def perform(self, node, inputs, out_storage):
y, alpha, A, x, beta = inputs
inplace = self.inplace
if inplace and y.strides[0] < 0:
inplace = False
out_storage[0][0] = blas.gemv(alpha, A, x, beta, y,
overwrite_y=inplace)
def perform(self, node, inputs, out_storage):
y, alpha, A, x, beta = inputs
inplace = self.inplace
if inplace and y.strides[0] < 0:
inplace = False
out_storage[0][0] = blas.gemv(alpha, A, x, beta, y,
overwrite_y=inplace)
def gemv(shp, dtype, order, trans, offseted_i, sliced,
overwrite, init_y, alpha=1.0, beta=0.0):
cA, gA = gen_gpuarray(shp, dtype, order=order, offseted_inner=offseted_i,
sliced=sliced, ctx=context)
if trans:
shpX = (shp[0],)
shpY = (shp[1],)
else:
shpX = (shp[1],)
shpY = (shp[0],)
cX, gX = gen_gpuarray(shpX, dtype, offseted_inner=offseted_i,
sliced=sliced, ctx=context)
if init_y:
cY, gY = gen_gpuarray(shpY, dtype, ctx=context)
else:
cY, gY = None, None
if dtype == 'float32':
cr = fblas.sgemv(alpha, cA, cX, beta, cY, trans=trans,
overwrite_y=overwrite)
else:
cr = fblas.dgemv(alpha, cA, cX, beta, cY, trans=trans,
overwrite_y=overwrite)
gr = gblas.gemv(alpha, gA, gX, beta, gY, trans_a=trans,
overwrite_y=overwrite)
numpy.testing.assert_allclose(cr, numpy.asarray(gr), rtol=1e-6)
def gemv(shp, dtype, order, trans, offseted_i, sliced,
overwrite, init_y, alpha=1.0, beta=0.0):
cA, gA = gen_gpuarray(shp, dtype, order=order, offseted_inner=offseted_i,
sliced=sliced, ctx=context)
if trans:
shpX = (shp[0],)
shpY = (shp[1],)
else:
shpX = (shp[1],)
shpY = (shp[0],)
cX, gX = gen_gpuarray(shpX, dtype, offseted_inner=offseted_i,
sliced=sliced, ctx=context)
if init_y:
cY, gY = gen_gpuarray(shpY, dtype, ctx=context)
else:
cY, gY = None, None
if dtype == 'float32':
cr = fblas.sgemv(alpha, cA, cX, beta, cY, trans=trans,
overwrite_y=overwrite)
else:
cr = fblas.dgemv(alpha, cA, cX, beta, cY, trans=trans,
overwrite_y=overwrite)
gr = gblas.gemv(alpha, gA, gX, beta, gY, trans_a=trans,
overwrite_y=overwrite)
numpy.testing.assert_allclose(cr, numpy.asarray(gr), rtol=1e-6)
def perform(self, node, inputs, out_storage):
y, alpha, A, x, beta = inputs
out_storage[0][0] = blas.gemv(alpha,
A,
x,
beta,
y,
trans=False,
overwrite_y=self.inplace)
def perform(self, node, inputs, out_storage):
y, alpha, A, x, beta = inputs
out_storage[0][0] = blas.gemv(alpha, A, x, beta, y, trans=False,
overwrite_y=self.inplace)
