def time_simple(N, nargs, niter=100):
from gputools import OCLReductionKernel
map_exprs = ["%s*x[i]" % i for i in xrange(nargs)]
ks = [
OCLReductionKernel(np.float32,
neutral="0",
reduce_expr="a+b",
map_expr="%s*x[i]" % i,
arguments="__global float *x")
for i in xrange(len(map_exprs))
]
ins = [
OCLArray.from_array(np.ones(N, np.float32))
for _ in xrange(len(map_exprs))
]
outs = [OCLArray.empty(1, np.float32) for _ in xrange(len(map_exprs))]
from time import time
t = time()
for _ in xrange(niter):
for k, inn, out in zip(ks, ins, outs):
k(inn, out=out)
get_device().queue.finish()
t = (time() - t) / niter
print "simple reduction: result =", [float(out.get()) for out in outs]
print "simple reduction:\t\t%.2f ms" % (1000 * t)
return t
def setup(self, size, units, lam = .5, n0 = 1.,
use_fresnel_approx = False):
"""
sets up the internal variables e.g. propagators etc...
:param size: the size of the geometry in pixels (Nx,Ny,Nz)
:param units: the phyiscal units of each voxel in microns (dx,dy,dz)
:param lam: the wavelength of light in microns
:param n0: the refractive index of the surrounding media
:param use_fresnel_approx: if True, uses fresnel approximation for propagator
"""
Bpm3d_Base.setup(self,size, units, lam = lam, n0 = n0,
use_fresnel_approx = use_fresnel_approx)
#setting up the gpu buffers and kernels
self.program = OCLProgram(absPath("kernels/bpm_3d_kernels.cl"))
Nx, Ny = self.size[:2]
plan = fft_plan(())
self._H_g = OCLArray.from_array(self._H.astype(np.complex64))
self.scatter_weights_g = OCLArray.from_array(self.scatter_weights.astype(np.float32))
self.gfactor_weights_g = OCLArray.from_array(self.gfactor_weights.astype(np.float32))
self.scatter_cross_sec_g = OCLArray.zeros(Nz,"float32")
self.gfactor_g = OCLArray.zeros(Nz,"float32")
self.reduce_kernel = OCLReductionKernel(
np.float32, neutral="0",
reduce_expr="a+b",
map_expr="weights[i]*cfloat_abs(field[i]-(i==0)*plain)*cfloat_abs(field[i]-(i==0)*plain)",
arguments="__global cfloat_t *field, __global float * weights,cfloat_t plain")
from time import time
t = time()
for _ in range(niter):
for k,inn,out in zip(ks,ins,outs):
k(inn, out = out)
get_device().queue.finish()
t = (time()-t)/niter
print("simple reduction: result =", [float(out.get()) for out in outs])
print("simple reduction:\t\t%.2f ms"%(1000*t))
return t
from gputools import OCLReductionKernel
k1 = OCLReductionKernel(np.float32,
neutral="0", reduce_expr="a+b",
map_expr="x[i]",
arguments="__global float *x")
k2 = OCLMultiReductionKernel(np.float32,
neutral="0", reduce_expr="a+b",
map_exprs=["x[i]"],
arguments="__global float *x")
N = 32
a = OCLArray.from_array(np.ones((N,N),np.float32))
# print k1(a[0], out = out[0])
print(k2(a))