Change the order to Relay IR when converting to string

Questions
#1

For graphs with multi branches, there might be more than one execution order. For example, the following DAG can be either executed in d -> e -> f -> out or e -> d -> f -> out. However, it seems in relay IR, we cannot determine which order to display when serializing.

a = relay.var("a", shape=[1, 10])
b = relay.var("b", shape=[1, 10])
c = relay.var("c", shape=[1, 10])
d = a + b
e = a * c
f = b / e
out = d - f
fn = relay.Function([a,b,c], out)
print(fn)
'''
fn (%a: Tensor[(1, 10), float32], %b: Tensor[(1, 10), float32], %c: Tensor[(1, 10), float32]) {
  %0 = add(%a, %b) /* string */;
  %1 = multiply(%a, %c) /* string */;
  %2 = divide(%b, %1) /* string */;
  subtract(%0, %2) /* string */
}
'''
#wanted version
'''
fn (%a: Tensor[(1, 10), float32], %b: Tensor[(1, 10), float32], %c: Tensor[(1, 10), float32]) {
  %1 = multiply(%a, %c) /* string */;
  %0 = add(%a, %b) /* string */;
  %2 = divide(%b, %1) /* string */;
  subtract(%0, %2) /* string */
}
'''

Even we manually switch the order, it will still display the original one

code = '''
fn (%a: Tensor[(1, 10), float32], %b: Tensor[(1, 10), float32], %c: Tensor[(1, 10), float32]) {
  %0 = add(%a, %b) /* string */;
  %1 = multiply(%a, %c) /* string */;
  %2 = divide(%b, %1) /* string */;
  subtract(%0, %2) /* string */
}
'''
fn = tvm.parser.parse_expr(code)
print(fn)
'''
fn (%a: Tensor[(1, 10), float32], %b: Tensor[(1, 10), float32], %c: Tensor[(1, 10), float32]) {
  %0 = multiply(%a, %c);
  %1 = add(%a, %b);
  %2 = divide(%b, %0);
  subtract(%1, %2)
}
'''
1 Like
#2

The order of multiple branches in the DAG is unspecified. You might want to use A normal form if you need strict ordering.

1 Like