Typo in io and utils and Timings in meas

src/io.jl

@@ -127,7 +127,7 @@ function save_data()
         BDIO_start_record!(fb, BDIO_BIN_GENERIC, 1)
         BDIO_write!(fb, [convert(Int32, 4)])
         BDIO_write!(fb, [convert(Int32, lp.iL[i]) for i in 1:4])
-        BDIO_write!(fb, [dpar.m0, dpar.csw, params["Fermion"]["Cfl"]])
+        BDIO_write!(fb, [dpar.m0, dpar.csw])
         BDIO_write!(fb, [dpar.th[i] for i in 1:4])
         BDIO_write!(fb, [convert(Int32,length(flow_times))])
         length(flow_times) > 0 ? BDIO_write!(fb, flow_times) : nothing

src/meas.jl

@@ -88,13 +88,13 @@ function two_pt()
             end
         end
 
-        copyto!(U,U_CPU)
+        @timeit "CPU to GPU" copyto!(U,U_CPU)
     end
 
     pp_corr_t0 .= pp_corr_t0./prod(lp.iL[1:3])
     ap_corr_t0 .= ap_corr_t0./prod(lp.iL[1:3])
     pp_corr_t .= pp_corr_t./prod(lp.iL[1:3])
-    ap_corr_t .= pp_corr_t./prod(lp.iL[1:3])
+    ap_corr_t .= ap_corr_t./prod(lp.iL[1:3])
 
     println(log_file,"Finished measuring 2pt")
     flush(log_file)
@@ -112,7 +112,7 @@ The final result is 'Quark_cond' + C_fl*'Quark_cond_cfl'
 """
 function one_pt()
 
-    copyto!(U,U_CPU)
+    @timeit "CPU to GPU" copyto!(U,U_CPU)
 
     for fl in 1:length(flow_times)
         println(log_file,"Measuring 1pt at t = "*string(flow_times[fl])*"...")
@@ -124,20 +124,22 @@ function one_pt()
 
             backflow(psi, U, flow_times[fl], params["Measurements"]["Nsaves"], gp, dpar, lp, ymws, dws)
 
-            psi_CPU .= Array(psi)
+            @timeit "GPU to CPU" psi_CPU .= Array(psi)
 
-            dws.sp .= dmul.(Gamma{5},psi)
-            g5Dw!(psi,U,dws.sp,dpar,dws,lp)
+            @timeit "Inversion" begin
+                dws.sp .= dmul.(Gamma{5},psi)
+                g5Dw!(psi,U,dws.sp,dpar,dws,lp)
 
-            niter = CG!(psi,U,DwdagDw!,dpar,lp,dws,params["Measurements"]["maxiter"], params["Measurements"]["tolerance"])
+                niter = CG!(psi,U,DwdagDw!,dpar,lp,dws,params["Measurements"]["maxiter"], params["Measurements"]["tolerance"])
+            end
 
             println(log_file,"CG converged after ",niter," iterations with residue ",CUDA.mapreduce(x -> norm2(x), +, dws.sr))
 
-            copyto!(dws.sp,psi_CPU)
+            @timeit "CPU to GPU" copyto!(dws.sp,psi_CPU)
             Quark_cond[noi,fl] = sum(dot.(dws.sp,psi))
-            Quark_cond_cfl[noi,fl] = sum(-norm2.(dws.sp))
+            Quark_cond_cfl[noi,fl] = -sum(norm2.(dws.sp))
 
-            copyto!(U,U_CPU)
+            @timeit "CPU to GPU" copyto!(U,U_CPU)
 
         end
     end

src/utils.jl

@@ -32,7 +32,7 @@ function read_ff(name::String)
     TvecC = Vector{ComplexF64}(undef,iL[4]);
     CN = Vector{ComplexF64}(undef,1);
     RN = Vector{Float64}(undef,1);
-    ET = Vector{ComplexF64}(undef,nflow+1);
+    ET = Vector{ComplexF64}(undef,1+nflow);
 
     pp_corr = Array{Float64}(undef,1,iL[4],N_noise);
     ap_corr = Array{ComplexF64}(undef,1,iL[4],N_noise);
@@ -80,8 +80,10 @@ function read_ff(name::String)
                 Phat_t[1,noi,fl]= RN[1]
             end
         end
-        BDIO_read(file,ET)
-        Eoft[1,:] .= ET
+        for time in 1:iL[4]
+            BDIO_read(file,ET)
+            Eoft[1,:,time] .= ET
+        end
         Eoft_mc=[Eoft_mc;Eoft]
 
         pp_corr_mc=[pp_corr_mc;pp_corr]