void weighted_merge_planar_mmx(BYTE *p1, const BYTE *p2, int p1_pitch, int p2_pitch, int width, int height, int weight, int invweight) {
__m64 round_mask = _mm_set1_pi32(0x4000);
__m64 zero = _mm_setzero_si64();
__m64 mask = _mm_set_pi16(weight, invweight, weight, invweight);
int wMod8 = (width/8) * 8;
for (int y = 0; y < height; y++) {
for (int x = 0; x < wMod8; x += 8) {
__m64 px1 = *(reinterpret_cast<const __m64*>(p1+x)); //y7y6 y5y4 y3y2 y1y0
__m64 px2 = *(reinterpret_cast<const __m64*>(p2+x)); //Y7Y6 Y5Y4 Y3Y2 Y1Y0
__m64 p0123 = _mm_unpacklo_pi8(px1, px2); //Y3y3 Y2y2 Y1y1 Y0y0
__m64 p4567 = _mm_unpackhi_pi8(px1, px2); //Y7y7 Y6y6 Y5y5 Y4y4
__m64 p01 = _mm_unpacklo_pi8(p0123, zero); //00Y1 00y1 00Y0 00y0
__m64 p23 = _mm_unpackhi_pi8(p0123, zero); //00Y3 00y3 00Y2 00y2
__m64 p45 = _mm_unpacklo_pi8(p4567, zero); //00Y5 00y5 00Y4 00y4
__m64 p67 = _mm_unpackhi_pi8(p4567, zero); //00Y7 00y7 00Y6 00y6
p01 = _mm_madd_pi16(p01, mask);
p23 = _mm_madd_pi16(p23, mask);
p45 = _mm_madd_pi16(p45, mask);
p67 = _mm_madd_pi16(p67, mask);
p01 = _mm_add_pi32(p01, round_mask);
p23 = _mm_add_pi32(p23, round_mask);
p45 = _mm_add_pi32(p45, round_mask);
p67 = _mm_add_pi32(p67, round_mask);
p01 = _mm_srli_pi32(p01, 15);
p23 = _mm_srli_pi32(p23, 15);
p45 = _mm_srli_pi32(p45, 15);
p67 = _mm_srli_pi32(p67, 15);
p0123 = _mm_packs_pi32(p01, p23);
p4567 = _mm_packs_pi32(p45, p67);
__m64 result = _mm_packs_pu16(p0123, p4567);
*reinterpret_cast<__m64*>(p1+x) = result;
}
for (int x = wMod8; x < width; x++) {
p1[x] = (p1[x]*invweight + p2[x]*weight + 16384) >> 15;
}
p1 += p1_pitch;
p2 += p2_pitch;
}
_mm_empty();
}
static void weighted_merge_luma_yuy2_mmx(BYTE *src, const BYTE *luma, int pitch, int luma_pitch,int width, int height, int weight, int invweight)
{
__m64 round_mask = _mm_set1_pi32(0x4000);
__m64 mask = _mm_set_pi16(weight, invweight, weight, invweight);
__m64 luma_mask = _mm_set1_pi16(0x00FF);
#pragma warning(push)
#pragma warning(disable: 4309)
__m64 chroma_mask = _mm_set1_pi16(0xFF00);
#pragma warning(pop)
int wMod8 = (width/8) * 8;
for (int y = 0; y < height; y++) {
for (int x = 0; x < wMod8; x += 8) {
__m64 px1 = *reinterpret_cast<const __m64*>(src+x); //V1 Y3 U1 Y2 V0 Y1 U0 Y0
__m64 px2 = *reinterpret_cast<const __m64*>(luma+x); //v1 y3 u1 y2 v0 y1 u0 y0
__m64 src_lo = _mm_unpacklo_pi16(px1, px2); //v0 y1 V0 Y1 u0 y0 U0 Y0
__m64 src_hi = _mm_unpackhi_pi16(px1, px2);
src_lo = _mm_and_si64(src_lo, luma_mask); //00 v0 00 V0 00 u0 00 U0
src_hi = _mm_and_si64(src_hi, luma_mask);
src_lo = _mm_madd_pi16(src_lo, mask);
src_hi = _mm_madd_pi16(src_hi, mask);
src_lo = _mm_add_pi32(src_lo, round_mask);
src_hi = _mm_add_pi32(src_hi, round_mask);
src_lo = _mm_srli_pi32(src_lo, 15);
src_hi = _mm_srli_pi32(src_hi, 15);
__m64 result_luma = _mm_packs_pi32(src_lo, src_hi);
__m64 result_chroma = _mm_and_si64(px1, chroma_mask);
__m64 result = _mm_or_si64(result_chroma, result_luma);
*reinterpret_cast<__m64*>(src+x) = result;
}
for (int x = wMod8; x < width; x+=2) {
src[x] = (luma[x] * weight + src[x] * invweight + 16384) >> 15;
}
src += pitch;
luma += luma_pitch;
}
_mm_empty();
}
__m64 test49(__m64 a, __m64 b) {
// CHECK: pmaddwd
return _mm_madd_pi16(a, b);
}
__m64 test_mm_madd_pi16(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_madd_pi16
// CHECK: call x86_mmx @llvm.x86.mmx.pmadd.wd
return _mm_madd_pi16(a, b);
}