b[unzig[zig]] *= qt[zig]
	}
	idct(b)
	dst, stride := []byte(nil), 0
	if d.nComp == 1 {
		dst, stride = d.img1.Pix[8*(by*d.img1.Stride+bx):], d.img1.Stride
	} else {
		switch compIndex {
		case 0:
			dst, stride = d.img3.Y[8*(by*d.img3.YStride+bx):], d.img3.YStride
		case 1:
			dst, stride = d.img3.Cb[8*(by*d.img3.CStride+bx):], d.img3.CStride
		case 2:
			dst, stride = d.img3.Cr[8*(by*d.img3.CStride+bx):], d.img3.CStride

			m1 := m1.(*image.YCbCr)
			if err := check(m0.Bounds(), m0.Y, m1.Y, m0.YStride, m1.YStride); err != nil {
				t.Errorf("%s (Y): %v", tc, err)
				continue
			}
			if err := check(m0.Bounds(), m0.Cb, m1.Cb, m0.CStride, m1.CStride); err != nil {
				t.Errorf("%s (Cb): %v", tc, err)
				continue
			}
			if err := check(m0.Bounds(), m0.Cr, m1.Cr, m0.CStride, m1.CStride); err != nil {
				t.Errorf("%s (Cr): %v", tc, err)
				continue
			}

                            /*padding*/(GetAvgPoolOpPadAttr $old_val),
                            /*stride_h*/(GetI32At<0> (GetAsVectorAttr<"stride"> $attrs)),
                            /*stride_w*/(GetI32At<1> (GetAsVectorAttr<"stride"> $attrs)),
                            /*fused_activation_function*/TFL_AF_None,
                            (returnType (GetNhwcReturnTypeFromNchw $old_val))), 

		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

  // 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>

  // CHECK-DAG: %[[RESULT:.+]] = mhlo.convolution(%[[INPUT]], %[[NEW_FILTER]]) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {stride = [1, 1], pad = {{\[\[}}0, 0], [0, 0]], rhs_dilate = [1, 1]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x256x256x3xf32>, tensor<1x1x3x2xf32>) -> tensor<1x256x256x2xf32>

		}
	}
}

func TestDumper(t *testing.T) {
	var in [40]byte
	for i := range in {
		in[i] = byte(i + 30)
	}

	for stride := 1; stride < len(in); stride++ {
		var out bytes.Buffer
		dumper := Dumper(&out)
		done := 0
		for done < len(in) {
			todo := done + stride
			if todo > len(in) {
				todo = len(in)
			}
			dumper.Write(in[done:todo])
			done = todo
		}

		dumper.Close()

  Value stride_value = CreateScalarConstValue<int32_t>(builder, loc, stride);
  Value dilation_rate_value =
      CreateScalarConstValue<int32_t>(builder, loc, dilation_rate);

  Value effective_filter_size_op = scalar_add(
      scalar_mul(dilation_rate_value, scalar_sub(filter_size, one)), one);

  // output_size = (input_size + stride - 1) / stride
  Value output_size = scalar_div(

			} else if nrgba != nil {
				stride, pix = nrgba.Stride, nrgba.Pix
			}
			if stride != 0 {
				j0 := (y - b.Min.Y) * stride
				j1 := j0 + b.Dx()*4
				for j := j0; j < j1; j += 4 {
					cr0[i+0] = pix[j+0]
					cr0[i+1] = pix[j+1]
					cr0[i+2] = pix[j+2]
					i += 3
				}
			} else {
				for x := b.Min.X; x < b.Max.X; x++ {
					r, g, b, _ := m.At(x, y).RGBA()

  //
  // `axis` should be a valid dimension in `shape`.
  static int64_t StrideForAxis(ArrayRef<int64_t> shape, int64_t axis) {
    int64_t stride = 1;  // Start with the trailing dimension.
    for (int64_t dim = shape.size() - 1; dim > axis; --dim) {
      stride *= shape[dim];
    }
    return stride;
  }

  // Calculates how many values to skip across a 1-D contiguous array that holds