Merge pull request #413 from Xilinx/chaitany.backport_simplify_convtranspose_decomposition

Back port of simplify convtranspose decomposition

Author: chaitanyakamarapu

src/Dialect/ONNX/Transforms/Decompose.cpp

@@ -851,6 +851,10 @@ bool ShouldDecomposeConvTransposeOpToPhasedConvs(Value convTransposeResult,
   bool fourPhaseDecomposition = (stridesShape[0] == 2);
   bool ninePhaseDecomposition = (stridesShape[0] == 3);
   if (fourPhaseDecomposition) {
+    if (outputShape[0] != 1) {
+      // Currently support batch=1
+      return false;
+    }
     if (kernelShape[0] == 6 && padsShape[0] == 2 &&
         llvm::all_equal(padsShape)) {
       // Currently support only with pads [2, 2, 2, 2]
@@ -1713,102 +1717,56 @@ Value decomposeIntoPhasedConvs(PatternRewriter &rewriter, Location loc,
           startOnnxConstant, endOnnxConstant, axisOnnxConstant,
           stepOnnxConstant);
     }
-
-    // The four convOutputs are adjusted to add an extra dimension at the
-    // innermost level.
-    SmallVector<int64_t> outputShapePlusOneDim(convOutputShape);
-    outputShapePlusOneDim.push_back(1);
-    auto onnxConstForReshapeAddOneDim =
-        getONNXConstOpFromVector(rewriter, loc, outputShapePlusOneDim);
-
-    auto reshapeOutputType =
-        RankedTensorType::get(outputShapePlusOneDim, elementType);
-
-    auto reshapeOutputAddOneDimConv1 = rewriter.create<ONNXReshapeOp>(
-        loc, reshapeOutputType, conv1, onnxConstForReshapeAddOneDim);
-    auto reshapeOutputAddOneDimConv2 = rewriter.create<ONNXReshapeOp>(
-        loc, reshapeOutputType, conv2, onnxConstForReshapeAddOneDim);
-    auto reshapeOutputAddOneDimConv3 = rewriter.create<ONNXReshapeOp>(
-        loc, reshapeOutputType, conv3, onnxConstForReshapeAddOneDim);
-    auto reshapeOutputAddOneDimConv4 = rewriter.create<ONNXReshapeOp>(
-        loc, reshapeOutputType, conv4, onnxConstForReshapeAddOneDim);
-
-    SmallVector<int64_t> outputShapeLevel1Concat(outputShapePlusOneDim);
-    outputShapeLevel1Concat[outputShapeLevel1Concat.size() - 1] = 2;
-    auto level1ConcatOutputType =
-        RankedTensorType::get(outputShapeLevel1Concat, elementType);
+    // Four conv outputs are merged in channel dim
+    SmallVector<int64_t> outputShapeOfConcat = {
+        1, convOutputShape[1] * 4, convOutputShape[2], convOutputShape[3]};
+    auto concatOutputType =
+        RankedTensorType::get(outputShapeOfConcat, elementType);
     // for the case where convtranspose kernel is [4, 4] and with pads [1, 1, 1,
     // 1] The phased convs output are to be concatenated in the reverse order.
     // This is observed by looking at the phased conv outputs with respect to
     // convtranspose output.
     bool reverseConcatOrder = (needWeightsPadding || (kernelShape[0] == 4));
-    // Below concats result will have the innermost dim as 2.
+    // The concat output will have 4 times the channels of a single conv.
     auto firstConcat =
         (reverseConcatOrder)
-            ? rewriter.create<ONNXConcatOp>(loc, level1ConcatOutputType,
-                  ValueRange{
-                      reshapeOutputAddOneDimConv3, reshapeOutputAddOneDimConv1},
-                  -1)
-            : rewriter.create<ONNXConcatOp>(loc, level1ConcatOutputType,
-                  ValueRange{
-                      reshapeOutputAddOneDimConv1, reshapeOutputAddOneDimConv3},
-                  -1);
-    auto secondConcat =
-        (reverseConcatOrder)
-            ? rewriter.create<ONNXConcatOp>(loc, level1ConcatOutputType,
-                  ValueRange{
-                      reshapeOutputAddOneDimConv2, reshapeOutputAddOneDimConv4},
-                  -1)
-            : rewriter.create<ONNXConcatOp>(loc, level1ConcatOutputType,
-                  ValueRange{
-                      reshapeOutputAddOneDimConv4, reshapeOutputAddOneDimConv2},
-                  -1);
+            ? rewriter.create<ONNXConcatOp>(loc, concatOutputType,
+                  ValueRange{conv2, conv4, conv3, conv1}, 1)
+            : rewriter.create<ONNXConcatOp>(loc, concatOutputType,
+                  ValueRange{conv1, conv3, conv4, conv2}, 1);
 
-    // Reshaping to modify the two innermost levels,ensuring the second
-    // innermost level is set to 1
-    SmallVector<int64_t> outputShapeForDimAdjust(convOutputShape);
-    auto dimValueAtLastIndex = convOutputShape[convOutputShape.size() - 1] * 2;
-    outputShapeForDimAdjust[outputShapeForDimAdjust.size() - 1] = 1;
-    outputShapeForDimAdjust.push_back(dimValueAtLastIndex);
+    // Here we are reshaping the concatenated conv channels of 4*Conv_channels
+    // into groups of 2x2 channels. This can be visualized as
+    // H_chan(2) * W_Chan(2) * C_real,  then doing the transpose into
+    // Conv_channels H H_chan W W_chan. The adjecent H and H_chan will be merged
+    // into H, same way W and W_chan will be merged into W. This leads to
+    // doubling of the H and W. Keeping the channels same.
+
+    SmallVector<int64_t> outputShapeForDimAdjust = {
+        2, 2, convOutputShape[1], convOutputShape[2], convOutputShape[3]};
 
     auto onnxConstForReshapeDimAdjust =
         getONNXConstOpFromVector(rewriter, loc, outputShapeForDimAdjust);
 
     auto reshapeOutputForDimAdjustType =
         RankedTensorType::get(outputShapeForDimAdjust, elementType);
-    auto reshapeOutputDimAdjustOfFirstConcat =
+    auto reshapeOutputDimAdjust =
         rewriter.create<ONNXReshapeOp>(loc, reshapeOutputForDimAdjustType,
             firstConcat, onnxConstForReshapeDimAdjust);
-    auto reshapeOutputDimAdjustOfSecondConcat =
-        rewriter.create<ONNXReshapeOp>(loc, reshapeOutputForDimAdjustType,
-            secondConcat, onnxConstForReshapeDimAdjust);
 
-    SmallVector<int64_t> outputShapeForFinalConcat(outputShapeForDimAdjust);
-    outputShapeForFinalConcat[outputShapeForFinalConcat.size() - 2] = 2;
+    SmallVector<int64_t> transposeOuputShape = {
+        convOutputShape[1], convOutputShape[2], 2, convOutputShape[3], 2};
 
-    auto finalConcatOutputType =
-        RankedTensorType::get(outputShapeForFinalConcat, elementType);
+    auto transposeOutputType =
+        RankedTensorType::get(transposeOuputShape, elementType);
 
-    // Final Concat is performed on the two reshaped outputs at the
-    // second innermost level
-    auto finalConcat =
-        (reverseConcatOrder)
-            ? rewriter.create<ONNXConcatOp>(loc, finalConcatOutputType,
-                  ValueRange{reshapeOutputDimAdjustOfSecondConcat,
-                      reshapeOutputDimAdjustOfFirstConcat},
-                  -2)
-            : rewriter.create<ONNXConcatOp>(loc, finalConcatOutputType,
-                  ValueRange{reshapeOutputDimAdjustOfFirstConcat,
-                      reshapeOutputDimAdjustOfSecondConcat},
-                  -2);
+    auto permArrayAttr = rewriter.getI64ArrayAttr({2, 3, 0, 4, 1});
 
-    SmallVector<int64_t> outputShapeForResult(convOutputShape);
-    dimValueAtLastIndex = convOutputShape[convOutputShape.size() - 1] * 2;
-    auto dimValueAtSecondLastIndex =
-        convOutputShape[convOutputShape.size() - 2] * 2;
-    outputShapeForResult[outputShapeForResult.size() - 2] =
-        dimValueAtSecondLastIndex;
-    outputShapeForResult[outputShapeForResult.size() - 1] = dimValueAtLastIndex;
+    auto transpose = rewriter.create<ONNXTransposeOp>(
+        loc, transposeOutputType, reshapeOutputDimAdjust, permArrayAttr);
+
+    SmallVector<int64_t> outputShapeForResult = {
+        1, convOutputShape[1], convOutputShape[2] * 2, convOutputShape[3] * 2};
 
     auto onnxConstForLastReshape =
         getONNXConstOpFromVector(rewriter, loc, outputShapeForResult);
@@ -1818,7 +1776,7 @@ Value decomposeIntoPhasedConvs(PatternRewriter &rewriter, Location loc,
     // Result is reshaped back to match the original convtranspose output
     // dimensions
     auto finalOutput = rewriter.create<ONNXReshapeOp>(
-        loc, finalOutputType, finalConcat, onnxConstForLastReshape);
+        loc, finalOutputType, transpose, onnxConstForLastReshape);
     return finalOutput;
   }
   if (numPhases == 9) {