def kij_lu_decomposer_opt_gpu(M):
m = M.shape[0]
n = M.shape[1]
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
from skcuda.cublas import cublasCreate, cublasDaxpy, cublasDestroy
import skcuda.misc as misc
N_gpu = gpuarray.to_gpu(M)
h = cublasCreate()
for k in range(0,n):
for i in range(k+1,n):
N_gpu[i,k] = N_gpu[i,k] / N_gpu[k,k]
#N[i,k+1:] -= N[i,k] * N[k,k+1:]
cublasDaxpy(h, N_gpu[k,k+1:].size, -np.float64(N_gpu[i,k].get()), N_gpu[k,k+1:].gpudata, 1, N_gpu[i,k+1:].gpudata, 1)
#Move from GPU to CPU
N = N_gpu.get()
cublasDestroy(h)
return N
def jki_lu_decomposer_opt_gpu(M):
m = M.shape[0]
n = M.shape[1]
import pycuda.autoinit
import pycuda.gpuarray as gpuarray
from skcuda.cublas import cublasCreate, cublasDaxpy, cublasDscal, cublasDestroy
import skcuda.misc as misc
N_gpu = gpuarray.to_gpu(M)
h = cublasCreate()
for j in range(0,n):
for k in range(0,j):
#N[k+1:,j] = N[k+1:,j] - N[k+1:,k] * N[k,j]
cublasDaxpy(h, N_gpu[k+1:,k].size, -np.float64(N_gpu[k,j].get()), N_gpu[k+1:,k].gpudata, n, N_gpu[k+1:,j].gpudata, n)
#N[j+1:,j] /= N[j,j]
cublasDscal(h, N_gpu[j+1:,j].size, 1.0/np.float64(N_gpu[j,j].get()), N_gpu[j+1:,j].gpudata, n)
#Move from GPU to CPU
N = N_gpu.get()
cublasDestroy(h)
return N
def fun_b_k(self, b_k_gpu, k):
cublas.cublasDcopy(self.h, self.b_gpu.size, self.b_gpu.gpudata, 1,
b_k_gpu.gpudata, 1)
for i in range(len(self.Ax_gpu)):
if i != k:
cublas.cublasDaxpy(self.h, b_k_gpu.size, np.float64(-1),
self.Ax_gpu[i].gpudata, 1, b_k_gpu.gpudata,
1)
def test_cublasDaxpy(self):
alpha = np.float64(np.random.rand())
x = np.random.rand(5).astype(np.float64)
x_gpu = gpuarray.to_gpu(x)
y = np.random.rand(5).astype(np.float64)
y_gpu = gpuarray.to_gpu(y)
cublas.cublasDaxpy(self.cublas_handle, x_gpu.size, alpha, x_gpu.gpudata, 1,
y_gpu.gpudata, 1)
assert np.allclose(y_gpu.get(), alpha*x+y)
def _axpy(self, handle, alpha, x_gpu, y_gpu):
cublas.cublasDaxpy(handle, x_gpu.size, alpha, x_gpu.gpudata, 1,
y_gpu.gpudata, 1)
def _zaxpy(self, handle, z_gpu, alpha, x_gpu, y_gpu):
# copy y to z
cublas.cublasDcopy(handle, y_gpu.size, y_gpu.gpudata, 1, z_gpu.gpudata,
1)
cublas.cublasDaxpy(handle, x_gpu.size, alpha, x_gpu.gpudata, 1,
z_gpu.gpudata, 1)