[onnx] Lowerings from onnx.selu (#2634)

Lowerings for `selu` lowerings for ONNX to the corresponding torch implementations. Torch's `selu` implementation has fewer features so we use the a generalized `elu` with the input scale set to `1.0`.
llvm · Dec 14, 2023 · 4857606 · 4857606
1 parent 6ddeb1a
commit 4857606
Show file tree

Hide file tree

Showing 3 changed files with 65 additions and 1 deletion.
diff --git a/include/torch-mlir/Conversion/TorchOnnxToTorch/Patterns.h b/include/torch-mlir/Conversion/TorchOnnxToTorch/Patterns.h
@@ -113,6 +113,25 @@ struct OpBinder {
     return failure();
   }
 
+  ParseResult f32FloatAttr(float &value, StringRef nameSuffix,
+                           float defaultValue = 0.0f) {
+    SmallString<64> name("torch.onnx.");
+    name.append(nameSuffix);
+    auto attr = op->getAttr(name);
+    if (!attr) {
+      value = defaultValue;
+      return success();
+    }
+    if (auto floatAttr = dyn_cast<FloatAttr>(attr)) {
+      FloatType t = cast<FloatType>(floatAttr.getType());
+      if (t.getWidth() != 32)
+        return failure();
+      value = floatAttr.getValueAsDouble();
+      return success();
+    }
+    return failure();
+  }
+
   ParseResult customOpNameStringAttr(std::string &value, StringRef nameSuffix,
                              std::string defaultValue = "") {
     SmallString<64> name("torch.onnx.");

diff --git a/lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp b/lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp
@@ -26,4 +26,33 @@ using namespace mlir::torch::onnx_c;
 // results in a lot of ONNX test cases that all reduce to the exact same
 // thing here, so we simplify.
 void mlir::torch::onnx_c::populateDefaultDomainQtoZ(
-    OnnxCustomOpConversionPattern &patterns) {}
+    OnnxCustomOpConversionPattern &patterns) {
+
+  patterns.onOp(
+      "Selu", 6, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
+        Torch::ValueTensorType resultType;
+        float alpha, gamma;
+        Value operand;
+        if (binder.tensorOperand(operand) ||
+            binder.f32FloatAttr(alpha, "alpha") ||
+            binder.f32FloatAttr(gamma, "gamma") ||
+            binder.tensorResultType(resultType))
+          return failure();
+
+        Value vAlpha = rewriter.create<Torch::ConstantFloatOp>(
+            binder.getLoc(), rewriter.getType<Torch::FloatType>(),
+            rewriter.getFloatAttr(rewriter.getF64Type(), alpha));
+
+        Value vScale = rewriter.create<Torch::ConstantFloatOp>(
+            binder.getLoc(), rewriter.getType<Torch::FloatType>(),
+            rewriter.getFloatAttr(rewriter.getF64Type(), gamma));
+
+        Value vInputScale = rewriter.create<Torch::ConstantFloatOp>(
+            binder.getLoc(), rewriter.getType<Torch::FloatType>(),
+            rewriter.getFloatAttr(rewriter.getF64Type(), 1.0));
+
+        rewriter.replaceOpWithNewOp<Torch::AtenEluOp>(
+            binder.op, resultType, operand, vAlpha, vScale, vInputScale);
+        return success();
+      });
+}
diff --git a/test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir b/test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir
@@ -0,0 +1,16 @@
+// RUN: torch-mlir-opt <%s --split-input-file -convert-torch-onnx-to-torch | FileCheck %s
+// Generally, the test cases accumulated here come from running the importer
+// over all included backend tests that involve simple ops with no model
+// level constants. This is a pragmatic choice which lets us have a lot
+// of tests in this file, whereas the others tend to be more bespoke.
+
+
+// CHECK-LABEL: func.func @test_selu
+func.func @test_selu(%arg0: !torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32> attributes {torch.onnx_meta.opset_version = 6 : si64} {
+  // CHECK-DAG: %[[F1:.+]] = torch.constant.float 1
+  // CHECK-DAG: %[[F2:.+]] = torch.constant.float 2
+  // CHECK-DAG: %[[F3:.+]] = torch.constant.float 3
+  // CHECK: %[[ELU:.+]] = torch.aten.elu %arg0, %[[F2]], %[[F3]], %[[F1]]
+  %0 = torch.operator "onnx.Selu"(%arg0) {torch.onnx.alpha = 2.000000e+00 : f32, torch.onnx.gamma = 3.000000e+00 : f32} : (!torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32>
+  return %0 : !torch.vtensor<[3,4,5],f32>
+}