GPU CUDA/HIP: Set warp size automatically, not hard-coded

Author: davidrohr

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.cu

@@ -247,13 +247,9 @@ int32_t GPUReconstructionCUDA::InitDevice_Runtime()
     if (deviceProp.warpSize != GPUCA_WARP_SIZE) {
       throw std::runtime_error("Invalid warp size on GPU");
     }
+    mWarpSize = deviceProp.warpSize;
     mBlockCount = deviceProp.multiProcessorCount;
     mMaxBackendThreads = std::max<int32_t>(mMaxBackendThreads, deviceProp.maxThreadsPerBlock * mBlockCount);
-#ifndef __HIPCC__ // CUDA
-    mWarpSize = 32;
-#else // HIP
-    mWarpSize = 64;
-#endif
     mDeviceName = deviceProp.name;
     mDeviceName += " (CUDA GPU)";
 

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDAGenRTC.cxx

@@ -73,7 +73,8 @@ int32_t GPUReconstructionCUDA::genRTC(std::string& filename, uint32_t& nCompile)
     }
     fclose(fp);
   }
-  const std::string launchBounds = o2::gpu::internal::GPUDefParametersExport(*mParDevice, true);
+  const std::string launchBounds = o2::gpu::internal::GPUDefParametersExport(*mParDevice, true) +
+                                   "#define GPUCA_WARP_SIZE " + std::to_string(mWarpSize) + "\n";
   if (mProcessingSettings.rtctech.printLaunchBounds || mProcessingSettings.debugLevel >= 3) {
     GPUInfo("RTC Launch Bounds:\n%s", launchBounds.c_str());
   }

GPU/GPUTracking/Definitions/GPUDefParametersDefaults.h

@@ -275,6 +275,9 @@
 
 #ifdef GPUCA_GPUCODE
   // Default settings for GPU, if not already set for selected GPU type
+  #ifndef GPUCA_WARP_SIZE
+    #define GPUCA_WARP_SIZE 32
+  #endif
   #ifndef GPUCA_THREAD_COUNT_DEFAULT
     #define GPUCA_THREAD_COUNT_DEFAULT 256
   #endif

GPU/GPUTracking/Definitions/GPUDefParametersWrapper.h

@@ -38,14 +38,6 @@
 #define GPUCA_MERGER_INTERPOLATION_ERROR_TYPE_A GPUCA_DETERMINISTIC_CODE(float, GPUCA_MERGER_INTERPOLATION_ERROR_TYPE)
 #define GPUCA_DEDX_STORAGE_TYPE_A GPUCA_DETERMINISTIC_CODE(float, GPUCA_DEDX_STORAGE_TYPE)
 
-#ifndef GPUCA_WARP_SIZE
-  #ifdef GPUCA_GPUCODE
-    #define GPUCA_WARP_SIZE 32
-  #else
-    #define GPUCA_WARP_SIZE 1
-  #endif
-#endif
-
 // #define GPUCA_TRACKLET_CONSTRUCTOR_DO_PROFILE                       // Output Profiling Data for Tracklet Constructor Tracklet Scheduling
 
 // #define GPUCA_KERNEL_DEBUGGER_OUTPUT