Hi @junrushao1994 ,
an over-simplified example from the industrial context is the following:
```python
...
B0 = tvm.compute((m,n), lambda i,j: A0[i,j] + 2*A1[i,j], name = "B0")
C0 = tvm.compute((m,n), lambda i,j: A0[i,j] + 2*A1[i,j], name = "C0")
D0 = tvm.compute((m,n), lambda i,j: B0[i,j] + 3*C0[i,j], name = "D0")
...
```
The customized TVM will schedule and use `compute_at` to the extreme, and
transform into something like
```cpp
...
for (i, 0, m) {
for (j, 0, n) {
B0[0] = (A0[((i*stride) + (j*stride))] + (2f*A1[((i*stride) +
(j*stride))]))
C0[0] = (A0[((i*stride) + (j*stride))] + (2f*A1[((i*stride) +
(j*stride))]))
D0[((i*stride) + (j*stride))] = (B0[0] + (2f*C0[0]))
}}
...
```
This gives our 'incomplete' CSE and Copy Propagation a chance to make the C0
assigned by B0 and replace C0’s appearance in D0 into B0 and make C0 dead (or
not? dependent on the future).
```cpp
...
for (i, 0, m) {
for (j, 0, n) {
B0[0] = (A0[((i*stride) + (j*stride))] + (2f*A1[((i*stride) +
(j*stride))]))
C0[0] = B0[0]
D0[((i*stride) + (j*stride))] = (B0[0] + (2f*B0[0]))
}}
...
```
The above ‘incomplete’ CSE and Copy Propagation pass can do things safely in a
straight-line code in a small range (without data-flow analysis), but the same
thing did not happen for dead code elimination – if we don’t know any live
information out of this for loop, we cannot just eliminate the assignment to
C0[0].
Generally speaking, dead code can arise after copy propagation and how dead
code arises in TVM is similar to how they arise in LLVM and traditional
compiler passes.
--
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/apache/incubator-tvm/issues/4468#issuecomment-562559678
