47 bind(c, name =
'cuda_masked_copy')
48 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
49 type(c_ptr),
value :: a_d, b_d, mask_d, strm
50 integer(c_int) :: n, n_mask
54 bind(c, name =
'cuda_masked_gather_copy')
55 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
56 type(c_ptr),
value :: a_d, b_d, mask_d, strm
57 integer(c_int) :: n, n_mask
61 strm) bind(c, name = 'cuda_masked_gather_copy_aligned')
62 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
63 type(c_ptr),
value :: a_d, b_d, mask_d, strm
64 integer(c_int) :: n, n_mask
68 n2, lx, ly, lz, n_mask, strm) &
69 bind(c, name =
'cuda_face_masked_gather_copy')
70 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
71 type(c_ptr),
value :: a_d, b_d, mask_d, facet_d, strm
72 integer(c_int) :: n1, n2, lx, ly, lz, n_mask
76 bind(c, name =
'cuda_masked_scatter_copy')
77 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
78 type(c_ptr),
value :: a_d, b_d, mask_d, strm
79 integer(c_int) :: n, n_mask
83 bind(c, name =
'cuda_masked_scatter_copy_aligned')
84 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
85 type(c_ptr),
value :: a_d, b_d, mask_d, strm
86 integer(c_int) :: n, n_mask
90 bind(c, name =
'cuda_masked_atomic_reduction')
91 use,
intrinsic :: iso_c_binding, only : c_ptr, c_int
92 type(c_ptr),
value :: a_d, b_d, mask_d, strm
93 integer(c_int) :: n, m
97 bind(c, name =
'cuda_cfill_mask')
98 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
100 type(c_ptr),
value :: a_d, strm
103 type(c_ptr),
value :: mask_d
104 integer(c_int) :: n_mask
199 bind(c, name =
'cuda_add4')
200 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
202 type(c_ptr),
value :: a_d, b_d, c_d, d_d, strm
234 bind(c, name =
'cuda_add3s2')
235 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
237 type(c_ptr),
value :: a_d, b_d, c_d, strm
243 bind(c, name =
'cuda_add4s3')
244 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
246 type(c_ptr),
value :: a_d, b_d, c_d, d_d, strm
247 real(c_rp) :: c1, c2, c3
251 subroutine cuda_add5s4(a_d, b_d, c_d, d_d, e_d, c1, c2, c3, c4, n, strm) &
252 bind(c, name =
'cuda_add5s4')
253 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
255 type(c_ptr),
value :: a_d, b_d, c_d, d_d, e_d, strm
256 real(c_rp) :: c1, c2, c3, c4
331 bind(c, name =
'cuda_addcol4')
332 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
333 type(c_ptr),
value :: a_d, b_d, c_d, d_d, strm
346 subroutine cuda_vdot3(dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, n, strm) &
347 bind(c, name =
'cuda_vdot3')
348 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
349 type(c_ptr),
value :: dot_d, u1_d, u2_d, u3_d, v1_d, v2_d, v3_d, strm
354 w1_d, w2_d, w3_d, n, strm) &
355 bind(c, name =
'cuda_vcross')
356 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
357 type(c_ptr),
value :: u1_d, u2_d, u3_d
358 type(c_ptr),
value :: v1_d, v2_d, v3_d
359 type(c_ptr),
value :: w1_d, w2_d, w3_d
360 type(c_ptr),
value :: strm
389 bind(c, name =
'cuda_glsc3_many')
390 use,
intrinsic :: iso_c_binding, only : c_int, c_ptr
392 type(c_ptr),
value :: w_d, v_d_d, mult_d, strm
393 integer(c_int) :: j, n