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 string | dialect string | source_benchmark string | source_id string | nl string | canonical_fn_name string | canonical_signature string | inputs list | result_type string | expected_output float64 | expected_stdout_regex string | memref_inputs list | memref_print list | scalar_inputs list | iree_inputs list | expected_output_pattern string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
F01_arith_add | arith+func | mlir_spec_150 | 001_add-two-ints | Write a function `add` that takes two i32 values and returns their sum. | add | (i32, i32) -> i32 | [
{
"type": "i32",
"value": 7
},
{
"type": "i32",
"value": 3
}
] | i32 | 10 | ^10\b | null | null | null | null | null |
F02_arith_mul_i64 | arith+func | mlir_spec_150 | 003_multiply-ints | Write a function `mul` that multiplies two i64 values and returns their product. | mul | (i64, i64) -> i64 | [
{
"type": "i64",
"value": 6
},
{
"type": "i64",
"value": 7
}
] | i64 | 42 | ^42\b | null | null | null | null | null |
F03_arith_addf_f32 | arith+func | mlir_spec_150 | 004_add-floats | Write a function `addf` that takes two f32 values and returns their sum. | addf | (f32, f32) -> f32 | [
{
"type": "f32",
"value": 1.5
},
{
"type": "f32",
"value": 2.25
}
] | f32 | 3.75 | ^3\.75\b | null | null | null | null | null |
F04_arith_const42 | arith+func | mlir_spec_150 | 005_integer-constant | Write a function that returns the constant integer 42 as an i32. | const42 | () -> i32 | [] | i32 | 42 | ^42\b | null | null | null | null | null |
F05_arith_chain_addmul | arith+func | mlir_spec_150 | 008_chain-add-then-multiply | Write a function that computes `(a + b) * c` on i32 inputs and returns the i32 result. | fma_like | (i32, i32, i32) -> i32 | [
{
"type": "i32",
"value": 2
},
{
"type": "i32",
"value": 3
},
{
"type": "i32",
"value": 4
}
] | i32 | 20 | ^20\b | null | null | null | null | null |
F06_arith_band | arith+func | mlir_spec_150 | 016_bitwise-and | Write a function that computes the bitwise AND of two i32 values. | band | (i32, i32) -> i32 | [
{
"type": "i32",
"value": 12
},
{
"type": "i32",
"value": 10
}
] | i32 | 8 | ^8\b | null | null | null | null | null |
F07_arith_inc | arith+func | mlir_spec_150 | 033_increment-by-one | Write a function that takes an i32 and returns it incremented by 1. | inc | (i32) -> i32 | [
{
"type": "i32",
"value": 41
}
] | i32 | 42 | ^42\b | null | null | null | null | null |
F08_arith_double_f32 | arith+func | mlir_spec_150 | 034_double-value | Write a function that doubles an f32 value. | dbl | (f32) -> f32 | [
{
"type": "f32",
"value": 3.5
}
] | f32 | 7 | ^7(\.0+)?\b | null | null | null | null | null |
F09_arith_square_i32 | arith+func | mlir_spec_150 | 040_square | Write a function that computes x squared for an i32 x. | square | (i32) -> i32 | [
{
"type": "i32",
"value": 7
}
] | i32 | 49 | ^49\b | null | null | null | null | null |
F10_arith_cube_f32 | arith+func | mlir_spec_150 | 041_cube | Write a function that computes x * x * x for an f32 x. | cube | (f32) -> f32 | [
{
"type": "f32",
"value": 2
}
] | f32 | 8 | ^8(\.0+)?\b | null | null | null | null | null |
F11_linalg_fill_zero_1d | linalg+memref | linalg_spec_30 | 06_fill-zero-1d | Write a function that fills a 1-D f32 memref with zeros. | f0 | (memref<?xf32>) -> () | null | null | null | data\s*=\s*\[\s*0(\.0+)?\s*,\s*0(\.0+)?\s*,\s*0(\.0+)?\s*,\s*0(\.0+)?\s*\] | [
{
"name": "m",
"shape": [
4
],
"dtype": "f32",
"init": "alloc_only"
}
] | [
{
"name": "m",
"shape": [
4
],
"dtype": "f32"
}
] | null | null | null |
F12_linalg_fill_value_2d | linalg+memref | linalg_spec_30 | 07_fill-value-param | Write a function that fills a 2-D f32 memref with a given f32 value passed as a parameter. | fp | (f32, memref<?x?xf32>) -> () | null | null | null | (5(\.0+)?[^\d]+){4} | [
{
"name": "m",
"shape": [
2,
2
],
"dtype": "f32",
"init": "alloc_only"
}
] | [
{
"name": "m",
"shape": [
2,
2
],
"dtype": "f32"
}
] | [
{
"type": "f32",
"value": 5
}
] | null | null |
F13_linalg_fill_i32_const | linalg+memref | linalg_spec_30 | 08_fill-i32 | Write a function that fills a 1-D i32 memref with the integer constant 7. | f7 | (memref<?xi32>) -> () | null | null | null | (7[^\d]+){4} | [
{
"name": "m",
"shape": [
4
],
"dtype": "i32",
"init": "alloc_only"
}
] | [
{
"name": "m",
"shape": [
4
],
"dtype": "i32"
}
] | null | null | null |
F14_linalg_copy_1d | linalg+memref | linalg_spec_30 | 09_copy-1d | Write a function that copies a 1-D f32 memref into another 1-D f32 memref of the same shape. | c1 | (memref<?xf32>, memref<?xf32>) -> () | null | null | null | (3\.5[^\d]+){4} | [
{
"name": "s",
"shape": [
4
],
"dtype": "f32",
"init": "fill",
"fill_value": 3.5
},
{
"name": "d",
"shape": [
4
],
"dtype": "f32",
"init": "alloc_only"
}
] | [
{
"name": "d",
"shape": [
4
],
"dtype": "f32"
}
] | null | null | null |
F15_linalg_copy_2d_static | linalg+memref | linalg_spec_30 | 10_copy-2d-static | Write a function that copies a 4x4xf32 static memref into another 4x4xf32 memref. | c2 | (memref<4x4xf32>, memref<4x4xf32>) -> () | null | null | null | (1\.25[^\d]+){4} | [
{
"name": "s",
"shape": [
4,
4
],
"dtype": "f32",
"init": "fill",
"fill_value": 1.25,
"static": true
},
{
"name": "d",
"shape": [
4,
4
],
"dtype": "f32",
"init": "alloc_only",
"static": true
}
] | [
{
"name": "d",
"shape": [
4,
4
],
"dtype": "f32"
}
] | null | null | null |
F16_linalg_add_elemwise | linalg+memref | linalg_spec_30 | 20_add-elemwise | linalg.add elementwise on two 1-D f32 memrefs | ae | (memref<?xf32>, memref<?xf32>, memref<?xf32>) -> () | null | null | null | (5(\.0+)?[^\d]+){4} | [
{
"name": "a",
"shape": [
4
],
"dtype": "f32",
"init": "fill",
"fill_value": 2
},
{
"name": "b",
"shape": [
4
],
"dtype": "f32",
"init": "fill",
"fill_value": 3
},
{
"name": "c",
"shape": [
4
],
"dtype": "f32",
"init": "al... | [
{
"name": "c",
"shape": [
4
],
"dtype": "f32"
}
] | null | null | null |
F17_linalg_mul_elemwise_2d | linalg+memref | linalg_spec_30 | 22_mul-elemwise-2d | linalg.mul elementwise on two 2-D f32 memrefs (2x2) | me | (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> () | null | null | null | (12(\.0+)?[^\d]+){4} | [
{
"name": "a",
"shape": [
2,
2
],
"dtype": "f32",
"init": "fill",
"fill_value": 3
},
{
"name": "b",
"shape": [
2,
2
],
"dtype": "f32",
"init": "fill",
"fill_value": 4
},
{
"name": "c",
"shape": [
2,
2
],
"dty... | [
{
"name": "c",
"shape": [
2,
2
],
"dtype": "f32"
}
] | null | null | null |
F18_linalg_matmul_2x2 | linalg+memref | linalg_spec_30 | 02_matmul-static-2x2 | Write a function that performs matrix multiplication of two 2x2 f32 memrefs and writes the result into the output 2x2 f32 memref. | mm22 | (memref<2x2xf32>, memref<2x2xf32>, memref<2x2xf32>) -> () | null | null | null | (2(\.0+)?[^\d]+){4} | [
{
"name": "A",
"shape": [
2,
2
],
"dtype": "f32",
"init": "fill",
"fill_value": 1,
"static": true
},
{
"name": "B",
"shape": [
2,
2
],
"dtype": "f32",
"init": "fill",
"fill_value": 1,
"static": true
},
{
"name": "C",
"sh... | [
{
"name": "C",
"shape": [
2,
2
],
"dtype": "f32"
}
] | null | null | null |
F19_linalg_exp_elemwise | linalg+memref | linalg_spec_30 | 24_exp-elemwise | linalg.exp elementwise on a 1-D f32 memref | ee | (memref<?xf32>, memref<?xf32>) -> () | null | null | null | (1(\.0+)?[^\d]+){4} | [
{
"name": "x",
"shape": [
4
],
"dtype": "f32",
"init": "fill",
"fill_value": 0
},
{
"name": "y",
"shape": [
4
],
"dtype": "f32",
"init": "alloc_only"
}
] | [
{
"name": "y",
"shape": [
4
],
"dtype": "f32"
}
] | null | null | null |
F20_linalg_abs_elemwise | linalg+memref | linalg_spec_30 | 26_abs-elemwise | linalg.abs elementwise on a 1-D f32 memref | ab | (memref<?xf32>, memref<?xf32>) -> () | null | null | null | (2\.5[^\d]+){4} | [
{
"name": "x",
"shape": [
4
],
"dtype": "f32",
"init": "fill",
"fill_value": -2.5
},
{
"name": "y",
"shape": [
4
],
"dtype": "f32",
"init": "alloc_only"
}
] | [
{
"name": "y",
"shape": [
4
],
"dtype": "f32"
}
] | null | null | null |
F21_stablehlo_add_1d_f32 | stablehlo | stablehlo_spec_30 | 01_add-1d | Write a function that adds two 1-D f32 tensors of 16 elements using stablehlo.add. | a | (tensor<16xf32>, tensor<16xf32>) -> tensor<16xf32> | null | null | null | null | null | null | null | [
"16xf32=1.0",
"16xf32=2.0"
] | 16xf32=3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 |
F22_stablehlo_multiply_2d | stablehlo | stablehlo_spec_30 | 04_multiply-2d | Write a function that multiplies two 4x4 f32 tensors elementwise using stablehlo.multiply. | mul | (tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4xf32> | null | null | null | null | null | null | null | [
"4x4xf32=3.0",
"4x4xf32=4.0"
] | 4x4xf32=\[12 12 12 12\]\[12 12 12 12\]\[12 12 12 12\]\[12 12 12 12\] |
F23_stablehlo_subtract_1d_i32 | stablehlo | stablehlo_spec_30 | 03_subtract-1d-i32 | Write a function that subtracts two 1-D i32 tensors of 16 elements using stablehlo.subtract. | s | (tensor<16xi32>, tensor<16xi32>) -> tensor<16xi32> | null | null | null | null | null | null | null | [
"16xi32=10",
"16xi32=4"
] | 16xi32=6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 |
F24_stablehlo_abs_f32 | stablehlo | stablehlo_spec_30 | 06_abs-f32 | Write a function that computes the elementwise absolute value of a 1-D f32 tensor. | ab | (tensor<16xf32>) -> tensor<16xf32> | null | null | null | null | null | null | null | [
"16xf32=-2.5"
] | 16xf32=2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 2.5 |
F25_stablehlo_exp_1d | stablehlo | stablehlo_spec_30 | 07_exp-1d | Write a function that computes the elementwise exponential of a 1-D f32 tensor. | ex | (tensor<16xf32>) -> tensor<16xf32> | null | null | null | null | null | null | null | [
"16xf32=0.0"
] | 16xf32=1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 |
F26_stablehlo_transpose_2d | stablehlo | stablehlo_spec_30 | 09_transpose-2d | Write a function that transposes a 4x8 f32 tensor producing an 8x4 tensor. | t | (tensor<4x8xf32>) -> tensor<8x4xf32> | null | null | null | null | null | null | null | [
"4x8xf32=1.5"
] | 8x4xf32= |
F27_stablehlo_add_2d_dyn | stablehlo | stablehlo_spec_30 | 02_add-2d-dynamic | Write a function that adds two 2x4 f32 tensors using stablehlo.add. | a2 | (tensor<2x4xf32>, tensor<2x4xf32>) -> tensor<2x4xf32> | null | null | null | null | null | null | null | [
"2x4xf32=2.0",
"2x4xf32=2.5"
] | 2x4xf32=\[4.5 4.5 4.5 4.5\]\[4.5 4.5 4.5 4.5\] |
F28_stablehlo_divide_f64 | stablehlo | stablehlo_spec_30 | 05_divide-f64 | Write a function that divides two 1-D f64 tensors of 8 elements using stablehlo.divide. | d | (tensor<8xf64>, tensor<8xf64>) -> tensor<8xf64> | null | null | null | null | null | null | null | [
"8xf64=12.0",
"8xf64=4.0"
] | 8xf64=3 3 3 3 3 3 3 3 |
F29_stablehlo_dot_matmul | stablehlo | stablehlo_spec_30 | 17_dot_general-matmul | Write a function that performs a matmul of two f32 tensors via stablehlo.dot_general (lhs 4x8, rhs 8x4). | dm | (tensor<4x8xf32>, tensor<8x4xf32>) -> tensor<4x4xf32> | null | null | null | null | null | null | null | [
"4x8xf32=1.0",
"8x4xf32=1.0"
] | 4x4xf32=\[8 8 8 8\]\[8 8 8 8\]\[8 8 8 8\]\[8 8 8 8\] |
F30_stablehlo_transpose_3d | stablehlo | stablehlo_spec_30 | 10_transpose-3d | Write a function that transposes a 2x3x4 f32 tensor with permutation [2, 0, 1] producing a 4x2x3 tensor. | t3 | (tensor<2x3x4xf32>) -> tensor<4x2x3xf32> | null | null | null | null | null | null | null | [
"2x3x4xf32=2.5"
] | 4x2x3xf32= |
MLIR-Functional-Reference-30
Accepted to NeurIPS 2026 (Evaluations & Datasets Track)
Paper (arXiv) · Code (GitHub) · All six datasets (collection)
Hand-authored functional-correctness reference set for arith, linalg+memref, and stablehlo (n=30, 10 per dialect).
This dataset is one of six NL→MLIR benchmarks 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). The full suite — MLIR-Spec-150, Linalg-Spec-30, StableHLO-Spec-30, StableHLO-Held-Out-200, StableHLO-OutOfGrammar-25, MLIR-Functional-Reference-30 — totals 465 instances across three MLIR dialects.
Composition
- Instances: 30
- Format: one JSON record per line in
data/test.jsonl - Schema: fields =
canonical_fn_name,canonical_signature,dialect,expected_output,expected_output_pattern,expected_stdout_regex,id,inputs,iree_inputs,memref_inputs,memref_print,nl,result_type,scalar_inputs,source_benchmark,source_id - Verifier: dialect-specific lowering pipelines + execution comparison; see
eval/functional/run_functional.pyfor the per-dialect runners - License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions.
Loading
from datasets import load_dataset
ds = load_dataset("plawanrath/MLIR-Functional-Reference-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 (references.json) in eval/functional/ 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.
- Downloads last month
- 18