GPU RTC: Add overrideWarpSize option

Author: davidrohr

GPU/GPUTracking/Base/cuda/GPUReconstructionCUDA.cu

@@ -113,7 +113,7 @@ int32_t GPUReconstructionCUDA::InitDevice_Runtime()
   constexpr int32_t reqVerMin = 0;
 #endif
   if (GetProcessingSettings().rtc.enable && GetProcessingSettings().rtctech.runTest == 2) {
-    mWarpSize = GPUCA_WARP_SIZE;
+    mWarpSize = GetProcessingSettings().rtc.overrideWarpSize != -1 ? GetProcessingSettings().rtc.overrideWarpSize : GPUCA_WARP_SIZE;
     genAndLoadRTC();
     exit(0);
   }
@@ -245,7 +245,7 @@ int32_t GPUReconstructionCUDA::InitDevice_Runtime()
       GPUInfo("\ttextureAlignment = %ld", (uint64_t)deviceProp.textureAlignment);
       GPUInfo(" ");
     }
-    if (deviceProp.warpSize != GPUCA_WARP_SIZE && !GetProcessingSettings().rtc.enable) {
+    if (GetProcessingSettings().rtc.enable ? (GetProcessingSettings().rtc.overrideWarpSize != -1 && deviceProp.warpSize != GetProcessingSettings().rtc.overrideWarpSize) : (deviceProp.warpSize != GPUCA_WARP_SIZE)) {
       throw std::runtime_error("Invalid warp size on GPU");
     }
     mWarpSize = deviceProp.warpSize;

GPU/GPUTracking/Definitions/GPUSettingsList.h

@@ -216,6 +216,7 @@ AddOption(optSpecialCode, int8_t, -1, "", 0, "Insert GPUCA_RTC_SPECIAL_CODE spec
 AddOption(deterministic, bool, false, "", 0, "Compile RTC in deterministic mode, with NO_FAST_MATH flags and GPUCA_DETERMINISTIC_MODE define")
 AddOption(compilePerKernel, bool, true, "", 0, "Run one RTC compilation per kernel")
 AddOption(enable, bool, false, "", 0, "Use RTC to optimize GPU code")
+AddOption(overrideWarpSize, int32_t, -1, "", 0, "Override the warp size to be used for RTC")
 AddHelp("help", 'h')
 EndConfig()