/// <summary>
/// Creates a signature for the actual kernel entry point.
/// </summary>
/// <param name="unit">The target unit.</param>
/// <param name="entryPoint">The target entry point.</param>
/// <param name="parameterOffset">The parameter offset for the actual kernel parameters.</param>
/// <returns>A signature for the actual kernel entry point.</returns>
private LLVMTypeRef CreatePTXKernelFunctionType(
CompileUnit unit,
EntryPoint entryPoint,
out int parameterOffset)
{
parameterOffset = entryPoint.IsGroupedIndexEntry ? 0 : 1;
var numUniformVariables = entryPoint.NumUniformVariables;
var argTypes = new LLVMTypeRef[parameterOffset + numUniformVariables + entryPoint.NumDynamicallySizedSharedMemoryVariables];
// Custom dispatch-size information for implicitly grouped kernels
if (parameterOffset > 0)
{
argTypes[0] = unit.GetType(entryPoint.UngroupedIndexType);
}
Debug.Assert(parameterOffset >= 0 && parameterOffset < 2);
for (int i = 0, e = numUniformVariables; i < e; ++i)
{
argTypes[i + parameterOffset] = unit.GetType(entryPoint.UniformVariables[i].VariableType);
}
// Attach length information to dynamically sized variables using runtime information
for (int i = 0, e = entryPoint.NumDynamicallySizedSharedMemoryVariables; i < e; ++i)
{
argTypes[i + parameterOffset + numUniformVariables] = unit.GetType(typeof(int));
}
return(FunctionType(Context.LLVMContext.VoidType, argTypes));
}
/// <summary>
/// Builds a warp-shuffle mask.
/// </summary>
/// <param name="unit">The current unit.</param>
/// <param name="builder">The current builder.</param>
/// <param name="width">The width that was passed by the user.</param>
/// <param name="addOrMask">True, to add an or mask consisting of (WarpSize - 1).</param>
/// <returns>A value that represents the desired warp-shuffle mask.</returns>
private LLVMValueRef BuildWarpShuffleMask(
CompileUnit unit,
LLVMBuilderRef builder,
LLVMValueRef width,
bool addOrMask)
{
var warpSize = MakeWarpSize(builder);
var warpDiff = BuildSub(builder, warpSize, width, string.Empty);
var result = BuildShl(
builder,
warpDiff,
ConstInt(unit.GetType(BasicValueType.Int32), 8, false),
string.Empty);
if (addOrMask)
{
var orMask = BuildSub(
builder,
warpSize,
ConstInt(unit.GetType(BasicValueType.Int32), 1, false),
string.Empty);
result = BuildOr(
builder,
result,
orMask,
string.Empty);
}
return(result);
}
/// <summary>
/// Constructs a default LLVM-based ABI specification.
/// </summary>
/// <param name="unit">The compile unit used for ABI generation.</param>
public DefaultLLVMABI(CompileUnit unit)
: base(unit)
{
var backend = unit.Backend as LLVMBackend;
if (backend == null)
{
throw new NotSupportedException(ErrorMessages.NotSupportedBackend);
}
LLVMTargetData = CreateTargetDataLayout(backend.LLVMTargetMachine);
foreach (var managedAlignment in ManagedAlignments)
{
var managedType = managedAlignment.Key;
var llvmType = unit.GetType(managedType);
var alignment = ABIAlignmentOfType(LLVMTargetData, llvmType);
// We need a special case for the builtin mapping of 64bit floats
// to 32bit floats since this mapping changes the alignment logic.
if (unit.Force32BitFloats && managedType == typeof(double))
{
managedType = typeof(float);
}
if (ManagedAlignments[managedType] != alignment)
{
throw new NotSupportedException(string.Format(
ErrorMessages.CustomABIImplementationRequired, managedAlignment.Key));
}
Alignments.Add(managedAlignment.Key, alignment);
}
AddNonBlittableTypes();
AddPtrAlignment(ABIAlignmentOfType(
LLVMTargetData,
unit.LLVMContext.VoidPtrType));
}
/// <summary>
/// Creates an <see cref="Index3"/> in the LLVM world containing the current group-thread indices.
/// </summary>
/// <param name="unit">The target unit.</param>
/// <param name="entryPoint">The entry point.</param>
/// <param name="builder">The LLVM builder.</param>
/// <param name="cudaDeviceFunctions">A reference to the cuda device functions.</param>
/// <returns>An <see cref="Index3"/> in the LLVM world containg the current group-thread indices.</returns>
private static LLVMValueRef CreateGroupIndexValue(
CompileUnit unit,
EntryPoint entryPoint,
LLVMBuilderRef builder,
PTXDeviceFunctions cudaDeviceFunctions)
{
var indexType = unit.GetType(entryPoint.UngroupedIndexType);
var threadIndexValue = GetUndef(indexType);
Debug.Assert(entryPoint.Type >= IndexType.Index1D);
var isGroupedIndex = entryPoint.IsGroupedIndexEntry;
threadIndexValue = BuildInsertValue(builder, threadIndexValue, BuildCall(
builder, cudaDeviceFunctions.GetThreadIdxX.Value), 0, "TIdx1");
if (entryPoint.Type >= IndexType.Index2D && !isGroupedIndex || entryPoint.Type >= IndexType.GroupedIndex2D)
{
threadIndexValue = BuildInsertValue(builder, threadIndexValue, BuildCall(
builder, cudaDeviceFunctions.GetThreadIdxY.Value), 1, "TIdx2");
}
if (entryPoint.Type >= IndexType.Index3D && !isGroupedIndex || entryPoint.Type >= IndexType.GroupedIndex3D)
{
threadIndexValue = BuildInsertValue(builder, threadIndexValue, BuildCall(
builder, cudaDeviceFunctions.GetThreadIdxZ.Value), 2, "TIdx3");
}
return(threadIndexValue);
}
/// <summary>
/// Creates an <see cref="Index3"/> in the LLVM world containing the current grid indices.
/// </summary>
/// <param name="unit">The target unit.</param>
/// <param name="entryPoint">The entry point.</param>
/// <param name="builder">The LLVM builder.</param>
/// <param name="cudaDeviceFunctions">A reference to the cuda device functions.</param>
/// <returns>An <see cref="Index3"/> in the LLVM world containg the current grid indices.</returns>
private static LLVMValueRef CreateIndexValue(
CompileUnit unit,
EntryPoint entryPoint,
LLVMBuilderRef builder,
PTXDeviceFunctions cudaDeviceFunctions)
{
var indexType = unit.GetType(entryPoint.UngroupedIndexType);
var indexValue = GetUndef(indexType);
Debug.Assert(entryPoint.Type >= IndexType.Index1D);
indexValue = BuildInsertValue(builder, indexValue, BuildCall(
builder, cudaDeviceFunctions.GetBlockIdxX.Value), 0, "Idx1");
if (entryPoint.Type >= IndexType.Index2D && entryPoint.Type <= IndexType.Index3D ||
entryPoint.Type >= IndexType.GroupedIndex2D)
{
indexValue = BuildInsertValue(builder, indexValue, BuildCall(
builder, cudaDeviceFunctions.GetBlockIdxY.Value), 1, "Idx2");
}
if (entryPoint.Type == IndexType.Index3D || entryPoint.Type == IndexType.GroupedIndex3D)
{
indexValue = BuildInsertValue(builder, indexValue, BuildCall(
builder, cudaDeviceFunctions.GetBlockIdxZ.Value), 2, "Idx3");
}
return(indexValue);
}
public void Run()
{
CompileUnit cu = _context.CompileUnit;
new XmlSerializer(cu.GetType()).Serialize(Console.Out, cu);
Console.WriteLine();
}
override public void Run()
{
CompileUnit cu = Context.CompileUnit;
new XmlSerializer(cu.GetType()).Serialize(OutputWriter, cu);
Console.WriteLine();
}
/// <summary>
/// Creates an <see cref="Index3"/> in the LLVM world containing the current global indices
/// (gridIdx * blockDim + blockIdx).
/// </summary>
/// <param name="unit">The target unit.</param>
/// <param name="entryPoint">The entry point.</param>
/// <param name="builder">The LLVM builder.</param>
/// <param name="cudaDeviceFunctions">A reference to the cuda device functions.</param>
/// <param name="indexValue">The current grid-index value (gridIdx).</param>
/// <param name="groupIndexValue">The current group-thread-index value (blockIdx).</param>
/// <returns>An <see cref="Index3"/> in the LLVM world containg the current global indices.</returns>
private static LLVMValueRef CreateGlobalIndexValue(
CompileUnit unit,
EntryPoint entryPoint,
LLVMBuilderRef builder,
PTXDeviceFunctions cudaDeviceFunctions,
LLVMValueRef indexValue,
LLVMValueRef groupIndexValue)
{
var indexType = unit.GetType(entryPoint.UngroupedIndexType);
var globalIndexValue = GetUndef(indexType);
Debug.Assert(entryPoint.Type >= IndexType.Index1D && entryPoint.Type < IndexType.GroupedIndex1D);
var blockDimensions = cudaDeviceFunctions.GetBlockDimensions;
for (int i = 0, e = (int)entryPoint.Type; i < e; ++i)
{
var globalGroupOffset = BuildMul(
builder,
BuildExtractValue(
builder,
indexValue,
i,
"GridIdx_" + i),
BuildCall(
builder,
blockDimensions[i].Value),
"GlobalGroupOffset_" + i);
var globalIdx = BuildAdd(
builder,
globalGroupOffset,
BuildExtractValue(
builder,
groupIndexValue,
i,
"GroupIdx_" + i),
"GlobalIdxVal_" + i);
globalIndexValue = BuildInsertValue(
builder,
globalIndexValue,
globalIdx,
i,
"GlobalIdx_" + i);
}
return(globalIndexValue);
}
/// <summary>
/// Creates an instance of an <see cref="IGroupedIndex{TIndex}"/> in the LLVM world.
/// </summary>
/// <param name="unit">The target unit.</param>
/// <param name="entryPoint">The entry point.</param>
/// <param name="builder">The LLVM builder.</param>
/// <param name="cudaDeviceFunctions">A reference to the cuda device functions.</param>
/// <param name="indexValue">The current index values (first part of a grouped index).</param>
/// <param name="groupIndexValue">The current group-index values (second part of a grouped index).</param>
/// <returns>An instance of an <see cref="IGroupedIndex{TIndex}"/> in the LLVM world.</returns>
private static LLVMValueRef CreateGroupedIndex(
CompileUnit unit,
EntryPoint entryPoint,
LLVMBuilderRef builder,
PTXDeviceFunctions cudaDeviceFunctions,
LLVMValueRef indexValue,
LLVMValueRef groupIndexValue)
{
Debug.Assert(entryPoint.Type >= IndexType.GroupedIndex1D);
// Create a new blocked index
var blockIndexValue = GetUndef(unit.GetType(entryPoint.KernelIndexType));
blockIndexValue = BuildInsertValue(builder, blockIndexValue, indexValue, 0, "GridIdx");
blockIndexValue = BuildInsertValue(builder, blockIndexValue, groupIndexValue, 1, "GroupIdx");
return(blockIndexValue);
}
/// <summary cref="ABISpecification.GetSizeOf(Type)"/>
public override int GetSizeOf(Type type)
{
return(GetSizeOf(CompileUnit.GetType(type)));
}
/// <summary cref="ABISpecification.GetAlignmentOf(Type)"/>
public override int GetAlignmentOf(Type type)
{
return(ABIAlignmentOfType(
LLVMTargetData,
CompileUnit.GetType(type)));
}
/// <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);
}
