speed a few kernels up and GC before testing

nnagyd · Aug 21, 2020 · fc55f01 · fc55f01
1 parent 709b933
commit fc55f01
Show file tree

Hide file tree

Showing 7 changed files with 32 additions and 30 deletions.
diff --git a/Keller_Miksis_RK45/CPU_Julia/CPU_KellerMiksis_julia.jl b/Keller_Miksis_RK45/CPU_Julia/CPU_KellerMiksis_julia.jl
@@ -1,4 +1,4 @@
-using DifferentialEquations, DelimitedFiles, Plots, CPUTime
+using DifferentialEquations, DelimitedFiles, Plots, CPUTime, MuladdMacro
 
 const numberOfRuns = 3
 const numberOfParameters = 256
@@ -63,7 +63,7 @@ end
 
 #ensemble problem
 function prob_func!(problem,i,repeat)
-    @inbounds begin
+    @inbounds @muladd begin
         #calculating indexes
         problem.u0[1] = initialValues[i,1]
         problem.u0[2] = initialValues[i,2]
@@ -110,6 +110,7 @@ res = solve(
 	dense = false,
 	maxiters = 1e10,
 	dtmin = 1e-10)
+GC.gc()
 
 #solving ODE 3x and measuring elapsed CPU time
 times = Vector{Float64}(undef,numberOfRuns)

diff --git a/Keller_Miksis_RK45/GPU_Julia/GPU_KellerMiksis_julia.jl b/Keller_Miksis_RK45/GPU_Julia/GPU_KellerMiksis_julia.jl
@@ -116,6 +116,7 @@ res = solve(
 	dense = false,
 	maxiters = 1e10,
 	dtmin = 1e-10)
+GC.gc()
 
 #solving ODE 3x and measuring elapsed CPU time
 times = Vector{Float64}(undef,numberOfRuns)

diff --git a/Lorenz_RK4/CPU_Julia/CPU_Lorenz_julia.jl b/Lorenz_RK4/CPU_Julia/CPU_Lorenz_julia.jl
@@ -1,4 +1,4 @@
-using DifferentialEquations, CPUTime, Statistics
+using DifferentialEquations, CPUTime, Statistics, MuladdMacro
 
 #settings
 const unroll = 128
@@ -17,12 +17,11 @@ end
 
 #ODE
 function lorenz!(du, u, p, t)
-  @inbounds for j in 1:unroll
-      index = 3(j-1)+1
-      du[index] = 10.0 * (u[index+1] - u[index])
-      du[index+1] = p[j] * u[index] - u[index+1] - u[index] * u[index+2]
-      du[index+2] = u[index] * u[index+1] - 2.66666666 * u[index+2]
-  end
+    @muladd @inbounds for j in 1:unroll
+      du[j] = 10.0 * (u[j+unroll] - u[j])
+      du[j+unroll] = p[j] * u[j] - u[j+unroll] - u[j] * u[j+2*unroll]
+      du[j+2*unroll] = u[j] * u[j+unroll] - 2.66666666 * u[j+2*unroll]
+   end
 end
 
 #compile
@@ -39,6 +38,7 @@ solve(
   dt = 0.01,
   unstable_check = nocheck
   )
+GC.gc()
 
 #simulation
 times = Vector{Float64}(undef,numberOfRuns)

diff --git a/Lorenz_RK4/CPU_Julia/CPU_Lorenz_julia_ensemble.jl b/Lorenz_RK4/CPU_Julia/CPU_Lorenz_julia_ensemble.jl
@@ -1,4 +1,4 @@
-using DifferentialEquations, CPUTime, Statistics
+using DifferentialEquations, CPUTime, Statistics, MuladdMacro
 
 #settings
 const unroll = 128
@@ -14,12 +14,11 @@ u0 = ones(3*unroll,1)*10
 
 #ODE
 function lorenz!(du, u, p, t)
-  @inbounds for j in 1:128
-    index = 3(j-1)+1
-    du[index] = 10.0 * (u[index+1] - u[index])
-    du[index+1] = p[j] * u[index] - u[index+1] - u[index] * u[index+2]
-    du[index+2] = u[index] * u[index+1] - 2.66666666 * u[index+2]
-  end
+    @muladd @inbounds for j in 1:unroll
+      du[j] = 10.0 * (u[j+unroll] - u[j])
+      du[j+unroll] = p[j] * u[j] - u[j+unroll] - u[j] * u[j+2*unroll]
+      du[j+2*unroll] = u[j] * u[j+unroll] - 2.66666666 * u[j+2*unroll]
+   end
 end
 
 #parameter change function
@@ -48,6 +47,7 @@ solve(
   dense = false,
   unstable_check = nocheck,
   )
+GC.gc()
 
 #simulation
 times = Vector{Float64}(undef,numberOfRuns)

diff --git a/Lorenz_RK4/CPU_Julia/CPU_Lorenz_simple_julia.jl b/Lorenz_RK4/CPU_Julia/CPU_Lorenz_simple_julia.jl
@@ -1,4 +1,4 @@
-using DifferentialEquations, CPUTime, Statistics, SimpleDiffEq
+using DifferentialEquations, CPUTime, Statistics, SimpleDiffEq, MuladdMacro
 
 #settings
 const unroll = 128
@@ -16,12 +16,11 @@ end
 
 #ODE
 function lorenz!(du, u, p, t)
-  @inbounds for j in 1:unroll
-      index = 3(j-1)+1
-      du[index] = 10.0 * (u[index+1] - u[index])
-      du[index+1] = p[j] * u[index] - u[index+1] - u[index] * u[index+2]
-      du[index+2] = u[index] * u[index+1] - 2.66666666 * u[index+2]
-  end
+    @muladd @inbounds for j in 1:unroll
+      du[j] = 10.0 * (u[j+unroll] - u[j])
+      du[j+unroll] = p[j] * u[j] - u[j+unroll] - u[j] * u[j+2*unroll]
+      du[j+2*unroll] = u[j] * u[j+unroll] - 2.66666666 * u[j+2*unroll]
+   end
 end
 
 #compile
@@ -35,9 +34,9 @@ solve(
   save_start = false,
   save_end = true,
   adaptive = false,
-  dt = 0.01,
-  unstable_check = nocheck
+  dt = 0.01
   )
+GC.gc()
 
 #simulation
 times = Vector{Float64}(undef,numberOfRuns)
@@ -56,8 +55,7 @@ for runs in 1:numberOfRuns
       save_start = false,
       save_end = true,
       adaptive = false,
-      dt = 0.01,
-      unstable_check = nocheck
+      dt = 0.01
       )
   end
 

diff --git a/Lorenz_RK4/GPU_Julia/GPU_Lorenz_julia.jl b/Lorenz_RK4/GPU_Julia/GPU_Lorenz_julia.jl
@@ -1,12 +1,12 @@
-using DifferentialEquations, DiffEqGPU, CUDA, CPUTime, Statistics, SimpleDiffEq
+using DifferentialEquations, DiffEqGPU, CUDA, CPUTime, Statistics, MuladdMacro
 
 #settings
 const numberOfParameters = 46080
 const unroll = 2
 const numberOfTrajectories = Int64(numberOfParameters/unroll)
 const batchSize = numberOfTrajectories
 const numberOfRuns = 2
-const gpuID = 0 #Nvidia titan black device
+const gpuID = 1 #Nvidia titan black device
 
 #select device
 CUDA.device!(gpuID)
@@ -29,7 +29,7 @@ end
 parameterList = range(0.0,stop = 21.0,length=numberOfParameters)
 p = zeros(unroll,1)
 tspan = (0.0,10.0)
-u0 = ones(unroll*3,1)
+u0 = ones(unroll*3)
 
 #parameter change function
 #parameterChange = (prob,i,repeat) -> remake(prob,p=parameterList[i]) #it has slightly more allocations
@@ -59,6 +59,7 @@ solve(
   dt = 0.01,
   dense = false
   )
+GC.gc()
 
 #simulation
 times = Vector{Float64}(undef,numberOfRuns)

diff --git a/Lorenz_RK4/GPU_Julia/GPU_Lorenz_simple_julia.jl b/Lorenz_RK4/GPU_Julia/GPU_Lorenz_simple_julia.jl
@@ -59,6 +59,7 @@ solve(
   dt = 0.01,
   dense = false
   )
+GC.gc()
 
 #simulation
 times = Vector{Float64}(undef,numberOfRuns)