feectools 0.1.0__py3-none-any.whl

feectools/linalg/kernels/matvec_kernels.py

@@ -0,0 +1,206 @@
+from typing import TypeVar
+
+T = TypeVar('T', float, complex)
+
+def matvec_1d(mat00:'T[:,:]', x0:'T[:]', out0:'T[:]', starts: 'int64[:]', nrows: 'int64[:]', nrows_extra: 'int64[:]',
+                  dm:'int64[:]', cm:'int64[:]', pad_imp:'int64[:]', ndiags:'int64[:]', gpads: 'int64[:]'):
+
+    nrows1   = nrows[0]
+    dstart1  = starts[0]
+    dshift1  = dm[0]
+    cshift1  = cm[0]
+    ndiags1  = ndiags[0]
+    dpads1   = gpads[0]
+    pad_imp1 = pad_imp[0]
+
+    pxm1 = dpads1 * cshift1
+
+    start_impact1 = dstart1 % dshift1
+
+    v00 = mat00[0, 0] - mat00[0, 0] + x0[0] - x0[0]
+
+    for i1 in range(nrows1):
+        v00 *= 0
+        x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+        for k1 in range(ndiags1):
+            v00 += mat00[pxm1 + i1, k1] * x0[k1 + x_min1]
+        out0[pxm1 + i1] = v00
+
+    if 0 < nrows_extra[0]:
+        pxm1          += nrows1
+        start_impact1 += nrows1
+        for i1 in range(nrows_extra[0]):
+            v00 *= 0
+            x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+            for k1 in range(ndiags1 - i1 - 1):
+                v00 += mat00[pxm1 + i1, k1] * x0[x_min1 + k1]
+            out0[pxm1 + i1] = v00
+
+
+
+def matvec_2d(mat00:'T[:,:,:,:]', x0:'T[:,:]', out0:'T[:,:]', starts:'int64[:]', nrows:'int64[:]', nrows_extra:'int64[:]',
+                  dm:'int64[:]', cm:'int64[:]', pad_imp:'int64[:]', ndiags:'int64[:]', gpads: 'int64[:]'):
+
+    nrows1   = nrows[0]
+    nrows2   = nrows[1]
+    dstart1  = starts[0]
+    dstart2  = starts[1]
+    dshift1  = dm[0]
+    dshift2  = dm[1]
+    cshift1  = cm[0]
+    cshift2  = cm[1]
+    ndiags1  = ndiags[0]
+    ndiags2  = ndiags[1]
+    dpads1   = gpads[0]
+    dpads2   = gpads[1]
+    pad_imp1 = pad_imp[0]
+    pad_imp2 = pad_imp[1]
+
+    pxm1 = dpads1 * cshift1
+    pxm2 = dpads2 * cshift2
+
+    start_impact1 = dstart1 % dshift1
+    start_impact2 = dstart2 % dshift2
+
+    v00 = mat00[0, 0, 0, 0] - mat00[0, 0, 0, 0] + x0[0, 0] - x0[0, 0]
+
+    for i1 in range(nrows1):
+        for i2 in range(nrows2):
+            v00 *= 0
+            x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+            x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+            for k1 in range(ndiags1):
+                for k2 in range(ndiags2):
+                    v00 += mat00[pxm1 + i1, pxm2 + i2, k1, k2] * x0[k1 + x_min1, k2 + x_min2]
+            out0[pxm1 + i1, pxm2 + i2] = v00
+
+    if 0 < nrows_extra[0]:
+        pxm1          += nrows1
+        start_impact1 += nrows1
+        for i1 in range(nrows_extra[0]):
+            for i2 in range(nrows2):
+                v00 *= 0
+                x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                for k1 in range(ndiags1 - i1 - 1):
+                    for k2 in range(ndiags2):
+                        v00 += mat00[pxm1 + i1, pxm2 + i2, k1, k2] * x0[x_min1 + k1, x_min2 + k2]
+                out0[pxm1 + i1,  pxm2 + i2] = v00
+
+    if 0 < nrows_extra[1]:
+        pxm1           = dpads1  * cshift1
+        start_impact1  = dstart1 % dshift1
+        pxm2          += nrows2
+        start_impact2 += nrows2
+        for i1 in range(nrows1 + nrows_extra[0]):
+            for i2 in range(nrows_extra[1]):
+                v00 *= 0
+                x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                for k1 in range(ndiags1 - max(0, i1 + 1 - nrows1)):
+                    for k2 in range(ndiags2 - i2 - 1):
+                        v00 += mat00[pxm1 + i1, pxm2 + i2, k1, k2] * x0[x_min1 + k1, x_min2 + k2]
+                out0[pxm1 + i1, pxm2 + i2] = v00
+
+
+def matvec_3d(mat00:'T[:,:,:,:,:,:]', x0:'T[:,:,:]', out0:'T[:,:,:]', starts:'int64[:]', nrows:'int64[:]', nrows_extra:'int64[:]',
+                  dm:'int64[:]', cm:'int64[:]', pad_imp:'int64[:]', ndiags:'int64[:]', gpads: 'int64[:]'):
+
+    nrows1   = nrows[0]
+    nrows2   = nrows[1]
+    nrows3   = nrows[2]
+    dstart1  = starts[0]
+    dstart2  = starts[1]
+    dstart3  = starts[2]
+    dshift1  = dm[0]
+    dshift2  = dm[1]
+    dshift3  = dm[2]
+    cshift1  = cm[0]
+    cshift2  = cm[1]
+    cshift3  = cm[2]
+    ndiags1  = ndiags[0]
+    ndiags2  = ndiags[1]
+    ndiags3  = ndiags[2]
+    dpads1   = gpads[0]
+    dpads2   = gpads[1]
+    dpads3   = gpads[2]
+    pad_imp1 = pad_imp[0]
+    pad_imp2 = pad_imp[1]
+    pad_imp3 = pad_imp[2]
+
+    pxm1 = dpads1 * cshift1
+    pxm2 = dpads2 * cshift2
+    pxm3 = dpads3 * cshift3
+
+    start_impact1 = dstart1 % dshift1
+    start_impact2 = dstart2 % dshift2
+    start_impact3 = dstart3 % dshift3
+
+    v00 = mat00[0, 0, 0, 0, 0, 0] - mat00[0, 0, 0, 0, 0, 0] + x0[0, 0, 0] - x0[0, 0, 0]
+
+    for i1 in range(nrows1):
+        for i2 in range(nrows2):
+            for i3 in range(nrows3):
+                v00 *= 0
+                x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                x_min3 = pad_imp3 + (i3 + start_impact3) // cshift3 * dshift3
+                for k1 in range(ndiags1):
+                    for k2 in range(ndiags2):
+                        for k3 in range(ndiags3):
+                            v00 += mat00[pxm1 + i1, pxm2 + i2, pxm3 + i3, k1, k2, k3] * x0[k1 + x_min1, k2 + x_min2, k3 + x_min3]
+                out0[pxm1 + i1, pxm2 + i2, pxm3 + i3] = v00
+
+    if 0 < nrows_extra[0]:
+        pxm1 += nrows1
+        start_impact1 += nrows1
+        for i1 in range(nrows_extra[0]):
+            for i2 in range(nrows2):
+                for i3 in range(nrows3):
+                    v00 *= 0
+                    x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                    x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                    x_min3 = pad_imp3 + (i3 + start_impact3) // cshift3 * dshift3
+                    for k1 in range(ndiags1 - i1 - 1):
+                        for k2 in range(ndiags2):
+                            for k3 in range(ndiags3):
+                                v00 += mat00[pxm1 + i1, pxm2 + i2, pxm3 + i3, k1, k2, k3] *  x0[x_min1 + k1, x_min2 + k2, x_min3 + k3]
+                    out0[pxm1 + i1,  pxm2 + i2,  pxm3 + i3] = v00
+
+    if 0 < nrows_extra[1]:
+        pxm1           = dpads1  * cshift1
+        start_impact1  = dstart1 % dshift1
+        pxm2          += nrows2
+        start_impact2 += nrows2
+        for i1 in range(nrows1 + nrows_extra[0]):
+            for i2 in range(nrows_extra[1]):
+                for i3 in range(nrows3):
+                    v00 *= 0
+                    x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                    x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                    x_min3 = pad_imp3 + (i3 + start_impact3) // cshift3 * dshift3
+                    for k1 in range(ndiags1 - max(0, i1 + 1 - nrows1)):
+                        for k2 in range(ndiags2 - i2 - 1):
+                            for k3 in range(ndiags3):
+                                v00 += mat00[pxm1 + i1, pxm2 + i2, pxm3 + i3, k1, k2, k3] * x0[x_min1 + k1, x_min2 + k2, x_min3 + k3]
+                    out0[pxm1 + i1, pxm2 + i2, pxm3 + i3] = v00
+
+    if 0 < nrows_extra[2]:
+        pxm1           = dpads1  * cshift1
+        pxm2           = dpads2  * cshift2
+        start_impact1  = dstart1 % dshift1
+        start_impact2  = dstart2 % dshift2
+        pxm3          += nrows3
+        start_impact3 += nrows3
+        for i1 in range(nrows1 + nrows_extra[0]):
+            for i2 in range(nrows2 + nrows_extra[1]):
+                for i3 in range(nrows_extra[2]):
+                    v00 *= 0
+                    x_min1 = pad_imp1 + (i1 + start_impact1) // cshift1 * dshift1
+                    x_min2 = pad_imp2 + (i2 + start_impact2) // cshift2 * dshift2
+                    x_min3 = pad_imp3 + (i3 + start_impact3) // cshift3 * dshift3
+                    for k1 in range(ndiags1 - max(0, i1 + 1 - nrows1)):
+                        for k2 in range(ndiags2 - max(0, i2 + 1 - nrows2)):
+                            for k3 in range(ndiags3 - i3 - 1):
+                                v00 += mat00[pxm1 + i1, pxm2 + i2, pxm3 + i3, k1, k2, k3] * x0[x_min1 + k1, x_min2 + k2, x_min3 + k3]
+                    out0[pxm1 + i1, pxm2 + i2, pxm3 + i3] = v00

feectools/linalg/kernels/stencil2IJV_kernels.py

@@ -0,0 +1,227 @@
+# coding: utf-8
+
+from typing import TypeVar
+
+T = TypeVar('T', float, complex)
+
+#========================================================================================================
+def stencil2IJV_1d_C(A:'T[:,:]', Ib:'int64[:]', Jb:'int64[:]', Vb:'T[:]', rowmapb:'int64[:]',
+                     cnl1:'int64', dng1:'int64', cs1:'int64', cp1:'int64', cm1:'int64',
+                     dsh:'int64[:]', csh:'int64[:]', dgs1:'int64[:]', dge1:'int64[:]', 
+                     cgs1:'int64[:]', cge1:'int64[:]', dnlb1:'int64[:]', cnlb1:'int64[:]'):
+
+    nnz = 0
+    nnz_rows = 0
+    gr1 = cp1 * cm1
+
+    stencil_size1 = A.shape[1] # pre-compute the size
+
+    for i1 in range(cnl1):
+        nnz_in_row = 0
+        i1_n = cs1 + i1
+
+        pr_i1 = 0
+        for k in range(cgs1.size):
+            if i1_n < cgs1[k] or i1_n > cge1[k]:
+                continue
+            pr_i1 = k
+
+        i_g = csh[pr_i1] + i1_n - cgs1[pr_i1]
+        # this is used in upstream psydac, doesn't work with --language c
+        # stencil_size1 = A[i1 + gr1].size
+
+        for k1 in range(stencil_size1):
+
+            j1_n = (i1_n + k1 - stencil_size1 // 2) % dng1 
+            value = A[i1 + gr1, k1]
+
+            if abs(value) == 0.0:
+                continue
+
+            pr_j1 = 0
+            for k in range(dgs1.size):
+                if j1_n < dgs1[k] or j1_n > dge1[k]:
+                    continue
+                pr_j1 = k            
+                         
+            j_g = dsh[pr_j1] + j1_n - dgs1[pr_j1]
+
+            if nnz_in_row == 0:
+                rowmapb[nnz_rows] = i_g  
+
+            Jb[nnz] = j_g           
+            Vb[nnz] = value  
+            nnz += 1
+            nnz_in_row += 1
+
+        if nnz_in_row > 0:
+            Ib[1 + nnz_rows] = Ib[nnz_rows] + nnz_in_row
+            nnz_rows += 1
+
+    return nnz_rows, nnz
+
+#========================================================================================================
+def stencil2IJV_2d_C(A:'T[:,:,:,:]', Ib:'int64[:]', Jb:'int64[:]', Vb:'T[:]', rowmapb:'int64[:]',
+                     cnl1:'int64', cnl2:'int64', dng1:'int64', dng2:'int64', cs1:'int64', 
+                     cs2:'int64', cp1:'int64', cp2:'int64', cm1:'int64', cm2:'int64',
+                     dsh:'int64[:]', csh:'int64[:]', dgs1:'int64[:]', dgs2:'int64[:]', 
+                     dge1:'int64[:]', dge2:'int64[:]', cgs1:'int64[:]', cgs2:'int64[:]', 
+                     cge1:'int64[:]', cge2:'int64[:]', dnlb1:'int64[:]', dnlb2:'int64[:]', 
+                     cnlb1:'int64[:]', cnlb2:'int64[:]'):
+
+    nnz = 0
+    nnz_rows = 0
+    gr1 = cp1 * cm1
+    gr2 = cp2 * cm2
+
+    for i1 in range(cnl1):
+        for i2 in range(cnl2):
+            nnz_in_row = 0
+            i1_n = cs1 + i1
+            i2_n = cs2 + i2
+
+            pr_i1 = 0
+            for k in range(cgs1.size):
+                if i1_n < cgs1[k] or i1_n > cge1[k]:
+                    continue
+                pr_i1 = k
+
+            pr_i2 = 0
+            for k in range(cgs2.size):
+                if i2_n < cgs2[k] or i2_n > cge2[k]:
+                    continue
+                pr_i2 = k
+
+            pr_i = pr_i2 + pr_i1 * cgs2.size
+            i_g = csh[pr_i] + i2_n - cgs2[pr_i2] + (i1_n - cgs1[pr_i1]) * cnlb2[pr_i]
+            stencil_size1, stencil_size2 = A.shape[2:]
+
+            for k1 in range(stencil_size1):
+                for k2 in range(stencil_size2):
+                    j1_n = (i1_n + k1 - stencil_size1 // 2) % dng1 
+                    j2_n = (i2_n + k2 - stencil_size2 // 2) % dng2
+
+                    value = A[i1 + gr1, i2 + gr2, k1, k2]
+                    if abs(value) == 0.0:
+                        continue
+
+                    pr_j1 = 0
+                    for k in range(dgs1.size):
+                        if j1_n < dgs1[k] or j1_n > dge1[k]:
+                            continue
+                        pr_j1 = k
+
+                    pr_j2 = 0
+                    for k in range(dgs2.size):
+                        if j2_n < dgs2[k] or j2_n > dge2[k]:
+                            continue
+                        pr_j2 = k
+
+                    pr_j = pr_j2 + pr_j1 * dgs2.size
+                    j_g = dsh[pr_j] + j2_n - dgs2[pr_j2] + (j1_n - dgs1[pr_j1]) * dnlb2[pr_j]
+
+                    if nnz_in_row == 0:
+                        rowmapb[nnz_rows] = i_g  
+
+                    Jb[nnz] = j_g           
+                    Vb[nnz] = value  
+                    nnz += 1
+                    nnz_in_row += 1
+
+            if nnz_in_row > 0:
+                Ib[1 + nnz_rows] = Ib[nnz_rows] + nnz_in_row
+                nnz_rows += 1
+
+    return nnz_rows, nnz
+
+#========================================================================================================
+def stencil2IJV_3d_C(A:'T[:,:,:,:,:,:]', Ib:'int64[:]', Jb:'int64[:]', Vb:'T[:]', rowmapb:'int64[:]',
+                     cnl1:'int64', cnl2:'int64', cnl3:'int64', dng1:'int64', dng2:'int64', dng3:'int64', 
+                     cs1:'int64', cs2:'int64', cs3:'int64', cp1:'int64', cp2:'int64', cp3:'int64', 
+                     cm1:'int64', cm2:'int64', cm3:'int64', dsh:'int64[:]', csh:'int64[:]',
+                     dgs1:'int64[:]', dgs2:'int64[:]', dgs3:'int64[:]', dge1:'int64[:]', dge2:'int64[:]', 
+                     dge3:'int64[:]', cgs1:'int64[:]', cgs2:'int64[:]', cgs3:'int64[:]', 
+                     cge1:'int64[:]', cge2:'int64[:]', cge3:'int64[:]', dnlb1:'int64[:]', dnlb2:'int64[:]', 
+                     dnlb3:'int64[:]', cnlb1:'int64[:]', cnlb2:'int64[:]', cnlb3:'int64[:]'):
+
+    nnz = 0
+    nnz_rows = 0
+    gr1 = cp1*cm1
+    gr2 = cp2*cm2
+    gr3 = cp3*cm3
+
+    for i1 in range(cnl1):
+        for i2 in range(cnl2):
+            for i3 in range(cnl3):
+                nnz_in_row = 0
+                i1_n = cs1 + i1
+                i2_n = cs2 + i2
+                i3_n = cs3 + i3
+
+                pr_i1 = 0
+                for k in range(cgs1.size):
+                    if i1_n < cgs1[k] or i1_n > cge1[k]:
+                        continue
+                    pr_i1 = k
+
+                pr_i2 = 0
+                for k in range(cgs2.size):
+                    if i2_n < cgs2[k] or i2_n > cge2[k]:
+                        continue
+                    pr_i2 = k
+
+                pr_i3 = 0
+                for k in range(cgs3.size):
+                    if i3_n < cgs3[k] or i3_n > cge3[k]:
+                        continue
+                    pr_i3 = k                    
+
+                pr_i = pr_i3 + pr_i2 * cgs3.size + pr_i1 * cgs2.size * cgs3.size
+                i_g = csh[pr_i] + i3_n - cgs3[pr_i3] + (i2_n - cgs2[pr_i2]) * cnlb3[pr_i] + (i1_n - cgs1[pr_i1]) * cnlb2[pr_i] * cnlb3[pr_i]
+                stencil_size1, stencil_size2, stencil_size3 = A.shape[3:]
+
+                for k1 in range(stencil_size1):
+                    for k2 in range(stencil_size2):
+                        for k3 in range(stencil_size3):
+                            j1_n = (i1_n + k1 - stencil_size1 // 2) % dng1 
+                            j2_n = (i2_n + k2 - stencil_size2 // 2) % dng2
+                            j3_n = (i3_n + k3 - stencil_size3 // 2) % dng3
+
+                            value = A[i1 + gr1, i2 + gr2, i3 + gr3, k1, k2, k3]
+                            if abs(value) == 0.0:
+                                continue
+
+                            pr_j1 = 0
+                            for k in range(dgs1.size):
+                                if j1_n < dgs1[k] or j1_n > dge1[k]:
+                                    continue
+                                pr_j1 = k
+
+                            pr_j2 = 0
+                            for k in range(dgs2.size):
+                                if j2_n < dgs2[k] or j2_n > dge2[k]:
+                                    continue
+                                pr_j2 = k
+
+                            pr_j3 = 0
+                            for k in range(dgs3.size):
+                                if j3_n < dgs3[k] or j3_n > dge3[k]:
+                                    continue
+                                pr_j3 = k  
+
+                            pr_j = pr_j3 + pr_j2 * dgs3.size + pr_j1 * dgs2.size * dgs3.size
+                            j_g = dsh[pr_j] + j3_n - dgs3[pr_j3] + (j2_n - dgs2[pr_j2]) * dnlb3[pr_j] + (j1_n - dgs1[pr_j1]) * dnlb2[pr_j] * dnlb3[pr_j]
+
+                            if nnz_in_row == 0:
+                                rowmapb[nnz_rows] = i_g  
+
+                            Jb[nnz] = j_g           
+                            Vb[nnz] = value  
+                            nnz += 1
+                            nnz_in_row += 1
+
+                if nnz_in_row > 0:
+                    Ib[1 + nnz_rows] = Ib[nnz_rows] + nnz_in_row
+                    nnz_rows += 1
+
+    return nnz_rows, nnz

feectools/linalg/kernels/stencil2coo_kernels.py

@@ -0,0 +1,179 @@
+# coding: utf-8
+
+#!!!!!!!!!!!!!
+#TODO avoid using The expensive modulo operator % in the non periodic case to make the methods faster
+#!!!!!!!!!!!!!
+
+from typing import TypeVar
+
+T = TypeVar('T', float, complex)
+
+#__all__ = ['stencil2coo_1d_C','stencil2coo_1d_F','stencil2coo_2d_C','stencil2coo_2d_F', 'stencil2coo_3d_C', 'stencil2coo_3d_F']
+
+#========================================================================================================
+def stencil2coo_1d_C(A:'T[:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]', nrl1:'int64', ncl1:'int64',
+                     s1:'int64', nr1:'int64', nc1:'int64', dm1:'int64', cm1:'int64', p1:'int64', dp1:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    for i1 in range(nrl1):
+        I = s1+i1
+        for j1 in range(ncl1):
+            value = A[i1+pp1,j1]
+            if abs(value) == 0.0:continue
+            J = ((I*dm1//cm1)+j1-dp1)%nc1
+            rows[nnz] = I
+            cols[nnz] = J
+            data[nnz] = value
+            nnz += 1
+
+    return nnz
+
+#========================================================================================================
+def stencil2coo_1d_F(A:'T[:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]', nrl1:'int64', ncl1:'int64',
+                     s1:'int64', nr1:'int64', nc1:'int64', dm1:'int64', cm1:'int64', p1:'int64', dp1:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    for i1 in range(nrl1):
+        I = s1+i1
+        for j1 in range(ncl1):
+            value = A[i1+pp1,j1]
+            if abs(value) == 0.0:continue
+            J = ((I*dm1//cm1)+j1-dp1)%nc1
+            rows[nnz] = I
+            cols[nnz] = J
+            data[nnz] = value
+            nnz += 1
+
+    return nnz
+
+#========================================================================================================
+def stencil2coo_2d_C(A:'T[:,:,:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]',
+                     nrl1:'int64', nrl2:'int64', ncl1:'int64', ncl2:'int64',
+                     s1:'int64', s2:'int64', nr1:'int64', nr2:'int64',
+                     nc1:'int64', nc2:'int64', dm1:'int64', dm2:'int64',
+                     cm1:'int64', cm2:'int64', p1:'int64', p2:'int64',
+                     dp1:'int64', dp2:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    pp2 = cm2*p2
+    for i1 in range(nrl1):
+        for i2 in range(nrl2):
+            ii1 = s1+i1
+            ii2 = s2+i2
+            I   = ii1*nr2 + ii2
+            for j1 in range(ncl1):
+                for j2 in range(ncl2):
+                    value = A[i1+pp1,i2+pp2,j1,j2]
+                    if abs(value) == 0.0:continue
+                    jj1 = ((ii1*dm1//cm1)+j1-dp1)%nc1
+                    jj2 = ((ii2*dm2//cm2)+j2-dp2)%nc2
+
+                    J   = jj1*nc2 + jj2
+
+                    rows[nnz] = I
+                    cols[nnz] = J
+                    data[nnz] = value
+                    nnz += 1
+    return nnz
+
+#========================================================================================================
+def stencil2coo_2d_F(A:'T[:,:,:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]',
+                     nrl1:'int64', nrl2:'int64', ncl1:'int64', ncl2:'int64',
+                     s1:'int64', s2:'int64', nr1:'int64', nr2:'int64',
+                     nc1:'int64', nc2:'int64', dm1:'int64', dm2:'int64',
+                     cm1:'int64', cm2:'int64', p1:'int64', p2:'int64',
+                     dp1:'int64', dp2:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    pp2 = cm2*p2
+    for i1 in range(nrl1):
+        for i2 in range(nrl2):
+            ii1 = s1+i1
+            ii2 = s2+i2
+            I   = ii2*nr1 + ii1
+            for j1 in range(ncl1):
+                for j2 in range(ncl2):
+                    value = A[i1+pp1,i2+pp2,j1,j2]
+                    if abs(value) == 0.0:continue
+                    jj1 = ((ii1*dm1//cm1)+j1-dp1)%nc1
+                    jj2 = ((ii2*dm2//cm2)+j2-dp2)%nc2
+
+                    J   = jj2*nc1 + jj1
+
+                    rows[nnz] = I
+                    cols[nnz] = J
+                    data[nnz] = value
+                    nnz += 1
+    return nnz
+
+#========================================================================================================
+def stencil2coo_3d_C(A:'T[:,:,:,:,:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]',
+                     nrl1:'int64', nrl2:'int64', nrl3:'int64', ncl1:'int64', ncl2:'int64', ncl3:'int64',
+                     s1:'int64', s2:'int64', s3:'int64', nr1:'int64', nr2:'int64', nr3:'int64',
+                     nc1:'int64', nc2:'int64', nc3:'int64', dm1:'int64', dm2:'int64', dm3:'int64',
+                     cm1:'int64', cm2:'int64', cm3:'int64', p1:'int64', p2:'int64', p3:'int64',
+                     dp1:'int64', dp2:'int64', dp3:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    pp2 = cm2*p2
+    pp3 = cm3*p3
+    for i1 in range(nrl1):
+        for i2 in range(nrl2):
+            for i3 in range(nrl3):
+                ii1 = s1+i1
+                ii2 = s2+i2
+                ii3 = s3+i3
+                I   = ii1*nr2*nr3 + ii2*nr3 + ii3
+                for j1 in range(ncl1):
+                    for j2 in range(ncl2):
+                        for j3 in range(ncl3):
+                            value = A[i1+pp1,i2+pp2,i3+pp3,j1,j2,j3]
+                            if abs(value) == 0.0:continue
+                            jj1 = ((ii1*dm1//cm1)+j1-dp1)%nc1
+                            jj2 = ((ii2*dm2//cm2)+j2-dp2)%nc2
+                            jj3 = ((ii3*dm3//cm3)+j3-dp3)%nc3
+
+                            J   = jj1*nc2*nc3 + jj2*nc3 + jj3
+
+                            rows[nnz] = I
+                            cols[nnz] = J
+                            data[nnz] = value
+                            nnz += 1
+
+    return nnz
+
+
+#========================================================================================================
+def stencil2coo_3d_F(A:'T[:,:,:,:,:,:]', data:'T[:]', rows:'int64[:]', cols:'int64[:]',
+                     nrl1:'int64', nrl2:'int64', nrl3:'int64', ncl1:'int64', ncl2:'int64', ncl3:'int64',
+                     s1:'int64', s2:'int64', s3:'int64', nr1:'int64', nr2:'int64', nr3:'int64',
+                     nc1:'int64', nc2:'int64', nc3:'int64', dm1:'int64', dm2:'int64', dm3:'int64',
+                     cm1:'int64', cm2:'int64', cm3:'int64', p1:'int64', p2:'int64', p3:'int64',
+                     dp1:'int64', dp2:'int64', dp3:'int64'):
+    nnz = 0
+    pp1 = cm1*p1
+    pp2 = cm2*p2
+    pp3 = cm3*p3
+    for i1 in range(nrl1):
+        for i2 in range(nrl2):
+            for i3 in range(nrl3):
+                ii1 = s1+i1
+                ii2 = s2+i2
+                ii3 = s3+i3
+                I   = ii3*nr1*nr2 + ii2*nr1 + ii1
+                for j1 in range(ncl1):
+                    for j2 in range(ncl2):
+                        for j3 in range(ncl3):
+                            value = A[i1+pp1,i2+pp2,i3+pp3,j1,j2,j3]
+                            if abs(value) == 0.0:continue
+                            jj1 = ((ii1*dm1//cm1)+j1-dp1)%nc1
+                            jj2 = ((ii2*dm2//cm2)+j2-dp2)%nc2
+                            jj3 = ((ii3*dm3//cm3)+j3-dp3)%nc3
+
+                            J   = jj3*nc1*nc2 + jj2*nc1 + jj1
+
+                            rows[nnz] = I
+                            cols[nnz] = J
+                            data[nnz] = value
+                            nnz += 1
+    return nnz