attributes

gpu.py

@@ -12,10 +12,10 @@ fun = ep_pontok_module.get_function(kers[0])
 def start_kernel(Cx, Cy, Dx, Dy, Dz, v, w):
     print(f"Cnt: {Cx.size}x{Dx.size}={Cx.size*Dx.size}")
     print("Res size (byte): ", Cx.size*Dx.size*4*4)
+    print(fun.attributes)
     lcm = compute_lcm(Cx.size, Dx.size)
-    #print(Cx.size, ",", Cy.size, ",", Dx.size, ",", Dy.size, ",", Dz.size)
     egyensulyi_mtx = cp.zeros((Cx.size*Dx.size, 4, 4), dtype=cp.int8)
-    numBlock = int((Cx.size*Dx.size + 256 - 1) / 256)
-    fun((numBlock,), (256,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, lcm, egyensulyi_mtx))
+    numBlock = int((Cx.size*Dx.size + fun.max_threads_per_block - 1) / fun.max_threads_per_block)
+    fun((numBlock,), (fun.max_threads_per_block,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, lcm, egyensulyi_mtx))
 
     return egyensulyi_mtx

measure3.sh

@@ -5,8 +5,10 @@
 #SBATCH --gres gpu               # melyik partícióból
 #SBATCH --time=99:00:00      # maximális id?~Q
 #SBATCH -o politopok.out       # kimeneti fájl
-#SBATCH --memory=0
+#SBATCH --mem=0
+
 module load anaconda3
 module load cuda11.0
+module load mpi-3.1
 
 srun python tetrarun.py -n 51 -v 50 -w 50 -o /gv0/karsa/poli_51_50_50.out
\ No newline at end of file

tetrarun.py

@@ -5,6 +5,7 @@ import os
 from genax import gen_angels_to_pick, angles_alap, angles_ratet
 from gpu import start_kernel
 from utils import convert, printresults, search, exact_one, exact_one_gpu, filter_gpu, writetofile, writetofile2
+from mpi import size, rank
 
 def main(argv):
    outputfile = None
@@ -61,25 +62,26 @@ def main(argv):
       Dz_cpu = np.zeros(Dz.size, dtype=np.float64)
       Dz.get(out=Dz_cpu)
 
-      del space
-
       res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
 
       mtx_cpu = np.zeros(res.size, dtype=np.int8).reshape(res.shape)
       res.get(out=mtx_cpu)
    
-      writetofile(outputfile+'.full', Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi)
+      #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)
 
    if outputfile and binary:
+      offset = ''
+      if mpi:
+         offset = f"R{rank}/{size}"
       res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
       os.mkdir(outputfile)
-      cp.save(outputfile+'/Cx.npy', Cx)
-      cp.save(outputfile+'/Cy.npy', Cy)
-      cp.save(outputfile+'/Dx.npy', Dx)
-      cp.save(outputfile+'/Dy.npy', Dy)
-      cp.save(outputfile+'/Dz.npy', Dz)
-      cp.save(outputfile+'/mtx.npy', res)
+      cp.save(f"{outputfile}/{offset}_Cx.npy", Cx)
+      cp.save(f"{outputfile}/{offset}_Cy.npy", Cy)
+      cp.save(f"{outputfile}/{offset}_Dx.npy", Dx)
+      cp.save(f"{outputfile}/{offset}_Dy.npy", Dy)
+      cp.save(f"{outputfile}/{offset}_Dz.npy", Dz)
+      cp.save(f"{outputfile}/{offset}_mtx.npy", res)
 
    #printresults(res)
    #print("Exact one 3-3")