From 93b94c8e15ccc7bbb210de49a090a258a526835e Mon Sep 17 00:00:00 2001
From: ctpeterson <curtistaylorpetersonwork@gmail.com>
Date: Thu, 12 Dec 2024 11:22:23 -0500
Subject: [PATCH] Added interface for HISQ smearing

---
 src/gauge/fat7lderiv.nim | 27 ++++++++++++--
 src/gauge/hisqsmear.nim  | 80 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 104 insertions(+), 3 deletions(-)
 create mode 100644 src/gauge/hisqsmear.nim

diff --git a/src/gauge/fat7lderiv.nim b/src/gauge/fat7lderiv.nim
index 5ba34fd..fa55165 100644
--- a/src/gauge/fat7lderiv.nim
+++ b/src/gauge/fat7lderiv.nim
@@ -328,10 +328,29 @@ proc fat7lDeriv(deriv: auto, gauge: auto, mid: auto, coef: Fat7lCoefs,
   perf.flops += nflops * gauge[0].l.localGeom.prod
   perf.secs += getElapsedTime()
 
-proc fat7lDeriv(deriv: auto, gauge: auto, mid: auto, coef: Fat7lCoefs,
-                perf: var PerfInfo) =
-  fat7lDeriv(deriv, gauge, mid, coef, deriv, gauge, mid, 0.0, perf)
+proc fat7lDeriv*(
+    mid: auto, 
+    deriv: auto, 
+    gauge: auto, 
+    coef: Fat7lCoefs,
+    llderiv: auto, 
+    llgauge: auto, 
+    naik: float,
+    perf: var PerfInfo
+  ) =
+  var (fx,fxl) = (newOneOf(deriv),newOneOf(llderiv))
+  fat7lderiv(fx,gauge,deriv,coef,fxl,llgauge,llderiv,naik,perf)
+  threads:
+    for mu in 0..<mid.len:
+      for s in mid[mu]: mid[mu][s] := fx[mu][s] + fxl[mu][s]
 
+proc fat7lDeriv*(
+    mid: auto, 
+    deriv: auto, 
+    gauge: auto, 
+    coef: Fat7lCoefs,
+    perf: var PerfInfo
+  ) = fat7lderiv(mid,gauge,deriv,coef,mid,gauge,deriv,0.0,perf)
 
 when isMainModule:
   import qex, physics/qcdTypes
@@ -366,6 +385,7 @@ when isMainModule:
     g2[mu] := g[mu] + dg[mu]
     fd[mu] := 0
 
+  #[
   echo g.plaq
   echo g2.plaq
   makeImpLinks(fl, g, coef, info)
@@ -386,6 +406,7 @@ when isMainModule:
     #echo fd[mu].norm2
     dfl[mu] -= fd[mu]
     echo dfl[mu].norm2
+  ]#
 
   for mu in 0..<dg.len:
     fd[mu] := 0
diff --git a/src/gauge/hisqsmear.nim b/src/gauge/hisqsmear.nim
new file mode 100644
index 0000000..cfced90
--- /dev/null
+++ b/src/gauge/hisqsmear.nim
@@ -0,0 +1,80 @@
+import qex
+import physics/[hisqLinks]
+import gauge
+import gauge/[fat7l,fat7lderiv]
+
+export hisqLinks
+
+proc newHISQ*(lepage: float = 0.0; naik: float = 1.0): HisqCoefs =
+  result = HisqCoefs(naik: -naik/24.0)
+  result.fat7first.setHisqFat7(lepage,0.0)
+  result.fat7second.setHisqFat7(2.0-lepage,naik)
+
+proc smearGetForce*[T](
+    self: HisqCoefs; 
+    u: T; 
+    su,sul: T;
+    displayPerformance: bool = false
+  ): proc(dsdu: var T; dsdsu,dsdsul: T) =
+  mixin projectU,projectUderiv
+  let
+    lo = u[0].l
+    fat7l1 = self.fat7first
+    fat7l2 = self.fat7second
+    naik = self.naik
+  var
+    v = newOneOf(u)
+    w = newOneOf(u)
+    info: PerfInfo
+  
+  # Smear
+  v.makeImpLinks(u,fat7l1,info) # First fat7
+  threads: # Unitary projection
+    for mu in 0..<w.len: 
+      for s in w[mu]: w[mu][s].projectU(v[mu][s])
+  makeImpLinks(su,w,fat7l2,sul,w,naik,info) # Second fat7
+
+  # Chain rule - retains a reference to u,su,sul
+  proc smearedForce(dsdu: var T; dsdsu,dsdsul: T) =
+    var 
+      dsdx_dxdw = newOneOf(dsdu)
+      dsdx_dxdw_dwdv = newOneOf(dsdu)
+    dsdx_dxdw.fat7lderiv(dsdsu,su,fat7l2,dsdsul,sul,naik,info) # Second fat7
+    threads: # Unitary projection
+      for mu in 0..<dsdx_dxdw_dwdv.len:
+        for s in dsdx_dxdw_dwdv[mu]:
+          dsdx_dxdw_dwdv[mu][s].projectUderiv(w[mu][s],v[mu][s],dsdx_dxdw[mu][s]) 
+    dsdu.fat7lderiv(dsdx_dxdw_dwdv,u,fat7l1,info) # First fat7
+  
+  if displayPerformance: echo $(info)
+  return smearedForce
+
+if isMainModule:
+  qexInit()
+  let 
+    defaultLat = @[8,8,8,8]
+    hisq = newHISQ()
+  defaultSetup()
+  var
+    sg = lo.newGauge()
+    sgl = lo.newGauge()
+    f = lo.newGauge()
+    ff = lo.newGauge()
+    ffl = lo.newGauge()
+  g.random
+  echo f.plaq
+  echo sg.plaq
+  echo sgl.plaq
+  hisq.smear(g,sg,sgl)
+  echo "--"
+  echo sg.plaq
+  echo sgl.plaq
+  ff.gaugeForce(sg)
+  ffl.gaugeForce(sgl)
+  var force = hisq.smearGetForce(g,sg,sgl)
+  f.force(ff,ffl)
+  echo "--"
+  echo f.plaq
+  echo sg.plaq
+  echo sgl.plaq
+  qexFinalize()
\ No newline at end of file