bugfix

epgpu.cu

@@ -263,6 +263,8 @@ __global__ void gpu_egyensulyi(int v, int w, const double* Cx_arr, const  double
     int pos = blockDim.x * blockIdx.x + threadIdx.x;
     int pos = blockDim.x * blockIdx.x + threadIdx.x;
     if (pos >= size_C*size_D)
     if (pos >= size_C*size_D)
         return;
         return;
+    if (pos == 0)
+        printf("%d", size_C*size_D);
     vec3 C(Cx_arr[pos % size_C], Cy_arr[pos % size_C], 0.0);
     vec3 C(Cx_arr[pos % size_C], Cy_arr[pos % size_C], 0.0);
     vec3 D(Dx_arr[(pos + pos / lcm) % size_D], Dy_arr[(pos + pos / lcm) % size_D], Dz_arr[(pos + pos / lcm) % size_D]);
     vec3 D(Dx_arr[(pos + pos / lcm) % size_D], Dy_arr[(pos + pos / lcm) % size_D], Dz_arr[(pos + pos / lcm) % size_D]);
 
 

gpu.py

@@ -16,6 +16,7 @@ def start_kernel(Cx, Cy, Dx, Dy, Dz, v, w):
     lcm = compute_lcm(Cx.size, Dx.size)
     lcm = compute_lcm(Cx.size, Dx.size)
     egyensulyi_mtx = cp.zeros((Cx.size*Dx.size, 4, 4), dtype=cp.int8)
     egyensulyi_mtx = cp.zeros((Cx.size*Dx.size, 4, 4), dtype=cp.int8)
     numBlock = int((Cx.size*Dx.size + fun.max_threads_per_block - 1) / fun.max_threads_per_block)
     numBlock = int((Cx.size*Dx.size + fun.max_threads_per_block - 1) / fun.max_threads_per_block)
+    print(f"{Cx.size}, {Cy.size}, {Dx.size}, {Dy.size}, {Dz.size}, {egyensulyi_mtx.shape}, {egyensulyi_mtx.nbytes}")
     fun((numBlock,), (fun.max_threads_per_block,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, lcm, egyensulyi_mtx))
     fun((numBlock,), (fun.max_threads_per_block,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, lcm, egyensulyi_mtx))
 
 
     return egyensulyi_mtx
     return egyensulyi_mtx

measure3.sh

 #!/bin/bash
 #!/bin/bash
 #SBATCH --job-name=gpgpu     # a job neve
 #SBATCH --job-name=gpgpu     # a job neve
-#SBATCH -N 1                 # hány node-ot szeretnénk használni
+#SBATCH -N 4                 # hány node-ot szeretnénk használni
 #SBATCH -p gpu               # melyik partícióból
 #SBATCH -p gpu               # melyik partícióból
-#SBATCH --gres gpu               # melyik partícióból
+#SBATCH --gres gpu           # melyik partícióból
 #SBATCH --time=99:00:00      # maximális id?~Q
 #SBATCH --time=99:00:00      # maximális id?~Q
-#SBATCH -o politopok.out       # kimeneti fájl
+#SBATCH -o politopok.out     # kimeneti fájl
 #SBATCH --mem=0
 #SBATCH --mem=0
 
 
 module load anaconda3
 module load anaconda3
 module load cuda11.0
 module load cuda11.0
 module load mpi-3.1
 module load mpi-3.1
 
 
-srun python tetrarun.py -n 51 -v 50 -w 50 -o /gv0/karsa/poli_51_50_50.out
+mpirun 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

@@ -68,12 +68,12 @@ def main(argv):
       res.get(out=mtx_cpu)
       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)
+      #writetofile2(outputfile, Cx_cpu, Cy_cpu, Dx_cpu, Dy_cpu, Dz_cpu, mtx_cpu, mpi)
 
 
    if outputfile and binary:
    if outputfile and binary:
       offset = ''
       offset = ''
       if mpi:
       if mpi:
-         offset = f"R{rank}/{size}"
+         offset = f"R{rank}-{size}"
       res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
       res = start_kernel(Cx, Cy, Dx, Dy, Dz, v, w)
       os.mkdir(outputfile)
       os.mkdir(outputfile)
       cp.save(f"{outputfile}/{offset}_Cx.npy", Cx)
       cp.save(f"{outputfile}/{offset}_Cx.npy", Cx)