/// <summary>
/// Constructs a new Cuda backend.
/// </summary>
/// <param name="context">The context to use.</param>
/// <param name="architecture">The target GPU architecture.</param>
/// <param name="instructionSet">The target GPU instruction set.</param>
public PTXBackend(
Context context,
PTXArchitecture architecture,
PTXInstructionSet instructionSet)
: base(
context,
BackendType.PTX,
BackendFlags.None,
new PTXArgumentMapper(context))
{
Architecture = architecture;
InstructionSet = instructionSet;
InitializeKernelTransformers(
Context.HasFlags(ContextFlags.EnableAssertions) ?
IntrinsicSpecializerFlags.EnableAssertions :
IntrinsicSpecializerFlags.None,
builder =>
{
var transformerBuilder = Transformer.CreateBuilder(
TransformerConfiguration.Empty);
transformerBuilder.AddBackendOptimizations(
new PTXAcceleratorSpecializer(),
context.OptimizationLevel);
builder.Add(transformerBuilder.ToTransformer());
});
}
/// <summary>
/// Resolves the minimum CUDA driver version for the PTX instruction set
/// </summary>
/// <param name="instructionSet">The PTX instruction set</param>
/// <returns>The minimum driver version</returns>
public static CudaDriverVersion GetMinimumDriverVersion(PTXInstructionSet instructionSet)
{
if (!InstructionSetLookup.TryGetValue(instructionSet, out var result))
{
throw new NotSupportedException(RuntimeErrorMessages.NotSupportedPTXInstructionSet);
}
return(result);
}
/// <summary>
/// Constructs a new Cuda backend.
/// </summary>
/// <param name="context">The context to use.</param>
/// <param name="architecture">The target gpu architecture.</param>
/// <param name="instructionSet">The target gpu instruction set.</param>
/// <param name="platform">The target platform.</param>
public PTXBackend(
Context context,
PTXArchitecture architecture,
PTXInstructionSet instructionSet,
TargetPlatform platform)
: base(
context,
BackendType.PTX,
BackendFlags.RequiresIntrinsicImplementations,
new PTXABI(context.TypeContext, platform),
_ => new PTXArgumentMapper(context))
{
Architecture = architecture;
InstructionSet = instructionSet;
InitializeKernelTransformers(
new IntrinsicSpecializerConfiguration(context.Flags),
builder =>
{
// Append further backend specific transformations in release mode
var transformerBuilder = Transformer.CreateBuilder(TransformerConfiguration.Empty);
if (context.OptimizationLevel == OptimizationLevel.Release)
{
var acceleratorConfiguration = new AcceleratorSpecializerConfiguration(ABI);
transformerBuilder.Add(
new AcceleratorSpecializer <AcceleratorSpecializerConfiguration>(
acceleratorConfiguration));
}
if (!context.HasFlags(ContextFlags.NoInlining))
{
transformerBuilder.Add(new Inliner());
}
if (context.OptimizationLevel == OptimizationLevel.Release)
{
transformerBuilder.Add(new SimplifyControlFlow());
}
var transformer = transformerBuilder.ToTransformer();
if (transformer.Length > 0)
{
builder.Add(transformer);
}
});
}
/// <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="architecture">The target GPU architecture.</param>
/// <param name="instructionSet">The target GPU instruction set.</param>
public PTXBackend(
Context context,
PTXArchitecture architecture,
PTXInstructionSet instructionSet)
: base(
context,
BackendType.PTX,
BackendFlags.None,
new PTXArgumentMapper(context))
{
Architecture = architecture;
InstructionSet = instructionSet;
InitializeKernelTransformers(
Context.HasFlags(ContextFlags.EnableAssertions) ?
IntrinsicSpecializerFlags.EnableAssertions :
IntrinsicSpecializerFlags.None,
builder =>
{
var transformerBuilder = Transformer.CreateBuilder(
TransformerConfiguration.Empty);
transformerBuilder.AddBackendOptimizations(
new PTXAcceleratorSpecializer(),
context.OptimizationLevel);
// Append further backend specific transformations in release mode
if (!context.HasFlags(ContextFlags.NoInlining))
{
transformerBuilder.Add(new Inliner());
}
if (context.OptimizationLevel > OptimizationLevel.O0)
{
transformerBuilder.Add(new SimplifyControlFlow());
}
builder.Add(transformerBuilder.ToTransformer());
});
}
