P4 to Git Change 1753035 by wchau@wchau_OCL_Linux on 2019/03/07 15:02:39

SWDEV-182054 - Allow building OpenCL-Runtime with COMGR enabled without OpenCL-Driver being present

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/build/Makefile.hip#12 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/build/Makefile.api#184 edit
... //depot/stg/opencl/drivers/opencl/compiler/Makefile#71 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.cpp#242 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/device.hpp#335 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.cpp#30 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/devprogram.hpp#18 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#126 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#87 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#120 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#51 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocprogram.cpp#100 edit

rocclr/runtime/device/device.cpp

@@ -148,7 +148,7 @@ bool Device::init() {
 // GPU stack. The order of initialization is signiicant and if changed
 // amd::Device::registerDevice() must be accordingly modified.
 #if defined(WITH_HSA_DEVICE)
-  // @todo remove IS_LIGHTNING check when PAL-LC builds will be deprecated 
+  // @todo remove IS_LIGHTNING check when PAL-LC builds will be deprecated
   if ((GPU_ENABLE_PAL != 1) || IS_LIGHTNING) {
     // Return value of roc::Device::init()
     // If returned false, error initializing HSA stack.
@@ -229,6 +229,7 @@ bool Device::ValidateComgr() {
 #if defined(USE_COMGR_LIBRARY)
   // Check if Lightning compiler was requested
   if (settings_->useLightning_) {
+    LogInfo("Loading COMGR library.");
     std::call_once(amd::Comgr::initialized, amd::Comgr::LoadLib);
     // Use Lightning only if it's available
     settings_->useLightning_ = amd::Comgr::IsReady();
@@ -413,7 +414,7 @@ char* Device::getExtensionString() {
   return result;
 }
 
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 CacheCompilation::CacheCompilation(std::string targetStr, std::string postfix, bool enableCache,
                                    bool resetCache)
     : codeCache_(targetStr, 0, AMD_PLATFORM_BUILD_NUMBER, postfix),
@@ -498,13 +499,11 @@ bool CacheCompilation::compileToLLVMBitcode(amd::opencl_driver::Compiler* C,
         StringCache::CachedData cachedData = {bc->Ptr(), bc->Size()};
         bcSet.push_back(cachedData);
       } else if (input->Type() == DT_CL_HEADER) {
-#if !defined(USE_COMGR_LIBRARY)
         FileReference* bcFile = reinterpret_cast<FileReference*>(input);
         std::string bc;
         bcFile->ReadToString(bc);
         StringCache::CachedData cachedData = {bc.c_str(), bc.size()};
         bcSet.push_back(cachedData);
-#endif // !defined(USE_COMGR_LIBRARY)
       } else {
         buildLog += "Error: unsupported bitcode type for checking cache.\n";
         checkCache = false;
@@ -582,7 +581,7 @@ bool CacheCompilation::compileAndLinkExecutable(amd::opencl_driver::Compiler* C,
 
   return true;
 }
-#endif // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#endif // defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 
 }  // namespace amd
 

rocclr/runtime/device/device.hpp

@@ -19,10 +19,10 @@
 #include "devkernel.hpp"
 #include "amdocl/cl_profile_amd.h"
 
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 #include "caching/cache.hpp"
 #include "driver/AmdCompiler.h"
-#endif // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#endif // defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 #include "acl.h"
 
 #include "hwdebug.hpp"
@@ -69,7 +69,9 @@ class SvmUnmapMemoryCommand;
 class TransferBufferFileCommand;
 class HwDebugManager;
 class Device;
+#ifndef USE_COMGR_LIBRARY
 class CacheCompilation;
+#endif
 struct KernelParameterDescriptor;
 struct Coord3D;
 
@@ -1328,7 +1330,7 @@ class Device : public RuntimeObject {
   // P2P devices that are accessible from the current device
   std::vector<cl_device_id> p2pDevices_;
 
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
   amd::CacheCompilation* cacheCompilation() const { return cacheCompilation_.get(); }
 #endif
 
@@ -1354,7 +1356,7 @@ class Device : public RuntimeObject {
   BlitProgram* blitProgram_;      //!< Blit program info
   static AppProfile appProfile_;  //!< application profile
   HwDebugManager* hwDebugMgr_;    //!< Hardware Debug manager
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
                                   //! Compilation with cache support
   std::unique_ptr<amd::CacheCompilation> cacheCompilation_;
 #endif
@@ -1372,7 +1374,7 @@ class Device : public RuntimeObject {
   std::map<uintptr_t, device::Memory*>* vaCacheMap_;  //!< VA cache map
 };
 
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 //! Compilation process with cache support.
 class CacheCompilation : public amd::HeapObject {
  public:

rocclr/runtime/device/devprogram.cpp

@@ -12,7 +12,9 @@
 #include "utils/libUtils.h"
 
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#ifndef USE_COMGR_LIBRARY
 #include "driver/AmdCompiler.h"
+#endif
 #include "libraries.amdgcn.inc"
 #include "opencl1.2-c.amdgcn.inc"
 #include "opencl2.0-c.amdgcn.inc"
@@ -161,6 +163,7 @@ static void checkLLVM_BIN() {
 }
 #endif  // defined(ATI_OS_LINUX)
 
+#if !defined(USE_COMGR_LIBRARY)
 std::unique_ptr<amd::opencl_driver::Compiler> Program::newCompilerInstance() {
 #if defined(ATI_OS_WIN)
   static INIT_ONCE initOnce;
@@ -178,13 +181,10 @@ std::unique_ptr<amd::opencl_driver::Compiler> Program::newCompilerInstance() {
   }
 #endif  // defined(DEBUG)
 
-#if !defined(USE_COMGR_LIBRARY)
   return std::unique_ptr<amd::opencl_driver::Compiler>(
     amd::opencl_driver::CompilerFactory().CreateAMDGPUCompiler(llvmBin_));
-#else
-  return std::unique_ptr<amd::opencl_driver::Compiler>(nullptr);
-#endif // !defined(USE_COMGR_LIBRARY)
 }
+#endif // !defined(USE_COMGR_LIBRARY)
 #endif // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
 
 // ================================================================================================
@@ -1082,8 +1082,6 @@ bool Program::linkImpl(const std::vector<device::Program*>& inputPrograms,
 #if defined(USE_COMGR_LIBRARY)
 bool Program::linkImplLC(const std::vector<Program*>& inputPrograms,
                          amd::option::Options* options, bool createLibrary) {
-  using namespace amd::opencl_driver;
-  std::unique_ptr<Compiler> C(newCompilerInstance());
 
   amd_comgr_data_set_t inputs;
 

rocclr/runtime/device/devprogram.hpp

@@ -11,7 +11,9 @@
 #include "comgrctx.hpp"
 
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#ifndef USE_COMGR_LIBRARY
 #include "driver/AmdCompiler.h"
+#endif
 //#include "llvm/Support/AMDGPUMetadata.h"
 
 namespace llvm {
@@ -276,7 +278,7 @@ class Program : public amd::HeapObject {
 
   void setType(type_t newType) { type_ = newType; }
 
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && !defined(USE_COMGR_LIBRARY)
   //! Return a new transient compiler instance.
   static std::unique_ptr<amd::opencl_driver::Compiler> newCompilerInstance();
 #endif // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)

rocclr/runtime/device/pal/paldevice.cpp

@@ -280,7 +280,7 @@ bool NullDevice::create(Pal::AsicRevision asicRevision, Pal::GfxIpLevel ipLevel,
   info_.wavefrontWidth_ = settings().enableWave32Mode_ ? 32 : 64;
 
   if (settings().useLightning_) {
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
     //  create compilation object with cache support
     int gfxipMajor = hwInfo_->gfxipVersionLC_ / 100;
     int gfxipMinor = hwInfo_->gfxipVersionLC_ / 10 % 10;
@@ -967,7 +967,7 @@ bool Device::create(Pal::IDevice* device) {
   }
 
   if (settings().useLightning_) {
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
     //  create compilation object with cache support
     int gfxipMajor = hwInfo()->gfxipVersionLC_ / 100;
     int gfxipMinor = hwInfo()->gfxipVersionLC_ / 10 % 10;

rocclr/runtime/device/pal/palprogram.cpp

@@ -20,8 +20,10 @@
 #include "hsa_ext_image.h"
 #include "amd_hsa_loader.hpp"
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
-#include "llvm/Support/AMDGPUMetadata.h"
+#ifndef USE_COMGR_LIBRARY
 #include "driver/AmdCompiler.h"
+#endif
+#include "llvm/Support/AMDGPUMetadata.h"
 #include "libraries.amdgcn.inc"
 #include "gelf.h"
 #endif  // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)

rocclr/runtime/device/rocm/rocdevice.cpp

@@ -19,9 +19,9 @@
 #include "device/rocm/rocblit.hpp"
 #include "device/rocm/rocvirtual.hpp"
 #include "device/rocm/rocprogram.hpp"
-#if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#if defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 #include "driver/AmdCompiler.h"
-#endif  // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#endif  // defined(WITH_LIGHTNING_COMPILER) && ! defined(USE_COMGR_LIBRARY)
 #include "device/rocm/rocmemory.hpp"
 #include "device/rocm/rocglinterop.hpp"
 #ifdef WITH_AMDGPU_PRO
@@ -652,6 +652,7 @@ bool Device::create(bool sramEccEnabled) {
   if (settings().useLightning_) {
     scheduler = sch.c_str();
   }
+#ifndef USE_COMGR_LIBRARY
   //  create compilation object with cache support
   int gfxipMajor = deviceInfo_.gfxipVersion_ / 100;
   int gfxipMinor = deviceInfo_.gfxipVersion_ / 10 % 10;
@@ -676,6 +677,7 @@ bool Device::create(bool sramEccEnabled) {
   }
 
   cacheCompilation_.reset(compObj);
+#endif  // USE_COMGR_LIBRARY
 #endif
 
   amd::Context::Info info = {0};

rocclr/runtime/device/rocm/rockernel.cpp

@@ -10,7 +10,9 @@
 #ifndef WITHOUT_HSA_BACKEND
 
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
+#ifndef USE_COMGR_LIBRARY
 #include "driver/AmdCompiler.h"
+#endif
 #include "llvm/Support/AMDGPUMetadata.h"
 
 typedef llvm::AMDGPU::HSAMD::Metadata CodeObjectMD;

rocclr/runtime/device/rocm/rocprogram.cpp

@@ -9,8 +9,10 @@
 #include "rockernel.hpp"
 #if defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
 #include <gelf.h>
-#include "driver/AmdCompiler.h"
 #include "libraries.amdgcn.inc"
+#ifndef USE_COMGR_LIBRARY
+#include "driver/AmdCompiler.h"
+#endif
 #endif  // defined(WITH_LIGHTNING_COMPILER) || defined(USE_COMGR_LIBRARY)
 
 #include "utils/bif_section_labels.hpp"