Add TMA unittest app

[William-An]

• Add TMA unittest apps
• Add GMMA ubench
• Enable dynamic linking of CUDA runtime by default for all apps
• Add support for parallel NVCC compilation while retaining embedded PTX code
• Drop parboil as it is using python2
• Also fix Error while generating traces for GPU_Microbenchmark accel-sim-framework#491

[tgrogers]

This all looks good, but should we put it in the uBench folder?

[William-An]

This CI build may take some time to build all the GEMM kernels for GMMA. Perhaps we should build a single kernel instead from the test source files (i.e., lat_gmma_test.cu and MaxFlops_gmma_test.cu)?

Or let the cluster build this?

[Copilot]

Pull request overview

This PR adds new GPU microbenchmarking capabilities for TMA and GMMA operations while fixing exit codes and enabling dynamic CUDA runtime linking. The changes modernize the build system to support parallel compilation and add comprehensive test coverage for Hopper architecture (SM90) matrix operations.

Key Changes:

• Corrected return values from 1 to 0 for successful execution across all microbenchmark programs
• Added GMMA (General Matrix Multiply-Accumulate) latency and throughput microbenchmarks with comprehensive test coverage for multiple data types
• Enabled parallel NVCC compilation while preserving embedded PTX code generation

Reviewed changes

Copilot reviewed 299 out of 331 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
Multiple *.cu files	Fixed incorrect return value from 1 to 0 for successful program completion
lat_gmma/ directory	Added comprehensive GMMA latency microbenchmark infrastructure with support for F32, F16, and INT32 accumulators
MaxFlops_gmma/ directory	Added GMMA throughput microbenchmark kernels for various data type combinations
lat_gmma_common.h	Introduced shared kernel templates and helper macros for GMMA latency testing
lat_gmma.h	Defined function declarations for 385 different GMMA test configurations
lat_gmma/Makefile	Configured build system for SM90a architecture with C++17 and parallel compilation
.gitignore	Added ignore patterns for build artifacts (.a and .ptx files)

Comments suppressed due to low confidence (1)

src/cuda/GPU_Microbenchmark/ubench/core/lat_gmma/lat_gmma_common.h:1

• Corrected grammar from 'Simple a test kernel' to 'Simple test kernel'.

/***************************************************************************************************

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[William-An]

CI is failing due to insufficient space on the runner.