From 41ee2ee6890685b24d432820511516d276bb464d Mon Sep 17 00:00:00 2001
From: Vikram <Vikram.Hegde@amd.com>
Date: Tue, 2 May 2023 01:28:59 -0400
Subject: [PATCH] SWDEV-378651 - add proper wavefrontsize link time option to
 hipRTC, in accordance with env variable GPU_ENABLE_WAVE32_MODE

This is a copy of http://gerrit-git.amd.com/c/compute/ec/hipamd/+/841299
Raising this here again since hipamd repo is frozen.

Change-Id: I8722ef55f78a9176e131fd5766fb11101b51f1b5


[ROCm/clr commit: cec579c0068d48c11598169636067d687cc6bb46]
---
 projects/clr/hipamd/src/CMakeLists.txt        |  1 +
 projects/clr/hipamd/src/hiprtc/CMakeLists.txt |  2 ++
 .../clr/hipamd/src/hiprtc/hiprtcInternal.cpp  | 21 +++++++++++++++++--
 3 files changed, 22 insertions(+), 2 deletions(-)

diff --git a/projects/clr/hipamd/src/CMakeLists.txt b/projects/clr/hipamd/src/CMakeLists.txt
index 8fe1c4523f..b1707bde55 100644
--- a/projects/clr/hipamd/src/CMakeLists.txt
+++ b/projects/clr/hipamd/src/CMakeLists.txt
@@ -188,6 +188,7 @@ endif()
 
 set(HIPRTC_OBJECTS)
 # Add hiprtc
+set(FIXME_DIR ${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles/amdhip64.dir)
 add_subdirectory(hiprtc)
 
 if(NOT WIN32)
diff --git a/projects/clr/hipamd/src/hiprtc/CMakeLists.txt b/projects/clr/hipamd/src/hiprtc/CMakeLists.txt
index 37ef8796fa..d2e500733a 100644
--- a/projects/clr/hipamd/src/hiprtc/CMakeLists.txt
+++ b/projects/clr/hipamd/src/hiprtc/CMakeLists.txt
@@ -96,6 +96,8 @@ if(BUILD_SHARED_LIBS)
 endif()
 
 if(WIN32)
+  target_link_libraries(hiprtc PRIVATE ${FIXME_DIR}/fixme.cpp${CMAKE_CXX_OUTPUT_EXTENSION})
+  add_dependencies(hiprtc amdhip64)
   target_link_libraries(hiprtc PRIVATE Dbghelp.lib)
 endif()
 
diff --git a/projects/clr/hipamd/src/hiprtc/hiprtcInternal.cpp b/projects/clr/hipamd/src/hiprtc/hiprtcInternal.cpp
index 08eba0a3a1..0d23a1c697 100644
--- a/projects/clr/hipamd/src/hiprtc/hiprtcInternal.cpp
+++ b/projects/clr/hipamd/src/hiprtc/hiprtcInternal.cpp
@@ -133,8 +133,6 @@ RTCCompileProgram::RTCCompileProgram(std::string name_) : RTCProgram(name_), fgp
 
   if (GPU_ENABLE_WGP_MODE) compile_options_.push_back("-mcumode");
 
-  if (!GPU_ENABLE_WAVE32_MODE) compile_options_.push_back("-mwavefrontsize64");
-
   compile_options_.push_back(hipVerOpt);
   compile_options_.push_back(hipVerMajor);
   compile_options_.push_back(hipVerMinor);
@@ -261,6 +259,14 @@ bool RTCCompileProgram::transformOptions(std::vector<std::string>& compile_optio
   return findIsa();
 }
 
+static inline uint getArchMajorVersion(std::string &isa) {
+  if (const amd::Isa *isaIter = amd::Isa::findIsa(isa.data())) {
+    return isaIter->versionMajor();
+  }
+
+  return static_cast<uint>(-1);
+}
+
 amd::Monitor RTCProgram::lock_("HIPRTC Program", true);
 
 bool RTCCompileProgram::compile(const std::vector<std::string>& options, bool fgpu_rdc) {
@@ -286,6 +292,17 @@ bool RTCCompileProgram::compile(const std::vector<std::string>& options, bool fg
     return false;
   }
 
+  // Decide whether to enable wave64 compilation
+  auto majorVer = getArchMajorVersion(isa_);
+  if (majorVer <= 9 || !GPU_ENABLE_WAVE32_MODE) {
+    if (majorVer > 9) {
+       LogWarning("Wavefront size 64 is experimental for gfx10 and above. Warp "
+                  "functions may not work");
+    }
+    compileOpts.push_back("-mwavefrontsize64");
+    link_options_.push_back("wavefrontsize64");
+  }
+
   if (!compileToBitCode(compile_input_, isa_, compileOpts, build_log_, LLVMBitcode_)) {
     LogError("Error in hiprtc: unable to compile source to bitcode");
     return false;