Skip to content

compatibility.md

Latest commit

 

History

History
90 lines (72 loc) · 3.88 KB

compatibility.md

File metadata and controls

90 lines (72 loc) · 3.88 KB
Raw

CUDA Compatibility Tests

The compatibility adapter compiles and runs CUDA Samples on NVIDIA and CUDA-compatible accelerator toolchains. It reports compilation, execution, failure, and waived-sample counts without treating a partial pass rate as an adapter execution error.

Test Input

{
  "run_id": "cuda_samples.nvidia.quick",
  "testcase": "compatibility.CudaSamples.PassRate",
  "config": {
    "platform": "nvidia",
    "sample_filter": ["vectorAdd", "matrixMul", "clock"],
    "timeout_per_sample": 180,
    "jobs": 4
  }
}

Initialize the bundled CUDA Samples revision before running compatibility tests:

git submodule update --init submodules/cuda-samples

The adapter uses submodules/cuda-samples by default. An optional cuda_samples_dir override must contain a Samples directory. The adapter recursively discovers Samples/<category>/<sample> directories containing a Makefile or a standalone CMake project. CMake grouping manifests that only aggregate child directories are excluded. A requested sample name that is not found, or an empty sample_filter, is a configuration error instead of a successful zero-sample result.

The bundled submodule pins the CMake-based master revision at 7b601789. The upstream batch_test branch provides the Makefiles used by the original compatibility workflow. The adapter supports both layouts; use build_system to select cmake, make, or the default auto detection.

CMake samples are configured through a temporary wrapper project. The wrapper sets CUDA_ARCHITECTURES on every generated target after the sample manifest has been evaluated, so manifests that set their own default architecture list cannot override the requested sms value.

Platform Toolchains

Platform aliases and compiler candidates are shared with the hardware adapter through infinibench.hardware.constants. Compatibility-only architecture values and Make arguments live in infinibench.common.constants.

Platform Default compiler Default architecture
NVIDIA nvcc 80
MetaX cucc (falls back to mxcc) 70
Iluvatar CoreX (BI-V150/TG150) /usr/local/corex/bin/clang++ ivcore11

The supported canonical platform names are cuda, metax, and corex. The existing aliases nvidia and iluvatar are also accepted.

The CoreX default targets BI-V150/TG150. Override sms and make_args together when testing a different Iluvatar architecture.

Set compiler, sms, or make_args in the input when the installed vendor SDK uses a wrapper or different target. Arguments are passed directly as an argument list; shell expansion is not performed. The resolved default compiler and architecture are added to the result config when they were not explicit in the input. For MetaX, the adapter also infers MACA_PATH from the resolved cucc or mxcc location when the variable is unset. It uses the cu-bridge cmake_maca wrapper when available. An explicit MACA_PATH is preserved.

For non-NVIDIA Makefile builds, the default arguments remove NVIDIA-only --threads, -gencode, and -m64 flags. Platform support is declared only after its compile and runtime workflow has been validated on target hardware.

Metrics

  • compile_passed and compile_failed cover all discovered samples.
  • run_passed and run_failed cover samples that produced an executable.
  • run_skipped counts CUDA Samples that explicitly return a waived result.
  • A sample that does not run because compilation failed has run_result: "not_run" and is not included in run_skipped.
  • run_pass_rate keeps the original end-to-end definition: run passes divided by all selected samples.
  • details records each sample path and the final compiler or runtime error.

result_code: 0 means the compatibility test completed and produced valid measurements. It does not mean every sample passed; use the pass-rate metrics for that decision.