[TOSA] Defer input transpose until guards pass

Lazily create the NHWC input transpose so we emit it only once the failure
guards in the transposed and depthwise convolution rewrite succeed.

Change-Id: Ia362deda898794397107f6da3c44cd89f219f58f

Author: Lallapallooza

lib/Conversion/TorchToTosa/TorchToTosa.cpp

@@ -2335,6 +2335,8 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
   auto bias = adaptor.getBias();
 
   if (isa<Torch::NoneType>(bias.getType())) {
+    // ConvTranspose weights use IOHW; the helper expects OIHW, so swap
+    // dims 0/1 before we synthesize the bias.
     SmallVector<int64_t, 4> biasWeightShape =
         transposed ? SmallVector<int64_t, 4>{weightShape[1], weightShape[0],
                                              weightShape[2], weightShape[3]}
@@ -2405,13 +2407,13 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
     transposedInputShape.push_back(inputShape[dim]);
   auto transposedInputType = RankedTensorType::get(
       makeShapeLLVMCompatible(transposedInputShape), inputElemTy);
-  auto transposedInput =
-      rewriter
-          .create<tosa::TransposeOp>(
-              op->getLoc(),
-              getTypeConverter()->convertType(transposedInputType), input,
-              rewriter.getDenseI32ArrayAttr(nchwToNhwcDims))
-          .getResult();
+  auto createTransposedInput = [&]() {
+    return rewriter
+        .create<tosa::TransposeOp>(
+            op->getLoc(), getTypeConverter()->convertType(transposedInputType),
+            input, rewriter.getDenseI32ArrayAttr(nchwToNhwcDims))
+        .getResult();
+  };
 
   if (transposed) {
     if (groups != 1)
@@ -2424,17 +2426,6 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
               "TOSA");
 
     SmallVector<int32_t> iohwToOhwi({1, 2, 3, 0});
-    SmallVector<int64_t, 4> ohwiWeightShape;
-    for (int32_t dim : iohwToOhwi)
-      ohwiWeightShape.push_back(weightShape[dim]);
-    auto ohwiWeightType = RankedTensorType::get(
-        makeShapeLLVMCompatible(ohwiWeightShape), weightElemTy);
-    Value transformedWeight =
-        rewriter
-            .create<tosa::TransposeOp>(
-                op->getLoc(), getTypeConverter()->convertType(ohwiWeightType),
-                weight, rewriter.getDenseI32ArrayAttr(iohwToOhwi))
-            .getResult();
 
     // TOSA 'out_pad' is a 4D array {top,bottom,left,right}.
     // Map from PyTorch's (padding, output_padding):
@@ -2456,6 +2447,19 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
     SmallVector<int64_t, 4> outPad(
         {outPadTop, outPadBottom, outPadLeft, outPadRight});
 
+    Value nhwcInput = createTransposedInput();
+    SmallVector<int64_t, 4> ohwiWeightShape;
+    for (int32_t dim : iohwToOhwi)
+      ohwiWeightShape.push_back(weightShape[dim]);
+    auto ohwiWeightType = RankedTensorType::get(
+        makeShapeLLVMCompatible(ohwiWeightShape), weightElemTy);
+    Value transformedWeight =
+        rewriter
+            .create<tosa::TransposeOp>(
+                op->getLoc(), getTypeConverter()->convertType(ohwiWeightType),
+                weight, rewriter.getDenseI32ArrayAttr(iohwToOhwi))
+            .getResult();
+
     // Result type is NHWC (we'll transpose back).
     auto outNCHW = makeShapeTorchCompatible(outputTy.getShape());
     SmallVector<int64_t, 4> outNHWC;
@@ -2479,7 +2483,7 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
         rewriter
             .create<tosa::TransposeConv2DOp>(
                 op->getLoc(), getTypeConverter()->convertType(transConvOpTy),
-                transposedInput, transformedWeight, bias, inputZp, weightZp,
+                nhwcInput, transformedWeight, bias, inputZp, weightZp,
                 rewriter.getDenseI64ArrayAttr(outPad),
                 rewriter.getDenseI64ArrayAttr(stride), accType)
             .getResult();
@@ -2535,6 +2539,15 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
     SmallVector<int32_t> transposedDims({2, 3, 0, 1});
     SmallVector<int64_t> transposedWeightShape = {
         weightShape[2], weightShape[3], weightShape[0], weightShape[1]};
+
+    // reshape: HWO(I/G) -> HWIM
+    outputCDim = makeShapeTorchCompatible(outputTy.getShape())[1];
+    if (outputCDim == kUnknownSize) {
+      return rewriter.notifyMatchFailure(
+          op, "number of output channels must be statically known for "
+              "depthwise convolutions");
+    }
+
     auto transposedWeightType = RankedTensorType::get(
         makeShapeLLVMCompatible(transposedWeightShape), weightElemTy);
     auto transposedWeight =
@@ -2545,13 +2558,6 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
                 rewriter.getDenseI32ArrayAttr(transposedDims))
             .getResult();
 
-    // reshape: HWO(I/G) -> HWIM
-    outputCDim = makeShapeTorchCompatible(outputTy.getShape())[1];
-    if (outputCDim == kUnknownSize) {
-      return rewriter.notifyMatchFailure(
-          op, "number of output channels must be statically known for "
-              "depthwise convolutions");
-    }
     transformedWeightShape = {
         transposedWeightShape[0],
         transposedWeightShape[1],
@@ -2573,6 +2579,8 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
     llvm_unreachable("Unhandled convolution type");
   }
 
+  Value transposedInput = createTransposedInput();
+
   int64_t outputHDim, outputWDim;
   int64_t inputHDim = inputShape[2];
   int64_t inputWDim = inputShape[3];