Fixed GRU quality issues exposed by e2e tests (#3753)

Issue: nod-ai/SHARK-ModelDev#856

Related tests:
![Screenshot 2024-10-01
175305](https://github.com/user-attachments/assets/0dc0901b-058f-427c-a596-9e806fd38836)

lib/Conversion/TorchOnnxToTorch/OnnxRecurrentLayerOpExpanders.cpp

@@ -1072,11 +1072,10 @@ LogicalResult OnnxGruExpander(OpBinder binder,
   Value cstNone = b.create<ConstantNoneOp>();
   Value cstZero = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(0));
   Value cstOne = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(1));
-  Value cstTwo = b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2));
 
   // Binding arguments
   ValueTensorType yTy, Y_hType;
-  if (binder.tensorResultTypeAtIndex(yTy, 0) ||
+  if (binder.tensorResultTypeAtIndex(yTy, 0) &&
       binder.tensorResultTypeAtIndex(Y_hType, 1)) {
     return rewriter.notifyMatchFailure(binder.op,
                                        "At least one output must be present");
@@ -1132,6 +1131,7 @@ LogicalResult OnnxGruExpander(OpBinder binder,
   // Validations
   auto XShape = xTy.getSizes();
   int64_t batch_size = (layout == 0) ? XShape[1] : XShape[0];
+  int64_t seq_len = (layout == 0) ? XShape[0] : XShape[1];
   int64_t input_size = XShape[2];
 
   std::ostringstream oss;
@@ -1173,6 +1173,10 @@ LogicalResult OnnxGruExpander(OpBinder binder,
     Value cstDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
     initial_h =
         b.create<AtenZerosOp>(hTy, hShape, cstDtype, cstNone, cstNone, cstNone);
+  } else {
+    if (layout == 1) {
+      initial_h = StaticTranspose(b, initial_h, 0, 1);
+    }
   }
 
   if (binder.tensorOperandAtIndex(sequence_lens, 4))
@@ -1192,10 +1196,10 @@ LogicalResult OnnxGruExpander(OpBinder binder,
   // fill in B
   Value cstXDtype = getDtypeIntValueForType(rewriter, loc, xTy.getDtype());
   if (B == nullptr) {
-    SmallVector<int64_t> BShape = {num_directions, 2 * hidden_size};
+    SmallVector<int64_t> BShape = {num_directions, 6 * hidden_size};
     SmallVector<Value> BShapeListContents = {
         b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(num_directions)),
-        b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(2 * hidden_size))};
+        b.create<ConstantIntOp>(intType, b.getI64IntegerAttr(6 * hidden_size))};
     Value BShapeList = b.create<PrimListConstructOp>(
         b.getType<ListType>(intType), BShapeListContents);
     auto BType = b.getType<ValueTensorType>(BShape, wTy.getDtype());
@@ -1256,51 +1260,47 @@ LogicalResult OnnxGruExpander(OpBinder binder,
                       B_slices[4], B_slices[5]);
 
   // Process inputs based on layout
-  Value X_processed, initial_h_processed;
-  ValueTensorType yTy_processed, Y_hType_processed;
-
-  if (layout == 0) {
-    X_processed = X;
-    initial_h_processed = initial_h_forward;
-    yTy_processed = yTy;
-    Y_hType_processed = Y_hType;
-  } else {
-    X_processed = b.create<AtenTransposeIntOp>(X.getType(), X, cstZero, cstOne);
-    initial_h_processed = b.create<AtenTransposeIntOp>(
-        initial_h.getType(), initial_h_forward, cstZero, cstOne);
-
-    auto yTySizes = yTy.getSizes();
-    auto Y_hTypeSizes = Y_hType.getSizes();
-
-    yTy_processed = b.getType<ValueTensorType>(
-        llvm::SmallVector<int64_t>{yTySizes[1], yTySizes[0], yTySizes[2],
-                                   yTySizes[3]},
-        yTy.getDtype());
-
-    Y_hType_processed = b.getType<ValueTensorType>(
-        llvm::SmallVector<int64_t>{Y_hTypeSizes[1], Y_hTypeSizes[0],
-                                   Y_hTypeSizes[2]},
-        Y_hType.getDtype());
+  if (layout == 1) {
+    X = StaticTranspose(b, X, 0, 1);
   }
 
   // Weights and biases ready. Calling GRU layer to insert the actual ops.
-  GruLayerOutput gruLayerOutput =
-      gru_layer(b, X_processed, initial_h_processed, weights, activations,
-                linear_before_reset);
+  GruLayerOutput gruLayerOutput = gru_layer(b, X, initial_h_forward, weights,
+                                            activations, linear_before_reset);
 
   // Process outputs based on layout
-  Value Y_final, Y_h_final;
-  if (layout == 0) {
-    Y_final = b.create<AtenUnsqueezeOp>(yTy, gruLayerOutput.Y, cstOne);
-    Y_h_final = b.create<AtenUnsqueezeOp>(Y_hType, gruLayerOutput.Y_h, cstZero);
+  Value Y_final;
+  if (binder.tensorResultTypeAtIndex(yTy, 0)) {
+    Y_final = cstNone;
   } else {
-    auto Y_transposed = b.create<AtenTransposeIntOp>(
-        gruLayerOutput.Y.getType(), gruLayerOutput.Y, cstZero, cstOne);
-    Y_final = b.create<AtenUnsqueezeOp>(yTy, Y_transposed, cstTwo);
+    if (layout == 0) {
+      Y_final = b.create<AtenUnsqueezeOp>(yTy, gruLayerOutput.Y, cstOne);
+    } else {
+      Type yTy_original = b.getType<ValueTensorType>(
+          llvm::SmallVector<int64_t>{seq_len, 1, batch_size, hidden_size},
+          yTy.getDtype());
+      Y_final =
+          b.create<AtenUnsqueezeOp>(yTy_original, gruLayerOutput.Y, cstOne);
+      Y_final = StaticTranspose(b, Y_final, 1, 2);
+      Y_final = StaticTranspose(b, Y_final, 0, 1);
+    }
+  }
 
-    auto Y_h_transposed = b.create<AtenTransposeIntOp>(
-        gruLayerOutput.Y_h.getType(), gruLayerOutput.Y_h, cstZero, cstOne);
-    Y_h_final = b.create<AtenUnsqueezeOp>(Y_hType, Y_h_transposed, cstZero);
+  Value Y_h_final;
+  if (binder.tensorResultTypeAtIndex(Y_hType, 1)) {
+    Y_h_final = cstNone;
+  } else {
+    if (layout == 0) {
+      Y_h_final =
+          b.create<AtenUnsqueezeOp>(Y_hType, gruLayerOutput.Y_h, cstZero);
+    } else {
+      Type y_hTy_original = b.getType<ValueTensorType>(
+          llvm::SmallVector<int64_t>{1, batch_size, hidden_size},
+          Y_hType.getDtype());
+      Y_h_final = b.create<AtenUnsqueezeOp>(y_hTy_original, gruLayerOutput.Y_h,
+                                            cstZero);
+      Y_h_final = StaticTranspose(b, Y_h_final, 0, 1);
+    }
   }
 
   rewriter.replaceOp(binder.op, mlir::ValueRange{Y_final, Y_h_final});