From 3f5c581a77ee813fc283936e88906497fbce2755 Mon Sep 17 00:00:00 2001
From: Tobias Gysi <tobias.gysi@inf.ethz.ch>
Date: Thu, 11 Jun 2020 10:31:06 +0200
Subject: [PATCH] some small updates

---
 CMakeLists.txt                                   | 16 ++++++++--------
 README.md                                        |  6 +++++-
 lib/Conversion/LoopsToCUDA/CMakeLists.txt        |  4 ++--
 .../LoopsToCUDA/ConvertKernelFuncToCubin.cpp     |  2 +-
 .../LoopsToCUDA/ConvertKernelFuncToHsaco.cpp     |  2 +-
 oec-opt/oec-opt.cpp                              |  4 ++--
 6 files changed, 19 insertions(+), 15 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 1fe1d7a..244b831 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -32,16 +32,16 @@ include_directories(${PROJECT_BINARY_DIR}/include)
 link_directories(${LLVM_BUILD_LIBRARY_DIR})
 add_definitions(${LLVM_DEFINITIONS})
 
-set(OEC_CUDA_BACKEND_ENABLED 1 CACHE BOOL "Enable building the oec CUDA backend")
-set(OEC_ROCM_BACKEND_ENABLED 0 CACHE BOOL "Enable building the oec ROCM backend")
-if(OEC_CUDA_BACKEND_ENABLED)
-  add_definitions(-DOEC_CUDA_BACKEND_ENABLED)
+set(CUDA_BACKEND_ENABLED 1 CACHE BOOL "Enable building the oec CUDA backend")
+set(ROCM_BACKEND_ENABLED 0 CACHE BOOL "Enable building the oec ROCM backend")
+if(CUDA_BACKEND_ENABLED)
+  add_definitions(-DCUDA_BACKEND_ENABLED)
 endif()
-if(OEC_ROCM_BACKEND_ENABLED)
-  add_definitions(-DOEC_ROCM_BACKEND_ENABLED)
+if(ROCM_BACKEND_ENABLED)
+  add_definitions(-DROCM_BACKEND_ENABLED)
 endif()
 
-if (OEC_CUDA_BACKEND_ENABLED)
+if (CUDA_BACKEND_ENABLED)
   if (NOT ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD))
     message(SEND_ERROR "Building the oec CUDA backend requires NVPTX")
   endif()
@@ -54,7 +54,7 @@ if (OEC_CUDA_BACKEND_ENABLED)
   endif()
   find_library(CUDA_RUNTIME_LIBRARY cuda)
 endif()
-if (OEC_ROCM_BACKEND_ENABLED)
+if (ROCM_BACKEND_ENABLED)
   if (NOT ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD))
     message(SEND_ERROR "Building the oec ROCM backend requires AMDGPU")
   endif()
diff --git a/README.md b/README.md
index 8da90d4..973f49d 100644
--- a/README.md
+++ b/README.md
@@ -11,6 +11,10 @@ mkdir build && cd build
 cmake -G Ninja .. -DMLIR_DIR=$PREFIX/lib/cmake/mlir -DLLVM_EXTERNAL_LIT=$BUILD_DIR/bin/llvm-lit
 cmake --build . --target check-oec-opt
 ```
+The ROCM_BACKEND_ENABLED flag enables the support for AMDGPU tragets. It requires an llvm build with lld and we need to set the path to lld using the following flag:
+```sh
+-DLLD_DIR=$PREFIX/lib/cmake/lld
+```
 To build the documentation from the TableGen description of the dialect operations, run
 ```sh
 cmake --build . --target mlir-doc
@@ -22,7 +26,7 @@ cmake --build . --target mlir-doc
 Cmake configuration for llvm
 
 ```
-cmake -G Ninja ../llvm -DLLVM_BUILD_EXAMPLES=OFF -DLLVM_TARGETS_TO_BUILD="host;NVPTX" -DCMAKE_INSTALL_PREFIX=<install_root> -DLLVM_ENABLE_PROJECTS='mlir' -DLLVM_OPTIMIZED_TABLEGEN=ON -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_BINDINGS=OFF -DLLVM_INSTALL_UTILS=ON -DMLIR_CUDA_RUNNER_ENABLED=ON -DCMAKE_CUDA_COMPILER=<path_to_nvcc> -DCMAKE_LINKER=<path_to_lld> -DLLVM_PARALLEL_LINK_JOBS=2
+cmake -G Ninja ../llvm -DLLVM_BUILD_EXAMPLES=OFF -DLLVM_TARGETS_TO_BUILD="host;NVPTX;AMDGPU" -DCMAKE_INSTALL_PREFIX=<install_root> -DLLVM_ENABLE_PROJECTS='mlir;lld' -DLLVM_OPTIMIZED_TABLEGEN=ON -DLLVM_ENABLE_OCAMLDOC=OFF -DLLVM_ENABLE_BINDINGS=OFF -DLLVM_INSTALL_UTILS=ON -DMLIR_CUDA_RUNNER_ENABLED=ON -DCMAKE_CUDA_COMPILER=<path_to_nvcc> -DCMAKE_LINKER=<path_to_lld> -DLLVM_PARALLEL_LINK_JOBS=2
 ```
 
 Do not forget to apply possible patches to llvm before compiling (patches located in stencil-dialect/patches).
diff --git a/lib/Conversion/LoopsToCUDA/CMakeLists.txt b/lib/Conversion/LoopsToCUDA/CMakeLists.txt
index 0af1c00..083db14 100644
--- a/lib/Conversion/LoopsToCUDA/CMakeLists.txt
+++ b/lib/Conversion/LoopsToCUDA/CMakeLists.txt
@@ -1,4 +1,4 @@
-if(OEC_CUDA_BACKEND_ENABLED)
+if(CUDA_BACKEND_ENABLED)
   set(NVPTX_LIB
     NVPTXCodeGen
     NVPTXDesc
@@ -9,7 +9,7 @@ if(OEC_CUDA_BACKEND_ENABLED)
     MLIRTargetNVVMIR
   )
 endif()
-if(OEC_ROCM_BACKEND_ENABLED)
+if(ROCM_BACKEND_ENABLED)
   set(AMDGPU_LIBS
     AMDGPUAsmParser
     AMDGPUCodeGen
diff --git a/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToCubin.cpp b/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToCubin.cpp
index dceccc3..8db2ce4 100644
--- a/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToCubin.cpp
+++ b/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToCubin.cpp
@@ -13,7 +13,7 @@
 #include "mlir/Transforms/Passes.h"
 #include "llvm/Support/TargetSelect.h"
 
-#ifdef OEC_CUDA_BACKEND_ENABLED
+#ifdef CUDA_BACKEND_ENABLED
 #include "cuda.h"
 
 using namespace mlir;
diff --git a/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToHsaco.cpp b/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToHsaco.cpp
index 29e3efa..0d5d849 100644
--- a/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToHsaco.cpp
+++ b/lib/Conversion/LoopsToCUDA/ConvertKernelFuncToHsaco.cpp
@@ -18,7 +18,7 @@
 #include "llvm/Support/TargetSelect.h"
 #include "llvm/Target/TargetMachine.h"
 
-#ifdef OEC_ROCM_BACKEND_ENABLED
+#ifdef ROCM_BACKEND_ENABLED
 
 // MC headers.
 #include "llvm/MC/MCAsmBackend.h"
diff --git a/oec-opt/oec-opt.cpp b/oec-opt/oec-opt.cpp
index 9b84d83..4a32e95 100644
--- a/oec-opt/oec-opt.cpp
+++ b/oec-opt/oec-opt.cpp
@@ -88,10 +88,10 @@ int main(int argc, char **argv) {
   registerDialect<stencil::StencilDialect>();
 
   // Register the stencil pipelines
-#ifdef OEC_CUDA_BACKEND_ENABLED
+#ifdef CUDA_BACKEND_ENABLED
   registerGPUToCUBINPipeline();
 #endif 
-#ifdef OEC_ROCM_BACKEND_ENABLED
+#ifdef ROCM_BACKEND_ENABLED
   registerGPUToHSACOPipeline();
 #endif