/// <summary>
/// Constructs a new Cuda accelerator.
/// </summary>
/// <param name="context">The ILGPU context.</param>
/// <param name="deviceId">The target device id.</param>
/// <param name="acceleratorFlags">The accelerator flags.</param>
public CudaAccelerator(
Context context,
int deviceId,
CudaAcceleratorFlags acceleratorFlags)
: base(context, AcceleratorType.Cuda)
{
CudaException.ThrowIfFailed(
CurrentAPI.CreateContext(out contextPtr, acceleratorFlags, deviceId));
DeviceId = deviceId;
SetupAccelerator();
}
/// <summary>
/// Constructs a new OpenCL accelerator.
/// </summary>
/// <param name="context">The ILGPU context.</param>
/// <param name="description">The accelerator description.</param>
internal CLAccelerator(Context context, CLDevice description)
: base(context, description)
{
Backends.Backend.EnsureRunningOnNativePlatform();
// Create new context
CLException.ThrowIfFailed(
CurrentAPI.CreateContext(DeviceId, out var contextPtr));
NativePtr = contextPtr;
Bind();
DefaultStream = CreateStreamInternal();
InitVendorFeatures();
InitSubGroupSupport(description);
Init(new CLBackend(Context, Capabilities, Vendor));
}
/// <summary>
/// Constructs a new OpenCL accelerator.
/// </summary>
/// <param name="context">The ILGPU context.</param>
/// <param name="description">The accelerator description.</param>
internal CLAccelerator(Context context, CLDevice description)
: base(context, description)
{
Backends.Backend.EnsureRunningOnNativePlatform();
if (!description.Capabilities.GenericAddressSpace)
{
throw CLCapabilityContext.GetNotSupportedGenericAddressSpaceException();
}
// Create new context
CLException.ThrowIfFailed(
CurrentAPI.CreateContext(DeviceId, out var contextPtr));
NativePtr = contextPtr;
Bind();
DefaultStream = CreateStreamInternal();
InitVendorFeatures();
InitSubGroupSupport(description);
Init(new CLBackend(Context, Capabilities, Vendor, CLStdVersion));
}
/// <summary>
/// Constructs a new OpenCL accelerator.
/// </summary>
/// <param name="context">The ILGPU context.</param>
/// <param name="acceleratorId">The accelerator id.</param>
public CLAccelerator(Context context, CLAcceleratorId acceleratorId)
: base(context, AcceleratorType.OpenCL)
{
if (acceleratorId == null)
{
throw new ArgumentNullException(nameof(acceleratorId));
}
PlatformId = acceleratorId.PlatformId;
DeviceId = acceleratorId.DeviceId;
CVersion = acceleratorId.CVersion;
PlatformName = CurrentAPI.GetPlatformInfo(
PlatformId,
CLPlatformInfoType.CL_PLATFORM_NAME);
VendorName = CurrentAPI.GetPlatformInfo(
PlatformId,
CLPlatformInfoType.CL_PLATFORM_VENDOR);
// Create new context
CLException.ThrowIfFailed(
CurrentAPI.CreateContext(DeviceId, out contextPtr));
// Resolve device info
Name = CurrentAPI.GetDeviceInfo(
DeviceId,
CLDeviceInfoType.CL_DEVICE_NAME);
MemorySize = CurrentAPI.GetDeviceInfo <long>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_GLOBAL_MEM_SIZE);
DeviceType = (CLDeviceType)CurrentAPI.GetDeviceInfo <long>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_TYPE);
// Max grid size
int workItemDimensions = IntrinsicMath.Max(CurrentAPI.GetDeviceInfo <int>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS), 3);
var workItemSizes = new IntPtr[workItemDimensions];
CurrentAPI.GetDeviceInfo(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_WORK_ITEM_SIZES,
workItemSizes);
MaxGridSize = new Index3(
workItemSizes[0].ToInt32(),
workItemSizes[1].ToInt32(),
workItemSizes[2].ToInt32());
// Resolve max threads per group
MaxNumThreadsPerGroup = CurrentAPI.GetDeviceInfo <IntPtr>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_WORK_GROUP_SIZE).ToInt32();
// Resolve max shared memory per block
MaxSharedMemoryPerGroup = (int)IntrinsicMath.Min(
CurrentAPI.GetDeviceInfo <long>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_LOCAL_MEM_SIZE),
int.MaxValue);
// Resolve total constant memory
MaxConstantMemory = (int)CurrentAPI.GetDeviceInfo <long>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_PARAMETER_SIZE);
// Resolve clock rate
ClockRate = CurrentAPI.GetDeviceInfo <int>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_CLOCK_FREQUENCY);
// Resolve number of multiprocessors
NumMultiprocessors = CurrentAPI.GetDeviceInfo <int>(
DeviceId,
CLDeviceInfoType.CL_DEVICE_MAX_COMPUTE_UNITS);
// Result max number of threads per multiprocessor
MaxNumThreadsPerMultiprocessor = MaxNumThreadsPerGroup;
InitVendorFeatures();
InitSubGroupSupport(acceleratorId);
Bind();
DefaultStream = CreateStreamInternal();
Init(new CLBackend(Context, Vendor));
}