feat: add CUDA Samples compatibility testing

docs/compatibility.md

@@ -0,0 +1,90 @@
+# 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
+
+```json
+{
+  "run_id": "cuda_samples.nvidia.quick",
+  "testcase": "compatibility.CudaSamples.PassRate",
+  "config": {
+    "platform": "nvidia",
+    "cuda_samples_dir": "/workspace/cuda-samples",
+    "build_system": "make",
+    "sample_filter": ["vectorAdd", "matrixMul", "clock"],
+    "timeout_per_sample": 180,
+    "jobs": 4
+  }
+}
+```
+
+`cuda_samples_dir` 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 InfiniPerf cuda-samples submodule pins the CMake-based `master` revision.
+Its `batch_test` branch provides the Makefiles used by the original InfiniPerf
+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 `infinimetrics.hardware.constants`. Compatibility-only architecture
+values and Make arguments remain in `infinimetrics.compatibility.constants`.
+This keeps hardware detection and compiler discovery consistent without
+coupling CUDA Samples to the hardware benchmark build script.
+
+| Platform | Default compiler | Default architecture |
+| --- | --- | --- |
+| NVIDIA | `nvcc` | `80` |
+| MetaX | `cucc` (falls back to `mxcc`) | `70` |
+| Iluvatar CoreX | `/usr/local/corex/bin/clang++` | `ivcore20` |
+
+The supported canonical platform names are `cuda`, `metax`, and `corex`.
+The existing aliases `nvidia` and `iluvatar` are also accepted.
+
+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. 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.
+
+Generate a readable report and optional CSV with:
+
+```bash
+python scripts/generate_compat_report.py result.json --csv result.csv
+```

infinimetrics/common/constants.py

@@ -31,6 +31,7 @@ class TestCategory(str, Enum):
     INFER = "infer"
     COMM = "comm"
     TRAIN = "train"
+    COMPATIBILITY = "compatibility"
 
 
 # Valid test categories (derived from TestCategory enum)

infinimetrics/compatibility/__init__.py

@@ -0,0 +1,2 @@
+#!/usr/bin/env python3
+"""Compatibility testing module for CUDA Samples compilation and execution."""