/pi/MATLAB_ws/<release>

Pi PIL runs over SSH (not serial). The support package uses DeviceAddress, Username, and Password to establish the SSH connection. BuildDir specifies where the compiled binary is deployed on the target; if omitted, defaults to /home/pi/MATLAB_ws/<release>/. Do not set PILInterface or PILCOMPort — those are for serial-connected bare-metal boards only.

6. Generate Code

cfg.TargetLang = "C";
codegen -config cfg -args {inputArgs} myEntryPoint

7. Verify with the MEX → SIL → PIL Progression

For confidence in deployment, follow this sequence:

1. MEX — verify on host, fast iteration
2. SIL (Software-in-the-Loop) — run generated code on host, compare to MATLAB
3. PIL (Processor-in-the-Loop) — run on actual hardware, compare to MATLAB

cfgSil = coder.config("lib", "ecoder", true);
cfgSil.VerificationMode = "SIL";
codegen -config cfgSil -args {inputArgs} myEntryPoint

Key Properties

Property	Values	Purpose
VerificationMode	"PIL", "SIL", "None"	Enable in-the-loop verification
Hardware	coder.hardware(boardName)	Select target board
Hardware.PILInterface	"Serial"	PIL communication type
Hardware.PILCOMPort	"COM4", "/dev/ttyACM0"	Serial port
EnableDynamicMemoryAllocation	true (default), false	Master switch for heap
DynamicMemoryAllocationThreshold	numeric (bytes), default 65536	Arrays above this use heap
LargeConstantGeneration	"KeepInSourceFiles", "WriteOnlyDNNConstantsToDataFiles"	Where to put large constants
StackUsageMax	numeric (bytes)	Stack limit for generated code (Simulink: MaxStackSize)
EnableOpenMP	boolean	OpenMP multi-threading (Simulink: MultiThreadedLoops)
CodeReplacementLibrary	"ARM Cortex-A CMSIS", "GCC ARM Cortex-A" (SIMD), "ARM Cortex-M", …	Vendor-optimized op replacements. GCC ARM Cortex-A is a SIMD CRL — do not pair with an explicit ISE. MATLAB warns it is not recommended; use "ARM Cortex-A CMSIS" instead
TargetLang	"C", "C++"	Output language

Simulink Coder path (slbuild): properties in the "Simulink:" column above are set via set_param(modelName, 'PropertyName', value) directly on the model name. See references/simulink-config.md for the full Simulink Coder property mapping and code examples.

Common Mistakes

Mistake	Why It's Wrong	Correct Approach
DynamicMemoryAllocation = "Off"	Wrong property name and type	EnableDynamicMemoryAllocation = false (boolean)
Skipping SIL before PIL	PIL failures on hardware are harder to debug	Always validate with SIL first
Not setting StackUsageMax	Default may exceed target SRAM	Set explicitly based on hardware constraints
Using cfg = coder.config("lib") without "ecoder", true	Creates a generic config, not ERT-based	Always pass "ecoder", true for embedded targets

Conventions

• Always: use coder.config("lib", "ecoder", true) for embedded targets
• Always: disable dynamic memory for bare-metal Cortex-M targets
• Always: follow MEX → SIL → PIL verification order
• Never: use DynamicMemoryAllocation (wrong property name — it's EnableDynamicMemoryAllocation)
• Prefer: TargetLang = "C" for Cortex-M targets (smaller code footprint)

References

• references/supported-hardware.md — board specs, support packages, and PIL interface details
• references/simulink-config.md — Simulink Coder property naming, set_param patterns, and slbuild code examples

See Also

• matlab-generate-code — generic codegen foundation: coder directives, screener, MEX generation, SIL, config tuning, SIMD/OpenMP for non-embedded targets
• matlab-deploy-ai-model — full AI model codegen pipeline (load, verify, generate MEX/lib)

---

Copyright 2026 The MathWorks, Inc.

---