[Clang][CUDA][HIP] Externalize static global texture var

[minstrel223]

Externalize static global texture variable for CUDA/HIP.
Reason:
CUDA/HIP runtime needs reference the texture symbol in device elf when program is running. If a texture var has internal linkage type a runtime error will occur when running. To sovle this problem, CUDA nvcc externalizes static global texture var. But clang doesn't do it right now.

[github-actions]

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-codegen

Author: None (guopsh-sugon)

Changes

Externalize static global texture variable in CUDA/HIP.
Reason:
CUDA/HIP runtime needs reference the texture symbol in device elf when program is running. If a texture var has internal linkage type a runtime error will occur when running. To sovle this problem, CUDA nvcc externalizes static global texture var. But clang doesn't do it right now.

---

Full diff: https://github.com/llvm/llvm-project/pull/115819.diff

3 Files Affected:

• (modified) clang/lib/CodeGen/CGCUDANV.cpp (+14)
• (modified) clang/lib/CodeGen/CodeGenModule.cpp (+15)
• (added) clang/test/CodeGenCUDA/static-global-texture-var.cu (+33)

diff --git a/clang/lib/CodeGen/CGCUDANV.cpp b/clang/lib/CodeGen/CGCUDANV.cpp index ae14d74f2d9151..333b06a80e0cc7 100644 --- a/clang/lib/CodeGen/CGCUDANV.cpp +++ b/clang/lib/CodeGen/CGCUDANV.cpp @@ -306,6 +306,20 @@ std::string CGNVCUDARuntime::getDeviceSideName(const NamedDecl *ND) { CGM.printPostfixForExternalizedDecl(Out, ND); DeviceSideName = std::string(Out.str()); } + + // Make unique name for static global tetxure variable for HIP/CUDA. + if (const VarDecl *VD = dyn_cast<VarDecl>(ND)) { + if (VD->getType()->isCUDADeviceBuiltinTextureType() && + VD->getStorageClass() == SC_Static && VD->hasGlobalStorage() && + !VD->isStaticDataMember()) { + SmallString<256> Buffer; + llvm::raw_svector_ostream Out(Buffer); + Out << DeviceSideName; + CGM.printPostfixForExternalizedDecl(Out, ND); + DeviceSideName = std::string(Out.str()); + } + } + return DeviceSideName; } diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp index ba376f9ecfacde..859f707741e23e 100644 --- a/clang/lib/CodeGen/CodeGenModule.cpp +++ b/clang/lib/CodeGen/CodeGenModule.cpp @@ -1952,6 +1952,15 @@ static std::string getMangledNameImpl(CodeGenModule &CGM, GlobalDecl GD, CGM.getLangOpts().CUDAIsDevice) CGM.printPostfixForExternalizedDecl(Out, ND); + // Make unique name for static global tetxure variable for HIP/CUDA. + if (const VarDecl *VD = dyn_cast<VarDecl>(ND)) { + if (VD->getType()->isCUDADeviceBuiltinTextureType() && + VD->getStorageClass() == SC_Static && VD->hasGlobalStorage() && + !VD->isStaticDataMember()) { + CGM.printPostfixForExternalizedDecl(Out, ND); + } + } + return std::string(Out.str()); } @@ -5608,6 +5617,12 @@ void CodeGenModule::EmitGlobalVarDefinition(const VarDecl *D, // Set the llvm linkage type as appropriate. llvm::GlobalValue::LinkageTypes Linkage = getLLVMLinkageVarDefinition(D); + // Make static global texture variable externally visible. + if (D->getType()->isCUDADeviceBuiltinTextureType() && + D->getStorageClass() == SC_Static && !D->isStaticDataMember()) { + Linkage = llvm::GlobalValue::ExternalLinkage; + } + // CUDA B.2.1 "The __device__ qualifier declares a variable that resides on // the device. [...]" // CUDA B.2.2 "The __constant__ qualifier, optionally used together with diff --git a/clang/test/CodeGenCUDA/static-global-texture-var.cu b/clang/test/CodeGenCUDA/static-global-texture-var.cu new file mode 100644 index 00000000000000..86b5fa8d68548b --- /dev/null +++ b/clang/test/CodeGenCUDA/static-global-texture-var.cu @@ -0,0 +1,33 @@ +// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fcuda-is-device -std=c++11 \ +// RUN: -fno-threadsafe-statics -emit-llvm -o - %s | FileCheck -check-prefixes=CUDA-DEVICE %s + +// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -x hip -fcuda-is-device -std=c++11 \ +// RUN: -fno-threadsafe-statics -emit-llvm -o - %s | FileCheck -check-prefixes=HIP-DEVICE %s + +struct textureReference { + int desc; +}; + +enum ReadMode { + ElementType = 0, + NormalizedFloat = 1 +}; + +template <typename T, int dim = 1, enum ReadMode mode = ElementType> +struct __attribute__((device_builtin_texture_type)) texture : public textureReference { +}; + +// Confirm static global texture is externally visible and has a unique name.  +static texture<float, 2, ElementType> texRef; +//CUDA-DEVICE: @_ZL6texRef__static__{{.*}} = addrspace(1) externally_initialized global i64 undef, align 4 +//HIP-DEVICE: @_ZL6texRef.static.{{.*}} = addrspace(1) externally_initialized global %struct.texture undef, align 4 + +struct v4f { + float x, y, z, w; +}; + +__attribute__((device)) v4f tex2d_ld(texture<float, 2, ElementType>, float, float) asm("llvm.nvvm.texRef.unified.2d.v4f32.f32"); + +__attribute__((device)) float foo(float x, float y) { + return tex2d_ld(texRef, x, y).x; +}