Complete demonstration of CUDA launch configuration tuning. This sample shows: 1. Device initialization with cuda.core.Device 2. Kernel compilation with cuda.core.Program 3. Benchmarking different thread block configurations 4. Finding optimal threads-per-block for various
()
| 332 | |
| 333 | |
| 334 | def main(): |
| 335 | """ |
| 336 | Complete demonstration of CUDA launch configuration tuning. |
| 337 | |
| 338 | This sample shows: |
| 339 | 1. Device initialization with cuda.core.Device |
| 340 | 2. Kernel compilation with cuda.core.Program |
| 341 | 3. Benchmarking different thread block configurations |
| 342 | 4. Finding optimal threads-per-block for various kernel types |
| 343 | """ |
| 344 | print("=" * 60) |
| 345 | print("Launch Configuration Tuning (cuda.core)") |
| 346 | print("Finding the Best Block Size for Your Kernel") |
| 347 | print("=" * 60) |
| 348 | |
| 349 | # Initialize CUDA device |
| 350 | device = Device(0) |
| 351 | device.set_current() |
| 352 | |
| 353 | # This sample builds host NumPy views over managed memory. Devices without |
| 354 | # concurrent managed access keep managed allocations GPU-exclusive while |
| 355 | # the GPU is active, so those host views fault. |
| 356 | if not device.properties.concurrent_managed_access: |
| 357 | print("Concurrent managed memory access is not supported on this platform.") |
| 358 | sys.exit(2) |
| 359 | |
| 360 | # Print GPU information |
| 361 | print_gpu_info(device) |
| 362 | |
| 363 | # Create stream and device-preferred memory resource |
| 364 | stream = device.create_stream() |
| 365 | mr_options = ManagedMemoryResourceOptions(preferred_location=device.device_id) |
| 366 | mr = ManagedMemoryResource(mr_options) |
| 367 | |
| 368 | try: |
| 369 | # Compile kernels |
| 370 | print("\nCompiling CUDA kernels with cuda.core.Program...") |
| 371 | arch = f"sm_{device.arch}" |
| 372 | print(f" Target architecture: {arch}") |
| 373 | |
| 374 | vec_add_kernel = compile_kernel(device, VECTOR_ADD_KERNEL, "vector_add") |
| 375 | print(" [OK] vector_add kernel compiled") |
| 376 | |
| 377 | reduction_kernel = compile_kernel(device, REDUCTION_KERNEL, "reduce_sum") |
| 378 | print(" [OK] reduce_sum kernel compiled") |
| 379 | |
| 380 | # Run demonstrations |
| 381 | demo_vector_add_tuning(device, stream, mr, vec_add_kernel) |
| 382 | demo_reduction_tuning(device, stream, mr, reduction_kernel) |
| 383 | |
| 384 | print("\n" + "=" * 60) |
| 385 | print("SAMPLE COMPLETE") |
| 386 | print("=" * 60) |
| 387 | print("\nKey Takeaway: The optimal thread configuration depends on your") |
| 388 | print("specific kernel characteristics. Always benchmark to find the best!") |
| 389 | print() |
| 390 | finally: |
| 391 | stream.close() |
no test coverage detected