hzfan commented on pull request #6771:
URL: https://github.com/apache/incubator-tvm/pull/6771#issuecomment-717639785


   @tqchen  The ir (before narrowing) is like
   ```
   for (ax0.ax1.fused.ax2.outer.fused: int64, 0i64, 42i64) "parallel" {
     for (n.oc_chunk.fused.oh.outer.fused: int64, 0i64, 
a_very_long_upper_bound) {
       ...
       ...
       for (oh.inner: int32, 0, 2) {
         for (ow.inner: int32, 0, 14) {
           for (oc_block: int32, 0, 16) "vectorized" {
             ...floormod(n.oc_chunk.fused.oh.outer.fused, cast(int64, 
floordiv(((cast(int32, ((floormod(ax0.ax1.fused.ax2.outer.fused, 7i64)*2i64) + 
1i64)) + 1) - cast(int32, (floormod(ax0.ax1.fused.ax2.outer.fused, 
7i64)*2i64))), 2)))...
           }
         }
       }
     }
   }
   ```
   Let `a = n.oc_chunk.fused.oh.outer.fused` (the first operand of floormod), 
`b = cast(int64, floordiv(((cast(int32, 
((floormod(ax0.ax1.fused.ax2.outer.fused, 7i64)*2i64) + 1i64)) + 1) - 
cast(int32, (floormod(ax0.ax1.fused.ax2.outer.fused, 7i64)*2i64))), 2))` 
(second operand of floormod)
   `const_int_bound` checker gives `0 <= a <= 6` and `-5 <= b <= 7`.
   
   In a simplified form, 
   ```
   b = floordiv(c * 2 + 2 - c * 2, 2), where
   c = floormod(ax0.ax1.fused.ax2.outer.fused, 7)
   ```
   
   I guess `c * 2 + 2 - c * 2` is not simplified as `2`, instead, its bound is 
analyzed directly, so the min becomes negative due to the minus operation.


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to