Device Management

Initialisation

Before calling any device API you must call cuda.init(). This loads the CUDA runtime and driver libraries dynamically. If neither library is present the call succeeds and the CPU fallback backend is activated.

try cuda.init();
defer cuda.deinit();

Counting Devices

const count = try cuda.device.count();
std.debug.print("CUDA devices: {d}\n", .{count});

Returns 0 when running in CPU fallback mode.

Selecting a Device

const dev = try cuda.Device.select(0); // index 0 = first GPU

select calls cudaSetDevice and validates the index. Selecting a device beyond the available count returns error.InvalidDevice.

Reading Properties

const props = try dev.properties();

std.debug.print("Name:           {s}\n",   .{props.name});
std.debug.print("Total VRAM:     {d} MiB\n", .{props.totalGlobalMem / (1024 * 1024)});
std.debug.print("SM count:       {d}\n",   .{props.multiProcessorCount});
std.debug.print("Compute cap:    {d}.{d}\n", .{props.major, props.minor});
std.debug.print("Warp size:      {d}\n",   .{props.warpSize});
std.debug.print("Max threads/SM: {d}\n",   .{props.maxThreadsPerMultiProcessor});
std.debug.print("L2 cache:       {d} KiB\n", .{props.l2CacheSize / 1024});

DeviceProperties mirrors cudaDeviceProp directly. All fields documented in the NVIDIA CUDA Runtime API reference are accessible.

Resetting a Device

try cuda.device.reset();

reset calls cudaDeviceReset, which destroys all allocations, streams, and contexts on the current device. Use only at shutdown.

Driver Version

const v = try cuda.version.driver();
std.debug.print("Driver API version: {d}\n", .{v});

Runtime Version

const v = try cuda.version.runtime();
std.debug.print("CUDA Runtime version: {d}\n", .{v});

Peer Access (Multi-GPU)

See Multi-GPU guide for cuda.peer.canAccess, cuda.peer.enable, and cross-device memcpy.