def calculate_z_numpy_gpu(q, maxiter, z):
"""Calculate z using numpy on the GPU via gpuarray"""
outputg = gpuarray.to_gpu(
nm.resize(nm.array(0, ), q.shape).astype(nm.int32))
zg = gpuarray.to_gpu(z.astype(nm.complex64))
qg = gpuarray.to_gpu(q.astype(nm.complex64))
# 2.0 as an array
twosg = gpuarray.to_gpu(nm.array([2.0] * zg.size).astype(nm.float32))
# 0+0j as an array
cmplx0sg = gpuarray.to_gpu(
nm.array([0 + 0j] * zg.size).astype(nm.complex64))
# for abs_zg > twosg result
comparison_result = gpuarray.to_gpu(
nm.array([False] * zg.size).astype(nm.bool))
# we'll add 1 to iterg after each iteration
iterg = gpuarray.to_gpu(nm.array([0] * zg.size).astype(nm.int32))
for iter in range(maxiter):
zg = zg * zg + qg
# abs returns a complex (rather than a float) from the complex
# input where the real component is the absolute value (which
# looks like a bug) so I take the .real after abs()
abs_zg = abs(zg).real
comparison_result = abs_zg > twosg
qg = gpuarray.if_positive(comparison_result, cmplx0sg, qg)
zg = gpuarray.if_positive(comparison_result, cmplx0sg, zg)
outputg = gpuarray.if_positive(comparison_result, iterg, outputg)
iterg = iterg + 1
output = outputg.get()
return output
def calculate_z_numpy_gpu(q, maxiter, z):
"""Calculate z using numpy on the GPU"""
outputg = gpuarray.to_gpu(nm.resize(nm.array(0,), q.shape))
zg = gpuarray.to_gpu(z.astype(nm.complex64))
qg = gpuarray.to_gpu(q.astype(nm.complex64))
# 2.0 as an array
twosg = gpuarray.to_gpu(nm.array([2.0]*zg.size).astype(numpy.float32))
# 0+0j as an array
cmplx0sg = gpuarray.to_gpu(nm.array([0+0j]*zg.size).astype(nm.complex64))
# for abs_zg > twosg result
comparison_result = gpuarray.to_gpu(nm.array([False]*zg.size).astype(nm.bool))
# we'll add 1 to iterg after each iteration
iterg = gpuarray.to_gpu(nm.array([0]*zg.size).astype(nm.int32))
for iter in range(maxiter):
zg = zg*zg + qg
# abs returns a complex (rather than a float) from the complex
# input where the real component is the absolute value (which
# looks like a bug) so I take the .real after abs()
abs_zg = abs(zg).real
comparison_result = abs_zg > twosg
qg = gpuarray.if_positive(comparison_result, cmplx0sg, qg)
zg = gpuarray.if_positive(comparison_result, cmplx0sg, zg)
outputg = gpuarray.if_positive(comparison_result, iterg, outputg)
iterg = iterg + 1
output = outputg.get()
return output
def calculate_z_asnumpy_gpu(q, maxiter, z):
"""Calculate z using numpy on the GPU"""
# convert complex128s (2*float64) to complex64 (2*float32) so they run
# on older CUDA cards like the one in my MacBook. To use float64 doubles
# just edit these two lines
complex_type = np.complex64 # or nm.complex128 on newer CUDA devices
float_type = np.float32 # or nm.float64 on newer CUDA devices
# create an output array on the gpu of int32 as one long vector
outputg = gpuarray.to_gpu(np.resize(np.array(0, ), q.shape))
# resize our z and g as necessary to longer or shorter float types
z = z.astype(complex_type)
q = q.astype(complex_type)
# create zg and qg on the gpu
zg = gpuarray.to_gpu(z)
qg = gpuarray.to_gpu(q)
# create 2.0 as an array
twosg = gpuarray.to_gpu(np.array([2.0] * zg.size).astype(float_type))
# create 0+0j as an array
cmplx0sg = gpuarray.to_gpu(
np.array([0 + 0j] * zg.size).astype(complex_type))
# create a bool array to hold the (for abs_zg > twosg) result later
comparison_result = gpuarray.to_gpu(
np.array([False] * zg.size).astype(np.bool))
# we'll add 1 to iterg after each iteration, create an array to hold the iteration count
iterg = gpuarray.to_gpu(np.array([0] * zg.size).astype(np.int32))
for iter in range(maxiter):
# multiply z on the gpu by itself, add q (on the gpu)
zg = zg * zg + qg
# abs returns a complex (rather than a float) from the complex
# input where the real component is the absolute value (which
# looks like a bug) so I take the .real after abs()
# the above bug relates to pyCUDA from mid2010, it might be fixed now...
abs_zg = abs(zg).real
# figure out if zg is > 2
comparison_result = abs_zg > twosg
# based on the result either take 0+0j for qg and zg or leave unchanged
qg = gpuarray.if_positive(comparison_result, cmplx0sg, qg)
zg = gpuarray.if_positive(comparison_result, cmplx0sg, zg)
# if the comparison is true then update the iterations count to outputg
# which we'll extract later
outputg = gpuarray.if_positive(comparison_result, iterg, outputg)
# increment the iteration counter
iterg = iterg + 1
# extract the result from the gpu back to the cpu
output = outputg.get()
return output
def sqrt_normalize_gpu(img):
global posr, negr, posa, nega, stream
rgb = gpuarray.to_gpu(img[:, :, :3].copy())
a = gpuarray.to_gpu(img[:, :, 3].copy())
if not posr:
posr = gpuarray.zeros_like(rgb) + 1
negr = gpuarray.zeros_like(rgb) - 1
posa = gpuarray.zeros_like(a) + 1
nega = gpuarray.zeros_like(a) - 1
rgb = cumath.sqrt(abs(rgb), stream=stream) * gpuarray.if_positive(
rgb, posr, negr, stream=stream)
a = cumath.sqrt(abs(a), stream=stream) * gpuarray.if_positive(
a, posa, nega, stream=stream)
return normalize_gpu(rgb, a)
def linear_corr_cuda(self, image, l):
N = image.shape[0]
nd = self.pairwise_difference(image, N)
C = (1 - (nd / l))
zeros = misc.zeros(C.shape, C.dtype)
C = gpuarray.if_positive(C, C, zeros)
return C.copy()
def step_1(matrix_color, matrix_suma):
#La función gpuarray.if_positive evalua cada posición de la matriz
#Y de acuerdo a su valor realiza la primer operación o la segunda constatando una sentencia If Else
matrix_1 = gpuarray.if_positive(matrix_suma,
(3 * matrix_color) / matrix_suma,
matrix_suma)
return matrix_1
def calculate_z_asnumpy_gpu(q, maxiter, z):
"""Calculate z using numpy on the GPU"""
# convert complex128s (2*float64) to complex64 (2*float32) so they run
# on older CUDA cards like the one in my MacBook. To use float64 doubles
# just edit these two lines
complex_type = np.complex64 # or nm.complex128 on newer CUDA devices
float_type = np.float32 # or nm.float64 on newer CUDA devices
# create an output array on the gpu of int32 as one long vector
outputg = gpuarray.to_gpu(np.resize(np.array(0), q.shape))
# resize our z and g as necessary to longer or shorter float types
z = z.astype(complex_type)
q = q.astype(complex_type)
# create zg and qg on the gpu
zg = gpuarray.to_gpu(z)
qg = gpuarray.to_gpu(q)
# create 2.0 as an array
twosg = gpuarray.to_gpu(np.array([2.0] * zg.size).astype(float_type))
# create 0+0j as an array
cmplx0sg = gpuarray.to_gpu(np.array([0 + 0j] * zg.size).astype(complex_type))
# create a bool array to hold the (for abs_zg > twosg) result later
comparison_result = gpuarray.to_gpu(np.array([False] * zg.size).astype(np.bool))
# we'll add 1 to iterg after each iteration, create an array to hold the iteration count
iterg = gpuarray.to_gpu(np.array([0] * zg.size).astype(np.int32))
for iter in range(maxiter):
# multiply z on the gpu by itself, add q (on the gpu)
zg = zg * zg + qg
# abs returns a complex (rather than a float) from the complex
# input where the real component is the absolute value (which
# looks like a bug) so I take the .real after abs()
# the above bug relates to pyCUDA from mid2010, it might be fixed now...
abs_zg = abs(zg).real
# figure out if zg is > 2
comparison_result = abs_zg > twosg
# based on the result either take 0+0j for qg and zg or leave unchanged
qg = gpuarray.if_positive(comparison_result, cmplx0sg, qg)
zg = gpuarray.if_positive(comparison_result, cmplx0sg, zg)
# if the comparison is true then update the iterations count to outputg
# which we'll extract later
outputg = gpuarray.if_positive(comparison_result, iterg, outputg)
# increment the iteration counter
iterg = iterg + 1
# extract the result from the gpu back to the cpu
output = outputg.get()
return output
def make_sample_data(set_: int):
np.random.seed(set_ * 4347)
if set_ == 1: # Uniform distribution
data = np.random.uniform(0, 1, size=(samples, num_features))
if set_ == 2: # 3 Gaussian distribution
data = multi_gauss_clusters(n_clusters=3)
if set_ == 3: # 10 Gaussian distribution
data = multi_gauss_clusters(n_clusters=10)
df = pd.DataFrame()
np.random.shuffle(data)
df['vec'] = data.tolist()
# find nearest neighbours
from sklearn.neighbors import NearestNeighbors
nbrs = NearestNeighbors(n_neighbors=51,
algorithm='ball_tree',
leaf_size=30).fit(data)
_, nbrs_indices = nbrs.kneighbors(data)
for n_nbr in range(10, 51, 5):
df[f"known_neighbours_{n_nbr}"] = [
x[1:(n_nbr + 1)] for x in nbrs_indices
]
# hash using random hyperplane LSH
import pycuda.gpuarray as gpuarray
import skcuda.linalg as linalg
import pycuda.autoinit
linalg.init()
os.environ['CUDA_HOME'] = "/opt/cuda/"
vec_np = np.array(df['vec'].values.tolist(), dtype=np.float32)
LSH = LSHBias(feature_dim=num_features, bits=LSH_NUM_BITS)
W = np.array(LSH.W, dtype=np.float32)
b_gpu = gpuarray.to_gpu(W)
ones = np.ones(shape=(vec_np.shape[0], 1), dtype=np.float32)
X = np.concatenate((vec_np, ones), axis=1)
# do the matrix multiplication
a_gpu = gpuarray.to_gpu(X)
mul = linalg.mdot(a_gpu, b_gpu)
# get binary: 1 if value >= 0, else 0
res = gpuarray.if_positive(
mul >= gpuarray.zeros(mul.shape, dtype=np.float32),
then_=gpuarray.ones_like(mul),
else_=gpuarray.zeros_like(mul))
res = np.array(res.get(), dtype=np.uint32)
# convert grouped bits to integers
res = np_array_binary_to_grouped_integers(res)
df[f"hash_{LSH_NUM_BITS}_bits"] = [x for x in res]
df.to_parquet(f"{config.CUDA_neighbour_search_df_dir}df-{set_}.parquet",
index=False)
print("created test-data")
def wsparsify(w_gpu, percentage): """ Keeps only as many entries nonzero as specified by percentage. """ w = w_gpu.get() vals = sort(w)[::-1] idx = floor(prod(w.shape()) * percentage/100) zw_gpu = cua.zeros_like(w_gpu) # gpu array filled with zeros tw_gpu = cua.empty_like(w_gpu) # gpu array containing threshold tw_gpu.fill(vals[idx]) w_gpu = cua.if_positive(w_gpu > tw_gpu, w_gpu, zw_gpu) del zw_gpu del tw_gpu return w_gpu
def wsparsify(w_gpu, percentage): """ Keeps only as many entries nonzero as specified by percentage. """ w = w_gpu.get() vals = sort(w)[::-1] idx = floor(prod(w.shape()) * percentage/100) zw_gpu = cua.zeros_like(w_gpu) # gpu array filled with zeros tw_gpu = cua.empty_like(w_gpu) # gpu array containing threshold tw_gpu.fill(vals[idx]) w_gpu = cua.if_positive(w_gpu > tw_gpu, w_gpu, zw_gpu) del zw_gpu del tw_gpu return w_gpu
def run_gpu(self, Niters):
"""
Run G-S on GPU. The result is overwritten on the attribute "self.wdata"
containing a pycuda array.
"""
Nz, Ny, Nx = self.shape
# Allocate output data
wdata = gpuarray.empty((Ny, Nx), np.complex64)
sim = gpuarray.empty((Nz, Ny, Nx), np.complex64)
Isim = gpuarray.empty((Nz, Ny, Nx), np.complex64)
for io in trange(Niters):
# Propagate the initial wave to simulate defocused waves
# Psi(x,y,z) = convolve[Psi(x,y,0), CTF(x,y,z)]
cu_fft.fft(self.wdata, wdata, self.pft2dcc)
for kk in range(Nz):
sim[kk, :, :] = self.ctfd[kk, :, :] * wdata
cu_fft.ifft(sim, sim, self.pft3dcc, True)
if hasattr(self, 'Esdata'):
# Use the intensities, Isim = |Psi|**2
# Convolve with spatial-coherence envelope
# Isim = convolve[Isim, Es]
Isim = sim * sim.conj()
cu_fft.fft(Isim, Isim, self.pft3dcc)
cu_fft.ifft(Isim * self.Esdata, Isim, self.pft3dcc, True)
# Combine experimental and simulated amplitudes with simulated phase
# Psi' = [abs(Psi)+sqrt(Iexp)-sqrt(Isim)]*exp[i*arg(Psi)]
self.cuwave(self.Iexp, sim, Isim.real, Isim)
else:
# Combine experimental amplitudes with simulated phase
# Psi' = [sqrt(Iexp)]*exp[i*arg(Psi)]
self.cuwave(self.Iexp, sim, Isim)
sim = gpuarray.if_positive(self.mask, Isim, sim)
# then back-propagate to the exit plane and take average
# Psi(x,y,0) = < convolve[Psi, CTF*] >_z
cu_fft.fft(sim, sim, self.pft3dcc)
sim = sim * self.ctfd.conj()
cu_fft.ifft(sim, sim, self.pft3dcc, True)
wdata = misc.mean(sim.reshape(Nz, Nx * Ny), 0).reshape(Ny, Nx)
# update phase and wave
self.cuphase(wdata, self.wdata, self.phase_data)
self.wdata = wdata.copy()
def isgreater_gpu(x_gpu, y_gpu):
"""
Computes if x_gpu > y_gpu and gives back a mask with 0s and 1s.
Note, that y_gpu can be a scalar value as well.
"""
if ((y_gpu.__class__ == np.float) or
(y_gpu.__class__ == np.float32) or
(y_gpu.__class__ == np.float64)):
val = np.float32(y_gpu)
y_gpu = cua.empty_like(x_gpu) # gpu array containing threshold
y_gpu.fill(val)
zeros_gpu = cua.zeros_like(x_gpu) # gpu array filled with zeros
ones_gpu = cua.empty_like(x_gpu) # gpu array containing threshold
ones_gpu.fill(np.float32(1.))
mask_gpu = cua.if_positive(x_gpu > y_gpu, ones_gpu, zeros_gpu)
del zeros_gpu
del ones_gpu
return mask_gpu
def isgreater_gpu(x_gpu, y_gpu):
"""
Computes if x_gpu > y_gpu and gives back a mask with 0s and 1s.
Note, that y_gpu can be a scalar value as well.
"""
if ((y_gpu.__class__ == np.float) or
(y_gpu.__class__ == np.float32) or
(y_gpu.__class__ == np.float64)):
val = np.float32(y_gpu)
y_gpu = cua.empty_like(x_gpu) # gpu array containing threshold
y_gpu.fill(val)
zeros_gpu = cua.zeros_like(x_gpu) # gpu array filled with zeros
ones_gpu = cua.empty_like(x_gpu) # gpu array containing threshold
ones_gpu.fill(np.float32(1.))
mask_gpu = cua.if_positive(x_gpu > y_gpu, ones_gpu, zeros_gpu)
del zeros_gpu
del ones_gpu
return mask_gpu
# Read image. BW images have R=G=B so extract the R-value
image = img.imread(in_file_name)[:,:,0]
height, width = np.int32(image.shape)
area = height*width
print "Processing %d x %d image" % (width, height)
blocksize = (32,32,1)
gridsize = (int(width/32),int(height/32))
ones = np.empty([height,width])
ones[:,:] = 1
im_d = gpu.to_gpu(np.float32(np.array(image)))
ones_d = gpu.to_gpu(np.int32(np.array(ones)))
zero_d = gpu.to_gpu(np.int32(np.zeros([height,width])))
threshold_d = gpu.if_positive(im_d - seed_threshold, zero_d, ones_d)
new_d = gpu.to_gpu(np.int32(np.zeros([height,width])))
old_flags = 0
new_flags = gpu.sum(threshold_d).get()
while (new_flags - old_flags) != 0:
old_flags = new_flags
# run filter kernel
filter_kernel(threshold_d, im_d, new_d, width, height, threshold, block=blocksize, grid=gridsize)
new_flags = gpu.sum(new_d).get()
# transfer output to input
threshold_d = new_d
# hashing different .orc DataFrames
for filename in tqdm(glob(basepath + "part-*.orc")):
df = pd.read_orc(filename)
df = df.rename(columns={"FeatureVector_all_features": "vec"})
count += 1
vec_np = np.array(df['vec'].values.tolist(), dtype=np.float32)
# add bias term
ones = np.ones(shape=(vec_np.shape[0], 1), dtype=np.float32)
X = np.concatenate((vec_np, ones), axis=1)
# do the matrix multiplication
a_gpu = gpuarray.to_gpu(X)
mul = linalg.mdot(a_gpu, b_gpu)
# get binary: 1 if value >= 0, else 0
res = gpuarray.if_positive(
mul >= gpuarray.zeros(mul.shape, dtype=np.float32),
then_=gpuarray.ones_like(mul),
else_=gpuarray.zeros_like(mul))
res = np.array(res.get(), dtype=np.uint32)
# convert grouped bits to integers
res = np_array_binary_to_grouped_integers(res)
df[f"hash_{LSH_NUM_BITS}_bits"] = [x for x in res]
df = df[["rec_MBID", f"hash_{LSH_NUM_BITS}_bits"]]
df.to_parquet(f"{config.ABz_GPU_hashed_output_dir}{count}.parquet",
index=False)
# save as a single parquet file
spark = SparkSession \
.builder \
.appName("hashed file coalesce") \
H = 240
cap.set(cv.CAP_PROP_FRAME_HEIGHT, H)
cap.set(cv.CAP_PROP_FRAME_WIDTH, W)
ret, frame = cap.read()
gray_a = cv.cvtColor(frame, cv.COLOR_RGB2GRAY)
img_ori_gpu = gpuarray.to_gpu(gray_a.astype(np.float32))
img_buf_gpu = gpuarray.empty_like(img_ori_gpu)
img_sub = gpuarray.ones_like(img_ori_gpu)
img_sub = 25 * img_sub
img_bgm = gpuarray.zeros_like(img_sub)
while True:
ret, frame = cap.read()
gray_buff = cv.cvtColor(frame, cv.COLOR_BGR2GRAY)
img_res_gpu = gpuarray.to_gpu(gray_buff.astype(np.float32))
img_buf_gpu = cmath.fabs(img_ori_gpu - img_res_gpu)
img_buf_gpu = img_buf_gpu - img_sub
img_ori_gpu = img_res_gpu.copy()
img_res_gpu = gpuarray.if_positive(img_buf_gpu, img_bgm, img_res_gpu)
gray_buff = img_res_gpu.get()
gray_buff = gray_buff.astype(np.uint8)
frame = cv.cvtColor(gray_buff, cv.COLOR_GRAY2RGB)
cv.imshow("Moving Detecting!", frame)
if cv.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
cv.destroyAllWindows()
def map_if_positive(self, expr):
crit = self.rec(expr.criterion)
then = self.rec(expr.then)
else_ = self.rec(expr.else_)
return gpuarray.if_positive(crit, then, else_)
# `python Homework5_script.py [number of iteration for MCMC]`
###############
import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule
import numpy as np
import pycuda.gpuarray as gpuarray
from pycuda import curandom
import sys
N = int(sys.argv[1]) #int(1e3)
print("Number of iterations: " + str(N))
np.random.seed(123)
draws = np.random.uniform(-1, 1, N)
index = np.linspace(-1, 1, N)
a_gpu = cuda.mem_alloc(draws.nbytes)
cuda.memcpy_htod(a_gpu, draws)
gen = pycuda.curandom.XORWOWRandomNumberGenerator()
xy = gen.gen_uniform((2, N), np.float32)
xy = sum(xy**2)**0.5
M = gpuarray.sum(gpuarray.if_positive(xy - 1, xy * 0, xy * 0 + 1))
pi = 4 * M / N
print("The estimated value of pi is: " + str(pi))
def map_if_positive(self, expr):
crit = self.rec(expr.criterion)
then = self.rec(expr.then)
else_ = self.rec(expr.else_)
return gpuarray.if_positive(crit, then, else_)
print('a:\n{0}\nshape={1}\n'.format(a.get(), a.shape))
stream = drv.Stream()
b = gpuarray.to_gpu_async(h_array, stream=stream)
print('b:\n{0}\nshape={1}\n'.format(b.get(), b.shape))
c = gpuarray.empty((100, 100), dtype=dtype)
print('c:\n{0}\nshape={1}\n'.format(c, c.shape))
d = gpuarray.zeros((100, 100), dtype=dtype)
print('d:\n{0}\nshape={1}\n'.format(d, d.shape))
e = gpuarray.arange(0.0, 100.0, 1.0, dtype=dtype)
print('e:\n{0}\nshape={1}\n'.format(e, e.shape))
f = gpuarray.if_positive(e < 50, e - 100, e + 100)
print('f:\n{0}\nshape={1}\n'.format(f, f.shape))
g = gpuarray.if_positive(e < 50, gpuarray.ones_like(e), gpuarray.zeros_like(e))
print('g:\n{0}\nshape={1}\n'.format(g, g.shape))
h = gpuarray.maximum(e, f)
print('h:\n{0}\nshape={1}\n'.format(h, h.shape))
i = gpuarray.minimum(e, f)
print('i:\n{0}\nshape={1}\n'.format(i, i.shape))
g = gpuarray.sum(a)
print(g, type(g))
k = gpuarray.max(a)
