GPU OpenCL: Fixes and workaround for undefined symbols that make the clang SPIRV code actually link correctly

GPU/Common/GPUCommonAlgorithm.h

@@ -338,8 +338,29 @@ GPUdi() void GPUCommonAlgorithm::swap(T& a, T& b)
 // Nothing to do, work_group functions available
 #pragma OPENCL EXTENSION cl_khr_subgroups : enable
 
-#define warp_scan_inclusive_add(v) sub_group_scan_inclusive_add(v)
-#define warp_broadcast(v, i) sub_group_broadcast(v, i)
+template <class T>
+GPUdi() T work_group_scan_inclusive_add_FUNC(T v)
+{
+  return sub_group_scan_inclusive_add(v);
+}
+template <> // FIXME: It seems OpenCL does not support 8 and 16 bit subgroup operations
+GPUdi() uint8_t work_group_scan_inclusive_add_FUNC<uint8_t>(uint8_t v)
+{
+  return sub_group_scan_inclusive_add((uint32_t)v);
+}
+template <class T>
+GPUdi() T work_group_broadcast_FUNC(T v, int32_t i)
+{
+  return sub_group_broadcast(v, i);
+}
+template <>
+GPUdi() uint8_t work_group_broadcast_FUNC<uint8_t>(uint8_t v, int32_t i)
+{
+  return sub_group_broadcast((uint32_t)v, i);
+}
+
+#define warp_scan_inclusive_add(v) work_group_scan_inclusive_add_FUNC(v)
+#define warp_broadcast(v, i) work_group_broadcast_FUNC(v, i)
 
 #elif (defined(__CUDACC__) || defined(__HIPCC__))
 // CUDA and HIP work the same way using cub, need just different header

GPU/GPUTracking/Base/GPUReconstructionIncludesDeviceAll.template.h

@@ -25,10 +25,6 @@ namespace gpu
 } // namespace GPUCA_NAMESPACE
 using namespace GPUCA_NAMESPACE::gpu;
 
-#if !defined(GPUCA_OPENCL1) && (!defined(GPUCA_ALIROOT_LIB) || !defined(GPUCA_GPUCODE))
-#define GPUCA_KRNL_NOOCL1
-#endif
-
 // clang-format off
 $<JOIN:$<LIST:TRANSFORM,$<LIST:TRANSFORM,$<LIST:REMOVE_DUPLICATES,$<TARGET_PROPERTY:O2_GPU_KERNELS,O2_GPU_KERNEL_FILES>>,APPEND,">,PREPEND,#include ">,
 >

GPU/GPUTracking/Base/opencl-common/GPUReconstructionOCL.cxx

@@ -115,16 +115,17 @@ int32_t GPUReconstructionOCL::InitDevice_Runtime()
         clGetPlatformInfo(mInternals->platforms[i_platform], CL_PLATFORM_VERSION, sizeof(platform_version), platform_version, nullptr);
         clGetPlatformInfo(mInternals->platforms[i_platform], CL_PLATFORM_NAME, sizeof(platform_name), platform_name, nullptr);
         clGetPlatformInfo(mInternals->platforms[i_platform], CL_PLATFORM_VENDOR, sizeof(platform_vendor), platform_vendor, nullptr);
-        if (mProcessingSettings.debugLevel >= 2) {
-          GPUInfo("Available Platform %d: (%s %s) %s %s", i_platform, platform_profile, platform_version, platform_vendor, platform_name);
-        }
+        const char* platformUsageInfo = "";
         if (!found && CheckPlatform(i_platform)) {
           found = true;
           mInternals->platform = mInternals->platforms[i_platform];
           if (mProcessingSettings.debugLevel >= 2) {
-            GPUInfo("    Using this platform");
+            platformUsageInfo = "    !!! Using this platform !!!";
           }
         }
+        if (mProcessingSettings.debugLevel >= 2) {
+          GPUInfo("Available Platform %d: (%s %s) %s %s%s", i_platform, platform_profile, platform_version, platform_vendor, platform_name, platformUsageInfo);
+        }
       }
     }
 

GPU/GPUTracking/Base/opencl2/CMakeLists.txt

@@ -53,7 +53,7 @@ if(OPENCL2_ENABLED_SPIRV) # BUILD OpenCL2 intermediate code for SPIR-V target
       MAIN_DEPENDENCY ${CL_SRC}
       IMPLICIT_DEPENDS CXX ${CL_SRC}
       COMMAND_EXPAND_LISTS
-      COMMENT "Compiling OpenCL2 CL source file ${CL_SRC} to SPIRV")
+      COMMENT "Compiling OpenCL2 CL source file ${CL_SRC} to SPIRV ${CL_BIN}.spirv")
 
   create_binary_resource(${CL_BIN}.spirv ${CMAKE_CURRENT_BINARY_DIR}/GPUReconstructionOCLCode.spirv.o)
   set(SRCS ${SRCS} ${CMAKE_CURRENT_BINARY_DIR}/GPUReconstructionOCLCode.spirv.o)
@@ -64,12 +64,14 @@ if(OPENCL2_ENABLED) # BUILD OpenCL2 source code for runtime compilation target
   add_custom_command(
       OUTPUT ${CL_BIN}.src
       COMMAND ${LLVM_CLANG}
-              ${OCL_DEFINECL} -cl-no-stdinc
+              ${OCL_FLAGS}
+              ${OCL_DEFINECL}
+              -cl-no-stdinc
               -E ${CL_SRC} > ${CL_BIN}.src
       MAIN_DEPENDENCY ${CL_SRC}
       IMPLICIT_DEPENDS CXX ${CL_SRC}
       COMMAND_EXPAND_LISTS
-      COMMENT "Preparing OpenCL2 CL source file for run time compilation")
+      COMMENT "Preparing OpenCL2 CL source file for run time compilation ${CL_BIN}.src")
 
   create_binary_resource(${CL_BIN}.src ${CMAKE_CURRENT_BINARY_DIR}/GPUReconstructionOCLCode.src.o)
   set(SRCS ${SRCS} ${CMAKE_CURRENT_BINARY_DIR}/GPUReconstructionOCLCode.src.o)

GPU/GPUTracking/CMakeLists.txt

@@ -265,7 +265,7 @@ endif()
 file(MAKE_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/include_gpu_onthefly)
 file(GENERATE
      OUTPUT include_gpu_onthefly/GPUReconstructionKernelList.h
-     INPUT Base/GPUReconstructionKernels.template.h
+     INPUT Base/GPUReconstructionKernelList.template.h
 )
 file(GENERATE
      OUTPUT include_gpu_onthefly/GPUReconstructionKernelIncludes.h

GPU/GPUTracking/Definitions/GPUDef.h

@@ -25,7 +25,11 @@
 // Macros for masking ptrs in OpenCL kernel calls as uint64_t (The API only allows us to pass buffer objects)
 #ifdef __OPENCL__
   #define GPUPtr1(a, b) uint64_t b
-  #define GPUPtr2(a, b) ((__global a) (a) b)
+  #ifdef __OPENCLCPP__
+    #define GPUPtr2(a, b) ((__generic a) (a) b)
+  #else
+    #define GPUPtr2(a, b) ((__global a) (a) b)
+  #endif
 #else
   #define GPUPtr1(a, b) a b
   #define GPUPtr2(a, b) b

GPU/GPUTracking/Refit/GPUTrackingRefit.cxx

@@ -78,7 +78,7 @@ struct refitTrackTypes<TrackParCov> {
 } // anonymous namespace
 
 template <>
-GPUd() void GPUTrackingRefit::initProp<GPUTPCGMPropagator>(GPUTPCGMPropagator& prop)
+GPUd() void GPUTrackingRefit::initProp<GPUgeneric() GPUTPCGMPropagator>(GPUTPCGMPropagator& prop) // FIXME: GPUgeneric() needed to make the clang spirv output link correctly
 {
   prop.SetMaterialTPC();
   prop.SetMaxSinPhi(GPUCA_MAX_SIN_PHI);
@@ -91,7 +91,7 @@ GPUd() void GPUTrackingRefit::initProp<GPUTPCGMPropagator>(GPUTPCGMPropagator& p
 }
 
 template <>
-GPUd() void GPUTrackingRefit::initProp<const Propagator*>(const Propagator*& prop)
+GPUd() void GPUTrackingRefit::initProp<const Propagator * GPUgeneric()>(const Propagator*& prop) // FIXME: GPUgeneric() needed to make the clang spirv output link correctly
 {
   prop = mPpropagator;
 }

GPU/GPUTracking/SliceTracker/GPUTPCTrackletConstructor.cxx

@@ -585,8 +585,8 @@ GPUd() int32_t GPUTPCTrackletConstructor::FetchTracklet(GPUconstantref() MEM_GLO
 #endif // GPUCA_GPUCODE
 
 #if !defined(__OPENCL1__)
-template <>
-GPUd() int32_t GPUTPCTrackletConstructor::GPUTPCTrackletConstructorGlobalTracking<GPUTPCGlobalTracking::GPUSharedMemory>(GPUconstantref() MEM_GLOBAL(GPUTPCTracker) & GPUrestrict() tracker, GPUsharedref() GPUTPCGlobalTracking::GPUSharedMemory& sMem, MEM_LG(GPUTPCTrackParam) & GPUrestrict() tParam, int32_t row, int32_t increment, int32_t iTracklet, calink* rowHits)
+template <> // FIXME: GPUgeneric() needed to make the clang spirv output link correctly
+GPUd() int32_t GPUTPCTrackletConstructor::GPUTPCTrackletConstructorGlobalTracking<GPUgeneric() GPUTPCGlobalTracking::GPUSharedMemory>(GPUconstantref() MEM_GLOBAL(GPUTPCTracker) & GPUrestrict() tracker, GPUsharedref() GPUTPCGlobalTracking::GPUSharedMemory& sMem, MEM_LG(GPUTPCTrackParam) & GPUrestrict() tParam, int32_t row, int32_t increment, int32_t iTracklet, calink* rowHits)
 {
   GPUTPCThreadMemory rMem;
   rMem.mISH = iTracklet;