可以将文章内容翻译成中文,广告屏蔽插件可能会导致该功能失效(如失效，请关闭广告屏蔽插件后再试):

问题:

I am using the following command line for clang:

clang -Dcl_clang_storage_class_specifiers -isystem $LIBCLC/generic/include -include clc/clc.h -target nvptx--nvidiacl -x cl some_kernel.cl -emit-llvm -S -o some_kernel.ll

the result is:

; ModuleID = 'kernel.cl' target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64" target triple = "nvptx--nvidiacl"  ; Function Attrs: noinline nounwind define void @vector_add(float addrspace(1)* nocapture %vec1, float addrspace(1)* nocapture %vec2, float addrspace(1)* nocapture %vec3) #0 {   %1 = tail call i32 @get_global_id(i32 0) #2   %2 = getelementptr inbounds float addrspace(1)* %vec1, i32 1   %3 = load float addrspace(1)* %2, align 4, !tbaa !2   %4 = getelementptr inbounds float addrspace(1)* %vec2, i32 2   %5 = load float addrspace(1)* %4, align 4, !tbaa !2   %6 = fadd float %3, %5   %7 = fmul float %6, 5.000000e+00   %8 = getelementptr inbounds float addrspace(1)* %vec3, i32 %1   store float %7, float addrspace(1)* %8, align 4, !tbaa !2   ret void }  declare i32 @get_global_id(i32) #1  attributes #0 = { noinline nounwind "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" } attributes #1 = { "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" } attributes #2 = { nounwind }  !opencl.kernels = !{!0} !nvvm.annotations = !{!1}  !0 = metadata !{void (float addrspace(1)*, float addrspace(1)*, float addrspace(1)*)* @vector_add} !1 = metadata !{void (float addrspace(1)*, float addrspace(1)*, float addrspace(1)*)* @vector_add, metadata !"kernel", i32 1} !2 = metadata !{metadata !"float", metadata !3} !3 = metadata !{metadata !"omnipotent char", metadata !4} !4 = metadata !{metadata !"Simple C/C++ TBAA"}

As you may see in the parameter list it generates a nocapture ABI.

But when I try to do this with this code:

/*Create a diagnostics*/ mCompilerInst.createDiagnostics();   /*Creating NVIDIA-CL*/ clang::TargetOptions  TO; TO.Triple = "nvptx--nvidiacl"; TargetInfo* feTarget =     clang::TargetInfo::CreateTargetInfo(mCompilerInst.getDiagnostics(), &TO); mCompilerInst.setTarget(feTarget);  /*Commandline Parameters*/ std::vector<const char*> args; args.push_back("-Dcl_clang_storage_class_specifiers"); args.push_back("-include"); args.push_back("clc/clc.h"); args.push_back("-xcl"); args.push_back("-S"); /* Create the compiler invocation */ clang::CompilerInvocation::CreateFromArgs(mCompilerInst.getInvocation(), &args[0], &args[0] + args.size(), mCompilerInst.getDiagnostics());  CodeGenerator* llvmCodeGen = clang::CreateLLVMCodeGen(     mCompilerInst.getDiagnostics(), "kernel.cl", mCompilerInst.getCodeGenOpts(),mCompilerInst.getTargetOpts(),ctx); ParseAST(mCompilerInst.getPreprocessor(),llvmCodeGen,mCompilerInst.getASTContext()); mCompilerInst.getDiagnosticClient().EndSourceFile();   if(mCompilerInst.getDiagnostics().hasErrorOccurred()) {     llvm::outs() << "CAN'T COMPILE THE CODE"; }  /*GOING FOR THE LLVM BYTE_CODE*/ llvm::Module* module = llvmCodeGen->GetModule();  std::ofstream myfile; myfile.open ("result",(std::ios::binary|std::ios::out)); llvm::AssemblyAnnotationWriter writer; llvm::raw_os_ostream raw_stream(myfile); module->print(raw_stream, &writer); myfile.flush(); myfile.close();

what is being outputted is:

; ModuleID = 'kernel.cl' target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64" target triple = "nvptx--nvidiacl"  ; Function Attrs: noinline nounwind define void @mat_ex(float addrspace(1)* %vec1, float addrspace(1)* %vec2, float addrspace(1)* %vec3) #0 {   %1 = alloca float addrspace(1)*, align 4   %2 = alloca float addrspace(1)*, align 4   %3 = alloca float addrspace(1)*, align 4   %den = alloca i32, align 4   %offset = alloca i32, align 4   %i = alloca i32, align 4   store float addrspace(1)* %vec1, float addrspace(1)** %1, align 4, !tbaa !2   store float addrspace(1)* %vec2, float addrspace(1)** %2, align 4, !tbaa !2   store float addrspace(1)* %vec3, float addrspace(1)** %3, align 4, !tbaa !2   %4 = call i32 @get_global_id(i32 0)   store i32 %4, i32* %i, align 4, !tbaa !5   store i32 4, i32* %den, align 4, !tbaa !5   %5 = load i32* %den, align 4, !tbaa !5   %6 = sdiv i32 2, %5   %7 = mul nsw i32 %6, 4   %8 = add nsw i32 5, %7   store i32 %8, i32* %offset, align 4, !tbaa !5   %9 = load i32* %i, align 4, !tbaa !5   %10 = load float addrspace(1)** %1, align 4, !tbaa !2   %11 = getelementptr inbounds float addrspace(1)* %10, i32 %9   %12 = load float addrspace(1)* %11, align 4, !tbaa !6   %13 = load i32* %i, align 4, !tbaa !5   %14 = load float addrspace(1)** %2, align 4, !tbaa !2   %15 = getelementptr inbounds float addrspace(1)* %14, i32 %13   %16 = load float addrspace(1)* %15, align 4, !tbaa !6   %17 = fadd float %12, %16   %18 = load i32* %offset, align 4, !tbaa !5   %19 = sitofp i32 %18 to float   %20 = fmul float %17, %19   %21 = load i32* %i, align 4, !tbaa !5   %22 = load float addrspace(1)** %3, align 4, !tbaa !2   %23 = getelementptr inbounds float addrspace(1)* %22, i32 %21   store float %20, float addrspace(1)* %23, align 4, !tbaa !6   ret void }  declare i32 @get_global_id(i32) #1  attributes #0 = { noinline nounwind "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-frame-pointer-elim-non-leaf"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" } attributes #1 = { "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-frame-pointer-elim-non-leaf"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "unsafe-fp-math"="false" "use-soft-float"="false" }  !opencl.kernels = !{!0} !nvvm.annotations = !{!1}  !0 = metadata !{void (float addrspace(1)*, float addrspace(1)*, float addrspace(1)*)* @mat_ex} !1 = metadata !{void (float addrspace(1)*, float addrspace(1)*, float addrspace(1)*)* @mat_ex, metadata !"kernel", i32 1} !2 = metadata !{metadata !"any pointer", metadata !3} !3 = metadata !{metadata !"omnipotent char", metadata !4} !4 = metadata !{metadata !"Simple C/C++ TBAA"} !5 = metadata !{metadata !"int", metadata !3} !6 = metadata !{metadata !"float", metadata !3}

basically it mismatches, and I can't figure out why. Why are there are no "nocaptures"? I think the extra code is generated due to missing "nocapture".

回答1:

The command line

clang ... -emit-llvm ...

runs the Clang driver, which first runs the Clang front-end to generate LLVM IR, then runs LLVM to process the IR, then emits the processed IR.

The code

ParseAST(mCompilerInst.getPreprocessor(),llvmCodeGen,mCompilerInst.getASTContext());

on the other hand, only parses the AST into the IR file and then nothing further is done with it. You haven't invoked LLVM at all, only the Clang front-end, and that's why you're seeing different code. It actually has the same semantic* - it's just unoptimized.

To solve this, you need to actually run some passes (or better yet, use a pass manager) on your module. You can take a look at chapter 4 of the kaleidoscope tutorial for help on this.

* Actually the two modules aren't equivalent. The 1st code is equivalent to:

kernel void vector_add(global float* vec1, global float* vec2, global float* vec3) {     int i = get_global_id(0);     vec3[i] = (vec1[1] + vec2[2]) * 5f; }

While the 2nd is to:

kernel void vector_add(global float* vec1, global float* vec2, global float* vec3) {     int i = get_global_id(0);     vec3[i] = (vec1[i] + vec2[i]) * 5f; }

Notice the different indices used (i vs 1 and 2) - I'm guessing the 2nd version is what you want. If it's not a simple copy-paste error, I suggest you re-check your flow to verify you are working on the correct file.

文章来源: OpenCL LLVM IR generation from Clang

标签

clang

llvm

opencl

kernel