[compile] Does tvm support DivToMul without indications

Development
1

Base on 除法转乘法 — TVM 开发指南, I get a test, which show the tvm can do DivToMul with relay.transform.DivToMul. But for a more common case, we may have a model which have many div operators, so it is not convenience to transform it one by one by adding relay.transform.DivToMul in source, and my question：Does tvm support DivToMul without indications，and transform all the div into mul for better performance ?

  • test:
import tvm
from tvm import relay
import numpy as np

for dtype, rtol in [("float16", 1e-3), ("float32", 1e-7), ("float64", 1e-12)]:
    x = relay.var("x", relay.TensorType((), dtype))
    y = relay.Constant(tvm.nd.array(np.array([1.5]).astype(dtype)))
    z = x / y
    mod = tvm.IRModule.from_expr(z)
    transformed = relay.transform.DivToMul()(mod)
    assert transformed["main"].body.op.name == "multiply"
    np.testing.assert_allclose(transformed["main"].body.args[1].data.numpy()[0], 1 / 1.5, rtol=rtol)
  • debug
(Pdb) n
> divtomul.py(11)<module>()
-> assert transformed["main"].body.op.name == "multiply"
(Pdb) p transformed
def @main(%x: float16 /* ty=float16 */) -> Tensor[(1), float16] {
  multiply(%x, meta[relay.Constant][0] /* ty=Tensor[(1), float16] */) /* ty=Tensor[(1), float16] */
}