Visualizing MLIR codes to generate DFGs through graphviz

test/visualize/README.md

@@ -0,0 +1,19 @@
+# MLIR Code Visualization Guide
+
+This guide explains how to visualize MLIR (Multi-Level Intermediate Representation) code using various tools and techniques.
+
+## Usage
+
+First, run `apt install graphviz` to install Graphviz. For an MLIR file, use the option `--view-op-graph` to generate a Graphviz visualization of a function. Then, use the `dot` command to create an image file from the visualization.
+
+Example:
+
+```bash
+mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns test.mlir > test_fused.mlir
+mlir-opt --allow-unregistered-dialect --view-op-graph test_fused.mlir 2> test.dot 
+dot -Tpng test.dot -o test.png
+```
+
+It will generate a `test.png` file in the current directory, as shown below:
+
+![test](test.png)

test/visualize/test.mlir

@@ -0,0 +1,105 @@
+// Test cases for FusePatternsPass
+// Tests both FAddFAdd and FMulFAdd fusion patterns
+
+// Test 1: Basic FAddFAdd fusion (LHS fadd)
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-FADD-FADD-LHS
+
+func.func @test_fadd_fadd_lhs(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.addf %a, %b : f32
+  %res = arith.addf %temp, %c : f32
+  // CHECK-FADD-FADD-LHS: neur1ielZH^1a.fadd_fadd
+  return %res : f32
+}
+
+// Test 2: Basic FAddFAdd fusion (RHS fadd)
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-FADD-FADD-RHS
+
+func.func @test_fadd_fadd_rhs(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.addf %b, %c : f32
+  %res = arith.addf %a, %temp : f32
+  // CHECK-FADD-FADD-RHS: neura.fadd_fadd
+  return %res : f32
+}
+
+// Test 3: Basic FMulFAdd fusion (LHS fmul)
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-FMUL-FADD-LHS
+
+func.func @test_fmul_fadd_lhs(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.mulf %a, %b : f32
+  %res = arith.addf %temp, %c : f32
+  // CHECK-FMUL-FADD-LHS: neura.fmul_fadd
+  return %res : f32
+}
+
+// Test 4: Basic FMulFAdd fusion (RHS fmul)
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-FMUL-FADD-RHS
+
+func.func @test_fmul_fadd_rhs(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.mulf %b, %c : f32
+  %res = arith.addf %a, %temp : f32
+  // CHECK-FMUL-FADD-RHS: neura.fmul_fadd
+  return %res : f32
+}
+
+// Test 5: No fusion when first fadd has multiple uses
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-NO-FUSION-MULTI-USE
+
+func.func @test_no_fusion_multi_use(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.addf %a, %b : f32
+  %res1 = arith.addf %temp, %c : f32
+  %res2 = arith.addf %temp, %c : f32  // Second use of %temp
+  // CHECK-NO-FUSION-MULTI-USE-NOT: neura.fadd_fadd
+  // CHECK-NO-FUSION-MULTI-USE: neura.fadd
+  return %res1 : f32
+}
+
+// Test 6: No fusion when first fmul has multiple uses
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-NO-FMUL-MULTI-USE
+
+func.func @test_no_fmul_fusion_multi_use(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp = arith.mulf %a, %b : f32
+  %res1 = arith.addf %temp, %c : f32
+  %res2 = arith.addf %temp, %c : f32  // Second use of %temp
+  // CHECK-NO-FMUL-MULTI-USE-NOT: neura.fmul_fadd
+  // CHECK-NO-FMUL-MULTI-USE: neura.fmul
+  return %res1 : f32
+}
+
+// Test 7: Complex chain with multiple fusion opportunities
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-COMPLEX-CHAIN
+
+func.func @test_complex_chain(%a: f32, %b: f32, %c: f32, %d: f32) -> f32 {
+  %temp1 = arith.addf %a, %b : f32
+  %temp2 = arith.addf %temp1, %c : f32
+  %temp3 = arith.mulf %temp2, %d : f32
+  %res = arith.addf %temp3, %a : f32
+  // CHECK-COMPLEX-CHAIN: neura.fadd_fadd
+  // CHECK-COMPLEX-CHAIN: neura.fmul_fadd
+  return %res : f32
+}
+
+// Test 8: No fusion when operands are not the expected operations
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-NO-FUSION-WRONG-OPS
+
+func.func @test_no_fusion_wrong_ops(%a: f32, %b: f32, %c: f32) -> f32 {
+  %temp1 = arith.subf %a, %b : f32  // Using subf instead of addf
+  %res = arith.addf %temp1, %c : f32
+  // CHECK-NO-FUSION-WRONG-OPS-NOT: neura.fadd_fadd
+  // CHECK-NO-FUSION-WRONG-OPS: neura.fsub
+  // CHECK-NO-FUSION-WRONG-OPS: neura.fadd
+  return %res : f32
+}
+
+// Test 9: Multiple fusion patterns in sequence
+// RUN: mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns %s | FileCheck %s --check-prefix=CHECK-MULTIPLE-FUSION
+
+func.func @test_multiple_fusion(%a: f32, %b: f32, %c: f32, %d: f32) -> f32 {
+  %temp1 = arith.addf %a, %b : f32
+  %temp2 = arith.addf %temp1, %c : f32
+  %temp3 = arith.mulf %temp2, %d : f32
+  %res = arith.addf %temp3, %a : f32
+  // CHECK-MULTIPLE-FUSION: neura.fadd_fadd
+  // CHECK-MULTIPLE-FUSION: neura.fmul_fadd
+  return %res : f32
+}
+

test/visualize/visualize.sh

@@ -0,0 +1,3 @@
+mlir-neura-opt --assign-accelerator --lower-arith-to-neura --fuse-patterns test.mlir > test_fused.mlir
+mlir-opt --allow-unregistered-dialect --view-op-graph test_fused.mlir 2> test.dot 
+dot -Tpng test.dot -o test.png