def computeAx(x_gpu, Xprime_gpu, X_gpu, XX_gpu, Yprime_gpu, Y_gpu, YY_gpu,
Zprime_gpu, zzero, freq, FREQ_gpu, c, Deltaxprime, Deltayprime,
Deltazprime, sizePartition, sizeOut):
numPartitions = np.int32(np.ceil(Xprime_gpu.size / sizePartition))
y_gpu = gpuarray.zeros((sizeOut, 1), dtype=np.complex64)
p1 = 0
p2 = sizePartition
for k in range(numPartitions):
currentA_gpu = computeA(Xprime_gpu[p1:p2], X_gpu, XX_gpu,
Yprime_gpu[p1:p2], Y_gpu, YY_gpu,
Zprime_gpu[p1:p2], zzero, freq, FREQ_gpu, c,
Deltaxprime, Deltayprime, Deltazprime)
#y_gpu = y_gpu + culinalg.dot(currentA_gpu, x_gpu[p1 : p2], 'N', 'N')
m, n = currentA_gpu.shape
cublas.cublasCgemv(cublasHandle, 't', n, m, np.complex64(1),
currentA_gpu.gpudata, n, x_gpu[p1:p2].gpudata, 1,
np.complex64(1), y_gpu.gpudata, 1)
p1 = p2
if (k == (numPartitions - 2)):
p2 = Xprime_gpu.size
else:
p2 = p2 + sizePartition
return y_gpu
def computeAdy(cublasHandle, y_gpu, Xprime_gpu, XX_gpu, Yprime_gpu, YY_gpu,
Zprime_gpu, zzero, FREQ_gpu, c, Deltaprime, sizePartition,
sizeOut):
numPartitions = np.int32(np.ceil(XX_gpu.size / sizePartition))
x_gpu = gpuarray.zeros((sizeOut, 1), dtype=np.complex64)
p1 = 0
p2 = sizePartition
for k in range(numPartitions):
currentAd_gpu = computeAd(Xprime_gpu, XX_gpu[p1:p2], Yprime_gpu,
YY_gpu[p1:p2], Zprime_gpu, zzero,
FREQ_gpu[p1:p2], c, Deltaprime,
sizePartition)
m, n = currentAd_gpu.shape
cublas.cublasCgemv(cublasHandle, 't', n, m, np.complex64(1),
currentAd_gpu.gpudata, n, y_gpu[p1:p2].gpudata, 1,
np.complex64(1), x_gpu.gpudata, 1)
#x_gpu = x_gpu + culinalg.dot(currentAd_gpu, y_gpu[p1 : p2], 'N', 'N', cublasHandle)
p1 = p2
if (k == (numPartitions - 2)):
p2 = XX_gpu.size
else:
p2 = p2 + sizePartition
return x_gpu
def test_cublasCgemv(self):
a = (np.random.rand(2, 3)+1j*np.random.rand(2, 3)).astype(np.complex64)
x = (np.random.rand(3, 1)+1j*np.random.rand(3, 1)).astype(np.complex64)
a_gpu = gpuarray.to_gpu(a.T.copy())
x_gpu = gpuarray.to_gpu(x)
y_gpu = gpuarray.empty((2, 1), np.complex64)
alpha = np.complex64(1.0)
beta = np.complex64(0.0)
cublas.cublasCgemv(self.cublas_handle, 'n', 2, 3, alpha,
a_gpu.gpudata, 2, x_gpu.gpudata,
1, beta, y_gpu.gpudata, 1)
assert np.allclose(y_gpu.get(), np.dot(a, x))