move simd pragmas to inner loops, add constexpr to some config condit…

…ionals
cms-sw · Feb 7, 2024 · 02ea183 · 02ea183
1 parent a893c4c
commit 02ea183
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 11 deletions.
diff --git a/RecoTracker/MkFitCore/src/MkBuilder.cc b/RecoTracker/MkFitCore/src/MkBuilder.cc
@@ -412,7 +412,7 @@ namespace mkfit {
   //------------------------------------------------------------------------------
 
   void MkBuilder::seed_post_cleaning(TrackVec &tv) {
-    if (Const::nan_n_silly_check_seeds) {
+    if constexpr (Const::nan_n_silly_check_seeds) {
       int count = 0;
 
       for (int i = 0; i < (int)tv.size(); ++i) {
@@ -421,7 +421,7 @@ namespace mkfit {
                                           "Post-cleaning seed silly value check and fix");
         if (silly) {
           ++count;
-          if (Const::nan_n_silly_remove_bad_seeds) {
+          if constexpr (Const::nan_n_silly_remove_bad_seeds) {
             // XXXX MT
             // Could do somethin smarter here: set as Stopped ?  check in seed cleaning ?
             tv.erase(tv.begin() + i);
@@ -669,7 +669,7 @@ namespace mkfit {
             seed_cand_vec.push_back(std::pair<int, int>(iseed, ic));
             ccand[ic].resetOverlaps();
 
-            if (Const::nan_n_silly_check_cands_every_layer) {
+            if constexpr (Const::nan_n_silly_check_cands_every_layer) {
               if (ccand[ic].hasSillyValues(Const::nan_n_silly_print_bad_cands_every_layer,
                                            Const::nan_n_silly_fixup_bad_cands_every_layer,
                                            "Per layer silly check"))
@@ -683,7 +683,7 @@ namespace mkfit {
       }
     }
 
-    if (Const::nan_n_silly_check_cands_every_layer && silly_count > 0) {
+    if constexpr (Const::nan_n_silly_check_cands_every_layer && silly_count > 0) {
       m_nan_n_silly_per_layer_count += silly_count;
     }
 
@@ -1110,7 +1110,7 @@ namespace mkfit {
         // mkfndr->copyOutParErr(eoccs.refCandidates_nc(), end - itrack, true);
 
         // For prop-to-plane propagate from the last hit, not layer center.
-        if (Config::usePropToPlane) {
+        if constexpr (Config::usePropToPlane) {
           mkfndr->inputTracksAndHitIdx(eoccs.refCandidates(), seed_cand_idx, itrack, end, false);
         }
 

diff --git a/RecoTracker/MkFitCore/src/MkFinder.cc b/RecoTracker/MkFitCore/src/MkFinder.cc
@@ -1313,7 +1313,7 @@ namespace mkfit {
       MPlexLV propPar;
       clearFailFlag();
 
-      if (Config::usePropToPlane) {
+      if constexpr (Config::usePropToPlane) {
         // Maybe could use 2 matriplex packers ... ModuleInfo has 3 * SVector3 and uint
         MPlexHV norm, dir;
         packModuleNormDir(layer_of_hits, hit_cnt, norm, dir, N_proc);
@@ -1582,7 +1582,7 @@ namespace mkfit {
       MPlexLV propPar;
       clearFailFlag();
 
-      if (Config::usePropToPlane) {
+      if constexpr (Config::usePropToPlane) {
         // Maybe could use 2 matriplex packers ... ModuleInfo has 3 * SVector3 and uint
         MPlexHV norm, dir;
         packModuleNormDir(layer_of_hits, hit_cnt, norm, dir, N_proc);
@@ -1759,7 +1759,7 @@ namespace mkfit {
     // See comment in MkBuilder::find_tracks_in_layer() about intra / inter flags used here
     // for propagation to the hit.
     clearFailFlag();
-    if (Config::usePropToPlane) {
+    if constexpr (Config::usePropToPlane) {
       MPlexHV norm, dir;
       packModuleNormDir(layer_of_hits, 0, norm, dir, N_proc);
       kalmanPropagateAndUpdatePlane(m_Err[iP],

diff --git a/RecoTracker/MkFitCore/src/PropagationMPlex.cc b/RecoTracker/MkFitCore/src/PropagationMPlex.cc
@@ -228,9 +228,11 @@ namespace {
 
   // this version does not assume to know which elements are 0 or 1, so it does the full multiplication
   void MultHelixPropFull(const MPlexLL& A, const MPlexLS& B, MPlexLL& C) {
-#pragma omp simd
     for (int n = 0; n < NN; ++n) {
       for (int i = 0; i < 6; ++i) {
+// optimization reports indicate only the inner two loops are good
+// candidates for vectorization
+#pragma omp simd
         for (int j = 0; j < 6; ++j) {
           C(n, i, j) = 0.;
           for (int k = 0; k < 6; ++k)
@@ -242,9 +244,11 @@ namespace {
 
   // this version does not assume to know which elements are 0 or 1, so it does the full mupltiplication
   void MultHelixPropTranspFull(const MPlexLL& A, const MPlexLL& B, MPlexLS& C) {
-#pragma omp simd
     for (int n = 0; n < NN; ++n) {
       for (int i = 0; i < 6; ++i) {
+// optimization reports indicate only the inner two loops are good
+// candidates for vectorization
+#pragma omp simd
         for (int j = 0; j < 6; ++j) {
           C(n, i, j) = 0.;
           for (int k = 0; k < 6; ++k)
@@ -1280,7 +1284,7 @@ namespace mkfit {
       // const float thetaMSC2 = thetaMSC*thetaMSC;
       const float thetaMSC = 0.0136f * (1.f + 0.038f * std::log(radL)) / (beta * p);  // eq 32.15
       const float thetaMSC2 = thetaMSC * thetaMSC * radL;
-      if (Config::usePtMultScat) {
+      if constexpr (Config::usePtMultScat) {
         outErr.At(n, 3, 3) += thetaMSC2 * pz * pz * ipt2 * ipt2;
         outErr.At(n, 3, 5) -= thetaMSC2 * pz * ipt2;
         outErr.At(n, 4, 4) += thetaMSC2 * p2 * ipt2;