diff --git a/src/gauge/fat7l.nim b/src/gauge/fat7l.nim
index c845968..2be6fb9 100644
--- a/src/gauge/fat7l.nim
+++ b/src/gauge/fat7l.nim
@@ -165,6 +165,7 @@ when isMainModule:
   import qex
   import physics/qcdTypes
   import gauge
+  import strformat
   qexInit()
   #var defaultGaugeFile = "l88.scidac"
   let defaultLat = @[8,8,8,8]
@@ -174,19 +175,50 @@ when isMainModule:
 
   var info: PerfInfo
   var coef: Fat7lCoefs
-  coef.oneLink = 1.0
-  coef.threeStaple = 1.0
-  coef.fiveStaple = 1.0
-  coef.sevenStaple = 1.0
-  coef.lepage = 1.0
   var naik = 1.0
 
   var fl = lo.newGauge()
   var ll = lo.newGauge()
   var g3 = g
 
+  proc checkfat1(name: string) =
+    makeImpLinks(fl, g, coef, info)
+    let p = fl.plaq
+    let c = coef.oneLink + 6*coef.threeStaple + 24*coef.fiveStaple +
+            48*coef.sevenStaple + 6*coef.lepage
+    let s = pow(c,4)/6.0
+    var e2 = 0.0
+    for x in p:
+      e2 += (x-s)^2
+    echo &"{name:12} relerr: {sqrt(e2)/s}"
+
+  coef.oneLink = 1.0
+  checkfat1("oneLink")
+  coef.oneLink = 0.0
+  coef.threeStaple = 1.0
+  checkfat1("threeStaple")
+  coef.threeStaple = 0.0
+  coef.fiveStaple = 1.0
+  checkfat1("fiveStaple")
+  coef.fiveStaple = 0.0
+  coef.sevenStaple = 1.0
+  checkfat1("sevenStaple")
+  coef.sevenStaple = 0.0
+  coef.lepage = 1.0
+  checkfat1("lepage")
+  coef.oneLink = 1.0
+  coef.threeStaple = 1.0
+  coef.fiveStaple = 1.0
+  coef.sevenStaple = 1.0
+  coef.lepage = 1.0
+  checkfat1("all")
+
+  makeImpLinks(fl, g, coef, ll, g3, naik, info)
+  echo g.plaq
+  echo fl.plaq
+  echo ll.plaq
+#[
   echo g.plaq
-  makeImpLinks(fl, g, coef, info)
   makeImpLinks(fl, g, coef, ll, g3, naik, info)
   echo fl.plaq
   echo ll.plaq
@@ -195,5 +227,5 @@ when isMainModule:
   echo pow(1.0+6.0+6.0*4.0,4)/6.0
   echo pow(1.0+6.0+6.0*4.0+6.0*4.0*2.0,4)/6.0
   echo pow(1.0+6.0+6.0*4.0+6.0*4.0*2.0+6.0,4)/6.0
-
+]#
   qexFinalize()
diff --git a/src/gauge/fat7lderiv.nim b/src/gauge/fat7lderiv.nim
index 5ba34fd..3722664 100644
--- a/src/gauge/fat7lderiv.nim
+++ b/src/gauge/fat7lderiv.nim
@@ -119,7 +119,7 @@ side_force(QDP_ColorMatrix *force, QDP_ColorMatrix *bot0, QDP_ColorMatrix *side0
 }
 ]#
 
-proc fat7lDeriv(deriv: auto, gauge: auto, mid: auto, coef: Fat7lCoefs,
+proc fat7lDeriv*(deriv: auto, gauge: auto, mid: auto, coef: Fat7lCoefs,
                 llderiv: auto, llgauge: auto, llmid: auto, naik: float,
                 perf: var PerfInfo) =
   tic("fat7lDeriv")
@@ -328,7 +328,23 @@ 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,
+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*(deriv: auto, gauge: auto, mid: auto, coef: Fat7lCoefs,
                 perf: var PerfInfo) =
   fat7lDeriv(deriv, gauge, mid, coef, deriv, gauge, mid, 0.0, perf)
 
@@ -339,54 +355,103 @@ when isMainModule:
   qexInit()
   #var defaultGaugeFile = "l88.scidac"
   let defaultLat = @[8,8,8,8]
-  defaultSetup()
-  for mu in 0..<g.len: g[mu] := 1
-  #g.random
+  #defaultSetup()
 
-  var info: PerfInfo
-  var coef: Fat7lCoefs
+  var
+    (lo, g, r) = setupLattice(defaultLat)
+    fl = lo.newGauge()
+    fl2 = lo.newGauge()
+    ll = lo.newGauge()
+    ll2 = lo.newGauge()
+    ch = lo.newGauge()
+    g2 = lo.newGauge()
+    dg = lo.newGauge()
+    fd = lo.newGauge()
+    ld = lo.newGauge()
+    gc = GaugeActionCoeffs(plaq:1.0)
+    info: PerfInfo
+    coef: Fat7lCoefs
+    naik = 1.0
+    eps = floatParam("eps", 1e-5)
   coef.oneLink = 1.0
+  coef.threeStaple = 0.0
+  coef.fiveStaple = 0.0
+  coef.sevenStaple = 0.0
+  coef.lepage = 0.0
+
+  #g.unit
+  g.random r
+  #dg.randomTAH r
+  dg.gaussian r
+  for mu in 0..<g.len:
+    for e in g[mu]:
+      dg[mu][e] *= eps
+      #g2[mu][e] := exp(dg[mu][e]) * g[mu][e]
+      g2[mu][e] := g[mu][e] + dg[mu][e]
+
+  proc check(da: float, tol: float) =
+    var ds = 0.0
+    for mu in 0..3:
+      ds += redot(dg[mu], fd[mu])
+    let r = (da-ds)/da
+    #echo "Checking ", name
+    echo &"  da {da}  ds {ds}  rel {r}"
+    if abs(r) > tol*eps:
+      echo &"> ERROR rel error |{r}| > {tol*eps}"
+
+  proc checkG =
+    echo "Checking GaugeForce"
+    for mu in 0..<fd.len:
+      fd[mu] := 0
+    let a = gc.gaugeAction2(g)
+    let a2 = gc.gaugeAction2(g2)
+    gc.gaugeDeriv2(g, fd)
+    check(a2-a, 1)
+
+  proc checkS(name: string, tol: float) =
+    echo "Checking ", name
+    for mu in 0..<fd.len:
+      ch[mu] := 0
+      fd[mu] := 0
+    makeImpLinks(fl, g, coef, info)
+    info.clear
+    resetTimers()
+    makeImpLinks(fl, g, coef, info)
+    echo "  ", info
+    makeImpLinks(fl2, g2, coef, info)
+    echo "  ", info
+    let a = gc.gaugeAction2(fl)
+    let a2 = gc.gaugeAction2(fl2)
+    gc.gaugeDeriv2(fl, ch)
+    fat7lderiv(fd, g, ch, coef, info)
+    check(a2-a, tol)
+
+  checkG()
+
+  checkS("oneLink", 1)
+  coef.oneLink = 0.0
   coef.threeStaple = 1.0
+  checkS("threeStaple", 2)
+  coef.threeStaple = 0.0
   coef.fiveStaple = 1.0
+  checkS("fiveStaple", 2)
+  coef.fiveStaple = 0.0
   coef.sevenStaple = 1.0
+  checkS("sevenStaple", 3)
+  coef.sevenStaple = 0.0
   coef.lepage = 1.0
-  var naik = 1.0
-
-  var fl = lo.newGauge()
-  var fl2 = lo.newGauge()
-  var ll = lo.newGauge()
-  var ll2 = lo.newGauge()
-  var dfl = lo.newGauge()
-  var g2 = lo.newGauge()
-  var dg = lo.newGauge()
-  var fd = lo.newGauge()
-  var ld = lo.newGauge()
-  for mu in 0..<dg.len:
-    dg[mu] := 0.00001 * g[mu]
-    g2[mu] := g[mu] + dg[mu]
-    fd[mu] := 0
-
-  echo g.plaq
-  echo g2.plaq
-  makeImpLinks(fl, g, coef, info)
-  info.clear
-  resetTimers()
-  makeImpLinks(fl, g, coef, info)
-  echo info
-  makeImpLinks(fl2, g2, coef, info)
-  echo info
-  echo fl.plaq
-  echo fl2.plaq
+  checkS("Lepage", 1)
+  coef.oneLink = 1.0
+  coef.threeStaple = 1.0
+  coef.fiveStaple = 1.0
+  coef.sevenStaple = 1.0
+  coef.lepage = 0.0
+  checkS("not Lepage", 60)
+  coef.lepage = 1.0
+  checkS("all", 60)
 
-  fat7lderiv(fd, g, dg, coef, info)
-  echo info
-  for mu in 0..3:
-    dfl[mu] := fl2[mu] - fl[mu]
-    #echo dfl[mu].norm2
-    #echo fd[mu].norm2
-    dfl[mu] -= fd[mu]
-    echo dfl[mu].norm2
 
+#[
   for mu in 0..<dg.len:
     fd[mu] := 0
     ld[mu] := 0
@@ -403,6 +468,7 @@ when isMainModule:
     dfl[mu] := ll2[mu] - ll[mu]
     dfl[mu] -= ld[mu]
     echo dfl[mu].norm2
+]#
 
   echoProf()
   qexFinalize()
diff --git a/src/gauge/hisqsmear.nim b/src/gauge/hisqsmear.nim
new file mode 100644
index 0000000..f9b9de2
--- /dev/null
+++ b/src/gauge/hisqsmear.nim
@@ -0,0 +1,94 @@
+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()
+    g2 = lo.newGauge()
+    sg2 = lo.newGauge()
+    sgl2 = lo.newGauge()
+    dg = lo.newGauge()
+    fd = lo.newGauge()
+  g.random
+  for mu in 0..<dg.len:
+    dg[mu] := 0.00001 * g[mu]
+    g2[mu] := g[mu] + dg[mu]
+    fd[mu] := 0
+
+  echo g.plaq
+  echo g2.plaq
+  #hisq.smear(g,sg,sgl)
+  var force = hisq.smearGetForce(g,sg,sgl)
+  f.force(dg,dg)
+  hisq.smear(g2,sg2,sgl2)
+
+  echo f.plaq
+  echo sg.plaq
+  echo sgl.plaq
+  echo "--"
+  echo sg.plaq
+  echo sgl.plaq
+  ff.gaugeForce(sg)
+  ffl.gaugeForce(sgl)
+  echo "--"
+  echo f.plaq
+  echo sg.plaq
+  echo sgl.plaq
+  qexFinalize()