/// <summary>
/// Constructs a new Cuda accelerator reference.
/// </summary>
/// <param name="deviceId">The Cuda device id.</param>
internal CudaDevice(int deviceId)
{
if (deviceId < 0)
{
throw new ArgumentOutOfRangeException(nameof(deviceId));
}
DeviceId = deviceId;
InitDeviceInfo();
InitArchitectureInfo();
InitGridInfo();
InitMemoryInfo();
InitMiscInfo();
InitPCIInfo();
Capabilities = new CudaCapabilityContext(
Architecture ?? CudaArchitecture.SM_30);
}
/// <summary>
/// Constructs a new Cuda accelerator reference.
/// </summary>
/// <param name="deviceId">The Cuda device id.</param>
internal CudaDevice(int deviceId)
{
if (deviceId < 0)
{
throw new ArgumentOutOfRangeException(nameof(deviceId));
}
Backend.EnsureRunningOnPlatform(TargetPlatform.X64);
DeviceId = deviceId;
InitDeviceInfo();
InitArchitectureInfo();
InitGridInfo();
InitMemoryInfo();
InitMiscInfo();
InitPCIInfo();
Capabilities = new CudaCapabilityContext(
Architecture ?? CudaArchitecture.SM_30);
}
/// <summary>
/// Constructs a new Cuda backend.
/// </summary>
/// <param name="context">The context to use.</param>
/// <param name="capabilities">The supported capabilities.</param>
/// <param name="architecture">The target GPU architecture.</param>
/// <param name="instructionSet">The target GPU instruction set.</param>
public PTXBackend(
Context context,
CudaCapabilityContext capabilities,
PTXArchitecture architecture,
PTXInstructionSet instructionSet)
: base(
context,
capabilities,
BackendType.PTX,
BackendFlags.None,
new PTXArgumentMapper(context))
{
Architecture = architecture;
InstructionSet = instructionSet;
InitIntrinsicProvider();
InitializeKernelTransformers(
Context.HasFlags(ContextFlags.EnableAssertions) ?
IntrinsicSpecializerFlags.EnableAssertions :
IntrinsicSpecializerFlags.None,
builder =>
{
var transformerBuilder = Transformer.CreateBuilder(
TransformerConfiguration.Empty);
transformerBuilder.AddBackendOptimizations(
new PTXAcceleratorSpecializer(PointerType),
context.Flags,
context.OptimizationLevel);
if (Context.HasFlags(ContextFlags.EnhancedPTXBackendFeatures))
{
// Create an optimized PTX assembler block schedule
transformerBuilder.Add(new PTXBlockScheduling());
transformerBuilder.Add(new DeadCodeElimination());
}
builder.Add(transformerBuilder.ToTransformer());
});
}
/// <summary>
/// Constructs a new Cuda backend.
/// </summary>
/// <param name="context">The context to use.</param>
/// <param name="capabilities">The supported capabilities.</param>
/// <param name="architecture">The target GPU architecture.</param>
/// <param name="instructionSet">The target GPU instruction set.</param>
/// <param name="nvvmAPI">Optional NVVM API instance.</param>
public PTXBackend(
Context context,
CudaCapabilityContext capabilities,
CudaArchitecture architecture,
CudaInstructionSet instructionSet,
NvvmAPI nvvmAPI)
: base(
context,
capabilities,
BackendType.PTX,
new PTXArgumentMapper(context))
{
Architecture = architecture;
InstructionSet = instructionSet;
NvvmAPI = nvvmAPI;
InitIntrinsicProvider();
InitializeKernelTransformers(builder =>
{
var transformerBuilder = Transformer.CreateBuilder(
TransformerConfiguration.Empty);
transformerBuilder.AddBackendOptimizations <CodePlacement.GroupOperands>(
new PTXAcceleratorSpecializer(
PointerType,
Context.Properties.EnableAssertions,
Context.Properties.EnableIOOperations),
context.Properties.InliningMode,
context.Properties.OptimizationLevel);
if (Context.Properties.GetPTXBackendMode() == PTXBackendMode.Enhanced)
{
// Create an optimized PTX assembler block schedule
transformerBuilder.Add(new PTXBlockScheduling());
transformerBuilder.Add(new DeadCodeElimination());
}
builder.Add(transformerBuilder.ToTransformer());
});
}
