improve benchmarking

src/base/globals.nim

@@ -44,3 +44,13 @@ macro setDefaultNc*(n: static[int]): untyped =
   result = newEmptyNode()
 macro getDefaultNc*(): untyped =
   return newLit(defaultNc)
+
+template getDefPrecStr:string =
+  const defPrec {.strdefine.} = "D"
+  defPrec
+
+var defPrec* {.compiletime.} = getDefPrecStr()
+macro setDefaultSingle* =
+  defPrec = "S"
+  echo "Default precision: ", defPrec
+static: echo "Default precision: ", defPrec

src/base/profile.nim

@@ -1,6 +1,6 @@
 import threading
 export threading
-import comms/comms, stdUtils, base/basicOps
+import comms/comms, stdUtils, base/[basicOps,params]
 import os, strutils, sequtils, std/monotimes, std/tables, std/algorithm, strformat
 export monotimes
 getOptimPragmas()
@@ -20,7 +20,7 @@ var
 
 ##[
 
-Each Tic starts a local timer.  Each Toc records the time difference
+Each tic() starts a local timer.  Each toc() records the time difference
 of the current time with the one in the local timer visible in the
 scope, and then update the timer with the current time.
 
@@ -516,49 +516,10 @@ template tocI(f: SomeNumber; s:SString = ""; n = -1) =
       localCode {.global.} = newList[CodePoint]()
       thisCode = CodePoint(-1)
   if threadNum==0:
-    when false:
-      #echo "==== begin toc ",s," ",ii
-      #echo "     rtiStack: ",indent($rtiStack,5)
-      #echo "     cpHeap: ",indent($cpHeap,5)
-      if unlikely VerboseTimer: echoToc(s,ii)
-      if prevRTI.int32 >= 0:
-        if restartTimer:
-          thawTimers()
-          restartTimer = false
-        if not timersFrozen():
-          let theTime = getTics()
-          when not cname:
-            for c in items(localCode):
-              if cpHeap[c.int].name.equal(s):
-                thisCode = c
-                break
-          if thisCode.isNil:
-            thisCode = newCodePoint(ii.addr, s)
-            when not cname:
-              localCode.add thisCode
-          let
-            ns = theTime-localTimer
-            thisRTI = record(localTic, prevRTI, thisCode, ns, float(f))
-          var oh = rtiStack[thisRTI.int].childrenOverhead
-          let c = rtiStack[thisRTI.int].children
-          for i in 0..<c.len:
-            if toDropTimer(c[i].prev):
-              oh -= c[i].childrenOverhead
-          if oh.float / ns.float > DropWasteTimerRatio:
-            # Signal stop if the overhead is too large.
-            dropTimer(prevRTI)
-          if toDropTimer(thisCode):
-            freezeTimers()
-            restartTimer = true
-          localTimer = getTics()
-          rtiStack[thisRTI.int].overhead = nsec(localTimer-theTime)
-          prevRTI = thisRTI
-      #echo "==== end toc ",s," ",ii
+    when cname:
+      tocSet(localTimer,prevRTI,restartTimer,thisCode,f,s,ii.addr,localTic,false)
     else:
-      when cname:
-        tocSet(localTimer,prevRTI,restartTimer,thisCode,f,s,ii.addr,localTic,false)
-      else:
-        tocSet(localTimer,prevRTI,restartTimer,thisCode,f,s,ii.addr,localTic,addr localCode)
+      tocSet(localTimer,prevRTI,restartTimer,thisCode,f,s,ii.addr,localTic,addr localCode)
 
 when noTicToc:
   template toc*() = discard
@@ -822,7 +783,7 @@ proc makeHotspotTable(lrti: List[RTInfoObj]): tuple[ns:int64,oh:int64] =
           t.children.add ri.children[i]
     do: # loc not found
       hs[loc] = ri
-    #let tot = makeHotSpotTable(List[RTInfoObj](ri.children))
+    let tot = makeHotSpotTable(List[RTInfoObj](ri.children))
   return (nstot, ohtot)
 
 proc echoHotspots* =
@@ -879,6 +840,12 @@ proc echoHotspots* =
         let tsnspct = 100.0 * tsns / nstot
         echo &"{pct:6.3f} {tsnspct:7.3f} {count} {mf} {nc} S {lc} {nm}"
 
+proc echoProf*(def = 0) =
+  case intParam("prof",def)
+  of 1: echoHotspots()
+  of 2: echoTimers()
+  else: discard
+
 when isMainModule:
   import os
   proc test =

src/bench/benchLinalg.nim

@@ -45,6 +45,7 @@ proc checkMem =
 # fps: flops per site
 # bps: bytes moved (load+store) per site
 # mm: memory footprint (bytes) per site
+var minNrep = int.high
 template bench(fps,bps,mm,eqn: untyped) {.dirty.} =
   block:
     let vol = lo.nSites.float
@@ -73,12 +74,19 @@ template bench(fps,bps,mm,eqn: untyped) {.dirty.} =
     inc nbench
     echo "bench: ",nbench|(-6), "secs: ", dt|(6,3), "  mf: ", mf|7, "  mb: ", mb|7, "  mem: ", mem, "  nrep: ", nrep
     echo exp2string(eqn), "\n"
+    minNrep = min(minNrep, nrep)
 template bench(fps,bps,eqn: untyped) =
   bench(fps,bps,0,eqn)
 
 proc test(lat:auto) =
-  let maxl = intParam("maxl",16)
-  if lat[0] > maxl: return
+  let minl = intParam("minl",0)
+  if lat[0] < minl: return
+  let maxl = intParam("maxl",0)
+  if maxl>0:
+    if lat[0] > maxl: return
+  else:
+    if minNrep == 1: return
+  minNrep = int.high
   var lo = newLayout(lat)
   template newCV: untyped = lo.ColorVector()
   template newCM: untyped = lo.ColorMatrix()

src/bench/benchStagProp.nim

@@ -63,7 +63,7 @@ threads:
   threadBarrier()
   echo r.norm2
 #echo v2
-echoTimers()
+echoProf()
 
 var g3:array[8,type(g[0])]
 for i in 0..3:
@@ -76,6 +76,6 @@ var s3 = newStag3(g3)
 s3.solve(v2, v1, mass, sp)
 resetTimers()
 s3.solve(v2, v1, mass, sp)
-echoTimers()
+echoProf()
 
 qexFinalize()

src/experimental/stagag.nim

@@ -2103,8 +2103,5 @@ if outfn != "":
   echo "Saving gauge field to file: ", outfn
   let err = g.saveGauge outfn
 
-case intParam("prof",0)
-of 1: echoTimers()
-of 2: echoHotspots()
-else: discard
+echoProf()
 qexFinalize()

src/physics/qcdTypes.nim

@@ -416,7 +416,6 @@ macro makeConstructors(x: untyped): untyped =
   result = newStmtList()
   result.add getAst mp(ident(f&"S"), ident("S"&r&"V"), ident"result")
   result.add getAst mp(ident(f&"D"), ident("D"&r&"V"), ident"result")
-  const defPrec {.strdefine.} = "D"
   result.add getAst mp(ident(f), ident(defPrec&r&"V"), ident"result")
   # non-Simd versions
   result.add getAst mp(ident(f&"S1"), ident("S"&r), ident"result")

tests/base/tshift.nim

@@ -126,18 +126,20 @@ template testS16(Smd: typedesc) =
 qexInit()
 
 template makeSimdArrayX(T,N,B: untyped) {.dirty.} =
-  makeSimdArray(`T X`, N, B)
-  type T = Simd[`T X`]
-  template toDoubleImpl(x: `T X`): untyped = x  # always already double
+  #makeSimdArray(`T X`, N, B)
+  #type T = Simd[`T X`]
+  #template toDoubleImpl(x: `T X`): untyped = x  # always already double
+  type T = Simd[SimdArrayObj[N,B]]
 
 #testS1(float)
+
 makeSimdArrayX(SD1, 1, float)
 testS1(SD1)
 when declared(SimdD1):
   testS1(SimdD1)
 
-#makeSimdArrayX(SS1, 1, float32)
-#testS1(SS1)
+makeSimdArrayX(SS1, 1, float32)
+testS1(SS1)
 when declared(SimdS1):
   testS1(SimdS1)
 
@@ -146,31 +148,39 @@ testS2(SD2)
 when declared(SimdD2):
   testS2(SimdD2)
 
-#makeSimdArrayX(SS2, 2, float32)
-#testS2(SS2)
+makeSimdArrayX(SS2, 2, float32)
+testS2(SS2)
 when declared(SimdS2):
   testS2(SimdS2)
 
 makeSimdArrayX(SD4, 4, float)
 testS4(SD4)
 when declared(SimdD4):
   testS4(SimdD4)
+
+makeSimdArrayX(SS4, 4, float32)
+testS4(SS4)
 when declared(SimdS4):
   testS4(SimdS4)
 
 makeSimdArrayX(SD8, 8, float)
 testS8(SD8)
 when declared(SimdD8):
   testS8(SimdD8)
+
+makeSimdArrayX(SS8, 8, float32)
+testS8(SS8)
 when declared(SimdS8):
   testS8(SimdS8)
 
 makeSimdArrayX(SD16, 16, float)
 testS16(SD16)
 when declared(SimdD16):
   testS16(SimdD16)
+
+makeSimdArrayX(SS16, 16, float32)
+testS16(SS16)
 when declared(SimdS16):
   testS16(SimdS16)
 
-
 qexFinalize()