s[5] = (x0 - x4) >> 8
		s[6] = (x3 - x2) >> 8
		s[7] = (x7 - x1) >> 8
	}

	// Vertical 1-D IDCT.
	for x := 0; x < 8; x++ {
		// Similar to the horizontal 1-D IDCT case, if all the AC components are zero, then the IDCT is trivial.
		// However, after performing the horizontal 1-D IDCT, there are typically non-zero AC components, so
		// we do not bother to check for the all-zero case.

			t.Errorf("i=%d: FDCT\nsrc\n%s\ngot\n%s\nwant\n%s\n", i, &b, &got, &want)
		}
	}

	// Check that the optimized and slow IDCT implementations agree.
	for i, b := range blocks {
		got, want := b, b
		idct(&got)
		slowIDCT(&want)
		if differ(&got, &want) {
			t.Errorf("i=%d: IDCT\nsrc\n%s\ngot\n%s\nwant\n%s\n", i, &b, &got, &want)
		}
	}
}

// differ reports whether any pair-wise elements in b0 and b1 differ by 2 or

// to the image.
func (d *decoder) reconstructBlock(b *block, bx, by, compIndex int) error {
	qt := &d.quant[d.comp[compIndex].tq]
	for zig := 0; zig < blockSize; zig++ {
		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:

        `tf.linalg.matvec` convenience function.
    *   `tf.einsum()`raises `ValueError` for unsupported equations like
        `"ii->"`.
    *   Add DCT-I and IDCT-I in `tf.signal.dct` and `tf.signal.idct`.
    *   Add LU decomposition op.
    *   Add quantile loss to gradient boosted trees in estimator.
    *   Add `round_mode` to `QuantizeAndDequantizeV2` op to select rounding
        algorithm.