void pix_offset :: processRGBAMMX(imageStruct &image)
{
char R = m_offset[chRed];
char G = m_offset[chGreen];
char B = m_offset[chBlue];
char A = m_offset[chAlpha];
register int pixsize = (image.ysize * image.xsize)>>1;
register __m64 offset_64 = _mm_setr_pi8(R, G, B, A, R, G, B, A);
register __m64*data_p= (__m64*)image.data;
_mm_empty();
if(m_saturate) {
while(pixsize--) {
data_p[0]=_mm_adds_pu8(data_p[0], offset_64);
data_p++;
}
} else {
while(pixsize--) {
data_p[0]=_mm_add_pi8(data_p[0], offset_64);
data_p++;
}
}
_mm_empty();
}
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();
}
__m64 test_mm_setr_pi8(char a, char b, char c, char d, char e, char f, char g, char h) {
// CHECK-LABEL: test_mm_setr_pi8
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
// CHECK: insertelement <8 x i8>
return _mm_setr_pi8(a, b, c, d, e, f, g, h);
}
void pix_offset :: processYUVMMX(imageStruct &image)
{
register int pixsize = (image.ysize * image.xsize)>>2;
register __m64 offset_64 = _mm_setr_pi8(U, Y, V, Y, U, Y, V, Y);
register __m64*data_p= (__m64*)image.data;
_mm_empty();
while(pixsize--) {
data_p[0]=_mm_add_pi8(data_p[0], offset_64);
data_p++;
}
_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_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_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();
}
