def test_shift(self):
shape = (55, 66, 77)
a = np.random.randint(65536, size=shape).astype(np.int32)
a = bf.asarray(a, space='cuda')
b = bf.empty_like(a)
for _ in range(3):
bf.map("b = a(_-a.shape()/2)", data={'a': a, 'b': b})
a = a.copy('system')
b = b.copy('system')
np.testing.assert_equal(b, np.fft.fftshift(a))
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
shape = idata.shape
ind_names = ['i%i' % i for i in xrange(idata.ndim)]
inds = list(ind_names)
for ax in self.axes:
inds[ax] = '-' + inds[ax]
inds = ','.join(inds)
bf.map("b = a(%s)" % inds, shape, *ind_names, a=idata, b=odata)
def test_manydim(self):
known_data = np.arange(3**8).reshape([3] * 8).astype(np.float32)
a = bf.asarray(known_data, space='cuda')
a = a[:, :, :, :, :2, :, :, :]
b = bf.empty_like(a)
for _ in range(3):
bf.map("b = a+1", data={'a': a, 'b': b})
a = a.copy('system')
b = b.copy('system')
np.testing.assert_equal(b, a + 1)
def test_scalar(self):
n = 7919
# Note: Python integer division rounds to -inf, while C rounds toward 0
# We avoid the problem here by using only positive values
x = np.random.randint(1, 256, size=n)
x = bf.asarray(x, space='cuda')
y = bf.empty_like(x)
bf.map("y = (x-m)/s", x=x, y=y, m=1, s=3)
x = x.copy('system')
y = y.copy('system')
np.testing.assert_equal(y, (x - 1) // 3)
def test_explicit_indexing(self):
shape = (55,66,77)
a = np.random.randint(65536, size=shape).astype(np.int32)
a = bf.asarray(a, space='cuda')
b = bf.empty((a.shape[2],a.shape[0], a.shape[1]), a.dtype, 'cuda')
for _ in xrange(3):
bf.map("b(i,j,k) = a(j,k,i)", shape=b.shape, axis_names=('i','j','k'),
data={'a': a, 'b': b}, block_shape=(64,4), block_axes=('i','k'))
a = a.copy('system')
b = b.copy('system')
np.testing.assert_equal(b, a.transpose([2,0,1]))
def test_scalar(self):
n = 7919
# Note: Python integer division rounds to -inf, while C rounds toward 0
# We avoid the problem here by using only positive values
x = np.random.randint(1, 256, size=n)
x = bf.asarray(x, space='cuda')
y = bf.empty_like(x)
for _ in xrange(3):
bf.map("y = (x-m)/s", data={'x': x, 'y': y, 'm': 1, 's': 3})
x = x.copy('system')
y = y.copy('system')
np.testing.assert_equal(y, (x - 1) // 3)
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
#print ospan.data.shape, ispan.data.shape
bf.map("b(i, j, k, l) = a(i, j, k, l)",
ospan.data.shape,
'i',
'j',
'k',
'l',
a=ispan.data,
b=ospan.data)
def test_broadcast(self):
n = 89
a = np.arange(n).astype(np.float32)
a = bf.asarray(a, space='cuda')
b = a[:, None]
c = bf.empty((a.shape[0], b.shape[0]), a.dtype,
'cuda') # TODO: Need way to compute broadcast shape
bf.map("c = a*b", a=a, b=b, c=c)
a = a.copy('system')
b = b.copy('system')
c = c.copy('system')
np.testing.assert_equal(c, a * b)
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
beta = 0. if self.frame_count == 0 else 1.
bf.map("b = beta * b + (b_type)a", a=idata, b=odata, beta=beta)
self.frame_count += 1
if self.frame_count == self.nframe:
ncommit = 1
self.frame_count = 0
else:
ncommit = 0
return ncommit
def test_custom_shape(self):
shape = (55,66,77)
a = np.random.randint(65536, size=shape).astype(np.int32)
a = bf.asarray(a, space='cuda')
b = bf.empty((a.shape[0],a.shape[2]), a.dtype, 'cuda')
j = 11
for _ in xrange(3):
bf.map("b(i,k) = a(i,j,k)", shape=b.shape, axis_names=('i','k'),
data={'a': a, 'b': b, 'j': j})
a = a.copy('system')
b = b.copy('system')
np.testing.assert_equal(b, a[:,j,:])
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
axes = []
for i in range(len(idata.shape)):
axes.append(chr(i + ord('i')))
bf.map('b(%s) = Complex<float>(a(%s))' %
(','.join(axes), ','.join(axes)), {
'a': idata,
'b': odata
},
axis_names=axes,
shape=idata.shape)
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
shape = idata.shape
ind_names = ['i%i' % i for i in xrange(idata.ndim)]
inds = list(ind_names)
for ax in self.axes:
if self.inverse:
inds[ax] += '-(a.shape(%i)-a.shape(%i)/2)' % (ax, ax)
else:
inds[ax] += '-a.shape(%i)/2' % ax
inds = ','.join(inds)
bf.map("b = a(%s)" % inds, shape, *ind_names, a=idata, b=odata)
def run_simple_test(self, x, funcstr, func):
x_orig = x
x = bf.asarray(x, 'cuda')
y = bf.empty_like(x)
x.flags['WRITEABLE'] = False
x.bf.immutable = True # TODO: Is this actually doing anything? (flags is, just not sure about bf.immutable)
bf.map(funcstr, x=x, y=y)
x = x.copy('system')
y = y.copy('system')
if isinstance(x_orig, bf.ndarray):
x_orig = x
# Note: Using func(x) is dangerous because bf.ndarray does things like
# lazy .conj(), which break when used as if it were np.ndarray.
np.testing.assert_equal(y, func(x_orig))
def run_simple_test(self, x, funcstr, func):
x_orig = x
x = bf.asarray(x, 'cuda_managed')
y = bf.empty_like(x)
x.flags['WRITEABLE'] = False
x.bf.immutable = True # TODO: Is this actually doing anything? (flags is, just not sure about bf.immutable)
for _ in range(3):
bf.map(funcstr, {'x': x, 'y': y})
stream_synchronize()
if isinstance(x_orig, bf.ndarray):
x_orig = x
# Note: Using func(x) is dangerous because bf.ndarray does things like
# lazy .conj(), which break when used as if it were np.ndarray.
np.testing.assert_equal(y, func(x_orig))
def test_complex_integer(self):
n = 7919
for in_dtype in ('ci4', 'ci8', 'ci16', 'ci32'):
a_orig = bf.ndarray(shape=(n, ), dtype=in_dtype, space='system')
try:
a_orig['re'] = np.random.randint(256, size=n)
a_orig['im'] = np.random.randint(256, size=n)
except ValueError:
# ci4 is different
a_orig['re_im'] = np.random.randint(256, size=n)
for out_dtype in (in_dtype, 'cf32'):
a = a_orig.copy(space='cuda')
b = bf.ndarray(shape=(n, ), dtype=out_dtype, space='cuda')
bf.map('b(i) = a(i)', {
'a': a,
'b': b
},
shape=a.shape,
axis_names=('i', ))
a = a.copy(space='system')
try:
a = a['re'] + 1j * a['im']
except ValueError:
# ci4 is different
a = np.int8(a['re_im'] & 0xF0) + 1j * np.int8(
(a['re_im'] & 0x0F) << 4)
a /= 16
b = b.copy(space='system')
try:
b = b['re'] + 1j * b['im']
except ValueError:
# ci4 is different
b = np.int8(b['re_im'] & 0xF0) + 1j * np.int8(
(b['re_im'] & 0x0F) << 4)
b /= 16
except IndexError:
# pass through cf32
pass
np.testing.assert_equal(a, b)
def test_polarisation_products(self):
n = 89
real = np.random.randint(-127, 128, size=(n,2)).astype(np.float32)
imag = np.random.randint(-127, 128, size=(n,2)).astype(np.float32)
a = real + 1j * imag
a_orig = a
a = bf.asarray(a, space='cuda')
b = bf.empty_like(a)
for _ in xrange(3):
bf.map('''
auto x = a(_,0);
auto y = a(_,1);
b(_,0).assign(x.mag2(), y.mag2());
b(_,1) = x*y.conj();
''', shape=b.shape[:-1], data={'a': a, 'b': b})
b = b.copy('system')
a = a_orig
gold = np.empty_like(a)
def mag2(x):
return x.real * x.real + x.imag * x.imag
gold[...,0] = mag2(a[...,0]) + 1j * mag2(a[...,1])
gold[...,1] = a[...,0] * a[...,1].conj()
np.testing.assert_equal(b, gold)
def on_data(self, ispan, ospan):
idata = ispan.data
odata = ospan.data
itype = DataType(idata.dtype)
otype = DataType(odata.dtype)
if self.ifmt == 'matrix' and self.ofmt == 'matrix':
# Make a full-matrix copy of the lower-only input matrix
# odata[t,c,i,p,j,q] = idata[t,c,i,p,j,q] (lower filled only)
shape_nopols = list(idata.shape)
del shape_nopols[5]
del shape_nopols[3]
idata = idata.view(itype.as_vector(2))
odata = odata.view(otype.as_vector(2))
bf.map('''
bool in_lower_triangle = (i > j);
if( in_lower_triangle ) {
odata(t,c,i,0,j,0) = idata(t,c,i,0,j,0);
odata(t,c,i,1,j,0) = idata(t,c,i,1,j,0);
} else {
auto x = idata(t,c,j,0,i,0);
auto y = idata(t,c,j,1,i,0);
auto x1 = x[1];
x[0] = x[0].conj();
x[1] = y[0].conj();
if( i != j ) {
y[0] = x1.conj();
}
y[1] = y[1].conj();
odata(t,c,i,0,j,0) = x;
odata(t,c,i,1,j,0) = y;
}
''',
shape=shape_nopols,
axis_names=['t', 'c', 'i', 'j'],
data={
'idata': idata,
'odata': odata
})
elif self.ifmt == 'matrix' and self.ofmt == 'storage':
assert (idata.shape[2] <= 2048)
idata = idata.view(itype.as_vector(2))
odata = odata.view(otype.as_vector(4))
# TODO: Support L/R as well as X/Y pols
bf.map('''
// TODO: This only works up to 2048 in single-precision
#define project_triangular(i, j) ((i)*((i)+1)/2 + (j))
int i = int((sqrt(8.f*(b)+1)-1)/2);
int j = b - project_triangular(i, 0);
auto x = idata(t,c,i,0,j,0);
auto y = idata(t,c,i,1,j,0);
if( i == j ) {
x[1] = y[0].conj();
}
idata_type::value_type eye(0, 1);
auto I = (x[0] + y[1]);
auto Q = (x[0] - y[1]);
auto U = (x[1] + y[0]);
auto V = (x[1] - y[0]) * eye;
odata(t,b,c,0) = odata_type(I,Q,U,V);
''',
shape=odata.shape[:-1],
axis_names=['t', 'b', 'c'],
data={
'idata': idata,
'odata': odata
},
block_shape=[64, 8]) # TODO: Tune this
#elif self.ifmt == 'matrix' and self.ofmt == 'triangular':
elif self.ifmt == 'storage' and self.ofmt == 'matrix':
oshape_nopols = list(odata.shape)
del oshape_nopols[5]
del oshape_nopols[3]
idata = idata.view(itype.as_vector(4))
odata = odata.view(otype.as_vector(2))
bf.map('''
bool in_upper_triangle = (i < j);
auto b = in_upper_triangle ? j*(j+1)/2 + i : i*(i+1)/2 + j;
auto IQUV = idata(t,b,c,0);
auto I = IQUV[0], Q = IQUV[1], U = IQUV[2], V = IQUV[3];
idata_type::value_type eye(0, 1);
auto xx = 0.5f*(I + Q);
auto xy = 0.5f*(U - V*eye);
auto yx = 0.5f*(U + V*eye);
auto yy = 0.5f*(I - Q);
if( i == j ) {
xy = yx.conj();
}
if( in_upper_triangle ) {
auto tmp_xy = xy;
xx = xx.conj();
xy = yx.conj();
yx = tmp_xy.conj();
yy = yy.conj();
}
odata(t,c,i,0,j,0) = odata_type(xx, xy);
odata(t,c,i,1,j,0) = odata_type(yx, yy);
''',
shape=oshape_nopols,
axis_names=['t', 'c', 'i', 'j'],
data={
'idata': idata,
'odata': odata
},
block_shape=[64, 8]) # TODO: Tune this
else:
raise NotImplementedError
