MCPcopy Create free account
hub / github.com/NVIDIA/cuda-samples / run_numba_kernel

Function run_numba_kernel

python/1_GettingStarted/simplePrint/simplePrint.py:188–243  ·  view source on GitHub ↗

Demonstrate printing from a Numba CUDA kernel. This example uses numba.cuda.grid(3) as the primary indexing mechanism (recommended modern style), and also prints the equivalent blockId / threadId used in the CUDA C++ printf sample for side-by-side comparison. Uses cuda.core AP

(device, test_value=10)

Source from the content-addressed store, hash-verified

186
187
188def run_numba_kernel(device, test_value=10):
189 """
190 Demonstrate printing from a Numba CUDA kernel.
191
192 This example uses numba.cuda.grid(3) as the primary indexing mechanism
193 (recommended modern style), and also prints the equivalent blockId /
194 threadId used in the CUDA C++ printf sample for side-by-side comparison.
195
196 Uses cuda.core APIs for stream management, demonstrating interoperability
197 between Numba CUDA kernels and cuda.core infrastructure.
198 """
199 print("\n")
200 print("=" * 70)
201 print("METHOD 2: Numba CUDA Kernel (Pythonic / modern indexing)")
202 print("=" * 70)
203 print("Advantage: Uses numba.cuda.grid(3) for global indexing,")
204 print(" while still showing classic CUDA C++ IDs for reference.")
205 print(" Uses cuda.core for stream management (interoperability).")
206 print()
207
208 # Same launch configuration as the C++ version
209 grid_x, grid_y = 2, 2
210 block_x, block_y, block_z = 2, 2, 2
211
212 print("Kernel configuration:")
213 print(f" Grid: ({grid_x}, {grid_y})")
214 print(f" Block: ({block_x}, {block_y}, {block_z})")
215 print(f" Total threads: {grid_x * grid_y * block_x * block_y * block_z}")
216 print()
217
218 # Use cuda.core stream (same as C++ example) instead of numba.cuda.stream()
219 stream = device.create_stream()
220
221 print(f"Launching Numba kernel (grid(3) + classic IDs) with value={test_value}:")
222 print("Uses numba.cuda.grid(3) to get global (x, y, z),")
223 print("and prints the corresponding blockId/threadId like the C++ sample.")
224 print("Stream managed by cuda.core for consistency with C++ example.\n")
225
226 try:
227 # Launch Numba kernel on cuda.core stream
228 numba_print_kernel[(grid_x, grid_y), (block_x, block_y, block_z), stream](
229 test_value
230 )
231
232 # Synchronize cuda.core stream (same as C++ example)
233 stream.sync()
234 print("\nNumba CUDA kernel execution complete.")
235 except Exception as e:
236 print(f"\nError during Numba kernel execution: {e}")
237 traceback.print_exc()
238 return 1
239 finally:
240 # Cleanup
241 stream.close()
242
243 return 0
244
245

Callers 1

mainFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected