mlir

Name: mlir
Author: mohitmishra786

141stars19forks

How to install mlir?

mohitmishra786/low-level-dev-skills/mlir

$npx -y skills add mohitmishra786/low-level-dev-skills --skill mlir

Installs into the current project.

Summary

MLIR skill for multi-level intermediate representation. Use when writing custom dialects, defining ops with ODS, writing lowering passes, running mlir-opt, or building ML compilers with Torch-MLIR/IREE. Activates on queries about MLIR, dialect, ODS, mlir-opt, linalg, lowering pas

Files · 1

View on GitHub

SKILL.md

1	# MLIR
2
3	## Purpose
4
5	Guide agents through MLIR (Multi-Level IR): ops, regions, blocks, and values; built-in dialects (arith, func, memref, affine, linalg); writing custom dialects with ODS; lowering passes with `ConversionPattern`; `mlir-opt` CLI; and ML compiler use cases (Torch-MLIR, IREE).
6
7	## When to Use
8
9	- Building a domain-specific compiler IR (graphics, ML, hardware DSL)
10	- Lowering high-level ops to LLVM or GPU dialects
11	- Writing progressive lowering pipelines (linalg → loops → LLVM)
12	- Integrating with IREE or Torch-MLIR for ML deployment
13	- Creating reusable transformation passes across dialects
14	- Prototyping compiler optimizations at the right abstraction level
15
16	## Workflow
17
18	### 1. MLIR structure
19
20	```
21	Module
22	└── func.func @main()
23	└── region
24	└── block ^bb0:
25	└── operations (ops) producing SSA values
26	```
27
28	Key concepts:
29	- Operation — instruction-like node (`arith.addi`, `memref.load`)
30	- Region — container of blocks (functions, control flow)
31	- Block — CFG node with ordered ops
32	- Value — SSA result of an op or block argument
33
34	### 2. Built-in dialects
35
36	\| Dialect \| Purpose \|
37	\|---------\|---------\|
38	\| `arith` \| Integer/float arithmetic \|
39	\| `func` \| Function definitions and calls \|
40	\| `memref` \| Buffer abstraction with shapes/strides \|
41	\| `affine` \| Affine loop nests, map/set constraints \|
42	\| `linalg` \| Structured linear algebra ops \|
43	\| `scf` \| Structured control flow (for, if) \|
44	\| `llvm` \| LLVM IR dialect for final lowering \|
45	\| `gpu` \| GPU kernel launches \|
46
47	```mlir
48	// example.mlir
49	func.func @add(%a: memref<4xf32>, %b: memref<4xf32>, %c: memref<4xf32>) {
50	%c0 = arith.constant 0 : index
51	%c4 = arith.constant 4 : index
52	scf.for %i = %c0 to %c4 step %c1 {
53	%av = memref.load %a[%i] : memref<4xf32>
54	%bv = memref.load %b[%i] : memref<4xf32>
55	%sum = arith.addf %av, %bv : f32
56	memref.store %sum, %c[%i] : memref<4xf32>
57	}
58	return
59	}
60	```
61
62	### 3. mlir-opt CLI
63
64	```bash
65	# Parse and print
66	mlir-opt example.mlir
67
68	# Run canonicalization
69	mlir-opt example.mlir -canonicalize
70
71	# Lower affine to scf
72	mlir-opt affine.mlir -lower-affine
73
74	# Full pipeline toward LLVM
75	mlir-opt input.mlir \
76	--linalg-bufferize \
77	--convert-linalg-to-loops \
78	--convert-scf-to-cf \
79	--convert-arith-to-llvm \
80	--convert-memref-to-llvm \
81	--convert-func-to-llvm \
82	-o llvm.mlir
83	```
84
85	### 4. ODS — Operation Definition Specification
86
87	```tablegen
88	// MyOps.td
89	include "mlir/IR/OpBase.td"
90
91	def My_Dialect : Dialect {
92	let name = "my";
93	let summary = "My custom dialect";
94	}
95
96	class My_Op<string mnemonic, list<Trait> traits = []> :
97	Op<My_Dialect, mnemonic, traits>;
98
99	def AddOp : My_Op<"add", [Pure]> {
100	let summary = "Add two values";
101	let arguments = (ins AnyType:$lhs, AnyType:$rhs);
102	let results = (outs AnyType:$result);
103	let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
104	}
105	```
106
107	```bash
108	# Generate C++ from TableGen
109	mlir-tblgen -gen-op-defs MyOps.td -I include/ -o MyOps.cpp.inc
110	```
111
112	### 5. Custom dialect C++ implementation
113
114	```cpp
115	#include "mlir/IR/DialectImplementation.h"
116	#include "MyDialect.h"
117
118	#include "MyOps.cpp.inc"
119
120	void MyDialect::initialize() {
121	addOperations<
122	#define GET_OP_LIST
123	#include "MyOps.cpp.inc"
124	>();
125	}
126
127	#define GET_OP_CLASSES
128	#include "MyOps.cpp.inc"
129	```
130
131	### 6. Lowering passes
132
133	```cpp
134	#include "mlir/Conversion/LLVMCommon/ConversionTarget.h"
135	#include "mlir/Transforms/DialectConversion.h"
136
137	struct AddOpLowering : OpConversionPattern<my::AddOp> {
138	using OpConversionPattern::OpConversionPattern;
139
140	LogicalResult matchAndRewrite(my::AddOp op, OpAdaptor adaptor,
141	ConversionPatternRewriter &rewriter) const override {
142	rewriter.replaceOpWithNewOp<arith::AddIOp>(op, adaptor.getLhs(), adaptor.getRhs());
143	return success();
144	}
145	};
146
147	void populateLoweringPatterns(RewritePatternSet &patterns) {
148	patterns.add<AddOpLowering>(patterns.getContext());
149	}
150
151	// In pass:
152	mlir::ConversionTarget target(*context);
153	target.addIllegalDialect<my::MyDialect>();
154	target.addLegalDialect<arith::ArithDialect>();
155
156	if (failed(applyPartialConversion(module, target, std::move(patterns))))
157	signalPassFailure();
158	```
159
160	### 7. linalg for ML compilers
161
162	```mlir
163	%0 = linalg.matmul ins(%A, %B : tensor<128x256xf32>, tensor<256x64xf32>)
164	outs(%C : tensor<128x64xf32>) -> tensor<128x64xf32>
165	```
166
167	Lowering path: `linalg` → `scf` loops → `affine` → `llvm`
168
169	### 8. Torch-MLIR and IREE
170
171	```bash
172	# Torch-MLIR: PyTorch → MLIR
173	python -m torch_mlir.tools.import-onnx --onnx-model model.onnx -o model.mlir
174
175	# IREE: MLIR → GPU/CPU executable
176	iree-compile --iree-hal-target-backends=llvm-cpu model.mlir -o model.vmfb
177	iree-run

People who liked this also viewed

6 picks

Type

mlir

141stars19forks

View on GitHub

How to install mlir?

mohitmishra786/low-level-dev-skills/mlir

$npx -y skills add mohitmishra786/low-level-dev-skills --skill mlir

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# MLIR
2
3	## Purpose
4
5	Guide agents through MLIR (Multi-Level IR): ops, regions, blocks, and values; built-in dialects (arith, func, memref, affine, linalg); writing custom dialects with ODS; lowering passes with `ConversionPattern`; `mlir-opt` CLI; and ML compiler use cases (Torch-MLIR, IREE).
6
7	## When to Use
8
9	- Building a domain-specific compiler IR (graphics, ML, hardware DSL)
10	- Lowering high-level ops to LLVM or GPU dialects
11	- Writing progressive lowering pipelines (linalg → loops → LLVM)
12	- Integrating with IREE or Torch-MLIR for ML deployment
13	- Creating reusable transformation passes across dialects
14	- Prototyping compiler optimizations at the right abstraction level
15
16	## Workflow
17
18	### 1. MLIR structure
19
20	```
21	Module
22	└── func.func @main()
23	└── region
24	└── block ^bb0:
25	└── operations (ops) producing SSA values
26	```
27
28	Key concepts:
29	- Operation — instruction-like node (`arith.addi`, `memref.load`)
30	- Region — container of blocks (functions, control flow)
31	- Block — CFG node with ordered ops
32	- Value — SSA result of an op or block argument
33
34	### 2. Built-in dialects
35
36	\| Dialect \| Purpose \|
37	\|---------\|---------\|
38	\| `arith` \| Integer/float arithmetic \|
39	\| `func` \| Function definitions and calls \|
40	\| `memref` \| Buffer abstraction with shapes/strides \|
41	\| `affine` \| Affine loop nests, map/set constraints \|
42	\| `linalg` \| Structured linear algebra ops \|
43	\| `scf` \| Structured control flow (for, if) \|
44	\| `llvm` \| LLVM IR dialect for final lowering \|
45	\| `gpu` \| GPU kernel launches \|
46
47	```mlir
48	// example.mlir
49	func.func @add(%a: memref<4xf32>, %b: memref<4xf32>, %c: memref<4xf32>) {
50	%c0 = arith.constant 0 : index
51	%c4 = arith.constant 4 : index
52	scf.for %i = %c0 to %c4 step %c1 {
53	%av = memref.load %a[%i] : memref<4xf32>
54	%bv = memref.load %b[%i] : memref<4xf32>
55	%sum = arith.addf %av, %bv : f32
56	memref.store %sum, %c[%i] : memref<4xf32>
57	}
58	return
59	}
60	```
61
62	### 3. mlir-opt CLI
63
64	```bash
65	# Parse and print
66	mlir-opt example.mlir
67
68	# Run canonicalization
69	mlir-opt example.mlir -canonicalize
70
71	# Lower affine to scf
72	mlir-opt affine.mlir -lower-affine
73
74	# Full pipeline toward LLVM
75	mlir-opt input.mlir \
76	--linalg-bufferize \
77	--convert-linalg-to-loops \
78	--convert-scf-to-cf \
79	--convert-arith-to-llvm \
80	--convert-memref-to-llvm \
81	--convert-func-to-llvm \
82	-o llvm.mlir
83	```
84
85	### 4. ODS — Operation Definition Specification
86
87	```tablegen
88	// MyOps.td
89	include "mlir/IR/OpBase.td"
90
91	def My_Dialect : Dialect {
92	let name = "my";
93	let summary = "My custom dialect";
94	}
95
96	class My_Op<string mnemonic, list<Trait> traits = []> :
97	Op<My_Dialect, mnemonic, traits>;
98
99	def AddOp : My_Op<"add", [Pure]> {
100	let summary = "Add two values";
101	let arguments = (ins AnyType:$lhs, AnyType:$rhs);
102	let results = (outs AnyType:$result);
103	let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
104	}
105	```
106
107	```bash
108	# Generate C++ from TableGen
109	mlir-tblgen -gen-op-defs MyOps.td -I include/ -o MyOps.cpp.inc
110	```
111
112	### 5. Custom dialect C++ implementation
113
114	```cpp
115	#include "mlir/IR/DialectImplementation.h"
116	#include "MyDialect.h"
117
118	#include "MyOps.cpp.inc"
119
120	void MyDialect::initialize() {
121	addOperations<
122	#define GET_OP_LIST
123	#include "MyOps.cpp.inc"
124	>();
125	}
126
127	#define GET_OP_CLASSES
128	#include "MyOps.cpp.inc"
129	```
130
131	### 6. Lowering passes
132
133	```cpp
134	#include "mlir/Conversion/LLVMCommon/ConversionTarget.h"
135	#include "mlir/Transforms/DialectConversion.h"
136
137	struct AddOpLowering : OpConversionPattern<my::AddOp> {
138	using OpConversionPattern::OpConversionPattern;
139
140	LogicalResult matchAndRewrite(my::AddOp op, OpAdaptor adaptor,
141	ConversionPatternRewriter &rewriter) const override {
142	rewriter.replaceOpWithNewOp<arith::AddIOp>(op, adaptor.getLhs(), adaptor.getRhs());
143	return success();
144	}
145	};
146
147	void populateLoweringPatterns(RewritePatternSet &patterns) {
148	patterns.add<AddOpLowering>(patterns.getContext());
149	}
150
151	// In pass:
152	mlir::ConversionTarget target(*context);
153	target.addIllegalDialect<my::MyDialect>();
154	target.addLegalDialect<arith::ArithDialect>();
155
156	if (failed(applyPartialConversion(module, target, std::move(patterns))))
157	signalPassFailure();
158	```
159
160	### 7. linalg for ML compilers
161
162	```mlir
163	%0 = linalg.matmul ins(%A, %B : tensor<128x256xf32>, tensor<256x64xf32>)
164	outs(%C : tensor<128x64xf32>) -> tensor<128x64xf32>
165	```
166
167	Lowering path: `linalg` → `scf` loops → `affine` → `llvm`
168
169	### 8. Torch-MLIR and IREE
170
171	```bash
172	# Torch-MLIR: PyTorch → MLIR
173	python -m torch_mlir.tools.import-onnx --onnx-model model.onnx -o model.mlir
174
175	# IREE: MLIR → GPU/CPU executable
176	iree-compile --iree-hal-target-backends=llvm-cpu model.mlir -o model.vmfb
177	iree-run

People who liked this also viewed

6 picks

Type