void pix_subtract :: processYUV_MMX (imageStruct &image, imageStruct &right){
int datasize = image.xsize * image.ysize * image.csize;
__m64*leftPix = (__m64*)image.data;
__m64*rightPix = (__m64*)right.data;
datasize=datasize/sizeof(__m64)+(datasize%sizeof(__m64)!=0);
__m64 null64 = _mm_setzero_si64();
__m64 offset = _mm_setr_pi16(0x80, 0x00, 0x80, 0x00);
__m64 l0, l1, r0, r1;
while (datasize--) {
l1=leftPix[datasize];
r1=rightPix[datasize];
l0=_mm_unpacklo_pi8 (l1, null64);
r0=_mm_unpacklo_pi8 (r1, null64);
l1=_mm_unpackhi_pi8 (l1, null64);
r1=_mm_unpackhi_pi8 (r1, null64);
l0=_mm_adds_pu16(l0, offset);
l1=_mm_adds_pu16(l1, offset);
l0=_mm_subs_pu16(l0, r0);
l1=_mm_subs_pu16(l1, r1);
leftPix[datasize]=_mm_packs_pu16(l0, l1);
}
_mm_empty();
}
/* do the processing for all colourspaces */
void pix_motionblur :: processMMX(imageStruct &image)
{
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
int pixsize=image.ysize*image.xsize*image.csize;
pixsize=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
__m64*pixels=(__m64*)image.data;
__m64*old=(__m64*)m_savedImage.data;
__m64 newGain = _mm_set1_pi16(static_cast<short>(m_blur0));
__m64 oldGain = _mm_set1_pi16(static_cast<short>(m_blur1));
__m64 null64 = _mm_setzero_si64();
__m64 newpix1, newpix2, oldpix1, oldpix2;
while(pixsize--) {
newpix1=pixels[pixsize];
oldpix1=old[pixsize];
newpix2 = _mm_unpackhi_pi8(newpix1, null64);
newpix1 = _mm_unpacklo_pi8(newpix1, null64);
oldpix2 = _mm_unpackhi_pi8(oldpix1, null64);
oldpix1 = _mm_unpacklo_pi8(oldpix1, null64);
newpix1 = _mm_mullo_pi16(newpix1, newGain);
newpix2 = _mm_mullo_pi16(newpix2, newGain);
oldpix1 = _mm_mullo_pi16(oldpix1, oldGain);
oldpix2 = _mm_mullo_pi16(oldpix2, oldGain);
newpix1 = _mm_adds_pu16 (newpix1, oldpix1);
newpix2 = _mm_adds_pu16 (newpix2, oldpix2);
newpix1 = _mm_srli_pi16(newpix1, 8);
newpix2 = _mm_srli_pi16(newpix2, 8);
newpix1 = _mm_packs_pu16(newpix1, newpix2);
pixels[pixsize]=newpix1;
old [pixsize]=newpix1;
}
_mm_empty();
}
__m64 test40(__m64 a, __m64 b) {
// CHECK: paddusw
return _mm_adds_pu16(a, b);
}
__m64 test_mm_adds_pu16(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_adds_pu16
// CHECK: call x86_mmx @llvm.x86.mmx.paddus.w
return _mm_adds_pu16(a, b);
}