void png_read_filter_row_paeth4_sse2(png_row_infop row_info, png_bytep row,
png_const_bytep prev)
{
/* Paeth tries to predict pixel d using the pixel to the left of it, a,
* and two pixels from the previous row, b and c:
* prev: c b
* row: a d
* The Paeth function predicts d to be whichever of a, b, or c is nearest to
* p=a+b-c.
*
* The first pixel has no left context, and so uses an Up filter, p = b.
* This works naturally with our main loop's p = a+b-c if we force a and c
* to zero.
* Here we zero b and d, which become c and a respectively at the start of
* the loop.
*/
png_debug(1, "in png_read_filter_row_paeth4_sse2");
const __m128i zero = _mm_setzero_si128();
__m128i c, b = zero,
a, d = zero;
int rb = row_info->rowbytes;
while (rb > 0) {
/* It's easiest to do this math (particularly, deal with pc) with 16-bit
* intermediates.
*/
c = b; b = _mm_unpacklo_epi8(load4(prev), zero);
a = d; d = _mm_unpacklo_epi8(load4(row ), zero);
/* (p-a) == (a+b-c - a) == (b-c) */
__m128i pa = _mm_sub_epi16(b,c);
/* (p-b) == (a+b-c - b) == (a-c) */
__m128i pb = _mm_sub_epi16(a,c);
/* (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c) */
__m128i pc = _mm_add_epi16(pa,pb);
pa = abs_i16(pa); /* |p-a| */
pb = abs_i16(pb); /* |p-b| */
pc = abs_i16(pc); /* |p-c| */
__m128i smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb));
/* Paeth breaks ties favoring a over b over c. */
__m128i nearest = if_then_else(_mm_cmpeq_epi16(smallest, pa), a,
if_then_else(_mm_cmpeq_epi16(smallest, pb), b,
c));
/* Note `_epi8`: we need addition to wrap modulo 255. */
d = _mm_add_epi8(d, nearest);
store4(row, _mm_packus_epi16(d,d));
prev += 4;
row += 4;
rb -= 4;
}
}
void sk_paeth_sse2(png_row_infop row_info, uint8_t* row, const uint8_t* prev) {
// Paeth tries to predict pixel d using the pixel to the left of it, a,
// and two pixels from the previous row, b and c:
// prev: c b
// row: a d
// The Paeth function predicts d to be whichever of a, b, or c is nearest to p=a+b-c.
// The first pixel has no left context, and so uses an Up filter, p = b.
// This works naturally with our main loop's p = a+b-c if we force a and c to zero.
// Here we zero b and d, which become c and a respectively at the start of the loop.
const __m128i zero = _mm_setzero_si128();
__m128i c, b = zero,
a, d = zero;
int rb = row_info->rowbytes;
while (rb > 0) {
// It's easiest to do this math (particularly, deal with pc) with 16-bit intermediates.
c = b;
b = _mm_unpacklo_epi8(load<bpp>(prev), zero);
a = d;
d = _mm_unpacklo_epi8(load<bpp>(row ), zero);
__m128i pa = _mm_sub_epi16(b,c), // (p-a) == (a+b-c - a) == (b-c)
pb = _mm_sub_epi16(a,c), // (p-b) == (a+b-c - b) == (a-c)
pc = _mm_add_epi16(pa,pb); // (p-c) == (a+b-c - c) == (a+b-c-c) == (b-c)+(a-c)
pa = abs_i16(pa); // |p-a|
pb = abs_i16(pb); // |p-b|
pc = abs_i16(pc); // |p-c|
__m128i smallest = _mm_min_epi16(pc, _mm_min_epi16(pa, pb));
// Paeth breaks ties favoring a over b over c.
__m128i nearest = if_then_else(_mm_cmpeq_epi16(smallest, pa), a,
if_then_else(_mm_cmpeq_epi16(smallest, pb), b,
c));
d = _mm_add_epi8(d, nearest); // Note `_epi8`: we need addition to wrap modulo 255.
store<bpp>(row, _mm_packus_epi16(d,d));
prev += bpp;
row += bpp;
rb -= bpp;
}
}
