private static async Task <LLVMModuleRef> CompileAsync(ClrAssembly assembly)
{
var desc = await CreateContentDescriptionAsync(assembly);
return(LlvmBackend.Compile(desc, assembly.Resolver.TypeEnvironment).Module);
}
private static async Task <CudaModule> CompileAsync(
IMethod method,
IEnumerable <ITypeMember> memberRoots,
IEnumerable <IType> typeRoots,
int threadIdParamIndex,
ClrAssembly assembly,
CudaContext context)
{
// Figure out which members we need to compile.
var desc = await CreateContentDescriptionAsync(method, memberRoots, typeRoots, assembly);
// Compile those members to LLVM IR. Use an Itanium name mangling scheme.
var mangler = new ItaniumMangler(assembly.Resolver.TypeEnvironment);
var moduleBuilder = LlvmBackend.Compile(desc, assembly.Resolver.TypeEnvironment);
var module = moduleBuilder.Module;
// Generate type metadata for all type roots.
foreach (var type in typeRoots)
{
moduleBuilder.Metadata.GetMetadata(type, moduleBuilder);
}
// Get the compiled kernel function.
var kernelFuncName = mangler.Mangle(method, true);
var kernelFunc = LLVM.GetNamedFunction(module, kernelFuncName);
if (threadIdParamIndex >= 0)
{
// If we have a thread ID parameter, then we need to generate a thunk
// kernel function that calls our actual kernel function. This thunk's
// responsibility is to determine the thread ID of the kernel.
var thunkKernelName = "kernel";
var thunkTargetType = kernelFunc.TypeOf().GetElementType();
var thunkParamTypes = new List <LLVMTypeRef>(thunkTargetType.GetParamTypes());
if (threadIdParamIndex < thunkParamTypes.Count)
{
thunkParamTypes.RemoveAt(threadIdParamIndex);
}
var thunkKernel = LLVM.AddFunction(
module,
thunkKernelName,
LLVM.FunctionType(
thunkTargetType.GetReturnType(),
thunkParamTypes.ToArray(),
thunkTargetType.IsFunctionVarArg));
using (var builder = new IRBuilder(moduleBuilder.Context))
{
builder.PositionBuilderAtEnd(thunkKernel.AppendBasicBlock("entry"));
var args = new List <LLVMValueRef>(thunkKernel.GetParams());
args.Insert(threadIdParamIndex, ComputeUniqueThreadId(builder, module));
var call = builder.CreateCall(kernelFunc, args.ToArray(), "");
if (call.TypeOf().TypeKind == LLVMTypeKind.LLVMVoidTypeKind)
{
builder.CreateRetVoid();
}
else
{
builder.CreateRet(call);
}
}
kernelFuncName = thunkKernelName;
kernelFunc = thunkKernel;
}
// Mark the compiled kernel as a kernel symbol.
LLVM.AddNamedMetadataOperand(
module,
"nvvm.annotations",
LLVM.MDNode(new LLVMValueRef[]
{
kernelFunc,
MDString("kernel"),
LLVM.ConstInt(LLVM.Int32TypeInContext(LLVM.GetModuleContext(module)), 1, false)
}));
// LLVM.DumpModule(module);
// Compile that LLVM IR down to PTX.
LLVMTargetMachineRef machine;
var ptx = CompileToPtx(module, context.GetDeviceComputeCapability(), out machine);
// Console.WriteLine(System.Text.Encoding.UTF8.GetString(ptx));
// Load the PTX kernel.
return(new CudaModule(assembly, moduleBuilder, machine, context.LoadModulePTX(ptx), kernelFuncName, context));
}