mlib_status
mlib_m_sconv3x3_16nw_1(
mlib_image *dst,
mlib_image *src,
mlib_s32 *hkernel,
mlib_s32 *vkernel,
mlib_s32 scalef_expon)
{
GET_SRC_DST_PARAMETERS(mlib_s16);
__m64 hker0, hker1, hker2, vker0, vker1, vker2;
__m64 s0, s1, s2, v0, v1, aa, bb, rr, rh, rl;
__m64 *sp0, *sp1, *sp2, *dp;
__m64 zero, _rnd;
mlib_s32 shift, kerh_sum;
mlib_s32 i, j;
width -= 2;
height -= 2;
width *= NCHAN;
dl += dll + NCHAN;
GET_KERN();
zero = _mm_setzero_si64();
for (j = 0; j < height; j++) {
sp0 = (__m64 *) sl;
sp1 = (__m64 *) (sl + sll);
sp2 = (__m64 *) (sl + 2 * sll);
dp = (__m64 *) dl;
PREP_V();
for (i = 0; i < width / 4; i++) {
CONV_3x3();
dp[i] = rr;
}
if (width & 3) {
__m64 mask =
((__m64 *) mlib_mask64_arr)[2 * (width & 3)];
CONV_3x3();
dp[i] =
_mm_or_si64(_mm_and_si64(mask, rr),
_mm_andnot_si64(mask, dp[i]));
}
sl += sll;
dl += dll;
}
_mm_empty();
return (MLIB_SUCCESS);
}
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 :: 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();
}
mlib_status
mlib_ImageMinFilter7x7_S16(
void *dst,
void *src,
mlib_s32 dlb,
mlib_s32 slb,
mlib_s32 wid,
mlib_s32 hgt)
#endif /* MAX_FILTER */
{
mlib_u8 *pbuff, *buff0, *buff1, *buff2, *buff3, *buff4, *buff5, *buffT;
mlib_u8 *sl, *sp0, *sp1, *sp2, *sp3, *sp4, *sp5, *sp6, *sp7, *dl;
__m64 *dp0, *dp1;
__m64 aa, bb, cc, dd, ee, ff, r0, r1;
__m64 g0, g1, g2, g3, g4, g5, g6, gg;
__m64 h0, h1, h2, h3, h4, h5, h6, hh;
__m64 e_mask;
mlib_s32 i, j, wid8, tail;
wid = (wid - KSIZE1) * SSIZE;
wid8 = (wid + 7) & ~7;
pbuff = mlib_malloc(KSIZE1 * wid8);
buff0 = pbuff;
buff1 = buff0 + wid8;
buff2 = buff1 + wid8;
buff3 = buff2 + wid8;
buff4 = buff3 + wid8;
buff5 = buff4 + wid8;
sl = (mlib_u8 *)src;
dl = (mlib_u8 *)dst + (KSIZE1 / 2) * (dlb + SSIZE);
tail = wid & 7;
e_mask = ((__m64 *) mlib_mask64_arr)[tail];
for (j = 0; j < 3; j++) {
sp0 = buff4;
sp1 = buff5;
sp6 = sl;
sp7 = sl + slb;
sl += 2 * slb;
for (i = 0; i < wid; i += 8) {
g0 = *(__m64 *) sp6;
g1 = *(__m64 *) (sp6 + SSIZE);
g2 = *(__m64 *) (sp6 + 2 * SSIZE);
g3 = *(__m64 *) (sp6 + 3 * SSIZE);
g4 = *(__m64 *) (sp6 + 4 * SSIZE);
g5 = *(__m64 *) (sp6 + 5 * SSIZE);
g6 = *(__m64 *) (sp6 + 6 * SSIZE);
h0 = *(__m64 *) sp7;
h1 = *(__m64 *) (sp7 + SSIZE);
h2 = *(__m64 *) (sp7 + 2 * SSIZE);
h3 = *(__m64 *) (sp7 + 3 * SSIZE);
h4 = *(__m64 *) (sp7 + 4 * SSIZE);
h5 = *(__m64 *) (sp7 + 5 * SSIZE);
h6 = *(__m64 *) (sp7 + 6 * SSIZE);
gg = C_COMP(g0, g1);
hh = C_COMP(h0, h1);
g2 = C_COMP(g2, g3);
h2 = C_COMP(h2, h3);
g4 = C_COMP(g4, g5);
h4 = C_COMP(h4, h5);
gg = C_COMP(gg, g2);
hh = C_COMP(hh, h2);
gg = C_COMP(gg, g4);
hh = C_COMP(hh, h4);
gg = C_COMP(gg, g6);
hh = C_COMP(hh, h6);
*(__m64 *) sp0 = gg;
*(__m64 *) sp1 = hh;
sp0 += 8;
sp1 += 8;
sp6 += 8;
sp7 += 8;
}
if (j < 2) {
buffT = buff0;
buff0 = buff2;
buff2 = buff4;
buff4 = buffT;
buffT = buff1;
buff1 = buff3;
buff3 = buff5;
buff5 = buffT;
}
}
for (j = 0; j <= (hgt - KSIZE1 - 2); j += 2) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff0;
sp1 = buff1;
sp2 = buff2;
sp3 = buff3;
sp4 = buff4;
sp5 = buff5;
sp6 = sl;
sp7 = sl + slb;
/*
* line0: aa
* line1: bb
* line2: cc
* line3: dd
* line4: ee
* line5: ff
* line4: g0 g1 g2 g3 g4 g5 g6
* line5: h0 h1 h2 h3 h4 h5 h6
*/
for (i = 0; i <= wid - 8; i += 8) {
g0 = *(__m64 *) sp6;
g1 = *(__m64 *) (sp6 + SSIZE);
g2 = *(__m64 *) (sp6 + 2 * SSIZE);
g3 = *(__m64 *) (sp6 + 3 * SSIZE);
g4 = *(__m64 *) (sp6 + 4 * SSIZE);
g5 = *(__m64 *) (sp6 + 5 * SSIZE);
g6 = *(__m64 *) (sp6 + 6 * SSIZE);
h0 = *(__m64 *) sp7;
h1 = *(__m64 *) (sp7 + SSIZE);
h2 = *(__m64 *) (sp7 + 2 * SSIZE);
h3 = *(__m64 *) (sp7 + 3 * SSIZE);
h4 = *(__m64 *) (sp7 + 4 * SSIZE);
h5 = *(__m64 *) (sp7 + 5 * SSIZE);
h6 = *(__m64 *) (sp7 + 6 * SSIZE);
gg = C_COMP(g0, g1);
hh = C_COMP(h0, h1);
g2 = C_COMP(g2, g3);
h2 = C_COMP(h2, h3);
g4 = C_COMP(g4, g5);
h4 = C_COMP(h4, h5);
gg = C_COMP(gg, g2);
hh = C_COMP(hh, h2);
gg = C_COMP(gg, g4);
hh = C_COMP(hh, h4);
gg = C_COMP(gg, g6);
hh = C_COMP(hh, h6);
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
ee = *(__m64 *) sp4;
ff = *(__m64 *) sp5;
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
ff = C_COMP(ff, gg);
bb = C_COMP(bb, dd);
bb = C_COMP(bb, ff);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, hh);
*(__m64 *) sp0 = gg;
*(__m64 *) sp1 = hh;
(*dp0++) = r0;
(*dp1++) = r1;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
sp4 += 8;
sp5 += 8;
sp6 += 8;
sp7 += 8;
}
if (tail) {
g0 = *(__m64 *) sp6;
g1 = *(__m64 *) (sp6 + SSIZE);
g2 = *(__m64 *) (sp6 + 2 * SSIZE);
g3 = *(__m64 *) (sp6 + 3 * SSIZE);
g4 = *(__m64 *) (sp6 + 4 * SSIZE);
g5 = *(__m64 *) (sp6 + 5 * SSIZE);
g6 = *(__m64 *) (sp6 + 6 * SSIZE);
h0 = *(__m64 *) sp7;
h1 = *(__m64 *) (sp7 + SSIZE);
h2 = *(__m64 *) (sp7 + 2 * SSIZE);
h3 = *(__m64 *) (sp7 + 3 * SSIZE);
h4 = *(__m64 *) (sp7 + 4 * SSIZE);
h5 = *(__m64 *) (sp7 + 5 * SSIZE);
h6 = *(__m64 *) (sp7 + 6 * SSIZE);
gg = C_COMP(g0, g1);
hh = C_COMP(h0, h1);
g2 = C_COMP(g2, g3);
h2 = C_COMP(h2, h3);
g4 = C_COMP(g4, g5);
h4 = C_COMP(h4, h5);
gg = C_COMP(gg, g2);
hh = C_COMP(hh, h2);
gg = C_COMP(gg, g4);
hh = C_COMP(hh, h4);
gg = C_COMP(gg, g6);
hh = C_COMP(hh, h6);
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
ee = *(__m64 *) sp4;
ff = *(__m64 *) sp5;
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
ff = C_COMP(ff, gg);
bb = C_COMP(bb, dd);
bb = C_COMP(bb, ff);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, hh);
*(__m64 *) sp0 = gg;
*(__m64 *) sp1 = hh;
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
*dp1 =
_mm_or_si64(_mm_and_si64(e_mask, r1),
_mm_andnot_si64(e_mask, *dp1));
}
buffT = buff0;
buff0 = buff2;
buff2 = buff4;
buff4 = buffT;
buffT = buff1;
buff1 = buff3;
buff3 = buff5;
buff5 = buffT;
sl += 2 * slb;
dl += 2 * dlb;
}
/* last line */
if (j == (hgt - KSIZE1 - 1)) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff0;
sp1 = buff1;
sp2 = buff2;
sp3 = buff3;
sp4 = buff4;
sp5 = buff5;
sp6 = sl;
for (i = 0; i <= wid - 8; i += 8) {
g0 = *(__m64 *) sp6;
g1 = *(__m64 *) (sp6 + SSIZE);
g2 = *(__m64 *) (sp6 + 2 * SSIZE);
g3 = *(__m64 *) (sp6 + 3 * SSIZE);
g4 = *(__m64 *) (sp6 + 4 * SSIZE);
g5 = *(__m64 *) (sp6 + 5 * SSIZE);
g6 = *(__m64 *) (sp6 + 6 * SSIZE);
gg = C_COMP(g0, g1);
g2 = C_COMP(g2, g3);
g4 = C_COMP(g4, g5);
gg = C_COMP(gg, g2);
gg = C_COMP(gg, g4);
gg = C_COMP(gg, g6);
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
ee = *(__m64 *) sp4;
ff = *(__m64 *) sp5;
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
ff = C_COMP(ff, gg);
bb = C_COMP(bb, dd);
bb = C_COMP(bb, ff);
r0 = C_COMP(aa, bb);
(*dp0++) = r0;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
sp4 += 8;
sp5 += 8;
sp6 += 8;
}
if (tail) {
g0 = *(__m64 *) sp6;
g1 = *(__m64 *) (sp6 + SSIZE);
g2 = *(__m64 *) (sp6 + 2 * SSIZE);
g3 = *(__m64 *) (sp6 + 3 * SSIZE);
g4 = *(__m64 *) (sp6 + 4 * SSIZE);
g5 = *(__m64 *) (sp6 + 5 * SSIZE);
g6 = *(__m64 *) (sp6 + 6 * SSIZE);
gg = C_COMP(g0, g1);
g2 = C_COMP(g2, g3);
g4 = C_COMP(g4, g5);
gg = C_COMP(gg, g2);
gg = C_COMP(gg, g4);
gg = C_COMP(gg, g6);
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
ee = *(__m64 *) sp4;
ff = *(__m64 *) sp5;
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
ff = C_COMP(ff, gg);
bb = C_COMP(bb, dd);
bb = C_COMP(bb, ff);
r0 = C_COMP(aa, bb);
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
}
}
_mm_empty();
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
mlib_status
mlib_m_conv5x5_u16nw_2(
mlib_image *dst,
mlib_image *src,
mlib_s32 *kern,
mlib_s32 scalef_expon)
{
__m64 *pbuff, *buff_arr[20], **pbuff_arr = buff_arr;
__m64 *buff0, *buff1, *buff2, *buff3;
GET_SRC_DST_PARAMETERS(mlib_s16);
__m64 ker[5][5];
__m64 d0, d1, d2, aa, bb, rr, tmpa, tmpb, ker_off, mask8000;
__m64 prev0h, prev1h, prev2h, prev3h, sum0h, sum1h, sum2h, sum3h, sum4h,
tmph;
__m64 prev0l, prev1l, prev2l, prev3l, sum0l, sum1l, sum2l, sum3l, sum4l,
tmpl;
__m64 *sp, *dp;
mlib_s32 shift, ind, ker_sum = 0;
mlib_s32 row, wid4, i, j;
width -= 4;
height -= 4;
width *= NCHAN;
dl += 2 * (dll + NCHAN);
wid4 = (width + 7) / 4;
pbuff = mlib_malloc(sizeof (__m64) * 20 * wid4);
GET_KERN();
for (i = 0; i < 10; i++) {
buff_arr[i] = pbuff + i * 2 * wid4;
}
ind = 0;
for (j = 1; j <= 4; j++) {
buff0 = buff_arr[ind];
buff1 = buff_arr[ind + 1];
buff2 = buff_arr[ind + 2];
buff3 = buff_arr[ind + 3];
sp = (__m64 *) sl;
d1 = (*sp++);
d1 = _mm_xor_si64(d1, mask8000);
d2 = (*sp++);
d2 = _mm_xor_si64(d2, mask8000);
for (i = 0; i < wid4; i++) {
PREP_5x5();
}
sl += sll;
ind += j;
}
for (row = 0; row < height; row++) {
sp = (__m64 *) sl;
dp = (__m64 *) dl;
buff0 = pbuff_arr[0];
buff1 = pbuff_arr[2];
buff2 = pbuff_arr[5];
buff3 = pbuff_arr[9];
d1 = (*sp++);
d1 = _mm_xor_si64(d1, mask8000);
d2 = (*sp++);
d2 = _mm_xor_si64(d2, mask8000);
for (i = 0; i < width / 4; i++) {
CONV_5x5(hi, i);
dp[i] = rr;
}
if (width & 3) {
__m64 mask =
((__m64 *) mlib_mask64_arr)[2 * (width & 3)];
CONV_5x5(hi, i);
dp[i] =
_mm_or_si64(_mm_and_si64(mask, rr),
_mm_andnot_si64(mask, dp[i]));
}
ind = (pbuff_arr == buff_arr) ? 10 : -10;
pbuff_arr[ind + 0] = pbuff_arr[1];
pbuff_arr[ind + 1] = pbuff_arr[3];
pbuff_arr[ind + 2] = pbuff_arr[4];
pbuff_arr[ind + 3] = pbuff_arr[6];
pbuff_arr[ind + 4] = pbuff_arr[7];
pbuff_arr[ind + 5] = pbuff_arr[8];
pbuff_arr[ind + 6] = pbuff_arr[0];
pbuff_arr[ind + 7] = pbuff_arr[2];
pbuff_arr[ind + 8] = pbuff_arr[5];
pbuff_arr[ind + 9] = pbuff_arr[9];
pbuff_arr += ind;
sl += sll;
dl += dll;
}
_mm_empty();
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
mlib_status
mlib_m_sconv5x5_8nw_2(
mlib_image *dst,
mlib_image *src,
mlib_s32 *hkernel,
mlib_s32 *vkernel,
mlib_s32 scalef_expon)
{
__m64 *pbuff, *buff_arr[5];
__m64 *buff0, *buff1, *buff2, *buff3, *buff4, *buffT;
GET_SRC_DST_PARAMETERS(mlib_u8);
__m64 hker0, hker1, hker2, hker3, hker4;
__m64 vker0, vker1, vker2, vker3, vker4;
__m64 s0, d0, d1, d2, prev0;
__m64 sum0, sum1, sum2, sum3, sum4, aa, bb, res_hi, res_lo;
__m64 zero = _m_zero;
mlib_s32 shift, ind;
mlib_s32 *sp;
mlib_s32 row, wid4, i, j;
width -= 4;
height -= 4;
width *= NCHAN;
dl += 2 * (dll + NCHAN);
wid4 = 2 * ((width + 7) / 8);
pbuff = mlib_malloc(sizeof (__m64) * 5 * wid4);
GET_KERN();
for (i = 0; i < 5; i++) {
buff_arr[i] = pbuff + i * wid4;
}
for (j = 0; j < 4; j++) {
buff4 = buff_arr[j];
sp = (mlib_s32 *)sl;
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d1, lo);
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d2, lo);
for (i = 0; i < wid4; i++) {
*(mlib_s32 *)&s0 = sp[i];
PREP_5x5(lo, i);
}
sl += sll;
ind++;
}
buff0 = buff_arr[0];
buff1 = buff_arr[1];
buff2 = buff_arr[2];
buff3 = buff_arr[3];
buff4 = buff_arr[4];
for (row = 0; row < height; row++) {
__m64 *sp = (__m64 *) sl;
__m64 *dp = (__m64 *) dl;
s0 = (*sp++);
UNPACK_SRC(d1, lo);
UNPACK_SRC(d2, hi);
for (i = 0; i < width / 8; i++) {
s0 = sp[i];
CONV_5x5(lo, 2 * i);
CONV_5x5(hi, 2 * i + 1);
dp[i] = _mm_packs_pu16(res_lo, res_hi);
}
if (width & 7) {
__m64 mask = ((__m64 *) mlib_mask64_arr)[width & 7];
s0 = sp[i];
CONV_5x5(lo, 2 * i);
CONV_5x5(hi, 2 * i + 1);
res_hi = _mm_packs_pu16(res_lo, res_hi);
dp[i] =
_mm_or_si64(_mm_and_si64(mask, res_hi),
_mm_andnot_si64(mask, dp[i]));
}
buffT = buff0;
buff0 = buff1;
buff1 = buff2;
buff2 = buff3;
buff3 = buff4;
buff4 = buffT;
sl += sll;
dl += dll;
}
_mm_empty();
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
__m64 test54(__m64 a, __m64 b) {
// CHECK: pandn
return _mm_andnot_si64(a, b);
}
void lines_scale2(const unsigned char *src, unsigned y, unsigned char *dst1, unsigned char *dst2, unsigned nPix)
{
const unsigned char
*u = src + ((y-1) & 7)*sc2lines_width,
*m = src + ((y+0) & 7)*sc2lines_width,
*l = src + ((y+1) & 7)*sc2lines_width;
for (unsigned i = 0; i < nPix; i += 4) {
if (*(unsigned*)(u+i) ^ *(unsigned*)(l+i)) {
__m64 mm = *(__m64*)(m+i-2);
__m64 uu = *(__m64*)(u+i-2);
__m64 ll = *(__m64*)(l+i-2);
__m64 md = _mm_slli_si64(mm,8);
__m64 mf = _mm_srli_si64(mm,8);
__m64 maskall = _mm_or_si64(_mm_cmpeq_pi8(md,mf), _mm_cmpeq_pi8(uu,ll));
__m64 e0, e1, v1, v2;
e0 = _mm_cmpeq_pi8(md,uu);
e0 = _mm_andnot_si64(maskall, e0);
e0 = _mm_srli_si64(e0,16);
e0 = _mm_unpacklo_pi8(e0, _mm_setzero_si64());
e1 = _mm_cmpeq_pi8(mf,uu);
e1 = _mm_andnot_si64(maskall, e1);
e1 = _mm_srli_si64(e1,16);
e1 = _mm_unpacklo_pi8(_mm_setzero_si64(), e1);
e0 = _mm_or_si64(e0, e1);
v1 = _m_from_int(*(unsigned*)(m+i));
v2 = _m_from_int(*(unsigned*)(u+i));
v1 = _mm_unpacklo_pi8(v1,v1);
v2 = _mm_unpacklo_pi8(v2,v2);
*(__m64*)(dst1 + 2*i) = _mm_or_si64( _mm_and_si64(e0,v2), _mm_andnot_si64(e0,v1) );
e0 = _mm_cmpeq_pi8(md,ll);
e0 = _mm_andnot_si64(maskall, e0);
e0 = _mm_srli_si64(e0,16);
e0 = _mm_unpacklo_pi8(e0, _mm_setzero_si64());
e1 = _mm_cmpeq_pi8(mf,ll);
e1 = _mm_andnot_si64(maskall, e1);
e1 = _mm_srli_si64(e1,16);
e1 = _mm_unpacklo_pi8(_mm_setzero_si64(), e1);
e0 = _mm_or_si64(e0, e1);
v1 = _m_from_int(*(unsigned*)(m+i));
v2 = _m_from_int(*(unsigned*)(l+i));
v1 = _mm_unpacklo_pi8(v1,v1);
v2 = _mm_unpacklo_pi8(v2,v2);
*(__m64*)(dst2 + 2*i) = _mm_or_si64( _mm_and_si64(e0,v2), _mm_andnot_si64(e0,v1) );
} else {
__m64 v1 = _m_from_int(*(unsigned*)(m+i));
v1 = _mm_unpacklo_pi8(v1,v1);
*(__m64*)(dst1 + 2*i) = v1;
*(__m64*)(dst2 + 2*i) = v1;
}
}
}
/* *********************************************************** */
mlib_status
mlib_m_sconv3x3_8nw_1(
mlib_image *dst,
mlib_image *src,
mlib_s32 *hkernel,
mlib_s32 *vkernel,
mlib_s32 scalef_expon)
{
__m64 buff_loc[3 * BUFF_LINE], *pbuff = buff_loc;
__m64 *buff0, *buff1, *buffT;
GET_SRC_DST_PARAMETERS(mlib_u8);
__m64 hker0, hker1, hker2, vker0, vker1, vker2;
__m64 s0, d0, d1, sum0, sum1, sum2, aa, bb, res_hi, res_lo;
__m64 zero = _m_zero;
mlib_s32 shift;
mlib_s32 *sp;
mlib_s32 row, wid4, i, j;
width -= 2;
height -= 2;
dl += dll + 1;
wid4 = (width + 7) / 4;
if (wid4 > BUFF_LINE) {
pbuff = mlib_malloc(sizeof (__m64) * 3 * wid4);
}
GET_KERN();
buff0 = pbuff;
buff1 = buff0 + wid4;
for (j = 0; j < 2; j++) {
sp = (mlib_s32 *)sl;
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d1, lo);
for (i = 0; i < wid4; i++) {
*(mlib_s32 *)&s0 = sp[i];
PREP_3x3_1ch(lo, i);
}
sl += sll;
buffT = buff1;
buff1 = buff0;
buff0 = buffT;
}
for (row = 0; row < height; row++) {
__m64 *sp = (__m64 *) sl;
__m64 *dp = (__m64 *) dl;
s0 = (*sp++);
UNPACK_SRC(d1, lo);
for (i = 0; i < width / 8; i++) {
CONV_3x3_1ch(hi, 2 * i);
s0 = sp[i];
CONV_3x3_1ch(lo, 2 * i + 1);
dp[i] = _mm_packs_pu16(res_hi, res_lo);
}
if (width & 7) {
__m64 mask;
mask = ((__m64 *) mlib_mask64_arr)[width & 7];
CONV_3x3_1ch(hi, 2 * i);
s0 = sp[i];
CONV_3x3_1ch(lo, 2 * i + 1);
res_hi = _mm_packs_pu16(res_hi, res_lo);
dp[i] =
_mm_or_si64(_mm_and_si64(mask, res_hi),
_mm_andnot_si64(mask, dp[i]));
}
buffT = buff1;
buff1 = buff0;
buff0 = buffT;
sl += sll;
dl += dll;
}
_mm_empty();
if (pbuff != buff_loc)
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
mlib_status
mlib_ImageErode4_U16(
void *dst,
void *src,
mlib_s32 dlb,
mlib_s32 slb,
mlib_s32 wid,
mlib_s32 hgt)
#endif /* DILATE_FILTER */
{
mlib_u8 *sl, *sp0, *sp1, *sp2, *sp3, *dl;
__m64 *dp0, *dp1;
__m64 a1, b0, b1, b2, c0, c1, c2, d1, vv, h0, h1, r0, r1;
__m64 mask, mask80;
mlib_s32 i, j, tail;
sl = (mlib_u8 *)src;
/* dst ptrs skip top j and left col */
dl = (mlib_u8 *)dst + dlb + SSIZE;
wid = (wid - 2) * SSIZE;
tail = wid & 7;
mask = ((__m64 *) mlib_mask64_arr)[tail];
mask80 = mmx_from_int_dup(0x80008000);
for (j = 0; j <= (hgt - 2 - 2); j += 2) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = sl;
sp1 = sp0 + slb;
sp2 = sp1 + slb;
sp3 = sp2 + slb;
/*
* line0: a1
* line1: b0 b1 b2
* line2: c0 c1 c2
* line3: d1
*/
for (i = 0; i <= wid - 8; i += 8) {
a1 = *(__m64 *) (sp0 + SSIZE);
b0 = *(__m64 *) (sp1);
b1 = *(__m64 *) (sp1 + SSIZE);
b2 = *(__m64 *) (sp1 + 2 * SSIZE);
c0 = *(__m64 *) (sp2);
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
d1 = *(__m64 *) (sp3 + SSIZE);
vv = C_COMP(b1, c1);
h0 = C_COMP(b0, b2);
h1 = C_COMP(c0, c2);
r0 = C_COMP(vv, a1);
r1 = C_COMP(vv, d1);
r0 = C_COMP(r0, h0);
r1 = C_COMP(r1, h1);
(*dp0++) = r0;
(*dp1++) = r1;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
}
if (tail) {
a1 = *(__m64 *) (sp0 + SSIZE);
b0 = *(__m64 *) (sp1);
b1 = *(__m64 *) (sp1 + SSIZE);
b2 = *(__m64 *) (sp1 + 2 * SSIZE);
c0 = *(__m64 *) (sp2);
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
d1 = *(__m64 *) (sp3 + SSIZE);
vv = C_COMP(b1, c1);
h0 = C_COMP(b0, b2);
h1 = C_COMP(c0, c2);
r0 = C_COMP(vv, a1);
r1 = C_COMP(vv, d1);
r0 = C_COMP(r0, h0);
r1 = C_COMP(r1, h1);
*dp0 =
_mm_or_si64(_mm_and_si64(mask, r0),
_mm_andnot_si64(mask, *dp0));
*dp1 =
_mm_or_si64(_mm_and_si64(mask, r1),
_mm_andnot_si64(mask, *dp1));
}
sl += 2 * slb;
dl += 2 * dlb;
}
/* last line */
if (j == (hgt - 3)) {
dp0 = (void *)dl;
sp0 = sl;
sp1 = sp0 + slb;
sp2 = sp1 + slb;
for (i = 0; i <= wid - 8; i += 8) {
a1 = *(__m64 *) (sp0 + SSIZE);
b0 = *(__m64 *) (sp1);
b1 = *(__m64 *) (sp1 + SSIZE);
b2 = *(__m64 *) (sp1 + 2 * SSIZE);
c1 = *(__m64 *) (sp2 + SSIZE);
vv = C_COMP(b1, c1);
h0 = C_COMP(b0, b2);
r0 = C_COMP(vv, a1);
r0 = C_COMP(r0, h0);
(*dp0++) = r0;
sp0 += 8;
sp1 += 8;
sp2 += 8;
}
if (tail) {
a1 = *(__m64 *) (sp0 + SSIZE);
b0 = *(__m64 *) (sp1);
b1 = *(__m64 *) (sp1 + SSIZE);
b2 = *(__m64 *) (sp1 + 2 * SSIZE);
c1 = *(__m64 *) (sp2 + SSIZE);
vv = C_COMP(b1, c1);
h0 = C_COMP(b0, b2);
r0 = C_COMP(vv, a1);
r0 = C_COMP(r0, h0);
*dp0 =
_mm_or_si64(_mm_and_si64(mask, r0),
_mm_andnot_si64(mask, *dp0));
}
}
_mm_empty();
return (MLIB_SUCCESS);
}
mlib_status
mlib_ImageMinFilter5x5_U8(
void *dst,
void *src,
mlib_s32 dlb,
mlib_s32 slb,
mlib_s32 wid,
mlib_s32 hgt)
#endif /* MAX_FILTER */
{
mlib_u8 *pbuff, *buff0, *buff1, *buff2, *buff3, *buffT;
mlib_u8 *sl, *sp0, *sp1, *sp2, *sp3, *sp4, *sp5, *dl;
__m64 *dp0, *dp1;
__m64 aa, bb, cc, dd, e0, e1, e2, e3, e4, ee, f0, f1, f2, f3, f4, ff,
r0, r1;
__m64 e_mask, mask8080;
mlib_s32 i, j, wid8, tail;
wid = (wid - KSIZE1) * SSIZE;
wid8 = (wid + 7) & ~7;
pbuff = mlib_malloc(4 * wid8);
buff0 = pbuff;
buff1 = buff0 + wid8;
buff2 = buff1 + wid8;
buff3 = buff2 + wid8;
sl = (mlib_u8 *)src;
dl = (mlib_u8 *)dst + 2 * (dlb + SSIZE);
tail = wid & 7;
e_mask = ((__m64 *) mlib_mask64_arr)[tail];
mask8080 = mmx_from_int_dup(0x80808080);
for (j = 0; j < 2; j++) {
sp0 = buff0;
sp1 = buff1;
sp4 = sl;
sp5 = sl + slb;
sl += 2 * slb;
for (i = 0; i < wid; i += 8) {
e0 = *(__m64 *) sp4;
e1 = *(__m64 *) (sp4 + SSIZE);
e2 = *(__m64 *) (sp4 + 2 * SSIZE);
e3 = *(__m64 *) (sp4 + 3 * SSIZE);
e4 = *(__m64 *) (sp4 + 4 * SSIZE);
f0 = *(__m64 *) sp5;
f1 = *(__m64 *) (sp5 + SSIZE);
f2 = *(__m64 *) (sp5 + 2 * SSIZE);
f3 = *(__m64 *) (sp5 + 3 * SSIZE);
f4 = *(__m64 *) (sp5 + 4 * SSIZE);
ee = C_COMP(e0, e1);
ff = C_COMP(f0, f1);
e2 = C_COMP(e2, e3);
f2 = C_COMP(f2, f3);
ee = C_COMP(ee, e4);
ff = C_COMP(ff, f4);
ee = C_COMP(ee, e2);
ff = C_COMP(ff, f2);
*(__m64 *) sp0 = ee;
*(__m64 *) sp1 = ff;
sp0 += 8;
sp1 += 8;
sp4 += 8;
sp5 += 8;
}
buffT = buff0;
buff0 = buff2;
buff2 = buffT;
buffT = buff1;
buff1 = buff3;
buff3 = buffT;
}
for (j = 0; j <= (hgt - KSIZE1 - 2); j += 2) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff0;
sp1 = buff1;
sp2 = buff2;
sp3 = buff3;
sp4 = sl;
sp5 = sl + slb;
/*
* line0: aa
* line1: bb
* line2: cc
* line3: dd
* line4: e0 e1 e2 e3 e4
* line5: f0 f1 f2 f3 f4
*/
for (i = 0; i <= wid - 8; i += 8) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
e0 = *(__m64 *) sp4;
e1 = *(__m64 *) (sp4 + SSIZE);
e2 = *(__m64 *) (sp4 + 2 * SSIZE);
e3 = *(__m64 *) (sp4 + 3 * SSIZE);
e4 = *(__m64 *) (sp4 + 4 * SSIZE);
f0 = *(__m64 *) sp5;
f1 = *(__m64 *) (sp5 + SSIZE);
f2 = *(__m64 *) (sp5 + 2 * SSIZE);
f3 = *(__m64 *) (sp5 + 3 * SSIZE);
f4 = *(__m64 *) (sp5 + 4 * SSIZE);
ee = C_COMP(e0, e1);
ff = C_COMP(f0, f1);
e2 = C_COMP(e2, e3);
f2 = C_COMP(f2, f3);
ee = C_COMP(ee, e4);
ff = C_COMP(ff, f4);
ee = C_COMP(ee, e2);
ff = C_COMP(ff, f2);
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
bb = C_COMP(bb, dd);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, ff);
*(__m64 *) sp0 = ee;
*(__m64 *) sp1 = ff;
(*dp0++) = r0;
(*dp1++) = r1;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
sp4 += 8;
sp5 += 8;
}
if (tail) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
e0 = *(__m64 *) sp4;
e1 = *(__m64 *) (sp4 + SSIZE);
e2 = *(__m64 *) (sp4 + 2 * SSIZE);
e3 = *(__m64 *) (sp4 + 3 * SSIZE);
e4 = *(__m64 *) (sp4 + 4 * SSIZE);
f0 = *(__m64 *) sp5;
f1 = *(__m64 *) (sp5 + SSIZE);
f2 = *(__m64 *) (sp5 + 2 * SSIZE);
f3 = *(__m64 *) (sp5 + 3 * SSIZE);
f4 = *(__m64 *) (sp5 + 4 * SSIZE);
ee = C_COMP(e0, e1);
ff = C_COMP(f0, f1);
e2 = C_COMP(e2, e3);
f2 = C_COMP(f2, f3);
ee = C_COMP(ee, e4);
ff = C_COMP(ff, f4);
ee = C_COMP(ee, e2);
ff = C_COMP(ff, f2);
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
bb = C_COMP(bb, dd);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, ff);
*(__m64 *) sp0 = ee;
*(__m64 *) sp1 = ff;
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
*dp1 =
_mm_or_si64(_mm_and_si64(e_mask, r1),
_mm_andnot_si64(e_mask, *dp1));
}
buffT = buff0;
buff0 = buff2;
buff2 = buffT;
buffT = buff1;
buff1 = buff3;
buff3 = buffT;
sl += 2 * slb;
dl += 2 * dlb;
}
/* last line */
if (j == (hgt - KSIZE1 - 1)) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff0;
sp1 = buff1;
sp2 = buff2;
sp3 = buff3;
sp4 = sl;
for (i = 0; i <= wid - 8; i += 8) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
e0 = *(__m64 *) sp4;
e1 = *(__m64 *) (sp4 + SSIZE);
e2 = *(__m64 *) (sp4 + 2 * SSIZE);
e3 = *(__m64 *) (sp4 + 3 * SSIZE);
e4 = *(__m64 *) (sp4 + 4 * SSIZE);
ee = C_COMP(e0, e1);
e2 = C_COMP(e2, e3);
ee = C_COMP(ee, e4);
ee = C_COMP(ee, e2);
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
bb = C_COMP(bb, dd);
r0 = C_COMP(aa, bb);
(*dp0++) = r0;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
sp4 += 8;
}
if (tail) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
cc = *(__m64 *) sp2;
dd = *(__m64 *) sp3;
e0 = *(__m64 *) sp4;
e1 = *(__m64 *) (sp4 + SSIZE);
e2 = *(__m64 *) (sp4 + 2 * SSIZE);
e3 = *(__m64 *) (sp4 + 3 * SSIZE);
e4 = *(__m64 *) (sp4 + 4 * SSIZE);
ee = C_COMP(e0, e1);
e2 = C_COMP(e2, e3);
ee = C_COMP(ee, e4);
ee = C_COMP(ee, e2);
bb = C_COMP(bb, cc);
dd = C_COMP(dd, ee);
bb = C_COMP(bb, dd);
r0 = C_COMP(aa, bb);
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
}
}
_mm_empty();
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos_ps(v4sfu *xptr, v4sfu *sptr, v4sfu *cptr) {
__m128 x=*((__m128 *)xptr), *s=(__m128 *)sptr, *c=(__m128 *)cptr, xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
#ifdef USE_SSE2
__m128i emm0, emm2, emm4;
#else
__m64 mm0, mm1, mm2, mm3, mm4, mm5;
#endif
sign_bit_sin = x;
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm_and_ps(sign_bit_sin, *(__m128*)_ps_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in emm2 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm4 = emm2;
/* get the swap sign flag for the sine */
emm0 = _mm_and_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
__m128 swap_sign_bit_sin = _mm_castsi128_ps(emm0);
/* get the polynom selection mask for the sine*/
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm2:mm3 */
xmm3 = _mm_movehl_ps(xmm3, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm3);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm4 = mm2;
mm5 = mm3;
/* get the swap sign flag for the sine */
mm0 = _mm_and_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_and_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
__m128 swap_sign_bit_sin;
COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit_sin);
/* get the polynom selection mask for the sine */
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 poly_mask;
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
#ifdef USE_SSE2
emm4 = _mm_sub_epi32(emm4, *(__m128i*)_pi32_2);
emm4 = _mm_andnot_si128(emm4, *(__m128i*)_pi32_4);
emm4 = _mm_slli_epi32(emm4, 29);
__m128 sign_bit_cos = _mm_castsi128_ps(emm4);
#else
/* get the sign flag for the cosine */
mm4 = _mm_sub_pi32(mm4, *(__m64*)_pi32_2);
mm5 = _mm_sub_pi32(mm5, *(__m64*)_pi32_2);
mm4 = _mm_andnot_si64(mm4, *(__m64*)_pi32_4);
mm5 = _mm_andnot_si64(mm5, *(__m64*)_pi32_4);
mm4 = _mm_slli_pi32(mm4, 29);
mm5 = _mm_slli_pi32(mm5, 29);
__m128 sign_bit_cos;
COPY_MM_TO_XMM(mm4, mm5, sign_bit_cos);
_mm_empty(); /* good-bye mmx */
#endif
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
__m128 z = _mm_mul_ps(x,x);
y = *(__m128*)_ps_coscof_p0;
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
__m128 ysin2 = _mm_and_ps(xmm3, y2);
__m128 ysin1 = _mm_andnot_ps(xmm3, y);
y2 = _mm_sub_ps(y2,ysin2);
y = _mm_sub_ps(y, ysin1);
xmm1 = _mm_add_ps(ysin1,ysin2);
xmm2 = _mm_add_ps(y,y2);
/* update the sign */
*s = _mm_xor_ps(xmm1, sign_bit_sin);
*c = _mm_xor_ps(xmm2, sign_bit_cos);
}
/* almost the same as sin_ps */
__m128 cos_ps(v4sfu *xPtr) { // any x
__m128 x=*((__m128 *)xPtr);
__m128 xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
#ifdef USE_SSE2
__m128i emm0, emm2;
#else
__m64 mm0, mm1, mm2, mm3;
#endif
/* take the absolute value */
x = _mm_and_ps(x, *(__m128*)_ps_inv_sign_mask);
/* scale by 4/Pi */
y = _mm_mul_ps(x, *(__m128*)_ps_cephes_FOPI);
#ifdef USE_SSE2
/* store the integer part of y in mm0 */
emm2 = _mm_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm_add_epi32(emm2, *(__m128i*)_pi32_1);
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_inv1);
y = _mm_cvtepi32_ps(emm2);
emm2 = _mm_sub_epi32(emm2, *(__m128i*)_pi32_2);
/* get the swap sign flag */
emm0 = _mm_andnot_si128(emm2, *(__m128i*)_pi32_4);
emm0 = _mm_slli_epi32(emm0, 29);
/* get the polynom selection mask */
emm2 = _mm_and_si128(emm2, *(__m128i*)_pi32_2);
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
__m128 sign_bit = _mm_castsi128_ps(emm0);
__m128 poly_mask = _mm_castsi128_ps(emm2);
#else
/* store the integer part of y in mm0:mm1 */
xmm2 = _mm_movehl_ps(xmm2, y);
mm2 = _mm_cvttps_pi32(y);
mm3 = _mm_cvttps_pi32(xmm2);
/* j=(j+1) & (~1) (see the cephes sources) */
mm2 = _mm_add_pi32(mm2, *(__m64*)_pi32_1);
mm3 = _mm_add_pi32(mm3, *(__m64*)_pi32_1);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_inv1);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_inv1);
y = _mm_cvtpi32x2_ps(mm2, mm3);
mm2 = _mm_sub_pi32(mm2, *(__m64*)_pi32_2);
mm3 = _mm_sub_pi32(mm3, *(__m64*)_pi32_2);
/* get the swap sign flag in mm0:mm1 and the
polynom selection mask in mm2:mm3 */
mm0 = _mm_andnot_si64(mm2, *(__m64*)_pi32_4);
mm1 = _mm_andnot_si64(mm3, *(__m64*)_pi32_4);
mm0 = _mm_slli_pi32(mm0, 29);
mm1 = _mm_slli_pi32(mm1, 29);
mm2 = _mm_and_si64(mm2, *(__m64*)_pi32_2);
mm3 = _mm_and_si64(mm3, *(__m64*)_pi32_2);
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
__m128 sign_bit, poly_mask;
COPY_MM_TO_XMM(mm0, mm1, sign_bit);
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
_mm_empty(); /* good-bye mmx */
#endif
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m128*)_ps_minus_cephes_DP1;
xmm2 = *(__m128*)_ps_minus_cephes_DP2;
xmm3 = *(__m128*)_ps_minus_cephes_DP3;
xmm1 = _mm_mul_ps(y, xmm1);
xmm2 = _mm_mul_ps(y, xmm2);
xmm3 = _mm_mul_ps(y, xmm3);
x = _mm_add_ps(x, xmm1);
x = _mm_add_ps(x, xmm2);
x = _mm_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(__m128*)_ps_coscof_p0;
__m128 z = _mm_mul_ps(x,x);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p1);
y = _mm_mul_ps(y, z);
y = _mm_add_ps(y, *(__m128*)_ps_coscof_p2);
y = _mm_mul_ps(y, z);
y = _mm_mul_ps(y, z);
__m128 tmp = _mm_mul_ps(z, *(__m128*)_ps_0p5);
y = _mm_sub_ps(y, tmp);
y = _mm_add_ps(y, *(__m128*)_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m128 y2 = *(__m128*)_ps_sincof_p0;
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p1);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_add_ps(y2, *(__m128*)_ps_sincof_p2);
y2 = _mm_mul_ps(y2, z);
y2 = _mm_mul_ps(y2, x);
y2 = _mm_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
y = _mm_andnot_ps(xmm3, y);
y = _mm_add_ps(y,y2);
/* update the sign */
y = _mm_xor_ps(y, sign_bit);
return y;
}
mlib_status
mlib_m_conv3x3_16nw_4(
mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kern,
mlib_s32 scalef_expon)
{
__m64 buff_loc[6 * BUFF_LINE], *pbuff = buff_loc;
__m64 *buff0, *buff1, *buff2, *buffT;
GET_SRC_DST_PARAMETERS(mlib_s16);
__m64 ker1, ker2, ker3, ker4, ker5, ker6, ker7, ker8, ker9;
__m64 d0, d1, d2, rr, tmpa, tmpb;
__m64 prev0h, prev1h, sum0h, sum1h, sum2h, tmph;
__m64 prev0l, prev1l, sum0l, sum1l, sum2l, tmpl;
__m64 *sp, *dp;
mlib_s32 shift;
mlib_s32 row, wid4, i, j;
width -= 2;
height -= 2;
width *= NCHAN;
dl += dll + NCHAN;
wid4 = (width + 3) / 4;
if (wid4 > BUFF_LINE) {
pbuff = mlib_malloc(sizeof (__m64) * 6 * wid4);
}
GET_KERN();
buff0 = pbuff;
buff1 = buff0 + 2 * wid4;
buff2 = buff1 + 2 * wid4;
for (j = 0; j < 2; j++) {
sp = (__m64 *) sl;
d1 = (*sp++);
d2 = (*sp++);
for (i = 0; i < wid4; i++) {
PREP_3x3(i);
}
sl += sll;
if (j == 0) {
buffT = buff1;
buff1 = buff0;
buff0 = buffT;
}
}
for (row = 0; row < height; row++) {
sp = (__m64 *) sl;
dp = (__m64 *) dl;
d1 = (*sp++);
d2 = (*sp++);
for (i = 0; i < width / 4; i++) {
CONV_3x3(i);
dp[i] = rr;
}
if (width & 3) {
__m64 mask =
((__m64 *) mlib_mask64_arr)[2 * (width & 3)];
CONV_3x3(i);
dp[i] =
_mm_or_si64(_mm_and_si64(mask, rr),
_mm_andnot_si64(mask, dp[i]));
}
buffT = buff1;
buff1 = buff0;
buff0 = buffT;
sl += sll;
dl += dll;
}
_mm_empty();
if (pbuff != buff_loc)
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
__m64 test_mm_andnot_si64(__m64 a, __m64 b) {
// CHECK-LABEL: test_mm_andnot_si64
// CHECK: call x86_mmx @llvm.x86.mmx.pandn
return _mm_andnot_si64(a, b);
}
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();
}
mlib_status
mlib_m_sconv5x5_u16nw_3(
mlib_image *dst,
mlib_image *src,
mlib_s32 *hkernel,
mlib_s32 *vkernel,
mlib_s32 scalef_expon)
{
GET_SRC_DST_PARAMETERS(mlib_s16);
__m64 hker0, hker1, hker2, hker3, hker4;
__m64 vker0, vker1, vker2, vker3, vker4;
__m64 s0, s1, s2, s3, s4, v0, v1, v2, v3, rr, rh, rl;
__m64 aa, bb, cc, zero, ker_off, mask8000;
__m64 *sp0, *sp1, *sp2, *sp3, *sp4, *dp;
mlib_s32 shift, ker_sum, kerh_sum = 0, kerv_sum = 0;
mlib_s32 i, j;
width -= 4;
height -= 4;
width *= NCHAN;
dl += 2 * (dll + NCHAN);
GET_KERN();
zero = _mm_setzero_si64();
for (j = 0; j < height; j++) {
sp0 = (__m64 *) sl;
sp1 = (__m64 *) (sl + sll);
sp2 = (__m64 *) (sl + 2 * sll);
sp3 = (__m64 *) (sl + 3 * sll);
sp4 = (__m64 *) (sl + 4 * sll);
dp = (__m64 *) dl;
PREP_V();
for (i = 0; i < width / 4; i++) {
CONV_5x5();
dp[i] = rr;
}
if (width & 3) {
__m64 mask =
((__m64 *) mlib_mask64_arr)[2 * (width & 3)];
CONV_5x5();
dp[i] =
_mm_or_si64(_mm_and_si64(mask, rr),
_mm_andnot_si64(mask, dp[i]));
}
sl += sll;
dl += dll;
}
_mm_empty();
return (MLIB_SUCCESS);
}
mlib_status
mlib_m_conv5x5_8nw_4(
mlib_image *dst,
mlib_image *src,
mlib_s32 *kern,
mlib_s32 scalef_expon)
{
__m64 *pbuff, *buff_arr[20], **pbuff_arr = buff_arr;
__m64 *buff0, *buff1, *buff2, *buff3;
GET_SRC_DST_PARAMETERS(mlib_u8);
__m64 ker[5][5];
__m64 s0, d0, d1, d2, d3, d4, prev0, prev1, prev2, prev3, aa, bb, cc;
__m64 sum0, sum1, sum2, sum3, sum4, res_hi, res_lo;
__m64 zero = _m_zero;
mlib_s32 shift, ind;
mlib_s32 *sp;
mlib_s32 row, wid4, i, j;
width -= (KSIZE - 1);
height -= (KSIZE - 1);
width *= NCHAN;
dl += ((KSIZE - 1) / 2) * (dll + NCHAN);
wid4 = (width + 7) / 4;
pbuff = mlib_malloc(sizeof (__m64) * 10 * wid4);
GET_KERN();
for (i = 0; i < 10; i++) {
buff_arr[i] = pbuff + i * wid4;
}
ind = 0;
for (j = 1; j <= 4; j++) {
buff0 = buff_arr[ind];
buff1 = buff_arr[ind + 1];
buff2 = buff_arr[ind + 2];
buff3 = buff_arr[ind + 3];
sp = (mlib_s32 *)sl;
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d1, lo);
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d2, lo);
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d3, lo);
*(mlib_s32 *)&s0 = (*sp++);
UNPACK_SRC(d4, lo);
for (i = 0; i < wid4; i++) {
*(mlib_s32 *)&s0 = sp[i];
PREP_5x5();
}
sl += sll;
ind += j;
}
for (row = 0; row < height; row++) {
__m64 *sp = (__m64 *) sl;
__m64 *dp = (__m64 *) dl;
buff0 = pbuff_arr[0];
buff1 = pbuff_arr[2];
buff2 = pbuff_arr[5];
buff3 = pbuff_arr[9];
s0 = (*sp++);
UNPACK_SRC(d1, lo);
UNPACK_SRC(d2, hi);
s0 = (*sp++);
UNPACK_SRC(d3, lo);
UNPACK_SRC(d4, hi);
for (i = 0; i < width / 8; i++) {
s0 = sp[i];
CONV_5x5(lo, 2 * i);
CONV_5x5(hi, 2 * i + 1);
dp[i] = _mm_packs_pu16(res_lo, res_hi);
}
if (width & 7) {
__m64 mask;
mask = ((__m64 *) mlib_mask64_arr)[width & 7];
s0 = sp[i];
CONV_5x5(lo, 2 * i);
CONV_5x5(hi, 2 * i + 1);
res_hi = _mm_packs_pu16(res_lo, res_hi);
dp[i] =
_mm_or_si64(_mm_and_si64(mask, res_hi),
_mm_andnot_si64(mask, dp[i]));
}
ind = (pbuff_arr == buff_arr) ? 10 : -10;
pbuff_arr[ind + 0] = pbuff_arr[1];
pbuff_arr[ind + 1] = pbuff_arr[3];
pbuff_arr[ind + 2] = pbuff_arr[4];
pbuff_arr[ind + 3] = pbuff_arr[6];
pbuff_arr[ind + 4] = pbuff_arr[7];
pbuff_arr[ind + 5] = pbuff_arr[8];
pbuff_arr[ind + 6] = pbuff_arr[0];
pbuff_arr[ind + 7] = pbuff_arr[2];
pbuff_arr[ind + 8] = pbuff_arr[5];
pbuff_arr[ind + 9] = pbuff_arr[9];
pbuff_arr += ind;
sl += sll;
dl += dll;
}
_mm_empty();
mlib_free(pbuff);
return (MLIB_SUCCESS);
}
mlib_status
mlib_m_sconv7x7_16nw_4(
mlib_image *dst,
mlib_image *src,
mlib_s32 *hkernel,
mlib_s32 *vkernel,
mlib_s32 scalef_expon)
{
GET_SRC_DST_PARAMETERS(mlib_s16);
__m64 hker0, hker1, hker2, hker3, hker4, hker5, hker6;
__m64 vker0, vker1, vker2, vker3, vker4, vker5, vker6;
__m64 s0, s1, s2, s3, s4, s5, s6, v0, v1, v2, v3, v4, v5, v6, rr, rh,
rl;
__m64 zero, _rnd;
__m64 *sp0, *sp1, *sp2, *sp3, *sp4, *sp5, *sp6, *dp;
mlib_s32 shift, kerh_sum;
mlib_s32 i, j;
width -= KSIZE1;
height -= KSIZE1;
width *= NCHAN;
dl += (KSIZE / 2) * (dll + NCHAN);
GET_KERN();
zero = _mm_setzero_si64();
for (j = 0; j < height; j++) {
sp0 = (__m64 *) sl;
sp1 = (__m64 *) (sl + sll);
sp2 = (__m64 *) (sl + 2 * sll);
sp3 = (__m64 *) (sl + 3 * sll);
sp4 = (__m64 *) (sl + 4 * sll);
sp5 = (__m64 *) (sl + 5 * sll);
sp6 = (__m64 *) (sl + 6 * sll);
dp = (__m64 *) dl;
PREP_V(v1);
PREP_V(v2);
PREP_V(v3);
PREP_V(v4);
PREP_V(v5);
PREP_V(v6);
for (i = 0; i < width / 4; i++) {
CONV_7x7();
dp[i] = rr;
}
if (width & 3) {
__m64 mask =
((__m64 *) mlib_mask64_arr)[2 * (width & 3)];
CONV_7x7();
dp[i] =
_mm_or_si64(_mm_and_si64(mask, rr),
_mm_andnot_si64(mask, dp[i]));
}
sl += sll;
dl += dll;
}
_mm_empty();
return (MLIB_SUCCESS);
}
mlib_status
mlib_ImageMinFilter3x3_S16(
void *dst,
void *src,
mlib_s32 dlb,
mlib_s32 slb,
mlib_s32 wid,
mlib_s32 hgt)
#endif /* MAX_FILTER */
{
mlib_u8 *buff, *buff1;
mlib_u8 *sl, *sp0, *sp1, *sp2, *sp3, *dl;
__m64 *dp0, *dp1;
__m64 aa, bb, c0, c1, c2, cc, d0, d1, d2, dd, r0, r1;
__m64 e_mask;
mlib_s32 i, j, wid8, tail;
wid = (wid - 2) * SSIZE;
wid8 = (wid + 7) & ~7;
buff = mlib_malloc(2 * wid8);
buff1 = buff + wid8;
sl = (mlib_u8 *)src;
/* dst ptrs skip top j and left col */
dl = (mlib_u8 *)dst + dlb + SSIZE;
tail = wid & 7;
e_mask = ((__m64 *) mlib_mask64_arr)[tail];
sp0 = buff;
sp1 = buff1;
sp2 = sl;
sp3 = sp2 + slb;
sl += 2 * slb;
for (i = 0; i < wid; i += 8) {
c0 = *(__m64 *) sp2;
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
d0 = *(__m64 *) sp3;
d1 = *(__m64 *) (sp3 + SSIZE);
d2 = *(__m64 *) (sp3 + 2 * SSIZE);
cc = C_COMP(c0, c1);
dd = C_COMP(d0, d1);
cc = C_COMP(cc, c2);
dd = C_COMP(dd, d2);
*(__m64 *) sp0 = cc;
*(__m64 *) sp1 = dd;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
}
for (j = 0; j <= (hgt - 2 - 2); j += 2) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff;
sp1 = buff1;
sp2 = sl;
sp3 = sp2 + slb;
/*
* line0: aa
* line1: bb
* line2: c0 c1 c2
* line3: d0 d1 d2
*/
for (i = 0; i <= wid - 8; i += 8) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
c0 = *(__m64 *) sp2;
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
d0 = *(__m64 *) sp3;
d1 = *(__m64 *) (sp3 + SSIZE);
d2 = *(__m64 *) (sp3 + 2 * SSIZE);
cc = C_COMP(c0, c1);
dd = C_COMP(d0, d1);
cc = C_COMP(cc, c2);
dd = C_COMP(dd, d2);
bb = C_COMP(bb, cc);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, dd);
*(__m64 *) sp0 = cc;
*(__m64 *) sp1 = dd;
(*dp0++) = r0;
(*dp1++) = r1;
sp0 += 8;
sp1 += 8;
sp2 += 8;
sp3 += 8;
}
if (tail) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
c0 = *(__m64 *) sp2;
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
d0 = *(__m64 *) sp3;
d1 = *(__m64 *) (sp3 + SSIZE);
d2 = *(__m64 *) (sp3 + 2 * SSIZE);
cc = C_COMP(c0, c1);
dd = C_COMP(d0, d1);
cc = C_COMP(cc, c2);
dd = C_COMP(dd, d2);
bb = C_COMP(bb, cc);
r0 = C_COMP(aa, bb);
r1 = C_COMP(bb, dd);
*(__m64 *) sp0 = cc;
*(__m64 *) sp1 = dd;
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
*dp1 =
_mm_or_si64(_mm_and_si64(e_mask, r1),
_mm_andnot_si64(e_mask, *dp1));
}
sl += 2 * slb;
dl += 2 * dlb;
}
/* last line */
if (j == (hgt - 3)) {
dp0 = (void *)dl;
dp1 = (void *)(dl + dlb);
sp0 = buff;
sp1 = buff1;
sp2 = sl;
for (i = 0; i <= wid - 8; i += 8) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
c0 = *(__m64 *) sp2;
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
cc = C_COMP(c0, c1);
cc = C_COMP(cc, c2);
r0 = C_COMP(aa, bb);
r0 = C_COMP(r0, cc);
(*dp0++) = r0;
sp0 += 8;
sp1 += 8;
sp2 += 8;
}
if (tail) {
aa = *(__m64 *) sp0;
bb = *(__m64 *) sp1;
c0 = *(__m64 *) sp2;
c1 = *(__m64 *) (sp2 + SSIZE);
c2 = *(__m64 *) (sp2 + 2 * SSIZE);
c1 = C_COMP(c0, c1);
cc = C_COMP(c1, c2);
r0 = C_COMP(aa, bb);
r0 = C_COMP(r0, cc);
*dp0 =
_mm_or_si64(_mm_and_si64(e_mask, r0),
_mm_andnot_si64(e_mask, *dp0));
}
}
_mm_empty();
mlib_free(buff);
return (MLIB_SUCCESS);
}
