%0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>
  %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], padding = "SAME", strides = [1, 1, 2, 1]}
       : (tensor<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16> loc("Model/conv2d")
  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>

    %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst_0) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<1x2x2x3xf32>
    %1 = "tf.BiasAdd"(%0, %cst) {data_format = "NHWC", device = ""} : (tensor<1x2x2x3xf32>, tensor<3xf32>) -> tensor<1x2x2x3xf32>

    %0 = "tf.Conv2D"(%arg0, %arg1) {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }

  pool_params.filter_height = pool_size[0];
  pool_params.filter_width = pool_size[1];

  // Retrieve strides.
  llvm::SmallVector<int32_t, 2> strides;
  if (failed(GetIntegerArraySafe(&func_, attrs, "strides", &strides, 2))) {
    return failure();
  }
  pool_params.stride_height = strides[0];
  pool_params.stride_width = strides[1];

  // Retrieves padding.
  auto padding = mlir::dyn_cast_or_null<StringAttr>(attrs.get("padding"));

  %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16>
  %2 = "tf.AvgPool"(%1) {data_format = "NHWC", device = "", ksize = [1, 1, 1, 1], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xbf16>

  // CHECK-DAG: %[[result:.*]] = chlo.broadcast_add %[[conv]], %[[combined]]
  // CHECK: return %[[result]]
  %0 = mhlo.convolution(%lhs, %rhs)
      dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f],
      window = {stride = [1, 1], pad = [[0, 0], [0, 0]],
      lhs_dilate = [1, 1], rhs_dilate = [1, 1]}
      {batch_group_count = 1 : i64, feature_group_count = 1 : i64}
    : (tensor<?x3x2x1xi8>, tensor<2x1x1x1xi8>) -> tensor<?x2x2x1xi32>

		ddelta = dst.Stride
		sdelta = src.Stride
		i0, i1, idelta = 0, dx*4, +4
	} else {
		// If the source start point is higher than the destination start point, or equal height but to the left,
		// then we compose the rows in right-to-left, bottom-up order instead of left-to-right, top-down.
		d0 += (dy - 1) * dst.Stride
		s0 += (dy - 1) * src.Stride
		ddelta = -dst.Stride
		sdelta = -src.Stride

		srcPix = src.(*image.Gray).Pix
		dstPix, stride, rect = target.Pix, target.Stride, target.Rect
		bytesPerPixel = 1
	case *image.Gray16:
		srcPix = src.(*image.Gray16).Pix
		dstPix, stride, rect = target.Pix, target.Stride, target.Rect
		bytesPerPixel = 2
	case *image.NRGBA:
		srcPix = src.(*image.NRGBA).Pix
		dstPix, stride, rect = target.Pix, target.Stride, target.Rect
		bytesPerPixel = 4
	case *image.NRGBA64:

  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NHWC",
         padding = "VALID",
         strides = [1, 1, 1, 1]
       } : (tensor<1x28x28x64xf16>, tensor<3x3x64x64xf16>)
        -> tensor<1x26x26x64xf16>

  func.return %0 : tensor<1x26x26x64xf16>
}

// CHECK-LABEL: func @transposeConv2DBackpropFilter_f16

			}
			if r <= range_.Hi {
				return range_.Stride == 1 || (r-range_.Lo)%range_.Stride == 0
			}
		}
		return false
	}

	// binary search over ranges
	lo := 0
	hi := len(ranges)
	for lo < hi {
		m := int(uint(lo+hi) >> 1)
		range_ := &ranges[m]
		if range_.Lo <= r && r <= range_.Hi {
			return range_.Stride == 1 || (r-range_.Lo)%range_.Stride == 0
		}
		if r < range_.Lo {
			hi = m