def run_test_r2c_dtype(self,
shape,
axes,
dtype=np.float32,
scale=1.,
misalign=0):
known_data = np.random.uniform(size=shape).astype(np.float32) * 2 - 1
known_data = (known_data * scale).astype(dtype)
# Force misaligned data
padded_shape = shape[:-1] + (shape[-1] + misalign, )
known_data = np.resize(known_data, padded_shape)
idata = bf.ndarray(known_data, space='cuda')
known_data = known_data[..., misalign:]
idata = idata[..., misalign:]
oshape = list(shape)
oshape[axes[-1]] = shape[axes[-1]] // 2 + 1
odata = bf.ndarray(shape=oshape, dtype='cf32', space='cuda')
fft = Fft()
fft.init(idata, odata, axes=axes)
fft.execute(idata, odata)
known_result = gold_rfftn(known_data.astype(np.float32) / scale,
axes=axes)
np.testing.assert_allclose(odata.copy('system'), known_result, RTOL,
ATOL)
def run_test_c2c_impl(self, shape, axes, inverse=False, fftshift=False):
shape = list(shape)
shape[-1] *= 2 # For complex
known_data = np.random.uniform(size=shape).astype(np.float32).view(
np.complex64)
idata = bf.ndarray(known_data, space='cuda')
odata = bf.empty_like(idata)
fft = Fft()
fft.init(idata, odata, axes=axes, apply_fftshift=fftshift)
fft.execute(idata, odata, inverse)
if inverse:
if fftshift:
known_data = np.fft.ifftshift(known_data, axes=axes)
# Note: Numpy applies normalization while CUFFT does not
norm = reduce(lambda a, b: a * b,
[known_data.shape[d] for d in axes])
known_result = gold_ifftn(known_data, axes=axes) * norm
else:
known_result = gold_fftn(known_data, axes=axes)
if fftshift:
known_result = np.fft.fftshift(known_result, axes=axes)
x = (np.abs(odata.copy('system') - known_result) / known_result >
RTOL).astype(np.int32)
a = odata.copy('system')
b = known_result
np.testing.assert_allclose(odata.copy('system'), known_result, RTOL,
ATOL)
def run_test_r2c_dtype(self,
shape,
axes,
dtype=np.float32,
scale=1.,
misalign=0):
known_data = np.random.normal(size=shape).astype(np.float32)
known_data = (known_data * scale).astype(dtype)
# Force misaligned data
padded_shape = shape[:-1] + (shape[-1] + misalign, )
known_data = np.resize(known_data, padded_shape)
idata = bf.ndarray(known_data, space='cuda_managed')
known_data = known_data[..., misalign:]
idata = idata[..., misalign:]
oshape = list(shape)
oshape[axes[-1]] = shape[axes[-1]] // 2 + 1
odata = bf.ndarray(shape=oshape, dtype='cf32', space='cuda_managed')
fft = Fft()
fft.init(idata, odata, axes=axes)
fft.execute(idata, odata)
stream_synchronize()
known_result = gold_rfftn(known_data.astype(np.float32) / scale,
axes=axes)
compare(odata, known_result)
def run_test_r2c(self, shape, axes):
known_data = np.random.uniform(size=shape).astype(np.float32)
idata = bf.ndarray(known_data, space='cuda')
oshape = list(shape)
oshape[axes[-1]] = shape[axes[-1]] // 2 + 1
odata = bf.ndarray(shape=oshape, dtype='cf32', space='cuda')
fft = Fft()
fft.init(idata, odata, axes=axes)
fft.execute(idata, odata)
known_result = gold_rfftn(known_data, axes=axes)
np.testing.assert_allclose(odata.copy('system'), known_result, RTOL,
ATOL)
def run_test_c2r(self, shape, axes):
ishape = list(shape)
ishape[axes[-1]] = shape[axes[-1]] // 2 + 1
ishape[-1] *= 2 # For complex
known_data = np.random.uniform(size=ishape).astype(np.float32).view(
np.complex64)
idata = bf.ndarray(known_data, space='cuda')
odata = bf.ndarray(shape=shape, dtype='f32', space='cuda')
fft = Fft()
fft.init(idata, odata, axes=axes)
fft.execute(idata, odata)
# Note: Numpy applies normalization while CUFFT does not
norm = reduce(lambda a, b: a * b, [shape[d] for d in axes])
known_result = gold_irfftn(known_data, axes=axes) * norm
np.testing.assert_allclose(odata.copy('system'), known_result, RTOL,
ATOL)
def run_test_c2r_impl(self, shape, axes, fftshift=False):
ishape = list(shape)
oshape = list(shape)
ishape[axes[-1]] = shape[axes[-1]] // 2 + 1
oshape[axes[-1]] = (ishape[axes[-1]] - 1) * 2
ishape[-1] *= 2 # For complex
known_data = np.random.normal(size=ishape).astype(np.float32).view(np.complex64)
idata = bf.ndarray(known_data, space='cuda')
odata = bf.ndarray(shape=oshape, dtype='f32', space='cuda')
fft = Fft()
fft.init(idata, odata, axes=axes, apply_fftshift=fftshift)
fft.execute(idata, odata)
# Note: Numpy applies normalization while CUFFT does not
norm = reduce(lambda a, b: a * b, [shape[d] for d in axes])
if fftshift:
known_data = np.fft.ifftshift(known_data, axes=axes)
known_result = gold_irfftn(known_data, axes=axes) * norm
compare(odata.copy('system'), known_result)
