class Measurement:
_kernel = cupy.ReductionKernel(
"uint64 target_mask, float64 p, T q_inout", "float64 b",
"(!(_j & target_mask)) * (q_inout * q_inout.conj()).real()", "a + b",
"""
q_inout = (!(_i & target_mask) == (a <= p)) * q_inout / sqrt((a <= p) * a + (a > b) * (1.0 - a))
""", "0.0")
no_cache = True
def __init__(self):
pass
def apply(self, helper, qubits, targets):
n_qubits = helper["n_qubits"]
i = helper["indices"]
p = random.random()
for target in slicing(targets, n_qubits):
target_mask = 1 << target
a = self._kernel(target_mask, p, qubits)
helper["cregs"][target] = int(a > p)
return qubits
def to_qasm(self, helper, targets):
n_qubits = helper["n_qubits"]
qasm = []
for target in slicing(targets, n_qubits):
qasm.append("measure q[{}] -> c[{}];".format(target, target))
return qasm
def forward_gpu(self, inputs):
if not self.gpu_optim:
return self.forward_cpu(inputs)
xp = cuda.get_array_module(*inputs)
x, gamma, beta = inputs
a = x - xp.mean(x, axis=1, keepdims=True)
assert len(a.shape) == 2
H = a.shape[1]
# inv_norm = inv_norm_comp(a/math.sqrt(H), axis=1, keepdims=True) # 1.0/xp.sqrt(xp.sum(a*a, axis=1, keepdims=True) + self.eps)
inv_norm = cp.ReductionKernel(
'T x', # input params
'T y', # output params
'x * x', # map
'a + b', # reduce
'y = 1.0/sqrt(a/%f + %f)' % (H, self.eps), # post-reduction map
'0', # identity value
'inv_norm_comp' # kernel name
)(a, axis=1, keepdims=True)
self.inv_norm = inv_norm
normalized, scaled = scale_output(a, inv_norm, gamma, beta)
self.normalized = normalized
return scaled,
def test_optimize_reduction_kernel(self):
my_sum = cupy.ReductionKernel(
'T x', 'T out', 'x', 'a + b', 'out = a', '0', 'my_sum')
x = testing.shaped_arange((3, 4), cupy)
y1 = my_sum(x, axis=1)
with cupyx.optimizing.optimize():
y2 = my_sum(x, axis=1)
testing.assert_array_equal(y1, y2)
def test_optimize_cache(self):
if (_accelerator.ACCELERATOR_CUB
in _accelerator.get_reduction_accelerators()):
pytest.skip('optimize cannot be mocked for CUB reduction')
target = cupyx.optimizing._optimize._optimize
target_full_name = '{}.{}'.format(target.__module__, target.__name__)
with mock.patch(target_full_name) as optimize_impl:
my_sum = cupy.ReductionKernel('T x', 'T out', 'x', 'a + b',
'out = a', '0', 'my_sum')
my_sum_ = cupy.ReductionKernel('T x', 'T out', 'x', 'a + b',
'out = a', '0', 'my_sum_')
x = testing.shaped_arange((3, 4), cupy)
x_ = testing.shaped_arange((3, 4), cupy)
y = testing.shaped_arange((4, 4), cupy)
z = testing.shaped_arange((3, 4), cupy)[::-1]
assert x.strides == y.strides
assert x.shape == z.shape
with cupyx.optimizing.optimize():
my_sum(x, axis=1)
assert optimize_impl.call_count == 1
my_sum(x, axis=1)
assert optimize_impl.call_count == 1
my_sum(x, axis=0)
assert optimize_impl.call_count == 2
my_sum(x_, axis=1)
assert optimize_impl.call_count == 2
my_sum(y, axis=1)
assert optimize_impl.call_count == 3
my_sum(z, axis=1)
assert optimize_impl.call_count == 4
my_sum_(x, axis=1)
assert optimize_impl.call_count == 5
with cupyx.optimizing.optimize(key='new_key'):
my_sum(x, axis=1)
assert optimize_impl.call_count == 6
with cupyx.optimizing.optimize(key=None):
my_sum(x, axis=1)
assert optimize_impl.call_count == 6
my_sum(x)
assert optimize_impl.call_count == 7
def test_invalid_kernel_name(self):
with self.assertRaisesRegex(ValueError, 'Invalid kernel name'):
cupy.ReductionKernel('T x',
'T y',
'x',
'a + b',
'y = a',
'0',
name='1')
def __get_count_votes_cupy_kernel(invert):
sign = '>' if invert else '<'
return cupy.ReductionKernel('X x',
'Y y',
'_raw_x[_in_ind.size()/2]{}x'.format(sign),
'a + b',
'y = a',
'0',
'lt',
reduce_type='int')
def cupy_threshold_local_mean(*args, **kwargs):
# Code snippet taken from https://github.com/cupy/cupy/issues/3909
my_mean = cupy.ReductionKernel(
'T x', # input params
'T y', # output params
'x', # map
'a + b', # reduce
'y = a / _in_ind.size()', # An undocumented variable and a hack
'0', # identity value
'mean' # kernel name
)
return my_mean
def reduce(in_params, out_params, map_expr, reduce_expr, post_map_expr,
identity, name, **kwargs):
"""Creates a global reduction kernel function.
This function uses :func:`~chainer.cuda.memoize` to cache the resulting
kernel object, i.e. the resulting kernel object is cached for each argument
combination and CUDA device.
The arguments are the same as those for
:class:`cupy.ReductionKernel`, except that the ``name`` argument is
mandatory.
"""
check_cuda_available()
return cupy.ReductionKernel(in_params, out_params, map_expr, reduce_expr,
post_map_expr, identity, name, **kwargs)
def test_optimize_cache_multi_gpus(self):
target = cupyx.optimizing._optimize._optimize
target_full_name = '{}.{}'.format(target.__module__, target.__name__)
with mock.patch(target_full_name) as optimize_impl:
my_sum = cupy.ReductionKernel(
'T x', 'T out', 'x', 'a + b', 'out = a', '0', 'my_sum')
with cupyx.optimizing.optimize():
with cupy.cuda.Device(0):
x = testing.shaped_arange((3, 4), cupy)
my_sum(x, axis=1)
assert optimize_impl.call_count == 1
with cupy.cuda.Device(1):
x = testing.shaped_arange((3, 4), cupy)
my_sum(x, axis=1)
assert optimize_impl.call_count == 2
def update_core_gpu(self, param):
grad = param.grad
if grad is None:
return
hp = self.hyperparam
p = self.state['p']
if HamiltonianExplicitRule._kernel_x is None:
HamiltonianExplicitRule._kernel_x = cp.ElementwiseKernel(
'T epsilon, T p, T denomp, T param', 'T x',
'x = param + epsilon * p / denomp', 'Hamiltonian_x')
if HamiltonianExplicitRule._kernel_r is None:
HamiltonianExplicitRule._kernel_r = cp.ReductionKernel(
'T p', 'T denomp', 'p * p', 'a + b', 'denomp = sqrt(a)',
'1', 'relativistic')
if hp.approx == 'first':
# p
if HamiltonianExplicitRule._kernel_p is None:
HamiltonianExplicitRule._kernel_p = cp.ElementwiseKernel(
'T delta, T epsilon, T grad, T p0', 'T p1',
'p1 = p0 * delta - epsilon * delta * grad',
'Hamiltonian_p')
p = HamiltonianExplicitRule._kernel_p(hp.delta, hp.epsilon,
grad, p)
# x
denomp = HamiltonianExplicitRule._kernel_r(p)
param.data = HamiltonianExplicitRule._kernel_x(
hp.epsilon, p, denomp, param.data)
else:
if HamiltonianExplicitRule._kernel_p is None:
HamiltonianExplicitRule._kernel_p = cp.ElementwiseKernel(
'T delta, T epsilon, T grad, T p0', 'T p1',
'p1 = p0 * (2.0 - (1.0 / delta)) - epsilon * grad',
'Hamiltonian_p')
else:
# p
p = HamiltonianExplicitRule._kernel_p(
hp.delta, hp.epsilon, grad, p)
# x
denomp = HamiltonianExplicitRule._kernel_r(p)
param.data = HamiltonianExplicitRule._kernel_x(
hp.epsilon, p, denomp, param.data)
def test_optimize_cache_multi_gpus(self):
if (_accelerator.ACCELERATOR_CUB
in _accelerator.get_reduction_accelerators()):
pytest.skip('optimize cannot be mocked for CUB reduction')
target = cupyx.optimizing._optimize._optimize
target_full_name = '{}.{}'.format(target.__module__, target.__name__)
with mock.patch(target_full_name) as optimize_impl:
my_sum = cupy.ReductionKernel('T x', 'T out', 'x', 'a + b',
'out = a', '0', 'my_sum')
with cupyx.optimizing.optimize():
with cupy.cuda.Device(0):
x = testing.shaped_arange((3, 4), cupy)
my_sum(x, axis=1)
assert optimize_impl.call_count == 1
with cupy.cuda.Device(1):
x = testing.shaped_arange((3, 4), cupy)
my_sum(x, axis=1)
assert optimize_impl.call_count == 2
def test_optimize_pickle(self):
my_sum = cupy.ReductionKernel(
'T x', 'T out', 'x', 'a + b', 'out = a', '0', 'my_sum')
x = testing.shaped_arange((3, 4), cupy)
with tempfile.TemporaryDirectory() as directory:
filepath = directory + '/optimize_params'
with cupyx.optimizing.optimize() as context:
my_sum(x, axis=1)
params_map = context._params_map
context.save(filepath)
cupy.core._optimize_config._clear_all_contexts_cache()
with cupyx.optimizing.optimize() as context:
assert params_map.keys() != context._params_map.keys()
context.load(filepath)
assert params_map.keys() == context._params_map.keys()
with cupyx.optimizing.optimize(key='other_key') as context:
with pytest.raises(ValueError):
context.load(filepath)
vert_norm[a*3+1] += yy;
vert_norm[a*3+2] += zz;
vert_norm[b*3] += xx;
vert_norm[b*3+1] += yy;
vert_norm[b*3+2] += zz;
vert_norm[c*3] += xx;
vert_norm[c*3+1] += yy;
vert_norm[c*3+2] += zz;
float area = sqrt(xx*xx + yy*yy + zz*zz) / 3;
vertex_weight[a] += area;
vertex_weight[b] += area;
vertex_weight[c] += area;
}
''', 'calc_vertex_norm')
calc_avg_vertex = cp.ReductionKernel('T x, T w', 'T y', 'x * w', 'a + b',
'y = a', '0', 'calc_avg_vertex')
class Obj3D(object):
def __init__(self, filename, lines=None):
if lines is None:
file = open(filename, 'r')
lines = file.readlines()
iter = 0
if lines[0].rstrip() == 'OFF':
self.numVertices, self.numFaces, self.numEdges = (
int(x) for x in str.split(lines[1]))
iter = 2
elif lines[0][0:3] == 'OFF':
self.numVertices, self.numFaces, self.numEdges = (
def get_sum_func(self):
return cupy.ReductionKernel(
'T x', 'T out', 'x', 'a + b', 'out = a', '0', 'my_sum')
import cupy
import numpy
add_kernel = cupy.ReductionKernel(
"T x",
"T m",
"x",
"a + b",
"m = a",
"0",
"avg",
)
def avg(x):
return add_kernel(x) / x.size
x = cupy.arange(10, dtype=numpy.float64)
print(avg(x), numpy.mean(x))
import cupy as cp
# Double precision
curesiduals = cp.ElementwiseKernel('float64 holo, float64 data, float64 noise',
'float64 residuals',
'residuals = (holo - data) / noise',
'curesiduals')
cuchisqr = cp.ReductionKernel(
'float64 holo, float64 data, float64 noise', 'float64 chisqr',
'((holo - data) / noise) * ((holo - data) / noise)', 'a + b', 'chisqr = a',
'0', 'cuchisqr')
cuabsolute = cp.ReductionKernel('float64 holo, float64 data, float64 noise',
'float64 s', 'abs((holo - data) / noise)',
'a + b', 's = a', '0', 'cuabsolute')
# Single precision
curesidualsf = cp.ElementwiseKernel(
'float32 holo, float32 data, float32 noise', 'float32 residuals',
'residuals = (holo - data) / noise', 'curesiduals')
cuchisqrf = cp.ReductionKernel(
'float32 holo, float32 data, float32 noise', 'float32 chisqr',
'((holo - data) / noise) * ((holo - data) / noise)', 'a + b', 'chisqr = a',
'0', 'cuchisqr')
cuabsolutef = cp.ReductionKernel('float32 holo, float32 data, float32 noise',
'float32 s', 'abs((holo - data) / noise)',
class TestFilterComplexFast(FilterTestCaseBase):
@testing.numpy_cupy_allclose(atol=1e-5, rtol=1e-5, scipy_name='scp')
@testing.with_requires('scipy>=1.6.0')
def test_filter(self, xp, scp):
return self._filter(xp, scp)
# Kernels and Functions for testing generic_filter
rms_raw = cupy.RawKernel(
'''extern "C" __global__
void rms(const double* x, int filter_size, double* y) {
double ss = 0;
for (int i = 0; i < filter_size; ++i) { ss += x[i]*x[i]; }
y[0] = ss/filter_size;
}''', 'rms')
rms_red = cupy.ReductionKernel('X x', 'Y y', 'x*x', 'a + b',
'y = a/_in_ind.size()', '0', 'rms')
def rms_pyfunc(x):
return (x * x).sum() / len(x)
lt_raw = cupy.RawKernel(
'''extern "C" __global__
void lt(const double* x, int filter_size, double* y) {
int n = 0;
double c = x[filter_size/2];
for (int i = 0; i < filter_size; ++i) { n += c>x[i]; }
y[0] = n;
}''', 'lt')
lt_red = cupy.ReductionKernel('X x',
def setUp(self):
self.my_sum = cupy.ReductionKernel('T x', 'T out', 'x', 'a + b',
'out = a', '0', 'my_sum')
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import cupy as cp
import numpy as np
if __name__ == '__main__':
x = cp.arange(10, dtype=np.float32).reshape(2, 5)
#
L2norm_kernel = cp.ReductionKernel(
'T x', # in_params:入力
'T y', # out_params:出力
' x * x', #map_expr:前処理
'a + b', #reduce_expr:リデュース
'y = sqrt(a)', #post_map_expr:後処理
'0', #identity:初期値
'l2norm') #name:名前
y = L2norm_kernel(x)
print(x)
print(y)
##from p.13 of https://www.slideshare.net/ryokuta/cupy
#import chainer
#import cupy as cp
##import numpy as np
#
#l2norm_kernel = cp.ReductionKernel(
# 'T x', #input
# 'T y', #output
# 'x * x', #preprocess
# 'a + b', #reduce ?
# 'y=sqrt(a)', #post-process
w_flat = w.reshape(-1, w.shape[2])
if w_flat_sq is None:
w_flat_sq = xp.power(w_flat, 2).sum(axis=1, keepdims=True)
x_sq = xp.power(x, 2).sum(axis=1, keepdims=True)
num = xp.dot(x, w_flat.T)
denum = xp.sqrt(x_sq * w_flat_sq.T)
similarity = xp.nan_to_num(num / denum)
return 1 - similarity
if _cupy_available:
_manhattan_distance_kernel = cp.ReductionKernel('T x, T w', 'T y',
'abs(x-w)', 'a+b', 'y = a',
'0', 'l1norm')
def manhattan_distance(x, w, xp=default_xp):
"""Calculate Manhattan distance
It is very slow (~10x) compared to euclidean distance
TODO: improve performance. Maybe a custom kernel is necessary
NB: result shape is (N,X*Y)
"""
if xp.__name__ == 'cupy':
d = _manhattan_distance_kernel(x[:, xp.newaxis, xp.newaxis, :],
w[xp.newaxis, :, :, :],
import math
import cupy as cp
import numpy as np
l2norm_kernel = cp.ReductionKernel(
'T x', # input params
'T y', # output params
'x * x', # map
'a + b', # reduce
'y = sqrt(a)', # post-reduction map
'0', # identity value
'l2norm' # kernel name
)
x = cp.arange(5, dtype=np.float32).reshape(1, 5)
print(x)
print(l2norm_kernel(x, axis=1))
print(math.sqrt(0 * 0 + 1 * 1 + 2 * 2 + 3 * 3 + 4 * 4))
import tabulate
import numpy as np
import cupy as cp
from timer import stopwatch
# define ReductionKernel
rd_kinetic_kernel = cp.ReductionKernel(
'T m, T v', # input
'T y', # output
'v * v', # pre-process
'a + b', # reduction
'y = 0.5 * m * a', # post-process
'0', # initial
'rd_kinetic') # name
@stopwatch
def kinetic_kernel(m, v):
return rd_kinetic_kernel(m, v)
@stopwatch
def kinetic(m, v):
return 0.5 * m * np.sum(v*v)
# test ReductionKernel
N = [1,10,1000000,10000000,100000000]
times_cpu = []
times_gpu = []
times_kernel = []
for n in N:
v = np.sin(np.linspace(-np.pi, np.pi, n)).astype(np.float32) / n
def setup(self, raster_dim, zone_dim, backend):
W = H = raster_dim
zW = zH = zone_dim
# Make sure that the raster dim is multiple of the zones dim
assert(W % zW == 0)
assert(H % zH == 0)
# initialize the values raster
self.values = get_xr_dataarray((H, W), backend)
# initialize the zones raster
zones = xr.DataArray(np.zeros((H, W)))
hstep = H//zH
wstep = W//zW
for i in range(zH):
for j in range(zW):
zones[i * hstep: (i+1)*hstep, j*wstep: (j+1)*wstep] = i*zW + j
''' zones now looks like this
>>> zones = np.array([
[0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
[0, 0, 0, 0, 0, 1, 1, 1, 1, 1],
[2, 2, 2, 2, 2, 3, 3, 3, 3, 3],
[2, 2, 2, 2, 2, 3, 3, 3, 3, 3],
[2, 2, 2, 2, 2, 3, 3, 3, 3, 3],
[2, 2, 2, 2, 2, 3, 3, 3, 3, 3],
[2, 2, 2, 2, 2, 3, 3, 3, 3, 3]])
'''
self.zones = create_arr(zones, backend=backend)
# Now setup the custom stat funcs
if backend == 'cupy':
import cupy
l2normKernel = cupy.ReductionKernel(
in_params='T x', out_params='float64 y',
map_expr='x*x', reduce_expr='a+b',
post_map_expr='y = sqrt(a)',
identity='0', name='l2normKernel'
)
self.custom_stats = {
'double_sum': lambda val: val.sum()*2,
'l2norm': lambda val: np.sqrt(cupy.sum(val * val)),
'l2normKernel': lambda val: l2normKernel(val)
}
else:
from xrspatial.utils import ngjit
@ngjit
def l2normKernel(arr):
acc = 0
for x in arr:
acc += x * x
return np.sqrt(acc)
self.custom_stats = {
'double_sum': lambda val: val.sum()*2,
'l2norm': lambda val: np.sqrt(np.sum(val * val)),
'l2normKernel': lambda val: l2normKernel(val)
}
import numpy as np
import cupy as cp
import time
"""
REDUCTION KERNEL
when processing the kernel into a simpler unit
Identity value: This value is used for the initial value of reduction.
Mapping expression: It is used for the pre-processing of each element to be reduced.
Reduction expression: It is an operator to reduce the multiple mapped values. The special variables a and b are used for its operands.
Post mapping expression: It is used to transform the resulting reduced values. The special variable a is used as its input. Output should be written to the output parameter.
"""
normalize_gpu = cp.ReductionKernel(
'float64 x', #input params
'float64 y', #output params
'x * x', # map
'a + b', # reduce
'y = sqrt(a)', # post reduction map
'0', # identity value
'normalize_gpu' # name
)
x = cp.array([0.0,6.0,8.0])
print(normalize_gpu(x))
@contextlib.contextmanager
def timer(message):
cupy.cuda.Stream.null.synchronize()
start = time.time()
yield
cupy.cuda.Stream.null.synchronize()
end = time.time()
print('%s: %f sec' % (message, end - start))
var_kernel = cupy.ElementwiseKernel(
'T x0, T x1, T c0, T c1', 'T out',
'out = (x0 - c0) * (x0 - c0) + (x1 - c1) * (x1 - c1)', 'var_kernel')
sum_kernel = cupy.ReductionKernel('T x, S mask', 'T out', 'mask ? x : 0',
'a + b', 'out = a', '0', 'sum_kernel')
count_kernel = cupy.ReductionKernel('T mask', 'float32 out',
'mask ? 1.0 : 0.0', 'a + b', 'out = a',
'0.0', 'count_kernel')
def fit_xp(X, n_clusters, max_iter):
assert X.ndim == 2
# Get NumPy or CuPy module from the supplied array.
xp = cupy.get_array_module(X)
n_samples = len(X)
# Make an array to store the labels indicating which cluster each sample is
# contained.
header[key.format(i)] = basename(f)
header['NCOMBINE'] = ldata
hdu = mkhdu(combined, header=header)
hdul = fits.HDUList([hdu] + exthdus)
hdul.writeto(name, **kwargs)
print('Combine: {:d} frames, Output: {}'.format(ldata, basename(name)))
return combined
check_sum = cp.ReductionKernel(in_params='T x',
out_params='uint64 z',
map_expr='(x!=0)',
reduce_expr='a+b',
post_map_expr='z=a',
identity='0',
name='check_sum')
weightedsum = cp.ReductionKernel(in_params='T x, T f',
out_params='T y',
map_expr='f * (f ? x : 0)',
reduce_expr='a+b',
post_map_expr='y=a',
identity='0',
name='weightedsum')
weightedvar = cp.ReductionKernel(in_params='T x, T m, T f',
out_params='T y',
map_expr='square(x,m,f)',
class KernelList:
ker_X = cp.ElementwiseKernel(
in_params="raw T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
""",
operation="""
y = x[i^mask];
""",
name="X",
)
ker_Y = cp.ElementwiseKernel(
in_params="raw T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> img = thrust::complex<double>(0,1);
""",
operation="""
if(i&mask) y = -img*x[i^mask];
else y = img*x[i^mask];
""",
name="Y",
)
ker_Z = cp.ElementwiseKernel(
in_params="T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
""",
operation="""
if(i&mask) y = -x;
else y = x;
""",
name="Z",
)
ker_H = cp.ElementwiseKernel(
in_params="raw T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> sq2 = thrust::complex<double>(1/sqrt(2.),0);
""",
operation="""
if(i&mask) y = (-x[i] + x[i^mask])*sq2;
else y = (x[i] + x[i^mask])*sq2;
""",
name="H",
)
ker_S = cp.ElementwiseKernel(
in_params="T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> img = thrust::complex<double>(0,1);
""",
operation="""
if(i&mask) y = img*x;
else y = x;
""",
name="S",
)
ker_T = cp.ElementwiseKernel(
in_params="S x, int32 k",
out_params="S y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> ph = thrust::complex<double>(1/sqrt(2.),1/sqrt(2.));
""",
operation="""
if(i&mask) y = ph*x;
else y = x;
""",
name="T",
)
ker_CX = cp.ElementwiseKernel(
in_params="raw T x, int32 k, int32 u",
out_params="T y",
loop_prep="""
int mask_k = 1<<k;
int mask_u = 1<<u;
""",
operation="""
if(i&mask_k) y = x[i^mask_u];
else y = x[i];
""",
name="CX",
)
ker_CZ = cp.ElementwiseKernel(
in_params="T x, int32 k, int32 u",
out_params="T y",
loop_prep="""
int mask_k = 1<<k;
int mask_u = 1<<u;
""",
operation="""
if((i&mask_k) && (i&mask_u)) y = -x;
else y = x;
""",
name="CZ",
)
ker_Toffoli = cp.ElementwiseKernel(
in_params="raw T x, int32 k, int32 u, int32 t",
out_params="T y",
loop_prep="""
int mask_k = 1<<k;
int mask_u = 1<<u;
int mask_t = 1<<t;
""",
operation="""
if((i&mask_k) && (i&mask_u)) y = x[i^mask_t];
else y = x[i];
""",
name="Toffoli",
)
ker_Xrot = cp.ElementwiseKernel(
in_params="raw T x, int32 k, float64 theta",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> c = thrust::complex<double>(cos(theta),0);
thrust::complex<double> s = thrust::complex<double>(0,sin(theta));
""",
operation="""
y = c*x[i]+s*x[i^mask];
""",
name="Xrot",
)
ker_Yrot = cp.ElementwiseKernel(
in_params="raw T x, int32 k, float64 theta",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> c = thrust::complex<double>(cos(theta),0);
thrust::complex<double> s = thrust::complex<double>(sin(theta),0);
""",
operation="""
if(i&mask) y = c*x[i]-s*x[i^mask];
else y = c*x[i]+s*x[i^mask];
""",
name="Yrot",
)
ker_Zrot = cp.ElementwiseKernel(
in_params="T x, int32 k, float64 theta",
out_params="T y",
loop_prep="""
int mask = 1<<k;
thrust::complex<double> c1 = thrust::complex<double>(cos(theta),sin(theta));
thrust::complex<double> c2 = thrust::complex<double>(cos(theta),-sin(theta));
""",
operation="""
if(i&mask) y = c2*x;
else y = c1*x;
""",
name="Zrot",
)
ker_XXrot = cp.ElementwiseKernel(
in_params="raw T x, int32 k, int32 u,float64 theta",
out_params="T y",
loop_prep="""
int mask_ku = (1<<k)+(1<<u);
thrust::complex<double> c = thrust::complex<double>(cos(theta),0);
thrust::complex<double> s = thrust::complex<double>(0,sin(theta));
""",
operation="""
y = c*x[i] + s*x[i^mask_ku];
""",
name="XXrot",
)
ker_MeasZ0 = cp.ElementwiseKernel(
in_params="T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
""",
operation="""
if(i&mask) y = 0;
else y = x;
""",
name="MeasZ0",
)
ker_MeasZ1 = cp.ElementwiseKernel(
in_params="T x, int32 k",
out_params="T y",
loop_prep="""
int mask = 1<<k;
""",
operation="""
if(i&mask) y = x;
else y = 0;
""",
name="MeasZ1",
)
ker_U = cp.ElementwiseKernel(
in_params="raw T x, int32 k, float64 t0, float64 t1, float64 t2",
out_params="T y",
loop_prep="""
int mask = 1<<k;
double t12p = (t1+t2)/2;
double t12m = (t1-t2)/2;
thrust::complex<double> u00 = thrust::complex<double>(cos(t12p),sin(-t12p)) * cos(t0/2);
thrust::complex<double> u01 = thrust::complex<double>(-cos(t12m),-sin(-t12m)) * sin(t0/2);
thrust::complex<double> u10 = thrust::complex<double>(cos(t12m),sin(t12m)) * sin(t0/2);
thrust::complex<double> u11 = thrust::complex<double>(cos(t12p),sin(t12p)) * cos(t0/2);
""",
operation="""
if(i&mask) y = u10*x[i^mask] + u11*x[i];
else y = u00*x[i] + u01*x[i^mask];
""",
name="U",
)
ker_trace = cp.ReductionKernel("T x", "T y", "x*thrust::conj(x)", "a+b",
"y = a", "0", "trace")
onePauli = [ker_X, ker_Y, ker_Z]
oneClifford = onePauli + [ker_H, ker_S]
oneGate = oneClifford + [ker_T]
twoPauli = []
twoClifford = [ker_CX, ker_CZ]
twoGate = twoPauli + twoClifford + []
threePauli = []
threeClifford = []
threeGate = [ker_Toffoli]
pauli = onePauli + twoPauli + threePauli
clifford = oneClifford + twoClifford
discreteGate = oneGate + twoGate + threeGate
oneRot = [ker_Xrot, ker_Yrot, ker_Zrot]
twoRot = [ker_XXrot]
matchgate = [ker_Zrot, ker_XXrot]
continuusGate = oneRot + twoRot
genericGate = [ker_U]
measurement = [ker_MeasZ0, ker_MeasZ1]
allGate = discreteGate + continuusGate + measurement + genericGate
allGateName = [g.name for g in allGate]
# require target
oneDiscrete = oneGate
# require control, target
twoDiscrete = twoGate
return xp.broadcast_to(x, shape)
else:
# numpy 1.9 doesn't support broadcast_to method
dummy = xp.empty(shape)
bx, _ = xp.broadcast_arrays(x, dummy)
return bx
try:
import cupy as cp
inv_norm_comp = cp.ReductionKernel(
'T x', # input params
'T y', # output params
'x * x', # map
'a + b', # reduce
'y = 1.0/sqrt(a + 1e-5)', # post-reduction map
'0', # identity value
'inv_norm_comp' # kernel name
)
scale_output = cp.ElementwiseKernel(
'T x, T inv_norm, T gamma, T beta', 'T normalized, T scaled', '''
normalized = x * inv_norm;
scaled = normalized * gamma + beta;
''', 'scale_output')
backprop_scale = cp.ElementwiseKernel(
'T inv_norm, T gy_centered, T normalized, T sc_prod', 'T z', '''
z = inv_norm *(gy_centered - normalized * sc_prod);
''', 'backprop_scale')
def calculate_trac(image_mat,
out: cp.array,
delay: int = 1,
order: int = 0) -> cp.array:
"""Temporal radiality auto-cumulant"""
frame_n, h, w = image_mat.shape[:3]
_trac2 = cp.ReductionKernel(
"T X1, T X2, T length",
"T out",
"X1 * X2",
"a + b",
"out = a / length",
"0",
"trac2",
)
_trac3 = cp.ReductionKernel(
"T X1, T X2, T X3, T length",
"T out",
"X1 * X2 * X3",
"a + b",
"out = a / length",
"0",
"trac3",
)
_trac4 = cp.ReductionKernel(
"T X1, T X2, T X3, T X4, T length",
"T out",
"X1 * X2 * X3 * X4",
"a + b",
"out = a / length",
"0",
"trac4",
)
_subtract_product = cp.ElementwiseKernel("T A, T B",
"T C",
"C -= A * B",
"subctract_product",
no_return=True)
if 2 * delay > frame_n and order != 0:
order = 0
warnings.warn(
"Total number of frames is too small to do TRAC, using TRA instead"
)
deltaRt = image_mat - cp.mean(image_mat, axis=0)
if order == 0:
# TRA(time average)
result = cp.mean(image_mat, axis=0)
elif order == 2:
# TRAC2
if delay == 0:
A = B = deltaRt
else:
A = deltaRt[:-delay]
B = deltaRt[delay:]
result = _trac2(A, B, frame_n, axis=0)
elif order == 3:
# TRAC3
if delay == 0:
A = B = C = deltaRt
else:
A = deltaRt[:-2 * delay]
B = deltaRt[delay:-delay]
C = deltaRt[2 * delay:]
result = _trac3(A, B, C, frame_n, axis=0)
elif order == 4:
# TRAC4
if delay == 0:
A = B = C = D = deltaRt
else:
A = deltaRt[:-3 * delay]
B = deltaRt[delay:-2 * delay]
C = deltaRt[2 * delay:-delay]
D = deltaRt[3 * delay:]
result = _trac4(A, B, C, D, frame_n, axis=0)
AB = _trac2(A, B, frame_n, axis=0)
CD = _trac2(C, D, frame_n, axis=0)
_subtract_product(AB * CD, result)
del AB, CD
AC = _trac2(A, C, frame_n, axis=0)
BD = _trac2(B, D, frame_n, axis=0)
_subtract_product(AC * BD, result)
del AC, BD
AD = _trac2(A, D, frame_n, axis=0)
BC = _trac2(B, C, frame_n, axis=0)
_subtract_product(AD * BC, result)
del AD, BC
else:
raise ValueError("sofi-order can only be 2, 3, 4 or 0!")
# result -= cp.min(result)
cp.abs(result, out=out)
bit_length = 64
size = 2**17
n_gpu = 8
print('bit', bit_length)
print('size', size)
print('n_gpu', n_gpu)
dtype_from_bit = {
8: 'uint8',
16: 'uint16',
32: 'uint32',
64: 'uint64',
}
parity_kernel = cupy.ReductionKernel('T x', 'T y', 'x', 'a ^ b', 'y = a', '0',
'parity')
def check(n):
count = 0
while n:
count += 1
n &= n - 1
return count
begin_gen = time()
matrix = numpy.random.randint(0, 1 << bit_length, (size, size),
dtype_from_bit[bit_length])
end_gen = time()
print('mem', matrix.__sizeof__())