Revert some changes to original state

Author: vivekkhandelwal1

lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp

@@ -497,7 +497,6 @@ void mlir::torch::onnx_c::populateDefaultDomainQtoZ(
         if (!isPerChannelQuantization)
           input = makePerTensor(input, inputScale, inputZp);
 
-        // TODO(suderman): insert convolution operator.
         llvm::SmallVector<Value> newOperands = {input, weight};
         if (bias)
           newOperands.push_back(bias);
@@ -524,6 +523,22 @@ void mlir::torch::onnx_c::populateDefaultDomainQtoZ(
                                newAttributes, binder.op->getRegions().size())
                            .getResult(0);
 
+        if (!isPerChannelQuantization) {
+          Value outScale = rewriter.create<Torch::AtenMulFloatOp>(
+              binder.getLoc(), rewriter.getType<Torch::FloatType>(), inputScale,
+              weightScale);
+          Value outZp = rewriter.create<Torch::ConstantIntOp>(
+              binder.getLoc(), rewriter.getType<Torch::IntType>(),
+              rewriter.getIntegerAttr(rewriter.getIntegerType(64), 0));
+          output = rewriter.create<Torch::Aten_MakePerTensorQuantizedTensorOp>(
+              binder.getLoc(), outputTy, output, outScale, outZp);
+          outputTy = rewriter.getType<Torch::ValueTensorType>(
+              resultType.getOptionalSizes(), rewriter.getF32Type());
+
+          output = rewriter.create<Torch::AtenDequantizeSelfOp>(
+              binder.getLoc(), outputTy, output);
+        }
+
         outputTy = getQTorchTypeFromTorchIntType(resultType);
         Value dtyVal = rewriter.create<Torch::ConstantIntOp>(
             binder.getLoc(), rewriter.getType<Torch::IntType>(),

test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir

@@ -90,8 +90,12 @@ func.func @test_qlinearconv_nobias(%arg0: !torch.vtensor<[1,1,7,7],ui8>, %arg1:
   // CHECK: %[[NONE:.+]] = torch.constant.none
   // CHECK: %[[INT1_5:.+]] = torch.constant.int 1
   // CHECK: %[[CONV:.+]] = torch.aten.convolution %[[A]], %[[B]], %[[NONE]], %[[DILATION]], %[[PAD]], %[[KERNEL]], %[[FALSE]], %[[STRIDE]], %[[INT1_5]] : !torch.vtensor<[1,1,7,7],!torch.quint8>, !torch.vtensor<[1,1,1,1],!torch.quint8>, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.qint32>
+  // CHECK: %[[convScale:.+]] = torch.aten.mul.float %[[aScale]], %[[bScale]] : !torch.float, !torch.float -> !torch.float
+  // CHECK: %[[INT0_6:.+]] = torch.constant.int 0
+  // CHECK: %[[C:.+]] = torch.aten._make_per_tensor_quantized_tensor %[[CONV]], %[[convScale]], %[[INT0_6]] : !torch.vtensor<[1,1,7,7],!torch.qint32>, !torch.float, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.qint32>
+  // CHECK: %[[DEQ:.+]] = torch.aten.dequantize.self %[[C]] : !torch.vtensor<[1,1,7,7],!torch.qint32> -> !torch.vtensor<[1,1,7,7],f32>
   // CHECK: %[[INT13:.+]] = torch.constant.int 13
-  // CHECK: %[[QUANT:.+]] = torch.aten.quantize_per_tensor %[[CONV]], %[[cScale]], %[[cZp]], %[[INT13]] : !torch.vtensor<[1,1,7,7],!torch.qint32>, !torch.float, !torch.int, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.quint8>
+  // CHECK: %[[QUANT:.+]] = torch.aten.quantize_per_tensor %[[DEQ]], %[[cScale]], %[[cZp]], %[[INT13]] : !torch.vtensor<[1,1,7,7],f32>, !torch.float, !torch.int, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.quint8>
   // CHECK: %[[INT:.+]] = torch.aten.int_repr %[[QUANT]] : !torch.vtensor<[1,1,7,7],!torch.quint8> -> !torch.vtensor<[1,1,7,7],ui8>
   // CHECK: return %[[INT]] : !torch.vtensor<[1,1,7,7],ui8>
   return %0 : !torch.vtensor<[1,1,7,7],ui8>
@@ -124,8 +128,12 @@ func.func @test_qlinearconv_bias(%arg0: !torch.vtensor<[1,1,7,7],ui8>, %arg1: !t
   // CHECK: %[[FALSE:.+]] = torch.constant.bool false
   // CHECK: %[[INT1_5:.+]] = torch.constant.int 1
   // CHECK: %[[CONV:.+]] = torch.aten.convolution %[[A]], %[[B]], %arg8, %[[DILATION]], %[[PAD]], %[[KERNEL]], %[[FALSE]], %[[STRIDE]], %[[INT1_5]] : !torch.vtensor<[1,1,7,7],!torch.quint8>, !torch.vtensor<[1,1,1,1],!torch.quint8>, !torch.vtensor<[7],si32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.qint32>
+  // CHECK: %[[convScale:.+]] = torch.aten.mul.float %[[aScale]], %[[bScale]] : !torch.float, !torch.float -> !torch.float
+  // CHECK: %[[INT0_6:.+]] = torch.constant.int 0
+  // CHECK: %[[C:.+]] = torch.aten._make_per_tensor_quantized_tensor %[[CONV]], %[[convScale]], %[[INT0_6]] : !torch.vtensor<[1,1,7,7],!torch.qint32>, !torch.float, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.qint32>
+  // CHECK: %[[DEQ:.+]] = torch.aten.dequantize.self %[[C]] : !torch.vtensor<[1,1,7,7],!torch.qint32> -> !torch.vtensor<[1,1,7,7],f32>
   // CHECK: %[[INT13:.+]] = torch.constant.int 13
-  // CHECK: %[[QUANT:.+]] = torch.aten.quantize_per_tensor %[[CONV]], %[[cScale]], %[[cZp]], %[[INT13]] : !torch.vtensor<[1,1,7,7],!torch.qint32>, !torch.float, !torch.int, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.quint8>
+  // CHECK: %[[QUANT:.+]] = torch.aten.quantize_per_tensor %[[DEQ]], %[[cScale]], %[[cZp]], %[[INT13]] : !torch.vtensor<[1,1,7,7],f32>, !torch.float, !torch.int, !torch.int -> !torch.vtensor<[1,1,7,7],!torch.quint8>
   // CHECK: %[[INT:.+]] = torch.aten.int_repr %[[QUANT]] : !torch.vtensor<[1,1,7,7],!torch.quint8> -> !torch.vtensor<[1,1,7,7],ui8>
   // CHECK: return %[[INT]] : !torch.vtensor<[1,1,7,7],ui8>
   return %0 : !torch.vtensor<[1,1,7,7],ui8>