/// <summary>
/// Resolves a PTX code generator for the given math-function configuration.
/// </summary>
/// <param name="minArchitecture">The target/minimum architecture.</param>
/// <returns>The resolved intrinsic representation.</returns>
private static PTXIntrinsic GetMathCodeGeneratorIntrinsic(
CudaArchitecture minArchitecture) =>
new PTXIntrinsic(
PTXMathType,
nameof(PTXMath.GenerateMathIntrinsic),
IntrinsicImplementationMode.GenerateCode,
minArchitecture);
/// <summary>
/// Constructs a new PTX intrinsic that can handle all architectures
/// newer or equal to <paramref name="minArchitecture"/>.
/// </summary>
/// <param name="handlerType">The associated target handler type.</param>
/// <param name="mode">The code-generation mode.</param>
/// <param name="minArchitecture">The target/minimum architecture.</param>
public PTXIntrinsic(
Type handlerType,
IntrinsicImplementationMode mode,
CudaArchitecture minArchitecture)
: this(handlerType, mode)
{
MinArchitecture = minArchitecture;
}
/// <summary>
/// Creates a new PTX intrinsic.
/// </summary>
/// <param name="name">The name of the intrinsic.</param>
/// <param name="mode">The implementation mode.</param>
/// <param name="minArchitecture">The minimum architecture.</param>
/// <param name="maxArchitecture">The maximum architecture.</param>
/// <returns>The created intrinsic.</returns>
private static PTXIntrinsic CreateIntrinsic(
string name,
IntrinsicImplementationMode mode,
CudaArchitecture?minArchitecture,
CudaArchitecture maxArchitecture) =>
new PTXIntrinsic(
PTXIntrinsicsType,
name,
mode,
minArchitecture,
maxArchitecture);
/// <summary>
/// Resolves the minimum CUDA driver version for the PTX architecture
/// </summary>
/// <param name="architecture">The PTX architecture</param>
/// <returns>The minimum driver version</returns>
public static CudaDriverVersion GetMinimumDriverVersion(
CudaArchitecture architecture)
{
if (ArchitectureLookup.TryGetValue(architecture, out var result))
{
return(result);
}
// If the architecture is unknown, return the highest driver version that
// we support. The user should already have a driver version higher than
// this, because they are most likely using a brand new graphics card.
return(ArchitectureLookup.OrderByDescending(x => x.Key).First().Value);
}
/// <summary>
/// Constructs a new PTX intrinsic.
/// </summary>
/// <param name="handlerType">The associated target handler type.</param>
/// <param name="methodName">The target method name (or null).</param>
/// <param name="mode">The code-generator mode.</param>
/// <param name="minArchitecture">The target/minimum architecture.</param>
public PTXIntrinsic(
Type handlerType,
string methodName,
IntrinsicImplementationMode mode,
CudaArchitecture minArchitecture)
: base(
BackendType.PTX,
handlerType,
methodName,
mode)
{
MinArchitecture = minArchitecture;
}
/// <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());
});
}
