[ONNX][TORCH-MLIR] LayerNorm (#2716)

Layer Normalization using the torch.aten.native_layer_norm nod-ai/SHARK-ModelDev#325
llvm · Jan 11, 2024 · 5862854 · 5862854
1 parent 0860c41
commit 5862854
Show file tree

Hide file tree

Showing 2 changed files with 56 additions and 0 deletions.
diff --git a/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp b/lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp
@@ -410,6 +410,49 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP(
         }
         return failure();
       });
+  patterns.onOp("LayerNormalization", 17,
+                [](OpBinder binder, ConversionPatternRewriter &rewriter) {
+                  Torch::ValueTensorType Y_type;
+                  Torch::ValueTensorType Mean_type;
+                  Torch::ValueTensorType InvStdDev_type;
+                  Value X;
+                  Value Scale;
+                  Value B;
+                  int64_t axis;
+		  float epsilon;
+                  int64_t stash_type;
+                  if (binder.tensorOperandAtIndex(X, 0) ||
+                      binder.tensorOperandAtIndex(Scale, 1) ||
+                      binder.tensorOperandAtIndex(B, 2) ||
+                      binder.tensorResultTypeAtIndex(Y_type, 0) || 
+		                  binder.tensorResultTypeAtIndex(Mean_type, 1) ||
+                      binder.tensorResultTypeAtIndex(InvStdDev_type, 2) || 
+		                  binder.s64IntegerAttr(axis, "axis", -1) ||
+                      binder.f32FloatAttr(epsilon, "epsilon", 0.00001) ||
+                      binder.s64IntegerAttr(stash_type, "stash_type", 1)) 
+                    return failure(); 
+                  Value constEpsilon = rewriter.create<Torch::ConstantFloatOp>(
+                    binder.getLoc(), rewriter.getType<Torch::FloatType>(),
+                    rewriter.getF64FloatAttr(epsilon));
+                  unsigned rank = 1;
+                  if(std::optional<unsigned> maybeRank = Torch::getTensorRank(X))
+                    rank = *maybeRank;
+                  SmallVector<Value> normalized;
+                  axis = Torch::toPositiveDim(axis, rank);
+                  auto X_type = X.getType().cast<Torch::ValueTensorType>();
+                  ArrayRef<int64_t> X_shape = X_type.getSizes();
+                  for (int64_t n = axis; n < rank ; n++) {                      
+                    normalized.push_back(rewriter.create<Torch::ConstantIntOp>(
+                    binder.getLoc(), rewriter.getI64IntegerAttr(X_shape[n])));
+                  }
+                  Value normalized_shape = rewriter.create<Torch::PrimListConstructOp>(
+                    binder.getLoc(),
+                    Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
+                    normalized);
+                  rewriter.replaceOpWithNewOp<Torch::AtenNativeLayerNormOp>(
+                      binder.op, Y_type, Mean_type, InvStdDev_type, X, normalized_shape, Scale, B, constEpsilon);
+                  return success();
+              });
   patterns.onOp("LeakyRelu", 1,
                 [](OpBinder binder, ConversionPatternRewriter &rewriter) {
                   Torch::ValueTensorType resultType;

diff --git a/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir b/test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir
@@ -116,6 +116,19 @@ func.func @test_gemm_alpha_beta(%arg0: !torch.vtensor<[3,5],f32>, %arg1: !torch.
 
 // -----
 
+// CHECK-LABEL : func.func @test_layer_norm
+func.func @test_layer_norm(%arg0: !torch.vtensor<[3,4],f32>, %arg1: !torch.vtensor<[3,4],f32>, %arg2: !torch.vtensor<[3,4],f32>) -> (!torch.vtensor<[3,4], f32>, !torch.vtensor<[1,1],f32>, !torch.vtensor<[1,1],f32>) 
+                           attributes {torch.onnx_meta.ir_version = 6 : si64, torch.onnx_meta.opset_version = 17 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} { 
+  // CHECK: %int3 = torch.constant.int 3
+  // CHECK: %int4 = torch.constant.int 4
+  // CHECK: %0 = torch.prim.ListConstruct %int3, %int4 : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %result0, %result1, %result2 = torch.aten.native_layer_norm %arg0, %0, %arg1, %arg2
+  %0:3 = torch.operator "onnx.LayerNormalization"(%arg0, %arg1, %arg2) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[3,4],f32>, !torch.vtensor<[3,4],f32>, !torch.vtensor<[3,4],f32>) -> (!torch.vtensor<[3,4],f32>, !torch.vtensor<[1,1],f32>, !torch.vtensor<[1,1],f32>)
+  return %0#0, %0#1, %0#2 : !torch.vtensor<[3,4],f32>, !torch.vtensor<[1,1],f32>, !torch.vtensor<[1,1],f32>
+}
+
+// -----
+
 // CHECK-LABEL: func.func @test_leaky_relu
 func.func @test_leaky_relu(%arg0: !torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32> attributes {torch.onnx_meta.opset_version = 16 : si64} {
   // CHECK-DAG: %[[F2:.+]] = torch.constant.float 2