| Issue |
243418
|
| Summary |
[MLIR][CSE] Repeated side-effect scans over overlapping intervals in SDXL CLIP
|
| Labels |
mlir
|
| Assignees |
|
| Reporter |
Michael-Chen-NJU
|
### Problem
While compiling a public [SDXL CLIP MLIR input](https://sharkpublic.blob.core.windows.net/sharkpublic/sai/sdxl-prompt-encoder/model.mlir) with IREE, CSE repeatedly scans overlapping intervals between equivalent reads.
In [`hasOtherSideEffectingOpInBetween`](https://github.com/llvm/llvm-project/blob/a7dcddcdd7fe0d6bed9f79e0090ac6438b05f794/mlir/lib/Transforms/Utils/CSE.cpp#L168), the existing cache reuses scan progress for the same `fromOp`, but does not share effect-query results across different starting operations. Queries with different starting operations can therefore repeatedly inspect overlapping intervals. One captured hotspot performs **2,590,482 intermediate-operation inspections for 695 queries**, including repeated recursive effect checks on `linalg.generic`. Across these queries, the read effects use `DefaultResource` and have no associated Value.
### Reproducer
[clip-cse-reproducer-minimal.zip](https://github.com/user-attachments/files/34965583/clip-cse-reproducer-minimal.zip)
The attachment contains the unchanged enclosing module captured immediately before a CSE invocation on `util.func @encode_prompts$async`. This intermediate IR requires IREE dialects:
```bash
"$IREE_BUILD/tools/iree-opt" clip-cse-fixed.mlir \
'--pass-pipeline=builtin.module(cse)' -o result.mlir
```
IREE revision: `5332c68297a68fafd596277d7686f91009ebf1cb`
Embedded LLVM: `a7dcddcdd7fe0d6bed9f79e0090ac6438b05f794`
Replay reproduces the hotspot's query sequence and inspection counts, although its module scope differs from the original function-pass scope. Input hashes and the full compilation command are in the attachment.
### Measurements
A block-local interval-summary prototype avoids rescanning intervals proven free of conflicting writes or unknown effects, with the original scan as fallback. For the identified 695-query hotspot, the prototype eliminates the 2,590,482 original interval-scan inspections, while performing 10,658 summary-building classifications. Summary classifications count root operations, not all recursively visited operations.
Release/O3, assertions off, AMD EPYC 9J14. Command timings use uninstrumented binaries; inspection counts come from separate diagnostic runs. Replay results are from the fixed-IR recheck; full-compilation results reuse the earlier seven-pair experiment. Both variants use the same input and build configuration. Each comparison has one warmup per variant followed by seven serial AB/BA pairs. CSE pass timings use a separate seven-pair run with standard `--mlir-timing`. Medians; changes computed before rounding:
| Measurement | Baseline | Prototype | Change |
|---|---:|---:|---:|
| Replay command, including parsing/output | 0.6995 s | 0.1953 s | **−72.08%** |
| CSE pass, separate `--mlir-timing` runs | 0.5448 s | 0.0382 s | **−92.99%** |
| Full CLIP compilation | 18.055 s | 17.839 s | −1.20% |
| Full ResNet50 compilation | 4.697 s | 4.687 s | −0.20% |
| Full Whisper tiny-en compilation | 3.352 s | 3.318 s | −1.00% |
The **72% replay-command reduction** is clear; end-to-end improvement is much smaller. Control-workload differences are within noise and do not exclude small regressions.
Full CLIP peak RSS medians increase **358.2→378.8 MiB (+5.75%)**, not yet fully attributed. Outputs match, and ResNet50 numerical tests pass; CLIP lacks complete numerical reference validation. A standalone upstream-MLIR reduction is not yet available. The prototype demonstrates avoidable work on this input; it is included for discussion rather than proposed as an upstream-ready fix.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs