/// <summary cref="LLVMBackend.CreateEntry(CompileUnit, EntryPoint, out string)"/>
internal override LLVMValueRef CreateEntry(CompileUnit unit, EntryPoint entryPoint, out string entryPointName)
{
if (!ptxDeviceFunctions.TryGetValue(unit, out PTXDeviceFunctions deviceFunctions))
{
throw new InvalidOperationException(ErrorMessages.NotSupportedCompileUnit);
}
entryPointName = unit.GetLLVMName(entryPoint.MethodInfo, CudaKernelCategory);
var context = unit.LLVMContext;
var module = unit.LLVMModule;
LLVMValueRef cudaEntryPoint = GetNamedFunction(module, entryPointName);
if (cudaEntryPoint.Pointer != IntPtr.Zero)
{
SetLinkage(cudaEntryPoint, LLVMLinkage.LLVMExternalLinkage);
return(cudaEntryPoint);
}
var entryPointType = CreatePTXKernelFunctionType(unit, entryPoint, out int parameterOffset);
cudaEntryPoint = AddFunction(module, entryPointName, entryPointType);
SetLinkage(cudaEntryPoint, LLVMLinkage.LLVMExternalLinkage);
var entryBlock = AppendBasicBlock(cudaEntryPoint, "Main");
var exitBlock = AppendBasicBlock(cudaEntryPoint, "Exit");
var builder = CreateBuilderInContext(unit.LLVMContext);
PositionBuilderAtEnd(builder, entryBlock);
// Create a proper entry point for the virtual entry point
var indexValue = CreateIndexValue(unit, entryPoint, builder, deviceFunctions);
var groupIndexValue = CreateGroupIndexValue(unit, entryPoint, builder, deviceFunctions);
if (!entryPoint.IsGroupedIndexEntry)
{
// We have to generate code for an implictly grouped kernel
// -> Compute the actual global idx
indexValue = CreateGlobalIndexValue(
unit,
entryPoint,
builder,
deviceFunctions,
indexValue,
groupIndexValue);
// Append a new main block that contains the actual body
var mainBlock = AppendBasicBlock(cudaEntryPoint, "Core");
// Emit the required check (custom dimension size is stored in parameter 0).
// This check is required to ensure that the index is always smaller than the
// specified user size. Otherwise, the index might be larger due to custom blocking!
Debug.Assert(parameterOffset > 0);
var rangeComparisonResult = CreateGlobalIndexRangeComparison(
unit,
entryPoint,
builder,
deviceFunctions,
indexValue,
GetParam(cudaEntryPoint, 0));
BuildCondBr(builder, rangeComparisonResult, mainBlock, exitBlock);
// Move builder to main block to emit the actual kernel body
PositionBuilderAtEnd(builder, mainBlock);
}
else
{
Debug.Assert(parameterOffset < 1);
indexValue = CreateGroupedIndex(
unit,
entryPoint,
builder,
deviceFunctions,
indexValue,
groupIndexValue);
}
// Call the virtual entry point
LLVMValueRef[] kernelValues = new LLVMValueRef[entryPoint.NumCustomParameters + 1];
kernelValues[0] = indexValue;
var kernelParameters = GetParams(cudaEntryPoint);
var uniformVariables = entryPoint.UniformVariables;
for (int i = 0, kernelParamIdx = parameterOffset, e = uniformVariables.Length; i < e; ++i, ++kernelParamIdx)
{
var variable = uniformVariables[i];
LLVMValueRef kernelParam;
var kernelValue = kernelParam = kernelParameters[kernelParamIdx];
if (variable.VariableType.IsPassedViaPtr())
{
// We have to generate a local alloca and store the current parameter value
kernelValue = BuildAlloca(builder, TypeOf(kernelParam), string.Empty);
BuildStore(builder, kernelParam, kernelValue);
}
kernelValues[variable.Index] = kernelValue;
}
var sharedMemoryVariables = entryPoint.SharedMemoryVariables;
foreach (var variable in sharedMemoryVariables)
{
// This type can be: ArrayType<T> or VariableType<T>
var variableType = unit.GetType(variable.Type);
var variableElementType = unit.GetType(variable.ElementType);
var sharedVariable = GetUndef(variableType);
if (variable.IsArray)
{
// However, ArrayType<T> encapsulates the type ArrayView<T, Index>
var genericArrayView = GetUndef(GetStructElementTypes(variableType)[0]);
var arrayType = ArrayType(variableElementType, variable.Count != null ? variable.Count.Value : 0);
var sharedMem = DeclareSharedMemoryVariable(unit, builder, arrayType);
genericArrayView = BuildInsertValue(builder, genericArrayView, sharedMem, 0, string.Empty);
LLVMValueRef intIndex;
if (variable.Count != null)
{
intIndex = ConstInt(context.Int32Type, variable.Count.Value, false);
}
else
{
// Attach the right length information that is given via a parameter
Debug.Assert(variable.SharedMemoryIndex >= 0);
intIndex = kernelParameters[uniformVariables.Length + variable.SharedMemoryIndex];
}
var indexInstance = GetUndef(unit.GetType(typeof(Index)));
indexInstance = BuildInsertValue(builder, indexInstance, intIndex, 0, string.Empty);
genericArrayView = BuildInsertValue(builder, genericArrayView, indexInstance, 1, string.Empty);
sharedVariable = BuildInsertValue(builder, sharedVariable, genericArrayView, 0, string.Empty);
}
else
{
var sharedMem = DeclareSharedMemoryVariable(unit, builder, variableElementType);
// Insert pointer into variable view
sharedVariable = BuildInsertValue(builder, sharedVariable, sharedMem, 0, string.Empty);
}
// Setup the pointer as generic pointer
kernelValues[variable.Index] = sharedVariable;
}
// Declare external entry point
var virtualEntryPoint = unit.GetMethod(entryPoint.MethodInfo);
BuildCall(builder, virtualEntryPoint.LLVMFunction, kernelValues);
// Verify method access in the scope of implicitly-grouped kernels
if (!entryPoint.IsGroupedIndexEntry)
{
virtualEntryPoint.VisitCalls((instruction, calledMethod) =>
{
CodeGenerator.VerifyAccessToMethodInImplicitlyGroupedKernel(
unit.CompilationContext,
calledMethod.MethodBase,
entryPoint);
});
}
// Jump to exit block
BuildBr(builder, exitBlock);
// Build exit block
PositionBuilderAtEnd(builder, exitBlock);
BuildRetVoid(builder);
unit.Optimize();
return(cudaEntryPoint);
}