Examples

The canonical, runnable examples for cuda.core live under samples/cuda_core/ in the cuda-python repository. Each sample is a self-contained directory with its own README.md, requirements.txt, and PEP 723 dependency block, and every sample is exercised as part of the cuda.core test suite.

Getting started

• vectorAdd compiles and launches a templated vector-add kernel and demonstrates Device.allocate() and name_expressions.

• deviceQuery enumerates every CUDA device with an nvidia-smi-style summary and an optional --verbose mode for the long-tail property list.

• systemInfo reports system-wide CUDA information (driver / runtime / NVML).

Kernels and launch configuration

• launchConfigTuning explores how launch-configuration choices affect kernel performance.

• threadBlockCluster demonstrates LaunchConfig(cluster=...) on Hopper-class GPUs.

• tmaTensorMap uses Tensor Memory Accelerator descriptors for bulk data movement.

• greenContext partitions SMs across kernels with green contexts.

• kernelNsysProfile annotates kernel launches with NVTX markers for Nsight Systems.

• host_launch.py orders a host function between two kernel launches on a stream and reads a pinned buffer from it.

Memory management and interop

• memoryResources covers DeviceMemoryResource, PinnedMemoryResource, ManagedMemoryResource, and GraphMemoryResource, plus configurable resource options.

• copyImageArraytoGPU stages host-to-device copies through PinnedMemoryResource.

• blurImageUnifiedMemory applies a stencil kernel over unified memory.

• stridedMemoryViewConstructors walks through the four StridedMemoryView.from_* constructors.

• stridedMemoryViewCpu dispatches to a JIT-compiled CPU function via cffi.

• stridedMemoryViewGpu dispatches to an NVRTC-compiled GPU kernel through the same decorator.

• ipcMemoryPool shares an IPC-enabled DeviceMemoryResource across processes.

• buffer_deallocation_stream.py transfers a buffer between streams and safely changes the stream that orders its deallocation.

CUDA graphs and linking

• cudaGraphs captures and replays a multi-kernel graph, then reuses it via Graph.update().

• jitLtoLinking links two device modules with PTX vs LTO and swaps in a runtime plug-in.

Framework interop and compute algorithms

• customPyTorchKernel launches a cuda.core kernel from a PyTorch autograd function.

• customTensorFlowKernel wires a cuda.core kernel into a TensorFlow custom op.

• numpyVsCupy compares NumPy and CuPy execution paths side-by-side.

• fftSignalAnalysis runs a CuPy FFT-based signal analysis pipeline.

• binarySearch, prefixSum, reduction, parallelReduction, reductionMultiBlockCG, parallelHistogram, blockwiseSum, matrixMulSharedMem, cudaComputeLambdas, pageRank – classic parallel-algorithm building blocks.

Multi-GPU and streams

• simpleMultiGpu runs independent kernels on two GPUs in the same process.

• simpleP2P demonstrates peer-to-peer memory access between GPUs.

• multiGPUGradientAverage averages gradients across GPUs with MPI.

• streamingCopyComputeOverlap overlaps copies and compute across multiple streams.

• simpleZeroCopy runs kernels against zero-copy mapped host memory.

• processCheckpoint checkpoints and restores CUDA process state.

Graphics interop

• glInteropPlasma writes CUDA-generated pixels into an OpenGL PBO with zero CPU round-trip.

Simple utilities

• simplePrint minimal kernel that prints from the device.