def test_against_fft_2d_mgpu(self):
from pyculib.fft.binding import Plan, CUFFT_R2C
rank = 64
rowsize = 64
N = rank * rowsize
x = np.arange(N, dtype=np.float32)
halfZ = rowsize // 2 + 1
xh = np.arange(rank * halfZ, dtype=np.complex64)
for j in range(rank):
for i in range(halfZ - 1):
ii = j * rowsize + 2 * i
r = x[ii]
if ii + 1 < N:
imag = x[ii + 1]
else:
imag = 0
xh[j * halfZ + i] = np.complex(r, imag)
xh[j * halfZ + halfZ - 1] = 0
x = x.reshape(rank, rowsize)
xh = xh.reshape(rank, halfZ)
xf = np.fft.fft2(x)
plan = Plan.many([rank, rowsize], CUFFT_R2C, 1, 2)
d_x_gpu = plan.to_device(xh)
xf_gpu = np.zeros(shape=(rank, halfZ), dtype=np.complex64)
#d_xf_gpu = plan.to_device(xf_gpu)
plan.forward(d_x_gpu, d_x_gpu) #Inplace
d_x_gpu.copy_to_host(xf_gpu)
self.assertTrue(np.allclose(xf[:, 0:halfZ], xf_gpu, atol=1e-6))
def test_against_fft_1d_mgpu(self):
return True
from pyculib.fft.binding import Plan, CUFFT_R2C, CUFFT_C2R
N = 32
x = np.arange(N, dtype=np.float32)
halfZ = N // 2 + 1
xh = np.arange(halfZ, dtype=np.complex64)
for i in range(halfZ - 1):
r = x[2 * i]
if 2 * i + 1 < N:
imag = x[2 * i + 1]
else:
imag = 0
xh[i] = np.complex(r, imag)
xh[halfZ - 1] = 0
print(x)
print(xh)
xf = np.fft.fft(x)
plan = Plan.many([N], CUFFT_R2C, 1, 2)
d_x_gpu = plan.to_device(xh)
xf_gpu = np.zeros(halfZ, dtype=np.complex64)
#d_xf_gpu = plan.to_device(xf_gpu)
plan.forward(d_x_gpu, d_x_gpu)
d_x_gpu.copy_to_host(xf_gpu)
self.assertTrue(np.allclose(xf[0:halfZ], xf_gpu, atol=1e-6))
def test_against_fft_1d(self):
from pyculib.fft.binding import Plan, CUFFT_R2C
N = 128
x = np.asarray(np.arange(N), dtype=np.float32)
xf = np.fft.fft(x)
d_x_gpu = cuda.to_device(x)
xf_gpu = np.zeros(N//2+1, np.complex64)
d_xf_gpu = cuda.to_device(xf_gpu)
plan = Plan.many(x.shape, CUFFT_R2C)
plan.forward(d_x_gpu, d_xf_gpu)
d_xf_gpu.copy_to_host(xf_gpu)
self.assertTrue( np.allclose(xf[0:N//2+1], xf_gpu,
atol=1e-6) )
def test_against_fft_2d(self):
from pyculib.fft.binding import Plan, CUFFT_R2C
rank = 2
rowsize = 128
N = rowsize * rank
x = np.arange(N, dtype=np.float32).reshape(rank, rowsize)
xf = np.fft.fft2(x)
d_x_gpu = cuda.to_device(x)
xf_gpu = np.zeros(shape=(rank, rowsize//2 + 1), dtype=np.complex64)
d_xf_gpu = cuda.to_device(xf_gpu)
plan = Plan.many(x.shape, CUFFT_R2C)
plan.forward(d_x_gpu, d_xf_gpu)
d_xf_gpu.copy_to_host(xf_gpu)
self.assertTrue(np.allclose(xf[:, 0:rowsize//2+1], xf_gpu, atol=1e-6))
def test_against_fft_3d(self):
from pyculib.fft.binding import Plan, CUFFT_R2C
depth = 2
colsize = 2
rowsize = 64
N = depth * colsize * rowsize
x = np.arange(N, dtype=np.float32).reshape(depth, colsize, rowsize)
xf = np.fft.fftn(x)
halfZ = rowsize // 2 + 1
plan = Plan.many(x.shape, CUFFT_R2C)
d_x_gpu = plan.to_device(x)
xf_gpu = np.zeros(shape=(depth, colsize, halfZ), dtype=np.complex64)
d_xf_gpu = plan.to_device(xf_gpu)
plan.forward(d_x_gpu, d_xf_gpu)
d_xf_gpu.copy_to_host(xf_gpu)
self.assertTrue(np.allclose(xf[:, :, 0:halfZ], xf_gpu, atol=1e-6))
def test_against_fft_3d_mgpu(self):
from pyculib.fft.binding import Plan, CUFFT_R2C
depth = 32
colsize = 32
rowsize = 32
N = depth * colsize * rowsize
x = np.arange(N, dtype=np.float32)
halfZ = rowsize // 2 + 1
xh = np.arange(depth * colsize * halfZ, dtype=np.complex64)
for k in range(depth):
for j in range(colsize):
for i in range(halfZ - 1):
ii = k * colsize * rowsize + j * rowsize + 2 * i
r = x[ii]
if ii + 1 < N:
imag = x[ii + 1]
else:
imag = 0
xh[k * colsize * halfZ + j * halfZ + i] = np.complex(
r, imag)
xh[k * colsize * halfZ + j * halfZ + halfZ - 1] = 0
x = x.reshape(depth, colsize, rowsize)
xh = xh.reshape(depth, colsize, halfZ)
xf = np.fft.fftn(x)
plan = Plan.many([depth, colsize, rowsize], CUFFT_R2C, 1, 2)
d_x_gpu = plan.to_device(xh)
xf_gpu = np.zeros(shape=(depth, colsize, halfZ), dtype=np.complex64)
#d_xf_gpu = plan.to_device(xf_gpu)
plan.forward(d_x_gpu, d_x_gpu)
d_x_gpu.copy_to_host(xf_gpu)
self.assertTrue(np.allclose(xf[:, :, 0:halfZ], xf_gpu, atol=1e-6))
# f = fft.FFTPlan(img_shape, np.complex64, np.complex64, 1, 0, fft.FFTPlan.MODE_FFTW_PADDING)
from pyculib.fft.binding import Plan, CUFFT_C2C
from pyculib import blas as cublas
n = (128 * 10)**2
data1 = np.arange(n, dtype=np.complex64).reshape(2, n // 2)
data = np.arange(n, dtype=np.complex64)
orig = data.copy()
d_data = cuda.to_device(data)
#s0 = cuda.stream()
# cuda.select_device(1)
# d_data1 = cuda.to_device(data)
#s1 = cuda.stream()
# fftplan = Plan.one(CUFFT_C2C, *data.shape)
# Plan.many()
fftplan1 = Plan.many(data.shape, CUFFT_C2C, 1500)
b = cublas.Blas()
rounds = 10000
start = time.clock()
for x in range(rounds):
# fft.fft_inplace(img)
# cuda.select_device(0)
# fftplan1.forward(d_data, d_data)
# fftplan1.inverse(d_data, d_data)
# cuda.select_device(1)
# fftplan1.forward(d_data1, d_data1)
#fftplan1.forward(d_data1, d_data1)
# fftplan.inverse(d_data, d_data)
# d_data = cuda.to_device(data)
# cublas.dot(d_data, d_data)