matlab-optimize-gpu-codegen

What it does

Iteratively optimize a MATLAB design file for GPU Coder: compile, benchmark, apply structural optimizations, profile with gpuPerformanceAnalyzer, fix diagnostics, and verify numerical equivalence at every step.

How it works

• Setup: creates a single session directory; uses two helpers in a scripts/ folder: benchmarkMex and extractDiagnostics; adjusts MATLAB path to include these scripts and the design file directory; writes artifacts to the session directory only.
• Step 1 — Codegen on Original: selects codegen config (default to mex), enables SIL for lib/dll targets if needed, resolves inputs, runs codegen -config cfg <designFile> -args {<inputs>}, saves the fixed file as <designFile>_v1.m in the session directory, and ensures a baseline MEX named <designFile>_v1_mex (or SIL-generated MEX with the same name for lib/dll).
• Step 2 — Baseline Benchmark: uses benchmarkMex("<designFile>_v1_mex", {<inputs>}) to obtain a MedianTime; prints baseline time. Always use benchmarkMex (or gputimeit) for GPU timing; avoid tic/toc.
• Step 3 — Structural Optimization Loop: iteratively creates <designFile>_v<N>.m with next optimization, verifies numerical equivalence against the original, runs codegen to generate a new MEX, benchmarks with benchmarkMex, and keeps the fastest passing version as the current best. It uses MATLAB-level restructuring patterns to reduce loop-invariant computations, replace explicit loops with implicit expansion, and reduce divergent branching; includes a rule-based exit when the performance target is met or no further structural optimizations are identifiable. An Exit criteria requires either the performance target is met or no further structural optimizations are found, limited to five iterations for this loop.
• Step 3b — Verify Numerical Equivalence: runs original and optimized versions on up to 5 input sets, generates inputs matching types/sizes, compares outputs with type-appropriate tolerance (double 1e-6, single 1e-3, half 1e-2, integer/logical exact), handles NaN consistency, and ensures outputs match for all cases.
• Step 4 — Profile with gpuPerformanceAnalyzer: runs PA on the current best version using the same cfg from Step 1, saves output to a dedicated paDir, and uses extractDiagnostics on the generated HTML report to obtain diagnostics. If no diagnostics are found, skips to Final Report. Diagnostics are given as a struct with id, message, file, and line.
• Step 5 — Fix Diagnostics: for each diagnostic, creates a new <designFile>_v<N+1>.m, re-runs codegen, re-verifies equivalence, and benchmarks. Provides a mapping from diagnostic IDs to fixes, such as adding coder.gpu.kernelfun, simplifying per-thread memory usage, converting inputs to gpuArray when flagged, increasing parallelism, or keeping results on GPU to avoid CPU-GPU memory copies.

When to use it

• User has a MATLAB function and wants faster GPU MEX or CUDA code; mentions GPU Coder, codegen, CUDA, gpuPerformanceAnalyzer; wants to profile generated GPU/CUDA code or GPU MEX; aims to reduce GPU memory transfers or improve kernel parallelism; rewrite MATLAB for CUDA; or run gpuPerformanceAnalyzer.

What it can touch

• Uses MATLAB codegen workflow with codegen, profiling via gpuPerformanceAnalyzer, and benchmarking via benchmarkMex.
• Interacts with inputs and outputs, potentially converts inputs to gpuArray when a diagnostic indicates GPU input usage; creates and writes <designFile>_v*.m variants in the session directory; relies on the scripts folder for helpers.

Caveats

• Target is limited to GPU codegen workflows; executable targets (exe) are not supported for benchmarking or equivalence checking; instructions require switching to mex or lib/dll if exe is specified.
• Activation depends on the user prompt mentioning GPU, codegen, CUDA, MEX, or profiling; otherwise, the skill remains inactive.
• The license is NOASSERTION; the skill relies on MathWorks tooling and may require appropriate licenses for gpuPerformanceAnalyzer and SIL when using lib/dll targets.