public CudaError LaunchKernelWithStreamBinding(
CudaStream stream,
CudaKernel kernel,
RuntimeKernelConfig config,
IntPtr args,
IntPtr kernelArgs)
{
var binding = stream.BindScoped();
var result = LaunchKernel(
kernel.FunctionPtr,
config.GridDim.X,
config.GridDim.Y,
config.GridDim.Z,
config.GroupDim.X,
config.GroupDim.Y,
config.GroupDim.Z,
config.SharedMemoryConfig.DynamicArraySize,
stream.StreamPtr,
args,
kernelArgs);
binding.Recover();
return(result);
}
public CudaError MemcpyDeviceToDevice(
IntPtr destinationDevice,
IntPtr sourceDevice,
IntPtr length,
AcceleratorStream stream)
{
CudaStream cudaStream = stream as CudaStream;
return(MemcpyDeviceToDevice(
destinationDevice,
sourceDevice,
length,
cudaStream?.StreamPtr ?? IntPtr.Zero));
}
public CudaError LaunchKernelWithStruct <T>(
CudaStream stream,
CudaKernel kernel,
RuntimeKernelConfig config,
ref T args,
int argsSizeInBytes)
where T : unmanaged
{
// Setup object size
var size = new IntPtr(argsSizeInBytes);
Debug.Assert(
argsSizeInBytes <= Interop.SizeOf <T>(),
"Invalid argument size");
// Pin object buffer in memory
fixed(T *pArgs = &args)
{
// Setup unmanaged launch configuration for the driver
var launchConfig = stackalloc void *[5];
launchConfig[0] = (void *)1; // CU_LAUNCH_PARAM_BUFFER_POINTER
launchConfig[1] = pArgs;
launchConfig[2] = (void *)2; // CU_LAUNCH_PARAM_BUFFER_SIZE
launchConfig[3] = &size;
launchConfig[4] = (void *)0; // CU_LAUNCH_PARAM_END
// Use existing launch configuration
return(LaunchKernelWithStreamBinding(
stream,
kernel,
config,
IntPtr.Zero,
new IntPtr(launchConfig)));
}
}
/// <summary>
/// Performs a Cuda memset operation.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <param name="stream">The Cuda stream to use (must not be null)</param>
/// <param name="value">The value to write into the buffer.</param>
/// <param name="targetView">The target view to write to.</param>
public static void CudaMemSet <T>(
CudaStream stream,
byte value,
in ArrayView <T> targetView)
/// <summary>
/// Setups all required settings.
/// </summary>
private void SetupAccelerator()
{
Bind();
CudaException.ThrowIfFailed(
CurrentAPI.GetDeviceName(out string name, DeviceId));
Name = name;
DefaultStream = new CudaStream(this, IntPtr.Zero);
CudaException.ThrowIfFailed(
CurrentAPI.GetTotalDeviceMemory(out long total, DeviceId));
MemorySize = total;
// Resolve max grid size
MaxGridSize = new Index3(
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z, DeviceId));
// Resolve max group size
MaxGroupSize = new Index3(
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z, DeviceId));
// Resolve max threads per group
MaxNumThreadsPerGroup = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, DeviceId);
// Resolve max shared memory per block
MaxSharedMemoryPerGroup = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK,
DeviceId);
// Resolve total constant memory
MaxConstantMemory = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY, DeviceId);
// Resolve clock rate
ClockRate = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_CLOCK_RATE, DeviceId);
// Resolve warp size
WarpSize = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_WARP_SIZE, DeviceId);
// Resolve number of multiprocessors
NumMultiprocessors = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, DeviceId);
// Result max number of threads per multiprocessor
MaxNumThreadsPerMultiprocessor = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
DeviceId);
// Resolve cache configuration
CudaException.ThrowIfFailed(
CurrentAPI.GetSharedMemoryConfig(out sharedMemoryConfiguration));
CudaException.ThrowIfFailed(
CurrentAPI.GetCacheConfig(out cacheConfiguration));
// Setup architecture and backend
CudaException.ThrowIfFailed(
CurrentAPI.GetDeviceComputeCapability(
out int major,
out int minor,
DeviceId));
Architecture = PTXArchitectureUtils.GetArchitecture(major, minor);
CudaException.ThrowIfFailed(
CurrentAPI.GetDriverVersion(out var driverVersion));
InstructionSet = GetInstructionSet(Architecture, driverVersion);
Init(new PTXBackend(
Context,
Architecture,
InstructionSet));
}
/// <summary>
/// Associates a CUDA stream with a cuFFT plan.
/// </summary>
public void SetStream(CudaStream cudaStream)
{
CuFFTException.ThrowIfFailed(
API.SetStream(PlanHandle, cudaStream));
}
/// <summary>
/// Setups all required settings.
/// </summary>
private void SetupAccelerator()
{
Bind();
CudaException.ThrowIfFailed(
CurrentAPI.GetDeviceName(out string name, DeviceId));
Name = name;
DefaultStream = new CudaStream(this, IntPtr.Zero, false);
CudaException.ThrowIfFailed(
CurrentAPI.GetTotalDeviceMemory(out long total, DeviceId));
MemorySize = total;
// Resolve max grid size
MaxGridSize = new Index3(
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z, DeviceId));
// Resolve max group size
MaxGroupSize = new Index3(
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y, DeviceId),
CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z, DeviceId));
// Resolve max threads per group
MaxNumThreadsPerGroup = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, DeviceId);
// Resolve max shared memory per block
MaxSharedMemoryPerGroup = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK,
DeviceId);
// Resolve total constant memory
MaxConstantMemory = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY, DeviceId);
// Resolve clock rate
ClockRate = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_CLOCK_RATE, DeviceId) / 1000;
// Resolve memory clock rate
MemoryClockRate = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE, DeviceId) / 1000;
// Resolve the bus width
MemoryBusWidth = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH, DeviceId);
// Resolve warp size
WarpSize = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_WARP_SIZE, DeviceId);
// Resolve number of multiprocessors
NumMultiprocessors = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, DeviceId);
// Result max number of threads per multiprocessor
MaxNumThreadsPerMultiprocessor = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
DeviceId);
// Resolve the L2 cache size
L2CacheSize = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE, DeviceId);
// Resolve the maximum amount of shared memory per multiprocessor
MaxSharedMemoryPerMultiprocessor = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR,
DeviceId);
// Resolve the total number of registers per multiprocessor
TotalNumRegistersPerMultiprocessor = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR,
DeviceId);
// Resolve the total number of registers per group
TotalNumRegistersPerGroup = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK, DeviceId);
// Resolve the max memory pitch
MaxMemoryPitch = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MAX_PITCH, DeviceId);
// Resolve the number of concurrent copy engines
NumConcurrentCopyEngines = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, DeviceId);
// Resolve whether this device has ECC support
HasECCSupport = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_ECC_ENABLED, DeviceId) != 0;
// Resolve whether this device supports managed memory
SupportsManagedMemory = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY, DeviceId) != 0;
// Resolve whether this device supports compute preemption
SupportsComputePreemption = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED,
DeviceId) != 0;
// Resolve the current driver mode
DriverMode = (DeviceDriverMode)CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_TCC_DRIVER,
DeviceId);
// Resolve the PCI domain id
PCIDomainId = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID,
DeviceId);
// Resolve the PCI device id
PCIBusId = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_BUS_ID,
DeviceId);
// Resolve the PCI device id
PCIDeviceId = CurrentAPI.GetDeviceAttribute(
DeviceAttribute.CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID,
DeviceId);
// Resolve cache configuration
CudaException.ThrowIfFailed(
CurrentAPI.GetSharedMemoryConfig(out sharedMemoryConfiguration));
CudaException.ThrowIfFailed(
CurrentAPI.GetCacheConfig(out cacheConfiguration));
// Setup architecture and backend
CudaException.ThrowIfFailed(
CurrentAPI.GetDeviceComputeCapability(
out int major,
out int minor,
DeviceId));
Architecture = PTXArchitectureUtils.GetArchitecture(major, minor);
CudaException.ThrowIfFailed(
CurrentAPI.GetDriverVersion(out var driverVersion));
DriverVersion = driverVersion;
InstructionSet = GetInstructionSet(Architecture, driverVersion);
base.Capabilities = new CudaCapabilityContext(Architecture);
Init(new PTXBackend(
Context,
Capabilities,
Architecture,
InstructionSet));
}
