set blockdim

genax.py

@@ -64,7 +64,7 @@ def angles_alap(anglestopick, plot = False):
     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)
     
     
-    parosit((1,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, cp.pi))
+    parosit((m,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, 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
@@ -100,7 +100,7 @@ def angles_ratet(anglestopick, plot = False):
     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)
     
     
-    parosit2((1,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, cp.pi))
+    parosit2((m,), (m,), (anglestopick, anglestopick, alpha_arr, beta_arr, m, 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

gpu.py

@@ -11,7 +11,7 @@ fun = ep_pontok_module.get_function(kers[0])
 def start_kernel(Cx, Cy, Dx, Dy, Dz, v, w):
 def start_kernel(Cx, Cy, Dx, Dy, Dz, v, w):
 
 
     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)
-
-    fun((1024,), (256,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, egyensulyi_mtx))
+    numBlock = int((Cx.size*Dx.size + 256 - 1) / 256)
+    fun((numBlock,), (256,), (v, w, Cx, Cy, Dx, Dy, Dz, Cx.size, Dx.size, egyensulyi_mtx))
 
 
     return egyensulyi_mtx
     return egyensulyi_mtx