Added OpenCL CIC code
This commit is contained in:
Guilhem Lavaux
parent
1e733f2318
commit
f6ad248f75
8 changed files with 297 additions and 78 deletions
|
|
@ -21,6 +21,7 @@ cdef extern from "project_tool.hpp" namespace "":
|
|||
|
||||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
@cython.wraparound(False)
|
||||
cdef int interp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
||||
DTYPE_t z,
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval) nogil:
|
||||
|
|
@ -45,6 +46,10 @@ cdef int interp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
|||
ry -= iy
|
||||
rz -= iz
|
||||
|
||||
ix = ix % Ngrid
|
||||
iy = iy % Ngrid
|
||||
iz = iz % Ngrid
|
||||
|
||||
jx = (ix+1)%Ngrid
|
||||
jy = (iy+1)%Ngrid
|
||||
jz = (iz+1)%Ngrid
|
||||
|
|
@ -53,13 +58,6 @@ cdef int interp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
|||
iy = iy%Ngrid
|
||||
iz = iz%Ngrid
|
||||
|
||||
if (ix < 0) or (jx >= Ngrid):
|
||||
return -1
|
||||
if (iy < 0) or (jy >= Ngrid):
|
||||
return -2
|
||||
if (iz < 0) or (jz >= Ngrid):
|
||||
return -3
|
||||
|
||||
f[0][0][0] = (1-rx)*(1-ry)*(1-rz)
|
||||
f[1][0][0] = ( rx)*(1-ry)*(1-rz)
|
||||
f[0][1][0] = (1-rx)*( ry)*(1-rz)
|
||||
|
|
@ -80,10 +78,9 @@ cdef int interp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
|||
d[ix ,jy ,jz ] * f[0][1][1] + \
|
||||
d[jx ,jy ,jz ] * f[1][1][1]
|
||||
|
||||
return 0
|
||||
|
||||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
@cython.wraparound(False)
|
||||
cdef int ngp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
||||
DTYPE_t z,
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval) nogil:
|
||||
|
|
@ -110,14 +107,13 @@ cdef int ngp3d_INTERNAL_periodic(DTYPE_t x, DTYPE_t y,
|
|||
|
||||
retval[0] = d[ix ,iy ,iz ]
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
@cython.wraparound(False)
|
||||
cdef int ngp3d_INTERNAL(DTYPE_t x, DTYPE_t y,
|
||||
DTYPE_t z,
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval) nogil:
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval, DTYPE_t inval) nogil:
|
||||
|
||||
cdef int Ngrid = d.shape[0]
|
||||
cdef DTYPE_t inv_delta = Ngrid/Lbox
|
||||
|
|
@ -134,23 +130,18 @@ cdef int ngp3d_INTERNAL(DTYPE_t x, DTYPE_t y,
|
|||
iy = int(round(ry))
|
||||
iz = int(round(rz))
|
||||
|
||||
if (ix < 0 or ix >= Ngrid):
|
||||
return -1
|
||||
if (iy < 0 or iy >= Ngrid):
|
||||
return -2
|
||||
if (iz < 0 or iz >= Ngrid):
|
||||
return -3
|
||||
if ((ix < 0) or (ix+1) >= Ngrid or (iy < 0) or (iy+1) >= Ngrid or (iz < 0) or (iz+1) >= Ngrid):
|
||||
retval[0] = inval
|
||||
|
||||
retval[0] = d[ix ,iy ,iz ]
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
@cython.wraparound(False)
|
||||
cdef int interp3d_INTERNAL(DTYPE_t x, DTYPE_t y,
|
||||
DTYPE_t z,
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval) nogil:
|
||||
DTYPE_t[:,:,:] d, DTYPE_t Lbox, DTYPE_t *retval, DTYPE_t inval) nogil:
|
||||
|
||||
cdef int Ngrid = d.shape[0]
|
||||
cdef DTYPE_t inv_delta = Ngrid/Lbox
|
||||
|
|
@ -170,14 +161,8 @@ cdef int interp3d_INTERNAL(DTYPE_t x, DTYPE_t y,
|
|||
ry -= iy
|
||||
rz -= iz
|
||||
|
||||
if ((ix < 0) or (ix+1) >= Ngrid):
|
||||
return -1
|
||||
|
||||
if ((iy < 0) or (iy+1) >= Ngrid):
|
||||
return -2
|
||||
|
||||
if ((iz < 0) or (iz+1) >= Ngrid):
|
||||
return -3
|
||||
if ((ix < 0) or (ix+1) >= Ngrid or (iy < 0) or (iy+1) >= Ngrid or (iz < 0) or (iz+1) >= Ngrid):
|
||||
retval[0] = inval
|
||||
# assert ((ix >= 0) and ((ix+1) < Ngrid))
|
||||
# assert ((iy >= 0) and ((iy+1) < Ngrid))
|
||||
# assert ((iz >= 0) and ((iz+1) < Ngrid))
|
||||
|
|
@ -202,13 +187,11 @@ cdef int interp3d_INTERNAL(DTYPE_t x, DTYPE_t y,
|
|||
d[ix ,iy+1,iz+1] * f[0][1][1] + \
|
||||
d[ix+1,iy+1,iz+1] * f[1][1][1]
|
||||
|
||||
return 0
|
||||
|
||||
@cython.boundscheck(False)
|
||||
def interp3d(x not None, y not None,
|
||||
z not None,
|
||||
npx.ndarray[DTYPE_t, ndim=3] d not None, DTYPE_t Lbox,
|
||||
bool periodic=False, bool centered=True, bool ngp=False):
|
||||
bool periodic=False, bool centered=True, bool ngp=False, DTYPE_t inval = 0):
|
||||
""" interp3d(x,y,z,d,Lbox,periodic=False,centered=True,ngp=False) -> interpolated values
|
||||
|
||||
Compute the tri-linear interpolation of the given field (d) at the given position (x,y,z). It assumes that they are box-centered coordinates. So (x,y,z) == (0,0,0) is equivalent to the pixel at (Nx/2,Ny/2,Nz/2) with Nx,Ny,Nz = d.shape. If periodic is set, it assumes the box is periodic
|
||||
|
|
@ -253,41 +236,29 @@ def interp3d(x not None, y not None,
|
|||
if not myngp:
|
||||
if myperiodic:
|
||||
for i in prange(Nelt):
|
||||
if interp3d_INTERNAL_periodic(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i]) < 0:
|
||||
with gil:
|
||||
raise ierror
|
||||
interp3d_INTERNAL_periodic(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i])
|
||||
else:
|
||||
for i in prange(Nelt):
|
||||
if interp3d_INTERNAL(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i]) < 0:
|
||||
with gil:
|
||||
raise ierror
|
||||
interp3d_INTERNAL(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i], inval)
|
||||
else:
|
||||
if myperiodic:
|
||||
for i in prange(Nelt):
|
||||
if ngp3d_INTERNAL_periodic(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i]) < 0:
|
||||
with gil:
|
||||
raise ierror
|
||||
ngp3d_INTERNAL_periodic(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i])
|
||||
else:
|
||||
for i in prange(Nelt):
|
||||
if ngp3d_INTERNAL(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i]) < 0:
|
||||
with gil:
|
||||
raise ierror
|
||||
ngp3d_INTERNAL(shifter+ax[i], shifter+ay[i], shifter+az[i], in_slice, Lbox, &out_slice[i], inval)
|
||||
return out
|
||||
else:
|
||||
if not myngp:
|
||||
if periodic:
|
||||
if interp3d_INTERNAL_periodic(shifter+x, shifter+y, shifter+z, d, Lbox, &retval) < 0:
|
||||
raise ierror
|
||||
interp3d_INTERNAL_periodic(shifter+x, shifter+y, shifter+z, d, Lbox, &retval)
|
||||
else:
|
||||
if interp3d_INTERNAL(shifter+x, shifter+y, shifter+z, d, Lbox, &retval) < 0:
|
||||
raise ierror
|
||||
interp3d_INTERNAL(shifter+x, shifter+y, shifter+z, d, Lbox, &retval, inval)
|
||||
else:
|
||||
if periodic:
|
||||
if ngp3d_INTERNAL_periodic(shifter+x, shifter+y, shifter+z, d, Lbox, &retval) < 0:
|
||||
raise ierror
|
||||
ngp3d_INTERNAL_periodic(shifter+x, shifter+y, shifter+z, d, Lbox, &retval)
|
||||
else:
|
||||
if ngp3d_INTERNAL(shifter+x, shifter+y, shifter+z, d, Lbox, &retval) < 0:
|
||||
raise ierror
|
||||
ngp3d_INTERNAL(shifter+x, shifter+y, shifter+z, d, Lbox, &retval, inval)
|
||||
return retval
|
||||
|
||||
@cython.boundscheck(False)
|
||||
|
|
@ -410,8 +381,7 @@ def interp2d(x not None, y not None,
|
|||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
cdef void INTERNAL_project_cic_no_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
||||
npx.ndarray[DTYPE_t, ndim=2] x, int Ngrid, float Lbox):
|
||||
cdef double half_Box = 0.5*Lbox
|
||||
npx.ndarray[DTYPE_t, ndim=2] x, int Ngrid, double Lbox, double shifter):
|
||||
cdef double delta_Box = Ngrid/Lbox
|
||||
cdef int i
|
||||
cdef double a[3], c[3]
|
||||
|
|
@ -422,7 +392,7 @@ cdef void INTERNAL_project_cic_no_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
|
||||
do_not_put = 0
|
||||
for j in range(3):
|
||||
a[j] = (x[i,j]+half_Box)*delta_Box
|
||||
a[j] = (x[i,j]+shifter)*delta_Box
|
||||
b[j] = int(floor(a[j]))
|
||||
a[j] -= b[j]
|
||||
c[j] = 1-a[j]
|
||||
|
|
@ -446,8 +416,7 @@ cdef void INTERNAL_project_cic_no_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
@cython.boundscheck(False)
|
||||
@cython.cdivision(True)
|
||||
cdef void INTERNAL_project_cic_no_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g,
|
||||
npx.ndarray[DTYPE_t, ndim=2] x, int Ngrid, double Lbox):
|
||||
cdef double half_Box = 0.5*Lbox
|
||||
npx.ndarray[DTYPE_t, ndim=2] x, int Ngrid, double Lbox, double shifter):
|
||||
cdef double delta_Box = Ngrid/Lbox
|
||||
cdef int i
|
||||
cdef double a[3], c[3]
|
||||
|
|
@ -463,7 +432,7 @@ cdef void INTERNAL_project_cic_no_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
|
||||
do_not_put = 0
|
||||
for j in range(3):
|
||||
a[j] = (ax[i,j]+half_Box)*delta_Box
|
||||
a[j] = (ax[i,j]+shifter)*delta_Box
|
||||
b[j] = int(floor(a[j]))
|
||||
b1[j] = (b[j]+1) % Ngrid
|
||||
|
||||
|
|
@ -488,8 +457,7 @@ cdef void INTERNAL_project_cic_no_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
cdef void INTERNAL_project_cic_with_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
||||
npx.ndarray[DTYPE_t, ndim=2] x,
|
||||
npx.ndarray[DTYPE_t, ndim=1] mass,
|
||||
int Ngrid, double Lbox):
|
||||
cdef double half_Box = 0.5*Lbox
|
||||
int Ngrid, double Lbox, double shifter):
|
||||
cdef double delta_Box = Ngrid/Lbox
|
||||
cdef int i
|
||||
cdef double a[3], c[3]
|
||||
|
|
@ -500,7 +468,7 @@ cdef void INTERNAL_project_cic_with_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
|
||||
do_not_put = False
|
||||
for j in range(3):
|
||||
a[j] = (x[i,j]+half_Box)*delta_Box
|
||||
a[j] = (x[i,j]+shifter)*delta_Box
|
||||
b[j] = int(a[j])
|
||||
a[j] -= b[j]
|
||||
c[j] = 1-a[j]
|
||||
|
|
@ -525,7 +493,7 @@ cdef void INTERNAL_project_cic_with_mass(npx.ndarray[DTYPE_t, ndim=3] g,
|
|||
cdef void INTERNAL_project_cic_with_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g,
|
||||
npx.ndarray[DTYPE_t, ndim=2] x,
|
||||
npx.ndarray[DTYPE_t, ndim=1] mass,
|
||||
int Ngrid, double Lbox):
|
||||
int Ngrid, double Lbox, double shifter):
|
||||
cdef double half_Box = 0.5*Lbox, m0
|
||||
cdef double delta_Box = Ngrid/Lbox
|
||||
cdef int i
|
||||
|
|
@ -535,7 +503,7 @@ cdef void INTERNAL_project_cic_with_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g
|
|||
for i in range(x.shape[0]):
|
||||
|
||||
for j in range(3):
|
||||
a[j] = (x[i,j]+half_Box)*delta_Box
|
||||
a[j] = (x[i,j]+shifter)*delta_Box
|
||||
b[j] = int(floor(a[j]))
|
||||
b1[j] = b[j]+1
|
||||
while b1[j] < 0:
|
||||
|
|
@ -559,14 +527,20 @@ cdef void INTERNAL_project_cic_with_mass_periodic(npx.ndarray[DTYPE_t, ndim=3] g
|
|||
|
||||
|
||||
def project_cic(npx.ndarray[DTYPE_t, ndim=2] x not None, npx.ndarray[DTYPE_t, ndim=1] mass, int Ngrid,
|
||||
double Lbox, bool periodic = False):
|
||||
double Lbox, bool periodic = False, centered=True):
|
||||
"""
|
||||
project_cic(x array (N,3), mass (may be None), Ngrid, Lbox, periodict)
|
||||
project_cic(x array (N,3), mass (may be None), Ngrid, Lbox, periodict, centered=True)
|
||||
|
||||
This function does a Cloud-In-Cell projection of a 3d unstructured dataset. First argument is a Nx3 array of coordinates.
|
||||
Second argument is an optinal mass. Ngrid is the size output grid and Lbox is the physical size of the grid.
|
||||
"""
|
||||
cdef npx.ndarray[DTYPE_t, ndim=3] g
|
||||
cdef double shifter
|
||||
|
||||
if centered:
|
||||
shifter = 0.5*Lbox
|
||||
else:
|
||||
shifter = 0
|
||||
|
||||
if x.shape[1] != 3:
|
||||
raise ValueError("Invalid shape for x array")
|
||||
|
|
@ -578,14 +552,14 @@ def project_cic(npx.ndarray[DTYPE_t, ndim=2] x not None, npx.ndarray[DTYPE_t, nd
|
|||
|
||||
if not periodic:
|
||||
if mass == None:
|
||||
INTERNAL_project_cic_no_mass(g, x, Ngrid, Lbox)
|
||||
INTERNAL_project_cic_no_mass(g, x, Ngrid, Lbox, shifter)
|
||||
else:
|
||||
INTERNAL_project_cic_with_mass(g, x, mass, Ngrid, Lbox)
|
||||
INTERNAL_project_cic_with_mass(g, x, mass, Ngrid, Lbox, shifter)
|
||||
else:
|
||||
if mass == None:
|
||||
INTERNAL_project_cic_no_mass_periodic(g, x, Ngrid, Lbox)
|
||||
INTERNAL_project_cic_no_mass_periodic(g, x, Ngrid, Lbox, shifter)
|
||||
else:
|
||||
INTERNAL_project_cic_with_mass_periodic(g, x, mass, Ngrid, Lbox)
|
||||
INTERNAL_project_cic_with_mass_periodic(g, x, mass, Ngrid, Lbox, shifter)
|
||||
|
||||
return g
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue