static WEBP_INLINE void HE4(uint8_t* dst, const uint8_t* top) { // horizontal
const int X = top[-1];
const int I = top[-2];
const int J = top[-3];
const int K = top[-4];
const int L = top[-5];
WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));
WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));
WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));
WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
}
static void HE4_C(uint8_t* dst) { // horizontal
const int A = dst[-1 - BPS];
const int B = dst[-1];
const int C = dst[-1 + BPS];
const int D = dst[-1 + 2 * BPS];
const int E = dst[-1 + 3 * BPS];
WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(A, B, C));
WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(B, C, D));
WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(C, D, E));
WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(D, E, E));
}
static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
const uint8_t* src) {
rescaler_t* frow = wrk->frow;
const rescaler_t* const frow_end = frow + wrk->dst_width * wrk->num_channels;
const int x_add = wrk->x_add;
int accum = x_add;
__m128i cur_pixels;
// SSE2 implementation only works with 16b signed arithmetic at max.
if (wrk->src_width < 8 || accum >= (1 << 15)) {
WebPRescalerImportRowExpand_C(wrk, src);
return;
}
assert(!WebPRescalerInputDone(wrk));
assert(wrk->x_expand);
if (wrk->num_channels == 4) {
LoadTwoPixels_SSE2(src, &cur_pixels);
src += 4;
while (1) {
const __m128i mult = _mm_set1_epi32(((x_add - accum) << 16) | accum);
const __m128i out = _mm_madd_epi16(cur_pixels, mult);
_mm_storeu_si128((__m128i*)frow, out);
frow += 4;
if (frow >= frow_end) break;
accum -= wrk->x_sub;
if (accum < 0) {
LoadTwoPixels_SSE2(src, &cur_pixels);
src += 4;
accum += x_add;
}
}
} else {
int left;
const uint8_t* const src_limit = src + wrk->src_width - 8;
LoadEightPixels_SSE2(src, &cur_pixels);
src += 7;
left = 7;
while (1) {
const __m128i mult = _mm_cvtsi32_si128(((x_add - accum) << 16) | accum);
const __m128i out = _mm_madd_epi16(cur_pixels, mult);
assert(sizeof(*frow) == sizeof(uint32_t));
WebPUint32ToMem((uint8_t*)frow, _mm_cvtsi128_si32(out));
frow += 1;
if (frow >= frow_end) break;
accum -= wrk->x_sub;
if (accum < 0) {
if (--left) {
cur_pixels = _mm_srli_si128(cur_pixels, 2);
} else if (src <= src_limit) {
LoadEightPixels_SSE2(src, &cur_pixels);
src += 7;
left = 7;
} else { // tail
cur_pixels = _mm_srli_si128(cur_pixels, 2);
cur_pixels = _mm_insert_epi16(cur_pixels, src[1], 1);
src += 1;
left = 1;
}
accum += x_add;
}
}
}
assert(accum == 0);
}