def ikj_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 i in range(0,n):
for k in range(0,i):
#N[i,k] = N[i,k] / N[k,k]
cublasDscal(h, N_gpu[i,k].size, 1.0/np.float64(N_gpu[k,k].get()), N_gpu[i,k].gpudata, 1)
#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 test_cublasDscal(self):
x = np.random.rand(5).astype(np.float64)
x_gpu = gpuarray.to_gpu(x)
alpha = np.float64(np.random.rand())
cublas.cublasDscal(self.cublas_handle, x_gpu.size, alpha,
x_gpu.gpudata, 1)
assert np.allclose(x_gpu.get(), alpha*x)
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