def sum(self, x, abs = False):
'''
sum(float* x, float* result, ulong x_size)
'''
if abs == False:
result = gpuarray.to_gpu(np.zeros(1).astype(np.float32))
self._sum_kernel(x.gpudata, result.gpudata, \
np.int64(x.size), \
block = (1024, 1, 1), \
grid = (int(x.size / 1024) + 1, 1) \
)
return result[0]
else:
return cublas.cublasSasum(self.handle, x.size, x.gpudata, 1)
def test_cublasSasum(self):
x = np.random.rand(5).astype(np.float32)
x_gpu = gpuarray.to_gpu(x)
result = cublas.cublasSasum(x_gpu.size, x_gpu.gpudata, 1)
assert np.allclose(result, np.sum(np.abs(x)))
def test_cublasSasum(self):
x = np.random.rand(5).astype(np.float32)
x_gpu = gpuarray.to_gpu(x)
result = cublas.cublasSasum(self.cublas_handle, x_gpu.size,
x_gpu.gpudata, 1)
assert np.allclose(result, np.sum(np.abs(x)))
