Cross-Dialect MLIR Benchmarks - NeurIPS 2026 (E&D)
Collection
NL→MLIR benchmarks (465 instances; arith/func/memref, linalg, StableHLO). Code: github.com/plawanrath/slm-to-mlir-constrained-emitter • 6 items • Updated
id stringlengths 9 24 | dialect stringclasses 2
values | difficulty stringclasses 3
values | nl stringlengths 54 155 | mlir stringlengths 131 318 | notes stringlengths 6 25 |
|---|---|---|---|---|---|
01_matmul-dynamic | linalg+memref+func | easy | Write a function that performs a matrix multiplication of two 2-D f32 memrefs with dynamic shapes and writes the result into a pre-allocated output memref. | module {
func.func @mm(%A: memref<?x?xf32>, %B: memref<?x?xf32>, %C: memref<?x?xf32>) {
linalg.matmul ins(%A, %B : memref<?x?xf32>, memref<?x?xf32>) outs(%C : memref<?x?xf32>)
return
}
}
| canonical linalg.matmul |
02_matmul-static-2x2 | linalg+memref+func | easy | Write a function that performs a 2x2 matrix multiplication of f32 matrices. All inputs are statically-shaped memrefs. | module {
func.func @mm22(%A: memref<2x2xf32>, %B: memref<2x2xf32>, %C: memref<2x2xf32>) {
linalg.matmul ins(%A, %B : memref<2x2xf32>, memref<2x2xf32>) outs(%C : memref<2x2xf32>)
return
}
}
| static-shape matmul |
03_matmul-rectangular | linalg+memref+func | medium | Write a function that multiplies a 4x8xf32 matrix by an 8x2xf32 matrix, writing to a 4x2xf32 result memref. | module {
func.func @mm482(%A: memref<4x8xf32>, %B: memref<8x2xf32>, %C: memref<4x2xf32>) {
linalg.matmul ins(%A, %B : memref<4x8xf32>, memref<8x2xf32>) outs(%C : memref<4x2xf32>)
return
}
}
| rectangular MxK * KxN |
04_matmul-f64 | linalg+memref+func | medium | Write a function that performs matmul on two f64 2-D memrefs into an f64 output memref. | module {
func.func @mmf64(%A: memref<?x?xf64>, %B: memref<?x?xf64>, %C: memref<?x?xf64>) {
linalg.matmul ins(%A, %B : memref<?x?xf64>, memref<?x?xf64>) outs(%C : memref<?x?xf64>)
return
}
}
| f64 matmul |
05_fill-matmul-chain | linalg+memref+arith+func | hard | Write a function that zeroes an f32 output memref using linalg.fill, then computes matmul of two input memrefs into it. | module {
func.func @fm(%A: memref<?x?xf32>, %B: memref<?x?xf32>, %C: memref<?x?xf32>) {
%z = arith.constant 0.0 : f32
linalg.fill ins(%z : f32) outs(%C : memref<?x?xf32>)
linalg.matmul ins(%A, %B : memref<?x?xf32>, memref<?x?xf32>) outs(%C : memref<?x?xf32>)
return
}
}
| zero-then-matmul idiom |
06_fill-zero-1d | linalg+memref+arith+func | easy | Write a function that fills a 1-D f32 memref with zeros. | module {
func.func @f0(%m: memref<?xf32>) {
%z = arith.constant 0.0 : f32
linalg.fill ins(%z : f32) outs(%m : memref<?xf32>)
return
}
}
| zero-fill 1D |
07_fill-value-param | linalg+memref+func | easy | Write a function that fills a 2-D f32 memref with a given f32 value passed as a parameter. | module {
func.func @fp(%v: f32, %m: memref<?x?xf32>) {
linalg.fill ins(%v : f32) outs(%m : memref<?x?xf32>)
return
}
}
| fill with parameter value |
08_fill-i32 | linalg+memref+arith+func | easy | Write a function that fills a 1-D i32 memref with the integer constant 7. | module {
func.func @f7(%m: memref<?xi32>) {
%c7 = arith.constant 7 : i32
linalg.fill ins(%c7 : i32) outs(%m : memref<?xi32>)
return
}
}
| integer-typed fill |
09_copy-1d | linalg+memref+func | easy | Write a function that copies a 1-D f32 memref into another 1-D f32 memref of the same shape. | module {
func.func @c1(%s: memref<?xf32>, %d: memref<?xf32>) {
linalg.copy ins(%s : memref<?xf32>) outs(%d : memref<?xf32>)
return
}
}
| 1D copy |
10_copy-2d-static | linalg+memref+func | easy | Write a function that copies a 4x4xf32 static memref into another 4x4xf32 memref. | module {
func.func @c2(%s: memref<4x4xf32>, %d: memref<4x4xf32>) {
linalg.copy ins(%s : memref<4x4xf32>) outs(%d : memref<4x4xf32>)
return
}
}
| 2D static copy |
11_copy-i32 | linalg+memref+func | easy | Write a function that copies a 1-D i32 memref into another 1-D i32 memref of the same shape. | module {
func.func @ci(%s: memref<?xi32>, %d: memref<?xi32>) {
linalg.copy ins(%s : memref<?xi32>) outs(%d : memref<?xi32>)
return
}
}
| i32 copy |
12_transpose-2d | linalg+memref+func | medium | Write a function that transposes a 2-D f32 memref into another 2-D f32 memref using permutation [1, 0]. | module {
func.func @t2(%s: memref<?x?xf32>, %d: memref<?x?xf32>) {
linalg.transpose ins(%s : memref<?x?xf32>) outs(%d : memref<?x?xf32>) permutation = [1, 0]
return
}
}
| 2D transpose |
13_transpose-3d | linalg+memref+func | hard | Write a function that transposes a 3-D f32 memref using permutation [0, 2, 1] (swap the last two dimensions). | module {
func.func @t3(%s: memref<?x?x?xf32>, %d: memref<?x?x?xf32>) {
linalg.transpose ins(%s : memref<?x?x?xf32>) outs(%d : memref<?x?x?xf32>) permutation = [0, 2, 1]
return
}
}
| 3D transpose |
14_transpose-static | linalg+memref+func | medium | Write a function that transposes a 3x5xf32 memref to a 5x3xf32 memref using permutation [1, 0]. | module {
func.func @ts(%s: memref<3x5xf32>, %d: memref<5x3xf32>) {
linalg.transpose ins(%s : memref<3x5xf32>) outs(%d : memref<5x3xf32>) permutation = [1, 0]
return
}
}
| static-shape transpose |
15_broadcast-1d-to-2d | linalg+memref+func | medium | Write a function that broadcasts a 1-D f32 memref to a 2-D f32 memref along the 0th dimension. | module {
func.func @b1(%s: memref<?xf32>, %d: memref<?x?xf32>) {
linalg.broadcast ins(%s : memref<?xf32>) outs(%d : memref<?x?xf32>) dimensions = [0]
return
}
}
| 1D to 2D broadcast, dim 0 |
16_broadcast-dim1 | linalg+memref+func | medium | Write a function that broadcasts a 1-D f32 memref to a 2-D f32 memref along the 1st dimension. | module {
func.func @b2(%s: memref<?xf32>, %d: memref<?x?xf32>) {
linalg.broadcast ins(%s : memref<?xf32>) outs(%d : memref<?x?xf32>) dimensions = [1]
return
}
}
| 1D to 2D broadcast, dim 1 |
17_matvec | linalg+memref+func | medium | Write a function that performs a matrix-vector multiplication: input is a 2-D f32 memref and a 1-D f32 memref; output is a 1-D f32 memref. | module {
func.func @mv(%A: memref<?x?xf32>, %x: memref<?xf32>, %y: memref<?xf32>) {
linalg.matvec ins(%A, %x : memref<?x?xf32>, memref<?xf32>) outs(%y : memref<?xf32>)
return
}
}
| canonical matvec |
18_matvec-static | linalg+memref+func | medium | Write a function that multiplies a 4x8xf32 matrix by an 8xf32 vector, writing to a 4xf32 result memref. | module {
func.func @mvs(%A: memref<4x8xf32>, %x: memref<8xf32>, %y: memref<4xf32>) {
linalg.matvec ins(%A, %x : memref<4x8xf32>, memref<8xf32>) outs(%y : memref<4xf32>)
return
}
}
| static matvec 4x8 by 8 |
19_matvec-f64 | linalg+memref+func | medium | Write a function that performs a matvec on f64 inputs. | module {
func.func @mvd(%A: memref<?x?xf64>, %x: memref<?xf64>, %y: memref<?xf64>) {
linalg.matvec ins(%A, %x : memref<?x?xf64>, memref<?xf64>) outs(%y : memref<?xf64>)
return
}
}
| f64 matvec |
20_add-elemwise | linalg+memref+func | easy | Write a function that adds two 1-D f32 memrefs element-wise into a third output memref. | module {
func.func @ae(%a: memref<?xf32>, %b: memref<?xf32>, %c: memref<?xf32>) {
linalg.add ins(%a, %b : memref<?xf32>, memref<?xf32>) outs(%c : memref<?xf32>)
return
}
}
| elemwise add |
21_sub-elemwise | linalg+memref+func | easy | Write a function that subtracts two 1-D f32 memrefs element-wise. | module {
func.func @se(%a: memref<?xf32>, %b: memref<?xf32>, %c: memref<?xf32>) {
linalg.sub ins(%a, %b : memref<?xf32>, memref<?xf32>) outs(%c : memref<?xf32>)
return
}
}
| elemwise sub |
22_mul-elemwise-2d | linalg+memref+func | medium | Write a function that multiplies two 2-D f32 memrefs element-wise (Hadamard product) into an output memref. | module {
func.func @me(%a: memref<?x?xf32>, %b: memref<?x?xf32>, %c: memref<?x?xf32>) {
linalg.mul ins(%a, %b : memref<?x?xf32>, memref<?x?xf32>) outs(%c : memref<?x?xf32>)
return
}
}
| Hadamard product |
23_div-elemwise | linalg+memref+func | medium | Write a function that divides two 1-D f32 memrefs element-wise. | module {
func.func @de(%a: memref<?xf32>, %b: memref<?xf32>, %c: memref<?xf32>) {
linalg.div ins(%a, %b : memref<?xf32>, memref<?xf32>) outs(%c : memref<?xf32>)
return
}
}
| elemwise div |
24_exp-elemwise | linalg+memref+func | easy | Write a function that applies the elementwise exponential to a 1-D f32 memref, writing the result into another memref. | module {
func.func @ee(%x: memref<?xf32>, %y: memref<?xf32>) {
linalg.exp ins(%x : memref<?xf32>) outs(%y : memref<?xf32>)
return
}
}
| elemwise exp |
25_exp-2d | linalg+memref+func | easy | Write a function that applies elementwise exp to a 2-D f32 memref, writing into an output memref of the same shape. | module {
func.func @e2(%x: memref<?x?xf32>, %y: memref<?x?xf32>) {
linalg.exp ins(%x : memref<?x?xf32>) outs(%y : memref<?x?xf32>)
return
}
}
| 2D exp |
26_abs-elemwise | linalg+memref+func | easy | Write a function that applies the elementwise absolute value to a 1-D f32 memref, writing the result into another memref. | module {
func.func @ae(%x: memref<?xf32>, %y: memref<?xf32>) {
linalg.abs ins(%x : memref<?xf32>) outs(%y : memref<?xf32>)
return
}
}
| elemwise abs |
27_abs-2d | linalg+memref+func | easy | Write a function that applies elementwise abs to a 2-D f32 memref. | module {
func.func @a2(%x: memref<?x?xf32>, %y: memref<?x?xf32>) {
linalg.abs ins(%x : memref<?x?xf32>) outs(%y : memref<?x?xf32>)
return
}
}
| 2D abs |
28_fill-then-copy | linalg+memref+arith+func | hard | Write a function that fills a 1-D f32 memref with 1.0 and then copies its contents into a second memref. | module {
func.func @fc(%m: memref<?xf32>, %d: memref<?xf32>) {
%one = arith.constant 1.0 : f32
linalg.fill ins(%one : f32) outs(%m : memref<?xf32>)
linalg.copy ins(%m : memref<?xf32>) outs(%d : memref<?xf32>)
return
}
}
| fill-then-copy chain |
29_add-then-exp | linalg+memref+func | hard | Write a function that adds two 1-D f32 memrefs element-wise into a temporary buffer, then applies exp to the buffer into the final output. | module {
func.func @ax(%a: memref<?xf32>, %b: memref<?xf32>, %t: memref<?xf32>, %y: memref<?xf32>) {
linalg.add ins(%a, %b : memref<?xf32>, memref<?xf32>) outs(%t : memref<?xf32>)
linalg.exp ins(%t : memref<?xf32>) outs(%y : memref<?xf32>)
return
}
}
| add-then-exp chain |
30_transpose-then-matmul | linalg+memref+func | hard | Write a function that transposes A (2-D f32) into a temp memref and then performs matmul of transposed A and B. | module {
func.func @tm(%A: memref<?x?xf32>, %At: memref<?x?xf32>, %B: memref<?x?xf32>, %C: memref<?x?xf32>) {
linalg.transpose ins(%A : memref<?x?xf32>) outs(%At : memref<?x?xf32>) permutation = [1, 0]
linalg.matmul ins(%At, %B : memref<?x?xf32>, memref<?x?xf32>) outs(%C : memref<?x?xf32>)
return
}
}
| transpose-then-matmul |
Linalg-Spec-30
Accepted to NeurIPS 2026 (Evaluations & Datasets Track)
Paper (arXiv) · Code (GitHub) · All six datasets (collection)
Hand-authored NL→MLIR pairs for linalg named ops under memref semantics (n=30).
This dataset is part of the suite released with the NeurIPS 2026 Evaluations & Datasets Track paper Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR (arXiv:2607.18254): four NL→MLIR benchmarks (MLIR-Spec-150, Linalg-Spec-30, StableHLO-Spec-30, StableHLO-Held-Out-200) and the StableHLO-OutOfGrammar-25 stress set, 435 benchmark instances across three MLIR dialects, plus MLIR-Functional-Reference-30, which attaches hand-authored inputs and expected outputs to 30 existing Spec prompts (no new prompts).
Composition
- Instances: 30
- Format: one JSON record per line in
data/test.jsonl - Schema: fields =
dialect,difficulty,id,mlir,nl,notes - Verifier:
mlir-opt --verify-diagnosticsagainst pinned LLVM 19.1.7 - License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions.
Loading
from datasets import load_dataset
ds = load_dataset("plawanrath/Linalg-Spec-30", split="test")
print(ds[0])
Each record is a self-contained natural-language→MLIR pair; verify-valid pass-rate under the dialect's verifier is the primary evaluation metric.
Source format
The JSONL file at data/test.jsonl is the canonical HuggingFace interface. It is exported 1-to-1 from the source records (examples/*.json) in eval/benchmarks/linalg_spec_30/ of the code repository.
Datasheet
Key points (full Gebru-style datasheet: docs/datasheets/datasheet.md):
- All reference MLIR programs are verifier-clean at the time of release.
- Hand-authored (no crowdsourcing, no LLM-authored references).
- Test-only — fine-tuning on these benchmarks contaminates future evaluation and is explicitly out of scope.
Croissant metadata
croissant.json in this repository is the dataset's MLCommons Croissant 1.1 metadata, including the Responsible AI (rai:) and provenance (prov:) fields required by the NeurIPS 2026 Evaluations & Datasets Track. (Hugging Face's auto-generated Croissant endpoint covers the core fields only.)
Code
Generation, verification and evaluation code: github.com/plawanrath/slm-to-mlir-constrained-emitter.
Citation
@inproceedings{rath2026crossdialect,
title = {Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for {MLIR}},
author = {Rath, Plawan Kumar},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
year = {2026},
eprint = {2607.18254},
archivePrefix = {arXiv}
}
License
Apache-2.0 (see LICENSE). Copyright 2026 Plawan Kumar Rath.
Disclaimer
Plawan Kumar Rath (@plawanrath). This work was conducted in the author's personal capacity. The views expressed here are those of the author and do not necessarily reflect the views of Meta.
- Downloads last month
- 104