mpi

genax.py

 import cupy as cp, matplotlib.pyplot as plt
 import cupy as cp, matplotlib.pyplot as plt
 from numba import cuda
 from numba import cuda
 from utils import expSpace
 from utils import expSpace
+from mpi import size, rank
 
 
 def gen_angels_to_pick(n, plot = False):
 def gen_angels_to_pick(n, plot = False):
     if n % 2 == 0:
     if n % 2 == 0:
@@ -67,7 +68,28 @@ void parosit2(const double* x1, const double* x2, double* a, double* b, const in
 }
 }
 ''', 'parosit2')
 ''', 'parosit2')
 
 
-def angles_alap(anglestopick, plot = False):
+parosit2_mpi = cp.RawKernel(r'''
+extern "C" 
+__global__ 
+void parosit2_mpi(const double* x1, const double* x2, double* a, double* b, const int m, const int m2, const double PI) {
+    int tid = blockDim.x * blockIdx.x + threadIdx.x;
+    if (m <= tid || m*m2 <= tid*m+m2-1)
+        return;
+    double alpha = x1[tid];
+    for (int i = 0; i < m2; i++) {
+        double betha = x2[i];
+        if ((alpha + betha) < PI && alpha > 0.0) {
+            a[tid*m+i] = alpha;
+            b[tid*m+i] = betha;
+        } else {
+            a[tid*m+i] = -1.0;
+            b[tid*m+i] = -1.0;
+        }
+    }
+}
+''', 'parosit2_mpi')
+
+def angles_alap(anglestopick):
     m = anglestopick.size
     m = anglestopick.size
     alpha_arr = cp.zeros((m, m), dtype=cp.float64)
     alpha_arr = cp.zeros((m, m), dtype=cp.float64)
     beta_arr = cp.zeros((m, m), dtype=cp.float64)
     beta_arr = cp.zeros((m, m), dtype=cp.float64)
@@ -103,12 +125,22 @@ def angles_alap(anglestopick, plot = False):
     return Cx, Cy
     return Cx, Cy
 
 
 
 
-def angles_ratet(anglestopick, plot = False):
+def angles_ratet(anglestopick, mpi):
     m = anglestopick.size
     m = anglestopick.size
-    alpha_arr = cp.zeros((m, m), dtype=cp.float64)
-    beta_arr = cp.zeros((m, m), dtype=cp.float64)
-    blocksize = int((m + 64 - 1) / 64)
-    parosit2((blocksize,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, cp.pi))
+    if not mpi:
+        alpha_arr = cp.zeros((m, m), dtype=cp.float64)
+        beta_arr = cp.zeros((m, m), dtype=cp.float64)
+        blocksize = int((m + 64 - 1) / 64)
+        parosit2((blocksize,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, cp.pi))
+    else:
+        also = int(rank/size * m)
+        felso = int((rank+1)/size * m)
+        anglestopick2 = anglestopick[also:felso]
+        m2 = anglestopick2.size
+        alpha_arr = cp.zeros((m, m2), dtype=cp.float64)
+        beta_arr = cp.zeros((m, m2), dtype=cp.float64)
+        blocksize = int((m + 64 - 1) / 64)
+        parosit2_mpi((blocksize,), (m,), (anglestopick, anglestopick2, alpha_arr, beta_arr, m, m2, cp.pi))
 
 
     tompa_beta_arr = beta_arr[beta_arr > cp.pi]
     tompa_beta_arr = beta_arr[beta_arr > cp.pi]
     tompa_beta_mpi_arr = tompa_beta_arr - cp.pi/2
     tompa_beta_mpi_arr = tompa_beta_arr - cp.pi/2

mpi.py

+try:
+    from mpi4py import MPI
+    comm = MPI.COMM_WORLD
+    rank = comm.Get_rank()
+    size = comm.Get_size()
+except:
+    comm = None
+    rank = None
+    size = None
\ No newline at end of file

tetrarun.py

@@ -10,9 +10,10 @@ def main(argv):
    v = 3
    v = 3
    w = 3
    w = 3
    PLOT = False
    PLOT = False
+   mpi = False
 
 
    try:
    try:
-      opts, args = getopt.getopt(argv,"hpn:v:w:d:o:")
+      opts, args = getopt.getopt(argv,"hpn:v:w:d:o:m:")
    except getopt.GetoptError as err:
    except getopt.GetoptError as err:
       print(err)
       print(err)
       print ('tetrarun.py -n <range division:int> -v <> -w <> -o <outputfile>')
       print ('tetrarun.py -n <range division:int> -v <> -w <> -o <outputfile>')
@@ -33,12 +34,14 @@ def main(argv):
          outputfile = arg
          outputfile = arg
       elif opt in ("-p", "--plot"):
       elif opt in ("-p", "--plot"):
          PLOT = True
          PLOT = True
+      elif opt in ("-m", "--mpi"):
+         mpi = True
 
 
    space = gen_angels_to_pick(n, PLOT)
    space = gen_angels_to_pick(n, PLOT)
 
 
-   Cx, Cy = angles_alap(space, PLOT)
+   Cx, Cy = angles_alap(space)
 
 
-   Dx, Dy, Dz = angles_ratet(space)
+   Dx, Dy, Dz = angles_ratet(space, mpi)
 
 
    res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
    res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
 
 
@@ -56,8 +59,8 @@ def main(argv):
    res.get(out=mtx_cpu)
    res.get(out=mtx_cpu)
    
    
    if outputfile:
    if outputfile:
-      writetofile(outputfile+'.full', Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu)
-      writetofile2(outputfile, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu)
+      writetofile(outputfile+'.full', Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi)
+      writetofile2(outputfile, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi)
 
 
    #printresults(res)
    #printresults(res)
    #print("Exact one 3-3")
    #print("Exact one 3-3")

utils.py

@@ -2,6 +2,7 @@ import cupy as cp
 from functools import wraps
 from functools import wraps
 import time
 import time
 import numpy as np
 import numpy as np
+from mpi import size, rank
 
 
 with open('filtering.cu') as f:
 with open('filtering.cu') as f:
     code = f.read()
     code = f.read()
@@ -56,12 +57,14 @@ def printresults(egyensulyi_mtx):
         print(np.resize(parok, (int(N/2), 2)))
         print(np.resize(parok, (int(N/2), 2)))
         print()
         print()
 
 
-def writetofile(filename, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu):
+def writetofile(filename, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi):
     size_C = Cx_cpu.size
     size_C = Cx_cpu.size
     size_D = Dx_cpu.size
     size_D = Dx_cpu.size
     lcm = compute_lcm(size_C, size_D)
     lcm = compute_lcm(size_C, size_D)
     pos = size_C * size_D
     pos = size_C * size_D
 
 
+    if mpi:
+        filename = f'R{rank}/{size}_' + filename
     f = open(filename, "w")
     f = open(filename, "w")
 
 
     for i in range(0, pos):
     for i in range(0, pos):
@@ -79,12 +82,15 @@ def writetofile(filename, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu):
 
 
     f.close()
     f.close()
 
 
-def writetofile2(filename, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu):
+def writetofile2(filename, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi):
     size_C = Cx_cpu.size
     size_C = Cx_cpu.size
     size_D = Dx_cpu.size
     size_D = Dx_cpu.size
     lcm = compute_lcm(size_C, size_D)
     lcm = compute_lcm(size_C, size_D)
     pos = size_C * size_D
     pos = size_C * size_D
 
 
+    if mpi:
+        filename = f'R{rank}/{size}_' + filename
+
     f = open(filename, "w")
     f = open(filename, "w")
     for i in range(0, pos):
     for i in range(0, pos):
         s = 0
         s = 0