def testElementwise(self):
t1 = ones((10000, 5000), chunk_size=500, gpu=True)
t2 = ones(5000, chunk_size=500, gpu=True)
t = (t1 - t2) / sqrt(t2 * (1 - t2) * len(t2))
g = t.build_graph(tiled=True)
RuntimeOptimizer(g, self.executor._engine).optimize([], False)
self.assertTrue(any(n.op.__class__.__name__ == 'TensorCpFuseChunk' for n in g))
c = next(n for n in g if n.op.__class__.__name__ == 'TensorCpFuseChunk')
self.assertGreater(len(_evaluate(c)), 1)
def testElementwise(self):
t1 = ones((10000, 5000), chunk_size=500, gpu=True)
t2 = ones(5000, chunk_size=500, gpu=True)
t = (t1 - t2) / sqrt(t2 * (1 - t2) * len(t2))
g = t.build_graph(tiled=True)
graph = self.executor._preprocess(g, [])
self.assertTrue(any(n.op.__class__.__name__ == 'TensorCpFuseChunk' for n in graph))
c = next(n for n in graph if n.op.__class__.__name__ == 'TensorCpFuseChunk')
print(_evaluate(c))