static void HE16_SSE41(uint8_t* dst) { // horizontal
int j;
const __m128i kShuffle3 = _mm_set1_epi8(3);
for (j = 16; j > 0; --j) {
const __m128i in = _mm_cvtsi32_si128(WebPMemToUint32(dst - 4));
const __m128i values = _mm_shuffle_epi8(in, kShuffle3);
_mm_storeu_si128((__m128i*)dst, values);
dst += BPS;
}
}
static void RescalerImportRowShrink_SSE2(WebPRescaler* const wrk,
const uint8_t* src) {
const int x_sub = wrk->x_sub;
int accum = 0;
const __m128i zero = _mm_setzero_si128();
const __m128i mult0 = _mm_set1_epi16(x_sub);
const __m128i mult1 = _mm_set1_epi32(wrk->fx_scale);
const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER);
__m128i sum = zero;
rescaler_t* frow = wrk->frow;
const rescaler_t* const frow_end = wrk->frow + 4 * wrk->dst_width;
if (wrk->num_channels != 4 || wrk->x_add > (x_sub << 7)) {
WebPRescalerImportRowShrink_C(wrk, src);
return;
}
assert(!WebPRescalerInputDone(wrk));
assert(!wrk->x_expand);
for (; frow < frow_end; frow += 4) {
__m128i base = zero;
accum += wrk->x_add;
while (accum > 0) {
const __m128i A = _mm_cvtsi32_si128(WebPMemToUint32(src));
src += 4;
base = _mm_unpacklo_epi8(A, zero);
// To avoid overflow, we need: base * x_add / x_sub < 32768
// => x_add < x_sub << 7. That's a 1/128 reduction ratio limit.
sum = _mm_add_epi16(sum, base);
accum -= x_sub;
}
{ // Emit next horizontal pixel.
const __m128i mult = _mm_set1_epi16(-accum);
const __m128i frac0 = _mm_mullo_epi16(base, mult); // 16b x 16b -> 32b
const __m128i frac1 = _mm_mulhi_epu16(base, mult);
const __m128i frac = _mm_unpacklo_epi16(frac0, frac1); // frac is 32b
const __m128i A0 = _mm_mullo_epi16(sum, mult0);
const __m128i A1 = _mm_mulhi_epu16(sum, mult0);
const __m128i B0 = _mm_unpacklo_epi16(A0, A1); // sum * x_sub
const __m128i frow_out = _mm_sub_epi32(B0, frac); // sum * x_sub - frac
const __m128i D0 = _mm_srli_epi64(frac, 32);
const __m128i D1 = _mm_mul_epu32(frac, mult1); // 32b x 16b -> 64b
const __m128i D2 = _mm_mul_epu32(D0, mult1);
const __m128i E1 = _mm_add_epi64(D1, rounder);
const __m128i E2 = _mm_add_epi64(D2, rounder);
const __m128i F1 = _mm_shuffle_epi32(E1, 1 | (3 << 2));
const __m128i F2 = _mm_shuffle_epi32(E2, 1 | (3 << 2));
const __m128i G = _mm_unpacklo_epi32(F1, F2);
sum = _mm_packs_epi32(G, zero);
_mm_storeu_si128((__m128i*)frow, frow_out);
}
}
assert(accum == 0);
}
void VP8LDoFillBitWindow(VP8LBitReader* const br) {
assert(br->bit_pos_ >= VP8L_WBITS);
// TODO(jzern): given the fixed read size it may be possible to force
// alignment in this block.
#if defined(VP8L_USE_UNALIGNED_LOAD)
if (br->pos_ + sizeof(br->val_) < br->len_) {
br->val_ >>= VP8L_WBITS;
br->bit_pos_ -= VP8L_WBITS;
// The expression below needs a little-endian arch to work correctly.
// This gives a large speedup for decoding speed.
br->val_ |= (vp8l_val_t)WebPMemToUint32(br->buf_ + br->pos_) <<
(VP8L_LBITS - VP8L_WBITS);
br->pos_ += VP8L_LOG8_WBITS;
return;
}
