GPU: Use half as storage type for interpolation errors on NVIDIA

davidrohr · davidrohr · commit 54d893afbfbe · 2020-11-09T09:16:50.000+01:00
diff --git a/GPU/Common/GPUDefGPUParameters.h b/GPU/Common/GPUDefGPUParameters.h
@@ -136,6 +136,7 @@
   #define GPUCA_NO_ATOMIC_PRECHECK 1
   #define GPUCA_COMP_GATHER_KERNEL 0
   #define GPUCA_DEDX_STORAGE_TYPE unsigned short
+  #define GPUCA_MERGER_INTERPOLATION_ERROR_TYPE half
   // #define GPUCA_USE_TEXTURES
 #elif defined(GPUCA_GPUTYPE_OPENCL)
 #elif defined(GPUCA_GPUCODE)
@@ -358,6 +359,9 @@
 #ifndef GPUCA_DEDX_STORAGE_TYPE
 #define GPUCA_DEDX_STORAGE_TYPE float
 #endif
+#ifndef GPUCA_MERGER_INTERPOLATION_ERROR_TYPE
+#define GPUCA_MERGER_INTERPOLATION_ERROR_TYPE float
+#endif
 
 #ifndef GPUCA_WARP_SIZE
 #ifdef GPUCA_GPUCODE
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMMergerTypes.h b/GPU/GPUTracking/Merger/GPUTPCGMMergerTypes.h
@@ -33,10 +33,8 @@ enum attachTypes { attachAttached = 0x40000000,
                    attachFlagMask = 0xFC000000 };
 
 struct InterpolationErrorHit {
-  float posY;
-  float errorY;
-  float posZ;
-  float errorZ;
+  float posY, posZ;
+  GPUCA_MERGER_INTERPOLATION_ERROR_TYPE errorY, errorZ;
 };
 
 struct InterpolationErrors {
diff --git a/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx b/GPU/GPUTracking/Merger/GPUTPCGMPropagator.cxx
@@ -642,7 +642,7 @@ GPUd() int GPUTPCGMPropagator::Update(float posY, float posZ, int iRow, const GP
   GetErr2(err2Y, err2Z, param, posZ, iRow, clusterState);
 
   if (rejectChi2 >= 2) {
-    if (rejectChi2 == 3 && inter->errorY < 0) {
+    if (rejectChi2 == 3 && inter->errorY < (GPUCA_MERGER_INTERPOLATION_ERROR_TYPE)0) {
       rejectChi2 = 1;
     } else {
       int retVal = InterpolateReject(posY, posZ, clusterState, rejectChi2, inter, err2Y, err2Z);
@@ -686,8 +686,8 @@ GPUd() int GPUTPCGMPropagator::InterpolateReject(float posY, float posZ, short c
     if (mFitInProjections || mT->NDF() <= 0) {
       const float Iz0 = inter->posY - mP[0];
       const float Iz1 = inter->posZ - mP[1];
-      const float Iw0 = 1.f / (mC[0] + inter->errorY);
-      const float Iw2 = 1.f / (mC[2] + inter->errorZ);
+      const float Iw0 = 1.f / (mC[0] + (float)inter->errorY);
+      const float Iw2 = 1.f / (mC[2] + (float)inter->errorZ);
       const float Ik00 = mC[0] * Iw0;
       const float Ik11 = mC[2] * Iw2;
       const float ImP0 = mP[0] + Ik00 * Iz0;
@@ -704,8 +704,8 @@ GPUd() int GPUTPCGMPropagator::InterpolateReject(float posY, float posZ, short c
     } else {
       const float Iz0 = inter->posY - mP[0];
       const float Iz1 = inter->posZ - mP[1];
-      float Iw0 = mC[2] + inter->errorZ;
-      float Iw2 = mC[0] + inter->errorY;
+      float Iw0 = mC[2] + (float)inter->errorZ;
+      float Iw2 = mC[0] + (float)inter->errorY;
       float Idet = CAMath::Max(1e-10f, Iw0 * Iw2 - mC[1] * mC[1]);
       Idet = 1.f / Idet;
       Iw0 *= Idet;
