Remove prefix so we can compare benchmarks

Author: danielfrg

cuda_bindings/benchmarks/README.md

@@ -32,26 +32,30 @@ sudo $(pixi run -e wheel -- which python) -m pyperf system tune
 To run the benchmarks combine the environment and task:
 
 ```bash
-
 # Run the Python benchmarks in the wheel environment
 pixi run -e wheel bench
 
 # Run the Python benchmarks in the source environment
 pixi run -e source bench
 
-# Run the C++ benchmarks (environment is irrelavant here)
+# Run the C++ benchmarks
 pixi run -e wheel bench-cpp
 ```
 
-## pyperf JSON
+Both runners automatically save results to JSON files in the benchmarks
+directory: `results-python.json` and `results-cpp.json`.
 
-The benchmarks are run using [pyperf](https://pyperf.readthedocs.io/en/latest/).
-The results are written to a JSON file in the format expected by pyperf.
+## Output JSON and analysis
 
-The C++ benchmarks also generate a valid JSON file, in the same format.
+The benchmarks are run using [pyperf](https://pyperf.readthedocs.io/en/latest/).
+Both Python and C++ results are saved in pyperf-compatible JSON format,
+which can be analyzed with pyperf commands:
 
-```
-pixi run -e wheel bench-cpp -0 cpp.json
+```bash
+# Show results and statistics
+pixi run -e wheel -- python -m pyperf stats results-python.json
+pixi run -e wheel -- python -m pyperf stats results-cpp.json
 
-pixi run -e wheel pyperf stats cpp.json
+# Compare C++ vs Python results
+pixi run -e wheel -- python -m pyperf compare_to results-cpp.json results-python.json
 ```

cuda_bindings/benchmarks/benchmarks/cpp/bench_ctx_device.cpp

@@ -23,7 +23,7 @@ static void check_cu(CUresult status, const char* message) {
 int main(int argc, char** argv) {
     bench::Options options = bench::parse_args(argc, argv);
     if (options.benchmark_name.empty()) {
-        options.benchmark_name = "cpp.ctx_device.ctx_get_current";
+        options.benchmark_name = "ctx_device.ctx_get_current";
     }
 
     // Setup: init CUDA and create a context

cuda_bindings/benchmarks/benchmarks/cpp/bench_pointer_attributes.cpp

@@ -23,7 +23,7 @@ static void check_cu(CUresult status, const char* message) {
 int main(int argc, char** argv) {
     bench::Options options = bench::parse_args(argc, argv);
     if (options.benchmark_name.empty()) {
-        options.benchmark_name = "cpp.pointer_attributes.pointer_get_attribute";
+        options.benchmark_name = "pointer_attributes.pointer_get_attribute";
     }
 
     // Setup: init CUDA, allocate memory

cuda_bindings/benchmarks/runner/main.py

@@ -30,7 +30,7 @@ def load_module(module_path: Path) -> ModuleType:
 def benchmark_id(module_name: str, function_name: str) -> str:
     module_suffix = module_name.removeprefix("bench_")
     suffix = function_name.removeprefix("bench_")
-    return f"bindings.{module_suffix}.{suffix}"
+    return f"{module_suffix}.{suffix}"
 
 
 def discover_benchmarks() -> dict[str, Callable[[int], float]]: