void pix_diff :: 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 mask = _mm_setr_pi8(0x40, 0x00, 0x40, 0x00,
0x40, 0x00, 0x40, 0x00);
__m64 l, r, b;
while (datasize--) {
l=leftPix[datasize];
r=rightPix[datasize];
l=_mm_adds_pu8(l, mask);
r=_mm_subs_pu8(r, mask);
b = l;
b = _mm_subs_pu8 (b, r);
r = _mm_subs_pu8 (r, l);
b = _mm_or_si64 (b, r);
leftPix[datasize]=b;
}
_mm_empty();
}
void pix_movement :: processGrayMMX(imageStruct &image)
{
// assume that the pix_size does not change !
bool doclear=(image.xsize*image.ysize != buffer.xsize*buffer.ysize);
buffer.xsize = image.xsize;
buffer.ysize = image.ysize;
buffer.reallocate();
if(doclear) {
buffer.setWhite();
}
buffer2.xsize = image.xsize;
buffer2.ysize = image.ysize;
buffer2.reallocate();
int pixsize = image.ysize * image.xsize / sizeof(__m64);
unsigned char thresh=threshold;
__m64*rp = (__m64*)image.data; // read pointer
__m64*wp = (__m64*)buffer.data; // write pointer to the copy
__m64*wp2= (__m64*)buffer2.data; // write pointer to the diff-image
__m64 m1, m2, grey;
__m64 thresh8=_mm_set_pi8(thresh,thresh,thresh,thresh,
thresh,thresh,thresh,thresh);
// there is still one problem with the threshold: is the cmpgt only for signed ?
while(pixsize--) {
grey = rp[pixsize]; // image.data
m2 = wp[pixsize]; // buffer.data
//m0 =_mm_cmpgt_pi8(grey, m2); // (grey>m2)
//m1 =_mm_subs_pu8 (grey, m2); // (grey-m2)
//m2 =_mm_subs_pu8 (m2, grey); // (m2-grey)
//m1 =_mm_and_si64 (m1, m0); // (m2-grey)&(grey>m2) ((??))
//m0 =_mm_andnot_si64(m0, m2); // !(grey>m2)&(grey-m2) ((??))
//m2 =_mm_or_si64 (m2, m0); // [(a-b)&(a>b)]|[(b-a)&!(a>b)]=abs(a-b)
// this is better: use saturated arithmetic!
m1 =_mm_subs_pu8 (grey, m2); // (grey-m2)
m2 =_mm_subs_pu8 (m2, grey); // (m2-grey)
wp[pixsize]=grey; // buffer.data
m2 = _mm_or_si64 (m2, m1); // |grey-m2|
m2 =_mm_subs_pu8 (m2, thresh8);
m2 =_mm_cmpgt_pi8(m2, _mm_setzero_si64());
wp2[pixsize]=m2; // output.data
}
_mm_empty();
image.data = buffer2.data;
}
void pix_background :: processRGBAMMX(imageStruct &image)
{
long i,pixsize;
pixsize = image.xsize * image.ysize * image.csize;
if(m_savedImage.xsize!=image.xsize ||
m_savedImage.ysize!=image.ysize ||
m_savedImage.format!=image.format)m_reset=1;
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
if (m_reset){
memcpy(m_savedImage.data,image.data,pixsize);
}
m_reset=0;
i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
__m64*data =(__m64*)image.data;
__m64*saved=(__m64*)m_savedImage.data;
const __m64 thresh=_mm_set_pi8(m_Yrange, m_Urange, m_Vrange, m_Arange,
m_Yrange, m_Urange, m_Vrange, m_Arange);
const __m64 offset=_mm_set_pi8(1, 1, 1, 1, 1, 1, 1, 1);
__m64 newpix, oldpix, m1;
while(i--){
/* 7ops, 3memops */
/* i have the feeling that this is not faster at all!
* even if i have the 3memops + ONLY 1 _mm_subs_pu8()
* i am equally slow as the generic code;
* adding the other instruction does not change much
*/
newpix=*data;
oldpix=*saved++;
m1 = newpix;
m1 = _mm_subs_pu8 (m1, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_adds_pu8 (m1, offset);
m1 = _mm_subs_pu8 (m1, thresh);
m1 = _mm_cmpeq_pi32 (m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh
m1 = _mm_andnot_si64(m1, newpix);
*data++ = m1;
}
_mm_empty();
}
void pix_compare :: processYUV_MMX(imageStruct &image, imageStruct &right)
{
long datasize = image.xsize * image.ysize * image.csize;
datasize=datasize/sizeof(__m64)+(datasize%sizeof(__m64)!=0);
__m64*leftPix = (__m64*)image.data;
__m64*rightPix = (__m64*)right.data;
__m64 l, r, b;
__m64 mask = _mm_setr_pi8((unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF);
__m64 zeros = _mm_set1_pi8((unsigned char)0x00);
//format is U Y V Y
if (m_direction) {
while(datasize--){
l=leftPix[datasize];
r=rightPix[datasize];
b=_mm_subs_pu8(l, r);
b=_mm_and_si64(b, mask);
b=_mm_cmpeq_pi32(b, zeros);
r=_mm_and_si64(r, b);
l=_mm_andnot_si64(b, l);
leftPix[datasize]=_mm_or_si64(l, r);
}
} else {
while(datasize--){
l=leftPix[datasize];
r=rightPix[datasize];
b=_mm_subs_pu8(r, l);
b=_mm_and_si64(b, mask);
b=_mm_cmpeq_pi32(b, zeros);
r=_mm_and_si64(r, b);
l=_mm_andnot_si64(b, l);
leftPix[datasize]=_mm_or_si64(l, r);
}
}
_mm_empty();
}
void pix_background :: processGrayMMX(imageStruct &image){
int i;
long pixsize;
pixsize = image.xsize * image.ysize * image.csize;
if(m_savedImage.xsize!=image.xsize ||
m_savedImage.ysize!=image.ysize ||
m_savedImage.format!=image.format)m_reset=1;
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
if (m_reset){
memcpy(m_savedImage.data,image.data,pixsize);
}
m_reset=0;
if(m_Yrange==0)return;
__m64*npixes=(__m64*)image.data;
__m64*opixes=(__m64*)m_savedImage.data;
__m64 newpix, oldpix, m1;
unsigned char thresh=m_Yrange-1;
__m64 thresh8=_mm_set_pi8(thresh,thresh,thresh,thresh,
thresh,thresh,thresh,thresh);
i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
while(i--){
newpix=npixes[i];
oldpix=opixes[i];
m1 = _mm_subs_pu8 (newpix, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_subs_pu8 (m1, thresh8);
m1 = _mm_cmpgt_pi8(m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh8
npixes[i] = _mm_and_si64(m1, newpix);
}
_mm_empty();
}
void pix_add :: processYUV_MMX (imageStruct &image, imageStruct &right){
int datasize = image.xsize * image.ysize * image.csize;
__m64*leftPix = reinterpret_cast<__m64*>(image.data);
__m64*rightPix = reinterpret_cast<__m64*>(right.data);
datasize=datasize/sizeof(__m64)+(datasize%sizeof(__m64)!=0);
__m64 mask = _mm_setr_pi8(0x40, 0x00, 0x40, 0x00,
0x40, 0x00, 0x40, 0x00);
__m64 l, r;
while (datasize--) {
l=leftPix[datasize];
r=rightPix[datasize];
l=_mm_subs_pu8(l, mask);
r=_mm_subs_pu8(r, mask);
l=_mm_adds_pu8(l,r);
leftPix[datasize]=l;
}
_mm_empty();
}
void pix_subtract :: processRGBA_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 l, r;
while (datasize--) {
l=leftPix[datasize];
r=rightPix[datasize];
leftPix[datasize]=_mm_subs_pu8(l,r);
}
_mm_empty();
}
void pix_background :: processYUVMMX(imageStruct &image)
{
long pixsize;
pixsize = image.xsize * image.ysize * image.csize;
if(m_savedImage.xsize!=image.xsize ||
m_savedImage.ysize!=image.ysize ||
m_savedImage.format!=image.format)m_reset=1;
m_savedImage.xsize=image.xsize;
m_savedImage.ysize=image.ysize;
m_savedImage.setCsizeByFormat(image.format);
m_savedImage.reallocate();
if (m_reset){
memcpy(m_savedImage.data,image.data,pixsize);
// return;
}
m_reset=0;
int i=pixsize/sizeof(__m64)+(pixsize%sizeof(__m64)!=0);
__m64*data =(__m64*)image.data;
__m64*saved=(__m64*)m_savedImage.data;
const __m64 thresh=_mm_set_pi8(m_Urange, m_Yrange, m_Vrange, m_Yrange,
m_Urange, m_Yrange, m_Vrange, m_Yrange);
const __m64 offset=_mm_set_pi8(1, 1, 1, 1, 1, 1, 1, 1);
const __m64 black =_mm_set_pi8((unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80,
(unsigned char)0x00,
(unsigned char)0x80);
__m64 newpix, oldpix, m1;
while(i--){
newpix=*data;
oldpix=*saved++;
m1 = newpix;
m1 = _mm_subs_pu8 (m1, oldpix);
oldpix= _mm_subs_pu8 (oldpix, newpix);
m1 = _mm_or_si64 (m1, oldpix); // |oldpix-newpix|
m1 = _mm_adds_pu8 (m1, offset); // to make thresh=0 work correctly
m1 = _mm_subs_pu8 (m1, thresh); // m1>thresh -> saturation -> 0
m1 = _mm_cmpeq_pi32 (m1, _mm_setzero_si64()); // |oldpix-newpix|>thresh
oldpix= black;
oldpix= _mm_and_si64 (oldpix, m1);
m1 = _mm_andnot_si64 (m1, newpix);
m1 = _mm_or_si64 (m1, oldpix);
*data++ = m1;
}
_mm_empty();
}
void pix_multiply :: 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 l0, r0, l1, r1;
__m64 mask= _mm_setr_pi8((unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00,
(unsigned char)0xFF,
(unsigned char)0x00);
__m64 yuvclamp0 = _mm_setr_pi8((unsigned char)0x00,
(unsigned char)0x10,
(unsigned char)0x00,
(unsigned char)0x10,
(unsigned char)0x00,
(unsigned char)0x10,
(unsigned char)0x00,
(unsigned char)0x10);
__m64 yuvclamp1 = _mm_setr_pi8((unsigned char)0x00,
(unsigned char)0x24,
(unsigned char)0x00,
(unsigned char)0x24,
(unsigned char)0x00,
(unsigned char)0x24,
(unsigned char)0x00,
(unsigned char)0x24);
__m64 yuvclamp2 = _mm_setr_pi8((unsigned char)0x00,
(unsigned char)0x14,
(unsigned char)0x00,
(unsigned char)0x14,
(unsigned char)0x00,
(unsigned char)0x14,
(unsigned char)0x00,
(unsigned char)0x14);
__m64 null64 = _mm_setzero_si64();
while(datasize--) {
r1=rightPix[datasize];
l1=leftPix [datasize];
r1=_mm_or_si64(r1, mask);
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_mullo_pi16 (l0, r0);
l1=_mm_mullo_pi16 (l1, r1);
l0=_mm_srli_pi16(l0, 8);
l1=_mm_srli_pi16(l1, 8);
l0=_mm_packs_pu16(l0, l1);
l0=_mm_subs_pu8(l0, yuvclamp0);
l0=_mm_adds_pu8(l0, yuvclamp1);
l0=_mm_subs_pu8(l0, yuvclamp2);
leftPix[datasize]=l0;
}
_mm_empty();
}
__m64 test47(__m64 a, __m64 b) {
// CHECK: psubusb
return _mm_subs_pu8(a, b);
}
test (__m64 s1, __m64 s2)
{
return _mm_subs_pu8 (s1, s2);
}
__m64 test_mm_subs_pu8(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_subs_pu8
// CHECK: call x86_mmx @llvm.x86.mmx.psubus.b
return _mm_subs_pu8(a, b);
}
