improve benchGauge

Author: jcosborn

src/base/flopcount.nim

@@ -0,0 +1,15 @@
+# flops counts
+# all names start with 'f' for flops
+
+# Complex number (fc...)
+template fcadd*:int = 2
+template fcmul*:int = 6
+template fcredot*:int = 3
+
+# Complex matrix (fcm...)
+template fcmadd*(n:int):int = n*n*fcadd
+template fcmmul*(n:int):int = n*n*(n*fcmul + (n-1)*fcadd)
+template fcmredot*(n:int):int = n*n*fcredot + (n*n-1)
+
+# single plaquette: redot(A*B,C*D)
+template fplaq*(n:int):int = 2*fcmmul(n) + fcmredot(n)

src/base/profile.nim

@@ -857,7 +857,8 @@ proc echoHotspots* =
       else:
         tsns += nk.ns
         let tsnspct = 100.0 * tsns / nstot
-        echo &"{pct:6.3f} {tsnspct:7.3f} {count} {mf} {nc} S {lc} {nm}"
+        #echo &"{pct:6.3f} {tsnspct:7.3f} {count} {mf} {nc} S {lc} {nm}"
+        echo &"{pct:6.3f} {tsnspct:7.3f} {count}         {nc} S {lc} {nm}"
 
 proc echoProf*(def = 0) =
   case intParam("prof",def)

src/bench/benchGauge.nim

@@ -6,6 +6,7 @@ import parseUtils
 import macros
 import gauge/hypsmear
 import commonBench
+import base/flopcount
 
 proc checkMem =
   #echo("mem: (used+free)/total: (", getOccupiedMem(), "+", getFreeMem(), ")/",
@@ -37,6 +38,9 @@ template bench(fps,bps:SomeNumber; eqn:untyped) =
   var t1 = epochTime()
   let dt = t1 - t0
 ]#
+  block:
+    eqn
+  resetTimers()
   let br = benchSingle:
     eqn
   let (nrep,dt) = (br.reps,br.secs)
@@ -45,6 +49,7 @@ template bench(fps,bps:SomeNumber; eqn:untyped) =
   let mb = (nrep.float*bytes)/(1e6*dt)
   echo "(", exp2string(eqn), ") secs: ", dt|(5,3), "  sec/n: ", dtn|(5,3),
       "  mf: ", mf.int, "  mb: ", mb.int
+  echoProf()
 
 proc test(lat: auto) =
   #var scale = 1
@@ -86,12 +91,16 @@ proc test(lat: auto) =
 
   resetTimers()
 
-  bench(np*(2*8*nc*nc*nc-1), nd*2*nc*nc*sizeof(numberType(g[0][0]))):
-    var pl = plaq(g)
-  bench(np*(2*8*nc*nc*nc-1), nd*2*nc*nc*sizeof(numberType(g[0][0]))):
-    var pl2 = plaq2(g)
-  bench(np*(2*8*nc*nc*nc-1), nd*2*nc*nc*sizeof(numberType(g[0][0]))):
-    var pl3 = plaq3(g)
+  block:
+    let flop = np*(fplaq(nc)+1)
+    #let mem = nd*2*nc*nc*sizeof(numberType(g[0][0]))
+    let mem = nd*sizeof(g[0][0]) div g[0][0][0,0].re.numNumbers
+    bench(flop, mem):
+      var pl = plaq(g)
+    bench(flop, mem):
+      var pl2 = plaq2(g)
+    bench(flop, mem):
+      var pl3 = plaq3(g)
 
   bench(np*(2*8*nc*nc*nc-1), nd*2*nc*nc*sizeof(numberType(g[0][0]))):
     var ga = gaugeAction1(g)

src/gauge/gaugeUtils.nim

@@ -287,7 +287,7 @@ proc plaq*[T](uu: openArray[T]): auto =
       rankSum(pl)
     toc("sum")
   result = pl
-  toc("end", flops=lo.nSites.float*float(np*(2*8*nc*nc*nc-1)))
+  toc("end", flops=lo.nSites.float*float(np*(fplaq(nc)+1)))
 
 template plaq*(g: Gauge): auto = plaq(g.u)
 
@@ -329,6 +329,7 @@ proc plaq2*[T](gg:openArray[T]):auto =
   let lo = g[0].l
   let nd = lo.nDim
   let nc = g[0][0].ncols
+  #let np = (nd*(nd-1)) div 2
   var m = lo.ColorMatrix()
   var s0 = lo.ColorMatrix()
   #var t0 = lo.ColorMatrix()
@@ -353,7 +354,7 @@ proc plaq2*[T](gg:openArray[T]):auto =
         m += (g[mu]*s0) * (g[nu]*s1).adj
         #m += (g[mu]*s0) * (g[nu]*s1)
         #echo mu, " ", nu, " ", trace(m)/nc
-        toc("mul")
+        toc("mul", flops=lo.nSites.float*float(3*fcmmul(nc)+fcmadd(nc)))
     toc("work")
     tr = trace(m)
     toc("trace")
@@ -379,13 +380,13 @@ proc plaq3*[T](g: seq[T]): auto =
         tic("plaq3 loop")
         #m += (t[mu]^*g[nu]) * (t[nu]^*g[mu]).adj
         discard t[mu]^*!g[nu]
-        toc("transport1")
+        toc("transport1", flops=lo.nSites.float*float(fcmmul(nc)))
         discard t[nu]^*!g[mu]
         threadBarrier()
-        toc("transport2")
+        toc("transport2", flops=lo.nSites.float*float(fcmmul(nc)))
         m += t[mu].field * t[nu].field.adj
         #echo mu, " ", nu, " ", trace(m)/nc
-        toc("mul")
+        toc("mul", flops=lo.nSites.float*float(fcmmul(nc)+fcmadd(nc)))
     toc("work")
     tr = trace(m)
     toc("trace")