def test_make_plan_many(self):
fft = cufft.CUFFT(self.ctx)
fft.auto_allocation = False
sz = fft.make_plan_many((256, 128), 8, cufft.CUFFT_C2C)
logging.debug(
"make_plan_many (default layout) for 256x128 x8 returned %d", sz)
logging.debug("size is %d", fft.size)
self.assertEqual(fft.execute, fft.exec_c2c)
fft = cufft.CUFFT(self.ctx)
fft.auto_allocation = False
sz = fft.make_plan_many((256, 128), 8, cufft.CUFFT_C2C, (256, 128), 1,
256 * 128, (256, 128), 1, 256 * 128)
logging.debug(
"make_plan_many (tight layout) for 256x128 x8 returned is %d", sz)
logging.debug("size is %d", fft.size)
def test_auto_allocation(self):
fft = cufft.CUFFT(self.ctx)
self.assertTrue(fft.auto_allocation)
fft.auto_allocation = False
self.assertFalse(fft.auto_allocation)
fft.auto_allocation = True
self.assertTrue(fft.auto_allocation)
def _test_exec_complex(self, dtype):
x = numpy.zeros([32, 64], dtype=dtype)
x.real = numpy.random.rand(x.size).reshape(x.shape) - 0.5
x.imag = numpy.random.rand(x.size).reshape(x.shape) - 0.5
y = numpy.ones_like(x)
x_gold = x.copy()
try:
y_gold = numpy.fft.fft2(x)
except TypeError:
y_gold = None # for pypy
xbuf = cu.MemAlloc(self.ctx, x)
ybuf = cu.MemAlloc(self.ctx, y)
# Forward transform
fft = cufft.CUFFT(self.ctx)
fft.auto_allocation = False
sz = fft.make_plan_many(x.shape, 1, {
numpy.complex64: cufft.CUFFT_C2C,
numpy.complex128: cufft.CUFFT_Z2Z
}[dtype])
tmp = cu.MemAlloc(self.ctx, sz)
fft.workarea = tmp
self.assertEqual(fft.workarea, tmp)
self.assertEqual(fft.execute, {
numpy.complex64: fft.exec_c2c,
numpy.complex128: fft.exec_z2z
}[dtype])
fft.execute(xbuf, ybuf, cufft.CUFFT_FORWARD)
ybuf.to_host(y)
if y_gold is not None:
delta = y - y_gold
max_diff = numpy.fabs(
numpy.sqrt(delta.real * delta.real +
delta.imag * delta.imag)).max()
logging.debug("Forward max_diff is %.6e", max_diff)
self.assertLess(max_diff, {
numpy.complex64: 1.0e-3,
numpy.complex128: 1.0e-6
}[dtype])
# Inverse transform
y /= x.size # correct scale before inverting
ybuf.to_device_async(y)
xbuf.memset32_async(0) # reset the resulting vector
fft.execute(ybuf, xbuf, cufft.CUFFT_INVERSE)
xbuf.to_host(x)
delta = x - x_gold
max_diff = numpy.fabs(
numpy.sqrt(delta.real * delta.real +
delta.imag * delta.imag)).max()
logging.debug("Inverse max_diff is %.6e", max_diff)
self.assertLess(max_diff, {
numpy.complex64: 1.0e-3,
numpy.complex128: 1.0e-6
}[dtype])
def test_version(self):
fft = cufft.CUFFT(self.ctx)
ver = fft.version
logging.debug("cuFFT version is %d", ver)
self.assertTrue(ver == int(ver))