[Bug] Unroll loop runs slowly

[KuiliangL]

Hello, These days I constructed some PrimFunc files and I found that the unroll loop ran more slower than the other forkind loop in some situations. I tested the For statement with different forkind and a big length of the loop. I’m curious about the reason why the unroll loop runs slower with the big extent.

Expected behavior

The different loop forkinds have the same speeds.

Actual behavior

The unroll loop ran slower than the other rorkinds loop.

Environment

TVM 0.10dev0 ,git commit ee319d9d23c80091da9c4fb764b1e6d49d462714

Steps to reproduce

import tvm
from tvm import tir
import time

c1=tir.const(32450,'uint32')
c2=tir.const(15000000,'uint32')
v1=tir.Var('v1','uint32')
for1=tir.For(v1,c1,c2,3,tir.Evaluate(1))
f1=tir.PrimFunc([],for1)
for2=tir.For(v1,c1,c2,1,tir.Evaluate(1))
f2=tir.PrimFunc([],for2)
for3=tir.For(v1,c1,c2,0,tir.Evaluate(1))
f3=tir.PrimFunc([],for3)

t0=time.time()
tvm.build(f2)
print(time.time()-t0)
t1=time.time()
tvm.build(f3)
print(time.time()-t1)
t2=time.time()
tvm.build(f1)
print(time.time()-t2)

Triage

• tir.op

• needs-triage

[KuiliangL]

If we use the UnrollLoop pass optimization, it also runs slowly.

import tvm
from tvm import tir
v1=tir.Var('v1','uint32')
c1=tir.const(905892654,'uint32')
c2=tir.const(174155511,'uint32')
f1=tir.Cast('float32',c2)
ceil1=tir.ceil(f1)
c3=tir.Cast('uint32',ceil1)
c4=tir.const(35,'uint32')

v2=tir.Var('v2','float32')
c5=tir.const(0.187,'float32')
let1=tir.LetStmt(v2,c5,tir.Evaluate(0))
for1=tir.For(v1,c1,c3,3,let1)
prim=tir.PrimFunc({},for1)
b1=tir.IntImm('bool',4582)
if1=tir.IfThenElse(tir.EQ(tir.And(b1,b1),tir.Sub(b1,b1)),for1,tir.Evaluate(tir.ret(c2)))
p2=tir.PrimFunc({},if1)
mod = tvm.IRModule({"main": p2})

mod=tir.transform.UnrollLoop()(mod)