[bug] Fix error with windows-clang compilation for cuda_runtime.cu (t…

…aichi-dev#7519)
quadpixels · May 13, 2023 · ed21306 · ed21306
1 parent 8bbef1e
commit ed21306
Show file tree

Hide file tree

Showing 5 changed files with 11 additions and 19 deletions.
diff --git a/.github/workflows/scripts/build.py b/.github/workflows/scripts/build.py
@@ -25,9 +25,9 @@ def setup_clang(as_compiler=True) -> None:
     if u.system == 'Linux':
         pass
     elif (u.system, u.machine) == ('Windows', 'AMD64'):
-        out = get_cache_home() / 'clang-15'
-        url = 'https://github.com/python3kgae/taichi_assets/releases/download/llvm15_vs2022_clang/clang-15.0.0-win.zip'
-        download_dep(url, out)
+        out = get_cache_home() / 'clang-15-v2'
+        url = 'https://github.com/taichi-dev/taichi_assets/releases/download/llvm15/clang-15.0.0-win-complete.zip'
+        download_dep(url, out, force=True)
         clang = str(out / 'bin' / 'clang++.exe').replace('\\', '\\\\')
         os.environ['TAICHI_CMAKE_ARGS'] += f' -DCLANG_EXECUTABLE={clang}'
 

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -127,6 +127,7 @@ endif()
 macro(check_clang_version)
   execute_process(COMMAND ${CLANG_EXECUTABLE} --version OUTPUT_VARIABLE CLANG_VERSION_OUTPUT)
   string(REGEX MATCH "([0-9]+)\\.[0-9]+(\\.[0-9]+)?" CLANG_VERSION "${CLANG_VERSION_OUTPUT}")
+
   message("${CLANG_EXECUTABLE} --version: ${CLANG_VERSION}")
 
   set(CLANG_VERSION_MAJOR "${CMAKE_MATCH_1}")

diff --git a/taichi/codegen/cuda/codegen_cuda.cpp b/taichi/codegen/cuda/codegen_cuda.cpp
@@ -305,7 +305,6 @@ class TaskCodeGenCUDA : public TaskCodeGenLLVM {
 
             old_val = Load(old_val_ptr)
       */
-
       // Allocate old_val_ptr to store the result of atomic_add
       auto char_type = llvm::Type::getInt8Ty(*tlctx->get_this_thread_context());
       auto half_type = llvm::Type::getHalfTy(*tlctx->get_this_thread_context());
@@ -335,12 +334,10 @@ class TaskCodeGenCUDA : public TaskCodeGenLLVM {
       builder->CreateStore(value1, value_ptr1);
       llvm::Value *value_half2_ptr =
           builder->CreateBitCast(value_ptr, ptr_type);
-
       // Defined in taichi/runtime/llvm/runtime_module/cuda_runtime.cu
       call("half2_atomic_add", dest_half2_ptr, old_val_ptr, value_half2_ptr);
 
-      llvm_val[atomic_stmt] = builder->CreateLoad(
-          old_val->getType()->getPointerElementType(), old_val);
+      llvm_val[atomic_stmt] = builder->CreateLoad(half_type, old_val);
       return;
     }
 

diff --git a/taichi/runtime/llvm/runtime_module/CMakeLists.txt b/taichi/runtime/llvm/runtime_module/CMakeLists.txt
@@ -30,15 +30,11 @@ foreach(arch IN LISTS HOST_ARCH CUDA_ARCH DX12_ARCH AMDGPU_ARCH)
   compile_llvm_runtime(${arch})
 endforeach()
 
-if(NOT WIN32)
-    # clang++ on Windows reports an internal error of "<built-in>:1:10: fatal error: '__clang_cuda_runtime_wrapper.h' file not found" when compiling for CUDA targets
-    # so we can't use hand-writen cuda libraries util this issue get resolved
-    if(CUDA_ARCH)
-        find_package(CUDAToolkit)
-        if(${CUDAToolkit_FOUND})
-          if(${CUDAToolkit_VERSION_MAJOR} GREATER_EQUAL 8)
-              compile_custom_cuda_library(${CUDAToolkit_TARGET_DIR})
-          endif()
-        endif()
+if(CUDA_ARCH)
+    find_package(CUDAToolkit)
+    if(${CUDAToolkit_FOUND})
+      if(${CUDAToolkit_VERSION_MAJOR} GREATER_EQUAL 8)
+          compile_custom_cuda_library(${CUDAToolkit_TARGET_DIR})
+      endif()
     endif()
 endif()
diff --git a/taichi/runtime/llvm/runtime_module/cuda_runtime.cu b/taichi/runtime/llvm/runtime_module/cuda_runtime.cu
@@ -1,5 +1,3 @@
-#pragma once
-
 #include <cuda_fp16.h>
 
 /*