[rocisa] Use comgr instead of calling amdclang++

Author: KKyang

tensilelite/CMakeLists.txt

@@ -32,15 +32,15 @@ cmake_minimum_required(VERSION 3.15)
 # This CMakeLists.txt supports two modes with setting DEVELOP_MODE ON or OFF:
 # 1. Script mode (DEVELOP_MODE=ON): It creates a script that can be run after the configuration
 #    When using mode, you can run it as:
-#        cmake -DTENSILE_BIN=Tensile -DDEVELOP_MODE=ON ..
+#        cmake -DCMAKE_CXX_COMPILER=/opt/rocm/lib/llvm/bin/amdclang++ -DTENSILE_BIN=Tensile -DDEVELOP_MODE=ON ..
 #    The script will be created in the build folder and will be named in Tensile.bat or Tensile.sh
 #    depending on the platform. You can then run the script as:
 #        Tensile.sh config.yaml ./
 #    or
-#        Tensile.bat config.yaml ./    
+#        Tensile.bat config.yaml ./
 # 2. Normal CMake mode: It sets the Python srguments at configuration stage with "BIN_ARGS"
 #    With this mode, you can run it as:
-#        cmake -DTENSILE_BIN=Tensile -DBIN_ARGS="config.yaml ./" .. && cmake --build .
+#        cmake -DCMAKE_CXX_COMPILER=/opt/rocm/lib/llvm/bin/amdclang++ -DTENSILE_BIN=Tensile -DBIN_ARGS="config.yaml ./" .. && cmake --build .
 #    The script will check if the TENSILE_BIN is valid and if the BIN_ARGS
 #    are provided. It will then run the specified Tensile binary with the
 #    provided arguments after the build step.
@@ -123,7 +123,7 @@ else()
     message(STATUS "Script set: ${TENSILE_BIN} ${BIN_ARGS_LIST}")
 
     # Ensure the Python script runs after the build
-    add_custom_target(RunPythonScript 
+    add_custom_target(RunPythonScript
         ALL
         COMMAND ${CMAKE_COMMAND} -E env PYTHONPATH=${PYTHONPATH} -- ${Python_EXECUTABLE} ${TENSILE_BIN_ROOT}/${TENSILE_BIN} ${BIN_ARGS_LIST}
         COMMENT "Running Python script ${TENSILE_BIN} ${BIN_ARGS_LIST}"

tensilelite/README.md

@@ -2,7 +2,7 @@
 
 ** How to Rebuild Object Codes Directly from Assembly
 
-During the tuning process, it is of interest to modify an assembly file/s and rebuild the corresponding object file/s and then relink the corresponding co file. Currently, we generate additional source files and a script to provide this workflow. 
+During the tuning process, it is of interest to modify an assembly file/s and rebuild the corresponding object file/s and then relink the corresponding co file. Currently, we generate additional source files and a script to provide this workflow.
 
 A new `Makefile` is added that manages rebuilding a co file during iterative development when tuning. One modifies an assembly file of interest, then runs `make` and make will detect what file/s changed and rebuild accordingly.
 
@@ -14,7 +14,7 @@ Assumptions:
 
 Example:
 
-```cmake -DTENSILE_BIN=Tensile -DDEVELOP_MODE=ON -S <path-to-tensilelite-root> -B <tensile-out>```
+```cmake -DCMAKE_CXX_COMPILER=/opt/rocm/lib/llvm/bin/amdclang++ -DTENSILE_BIN=Tensile -DDEVELOP_MODE=ON -S <path-to-tensilelite-root> -B <tensile-out>```
 
 The script will be created in the build folder and will be named in Tensile.bat or Tensile.sh depending on the platform. Then you can then run the script under the ``tensile-out`` folder as usual:
 

tensilelite/rocisa/CMakeLists.txt

@@ -26,6 +26,8 @@ set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
 
+find_package(amd_comgr REQUIRED CONFIG)
+
 if(DEFINED Python_EXECUTABLE AND Python_EXECUTABLE)
     message(STATUS "Manually set Python_EXECUTABLE to ${Python_EXECUTABLE}")
 endif()
@@ -73,5 +75,7 @@ nanobind_add_module(rocisa NOMINSIZE NB_SUPPRESS_WARNINGS
                         ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/macro.cpp
                         ${ROCISAINST_SOURCE}
                         ${ROCISAPASS_SOURCE})
+
 target_include_directories(rocisa PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/include)
+target_link_libraries(rocisa PRIVATE amd_comgr)
 set_target_properties(rocisa PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)

tensilelite/rocisa/rocisa/include/hardware_caps.hpp

@@ -38,27 +38,23 @@ inline bool tryAssembler(const IsaVersion&  isaVersion,
     if(isaVersion[0] >= 10)
         options.push_back("-mwavefrontsize64");
 
-    std::string isastr = getGfxNameTuple(isaVersion);
-
-    std::vector<std::string> cmd
-        = {assemblerPath, "-x", "assembler", "-target", "amdgcn-amdhsa", "-mcpu=" + isastr};
-    for(auto o : options)
+    const char** optionStr        = new const char*[options.size()];
+    std::string  optionOneLineStr = "";
+    for(size_t i = 0; i < options.size(); i++)
     {
-        cmd.push_back(o);
+        optionStr[i] = options[i].c_str();
+        optionOneLineStr += options[i] + " ";
     }
-    cmd.push_back("-");
-    std::vector<char*> args(cmd.size());
-    std::transform(cmd.begin(), cmd.end(), args.begin(), [](auto& str) { return &str[0]; });
-    args.push_back(nullptr);
-    auto [rcode, result] = run(args, asmString, debug);
+    std::string isastr = getGfxNameTuple(isaVersion);
+
+    auto [rcode, result] = runComgr(asmString, isastr, optionStr, options.size(), debug);
+    delete[] optionStr;
 
     if(debug)
     {
         std::string s;
-        for(auto c : cmd)
-            s += c + " ";
         std::cout << "isaVersion: " << isastr << std::endl;
-        std::cout << "asm_cmd: " << s << std::endl;
+        std::cout << "options: " << optionOneLineStr << std::endl;
         std::cout << "asmString: " << asmString << std::endl;
         std::cout << "result: " << result << std::endl;
         std::cout << "return code: " << rcode << std::endl;

tensilelite/rocisa/rocisa/include/helper.hpp

@@ -29,17 +29,20 @@
 
 using IsaVersion = std::array<int, 3>;
 
-std::pair<int, std::string>
-            run(const std::vector<char*>& cmd, const std::string& input, bool debug = false);
-std::string demangle(const char* name);
+std::pair<int, std::string> runComgr(const std::string& input,
+                                     const std::string& gfxStr,
+                                     const char**       optionStr,
+                                     const size_t       optionSize,
+                                     bool               debug);
+std::string                 demangle(const char* name);
 
 inline std::string getGfxNameTuple(const IsaVersion& isaVersion)
 {
     /*Converts an ISA version to a gfx architecture name.
 
     Args:
         arch: An object representing the major, minor, and step version of the ISA.
-    
+
     Returns:
         The name of the GPU architecture (e.g., 'gfx906').
     */