// with the value in Mul.
def FuseMulAndConv2D :
  Pat<(TF_MulOp:$mul (TF_Conv2DOp:$conv $input,
                          (Arith_ConstantOp:$filter F32ElementsAttr:$filter_value),
                          $strides, $use_cudnn, $padding, $explicit_padding,
                          IsDataFormatNHWC:$data_format, $dilations),
                     (Arith_ConstantOp:$multiplier F32ElementsAttr:$mul_value)),

    %0 = "tf.Conv2D"(%arg0, %cst_1) {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>
    // Dequantize added before BiasAdd.

  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  %1 = "tf.Transpose"(%arg0, %0) : (tensor<1x3x32x32xf32>, tensor<4xi32>) -> tensor<1x32x32x3xf32>

  // Compute in NHWC
  %2 = "tf.Conv2D"(%1, %arg1)
    {
      data_format = "NHWC",
      padding = "SAME",
      strides = [1, 1, 1, 1],
      dilations = [1, 1, 1, 1]
    } : (tensor<1x32x32x3xf32>, tensor<1x1x3x8xf32>) -> tensor<1x32x32x8xf32>

  // Convert result back: NHWC -> NCHW

	s[2] = c1.B
	s[3] = c1.A
}

func (p *slowestRGBA) PixOffset(x, y int) int {
	return (y-p.Rect.Min.Y)*p.Stride + (x-p.Rect.Min.X)*4
}

func convertToSlowestRGBA(m image.Image) *slowestRGBA {
	if rgba, ok := m.(*image.RGBA); ok {
		return &slowestRGBA{
			Pix:    append([]byte(nil), rgba.Pix...),
			Stride: rgba.Stride,
			Rect:   rgba.Rect,
		}
	}
	rgba := image.NewRGBA(m.Bounds())

    %0:2 = tf_executor.graph {
      %outputs:2, %control = tf_executor.island wraps "tf.MaxPoolWithArgmax"(%arg0) {T = f32, Targmax = i32, include_batch_in_index = false, ksize = [1, 2, 2, 1], padding = "SAME", strides = [1, 2, 2, 1]} : (tensor<2x4x2x2xf32>) -> (tensor<2x2x1x2xf32>, tensor<2x2x1x2xi32>)
      tf_executor.fetch %outputs#1, %outputs#0 : tensor<2x2x1x2xi32>, tensor<2x2x1x2xf32>
    }

  // must use NCHW data format with updated kernel size and strides.

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi32>}>
  // CHECK: %[[MAX_POOL:[0-9]*]] = "tf.MaxPool"(%arg0) <{data_format = "NCHW", ksize = [1, 1, 3, 3], padding = "SAME", strides = [1, 1, 2, 2]}> : (tensor<1x64x112x112xf32>) -> tensor<1x64x56x56xf32>

    %2 = "tf.Conv2D"(%1, %0) {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<2x3x3x2xf32>) -> tensor<*xf32>

	y0 := r.Min.Y - dst.Rect.Min.Y
	y1 := r.Max.Y - dst.Rect.Min.Y
	switch src.SubsampleRatio {

	case image.YCbCrSubsampleRatio444:
		for y, sy := y0, sp.Y; y != y1; y, sy = y+1, sy+1 {
			dpix := dst.Pix[y*dst.Stride:]
			yi := (sy-src.Rect.Min.Y)*src.YStride + (sp.X - src.Rect.Min.X)

			ci := (sy-src.Rect.Min.Y)*src.CStride + (sp.X - src.Rect.Min.X)
			for x := x0; x != x1; x, yi, ci = x+4, yi+1, ci+1 {

    %3 = "tf.Sub"(%2, %filter_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    %5 = "tf.Conv2D"(%1, %3) {
      padding = "VALID", strides = [1, 1, 1, 1],
      attr_map = "strides:0,use_cudnn_on_gpu:1,padding:2,explicit_paddings:3,dilations:4"
    } : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>
    func.return %5 : tensor<*xi32>
  }