[MMA_FROM_SMEM_IT_RES] Change how residuals are handled #393
Conversation
facebook-github-bot
added
the
CLA Signed
Codecov ReportBase: 91.14% // Head: 91.23% // Increases project coverage by
Additional details and impacted files@@ Coverage Diff @@
## it1 #393 +/- ##
==========================================
+ Coverage 91.14% 91.23% +0.09%
==========================================
Files 75 75
Lines 4358 4346 -12
==========================================
- Hits 3972 3965 -7
+ Misses 386 381 -5
Flags with carried forward coverage won't be shown. Click here to find out more.
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
danthe3rd
force-pushed
the
it2
branch
2 times, most recently
from
ffb63a7
to
a5aeb22
Compare
| @@ -337,6 +339,11 @@ class PredicatedTileIterator< | |||
| address_iterator_.clear_mask(enable); | |||
| } | |||
|
|
|||
| CUTLASS_HOST_DEVICE | |||
| void set_residual_tile(bool enable) { | |||
| address_iterator_.set_residual_tile(enable); | |||
There was a problem hiding this comment.
It's a pity that address_iterator is a private member, otherwise we could have just inherited from the base class and added this method to all the classes in this file.
| params_.inc_advance_ * LongIndex(tile_offset.contiguous() - 1); | ||
| pointer_ += Shape::kStrided * tile_offset.strided(); | ||
| } | ||
| if (!Gather) { |
There was a problem hiding this comment.
Same comment here, this is the only thing that we changed (in addition to the set_residual_tile) but we need to copy the whole file. :-/
Changes how we deal with residuals for the right-hand operand of MM from shared memory. This is used on the fw and bw pass, and enables to support arbitrary sequence lengths.
P100/V100 bw (causal)
P100/V100 bw
A100 bw
P100/V100 fw (causal)
P100/V100 fw
A100 fw
Stacked PR Chain: MMA_FROM_SMEM_IT_RES