[Bug] [opt] [doc] Fix OffsetAndExtractBitsStmt optimization and improve documentation on virtual/physical indices #1259
Conversation
taichi/ir/snode.cpp
Outdated
| int SNode::get_num_bits(int virtual_index) const { | ||
| int result = 0; | ||
| const SNode *snode = this; | ||
| auto physical_index = snode->physical_index_position[virtual_index]; | ||
| while (snode) { | ||
| result += extractors[physical_index].num_bits; | ||
| result += snode->extractors[physical_index].num_bits; | ||
| snode = snode->parent; | ||
| } |
There was a problem hiding this comment.
There are two issues here:
- We should use virtual indices as input and do a translation to physical indices within the function (see the doc updates)
snode->extractorsinstead ofextractors:-)
There was a problem hiding this comment.
For the second point, sorry for writing a bug here :)
For the first point, I think we're already using physical indices in LoopIndexStmts. See
taichi/taichi/transforms/lower_ast.cpp
Lines 286 to 287 in ca81fdf
There was a problem hiding this comment.
Ah you are right! I just now reverted that modification.
taichi/ir/snode.cpp
Outdated
| int SNode::get_num_bits(int virtual_index) const { | ||
| int result = 0; | ||
| const SNode *snode = this; | ||
| auto physical_index = snode->physical_index_position[virtual_index]; | ||
| while (snode) { | ||
| result += extractors[physical_index].num_bits; | ||
| result += snode->extractors[physical_index].num_bits; | ||
| snode = snode->parent; | ||
| } |
There was a problem hiding this comment.
For the second point, sorry for writing a bug here :)
For the first point, I think we're already using physical indices in LoopIndexStmts. See
taichi/taichi/transforms/lower_ast.cpp
Lines 286 to 287 in ca81fdf
Codecov Report
@@ Coverage Diff @@
## master #1259 +/- ##
==========================================
+ Coverage 84.78% 84.81% +0.02%
==========================================
Files 18 18
Lines 3267 3273 +6
Branches 613 615 +2
==========================================
+ Hits 2770 2776 +6
Misses 360 360
Partials 137 137
Continue to review full report at Codecov.
|
Related issue = #1178 #1229
[Click here for the format server]