xmc/ReplaceQDQResizePass: replace 1x1 spatial XFEResize with broadcas…#804
Open
nvunnam57128 wants to merge 2 commits into
Open
Conversation
…ting Add
Mirrors the IPU-specific 1x1-input optimisation in xcompiler's
ReplaceQDQResizePass (src/pass/passes/ReplaceQDQResizePass.cpp lines
200-282). When a quantised XFEResize takes a tensor of shape
[N, 1, 1, C] and upsamples it to [N, H, W, C] (NHWC, with H>1 or W>1),
the Resize is functionally a broadcast: there is only one source
pixel per (N, C) so bilinear/nearest collapse to replication.
The rewrite drops the Resize and emits
%zero = onnx.Constant of shape [N, H, W, C] in the output quant type,
storage value = output_zero_point (decodes to 0.0)
%out = onnx.Add(%resize_input, %zero)
ONNX Add broadcasts [N, 1, 1, C] + [N, H, W, C] -> [N, H, W, C],
producing the same numerical result as the original Resize. The
synthetic zero Add is then collapsed by downstream eltwise / const-fold
passes (xcompiler's pipeline does the same: ReplaceQDQResizePass tags
the eltwise with original_resize_opt=true, and a later fusion absorbs
the zero into a downstream skip-connection Add).
Match conditions:
* single-use XFEResize
* rank-4 static input AND output (NHWC)
* input_shape[1] == 1 && input_shape[2] == 1
* output_shape[1] > 1 || output_shape[2] > 1
* batch and channel dims match across input/output
* input and output are uniform quant types with matching scale/zp
This avoids the backend qlinear_resize kernel for the corner-case
1x1->HxW shape that often fails or runs sub-optimally on IPU; observed
on scene_parser_512_256_v2_int8 (Resize_173_8), PSO3, PSA2, PSA3 and
mep_v2/K2.
Placement: after ConvertToChannelLast (creates XFEResize), after the
5D->4D and transpose-optimisation passes (stable rank-4 NHWC), and
right before ReplaceQuantizedTileToAddPass (its analogue for Tile) so
the emitted onnx.Add is immediately lowered by ReplaceQDQEltwisePass.
Co-authored-by: Cursor <cursoragent@cursor.com>
nvunnam57128
force-pushed
the
nvunnam/replace-qdq-resize-1x1-to-broadcasting-add
branch
from
0910ff9 to
215fc3a
Compare
Apply repo .clang-format (LLVM + AlwaysBreakTemplateDeclarations: Yes + AlignAfterOpenBracket: DontAlign) to ReplaceQDQResizePass.cpp. Whitespace-only fix for the 4 violations reported by clang-format 20.0.0git on the previous commit (struct header break, two notifyMatchFailure continuation breaks, and the rewriter.create<ONNXConstantOp> continuation). Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.