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);
}
mlib_status
mlib_m_sconv5x5_u16nw_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;
__m64 vker0, vker1, vker2, vker3, vker4;
__m64 s0, s1, s2, s3, s4, v0, v1, v2, v3, v4, rr, rh, rl;
__m64 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;
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; i++) {
CONV_5x5();
dp[i] = rr;
}
sl += sll;
dl += dll;
}
_mm_empty();
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_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 CONV_FUNC_I(MxN)(mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 m,
mlib_s32 n,
mlib_s32 dm,
mlib_s32 dn,
mlib_s32 scale,
mlib_s32 cmask)
{
mlib_s32 buff[BUFF_SIZE], *buffd = buff;
mlib_s32 l, off, kw;
mlib_s32 d0, d1, shift1, shift2;
mlib_s32 k0, k1, k2, k3, k4, k5, k6;
mlib_s32 p0, p1, p2, p3, p4, p5, p6, p7;
DTYPE *adr_src, *sl, *sp = NULL;
DTYPE *adr_dst, *dl, *dp = NULL;
mlib_s32 wid, hgt, sll, dll;
mlib_s32 nchannel, chan1;
mlib_s32 i, j, c;
mlib_s32 chan2;
mlib_s32 k_locl[MAX_N*MAX_N], *k = k_locl;
GET_SRC_DST_PARAMETERS(DTYPE);
#if IMG_TYPE != 1
shift1 = 16;
#else
shift1 = 8;
#endif /* IMG_TYPE != 1 */
shift2 = scale - shift1;
chan1 = nchannel;
chan2 = chan1 + chan1;
wid -= (m - 1);
hgt -= (n - 1);
adr_dst += dn*dll + dm*nchannel;
if (wid > BUFF_SIZE) {
buffd = mlib_malloc(sizeof(mlib_s32)*wid);
if (buffd == NULL) return MLIB_FAILURE;
}
if (m*n > MAX_N*MAX_N) {
k = mlib_malloc(sizeof(mlib_s32)*(m*n));
if (k == NULL) {
if (buffd != buff) mlib_free(buffd);
return MLIB_FAILURE;
}
}
for (i = 0; i < m*n; i++) {
k[i] = kernel[i] >> shift1;
}
for (c = 0; c < nchannel; c++) {
if (!(cmask & (1 << (nchannel - 1 - c)))) continue;
sl = adr_src + c;
dl = adr_dst + c;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i < wid; i++) buffd[i] = 0;
for (j = 0; j < hgt; j++) {
mlib_s32 *pk = k;
for (l = 0; l < n; l++) {
DTYPE *sp0 = sl + l*sll;
for (off = 0; off < m;) {
sp = sp0 + off*chan1;
dp = dl;
kw = m - off;
if (kw > 2*MAX_KER) kw = MAX_KER; else
if (kw > MAX_KER) kw = kw/2;
off += kw;
p2 = sp[0]; p3 = sp[chan1]; p4 = sp[chan2];
p5 = sp[chan2 + chan1]; p6 = sp[chan2 + chan2]; p7 = sp[5*chan1];
k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
pk += kw;
sp += (kw - 1)*chan1;
if (kw == 7) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
p6 = sp[0];
p7 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
p6 = sp[0];
p7 = sp[chan1];
d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 6) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
p5 = sp[0];
p6 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
p5 = sp[0];
p6 = sp[chan1];
d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 5) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5;
p4 = sp[0];
p5 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5;
p4 = sp[0];
p5 = sp[chan1];
d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 4) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = sp[0];
p4 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = sp[0];
p4 = sp[chan1];
d0 = (p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 3) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3;
p2 = sp[0];
p3 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1 + p2*k2;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3;
p2 = sp[0];
p3 = sp[chan1];
d0 = (p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 2) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2;
p1 = sp[0];
p2 = sp[chan1];
buffd[i ] += p0*k0 + p1*k1;
buffd[i + 1] += p1*k0 + p2*k1;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2;
p1 = sp[0];
p2 = sp[chan1];
d0 = (p0*k0 + p1*k1 + buffd[i ]);
d1 = (p1*k0 + p2*k1 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
} else /*if (kw == 1)*/ {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = sp[0];
p1 = sp[chan1];
buffd[i ] += p0*k0;
buffd[i + 1] += p1*k0;
sp += chan2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = sp[0];
p1 = sp[chan1];
d0 = (p0*k0 + buffd[i ]);
d1 = (p1*k0 + buffd[i + 1]);
STORE_RES(dp[0 ], d0);
STORE_RES(dp[chan1], d1);
buffd[i ] = 0;
buffd[i + 1] = 0;
sp += chan2;
dp += chan2;
}
}
}
}
}
/* last pixels */
for (; i < wid; i++) {
mlib_s32 *pk = k, s = 0;
mlib_s32 x;
for (l = 0; l < n; l++) {
sp = sl + l*sll + i*chan1;
for (x = 0; x < m; x++) {
s += sp[0] * pk[0];
sp += chan1;
pk ++;
}
}
STORE_RES(dp[0], s);
sp += chan1;
dp += chan1;
}
sl += sll;
dl += dll;
}
}
if (buffd != buff) mlib_free(buffd);
if (k != k_locl) mlib_free(k);
return MLIB_SUCCESS;
}
mlib_status CONV_FUNC(MxN)(mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 m,
mlib_s32 n,
mlib_s32 dm,
mlib_s32 dn,
mlib_s32 scale,
mlib_s32 cmask)
{
FTYPE buff[BUFF_SIZE], *buffs_arr[2*(MAX_N + 1)];
FTYPE **buffs = buffs_arr, *buffd;
FTYPE akernel[256], *k = akernel, fscale = DSCALE;
mlib_s32 mn, l, off, kw, bsize, buff_ind;
mlib_s32 d0, d1;
FTYPE k0, k1, k2, k3, k4, k5, k6;
FTYPE p0, p1, p2, p3, p4, p5, p6, p7;
d64_2x32 dd;
DEF_VARS(DTYPE);
mlib_s32 chan2;
mlib_s32 *buffo, *buffi;
mlib_status status = MLIB_SUCCESS;
GET_SRC_DST_PARAMETERS(DTYPE);
if (scale > 30) {
fscale *= 1.0/(1 << 30);
scale -= 30;
}
fscale /= (1 << scale);
mn = m*n;
if (mn > 256) {
k = mlib_malloc(mn*sizeof(mlib_d64));
if (k == NULL) return MLIB_FAILURE;
}
for (i = 0; i < mn; i++) {
k[i] = kernel[i]*fscale;
}
if (m == 1) {
status = mlib_ImageConv1xN(dst, src, k, n, dn, cmask);
FREE_AND_RETURN_STATUS;
}
bsize = (n + 3)*wid;
if ((bsize > BUFF_SIZE) || (n > MAX_N)) {
pbuff = mlib_malloc(sizeof(FTYPE)*bsize + sizeof(FTYPE *)*2*(n + 1));
if (pbuff == NULL) {
status = MLIB_FAILURE;
FREE_AND_RETURN_STATUS;
}
buffs = (FTYPE **)(pbuff + bsize);
}
for (l = 0; l < (n + 1); l++) buffs[l] = pbuff + l*wid;
for (l = 0; l < (n + 1); l++) buffs[l + (n + 1)] = buffs[l];
buffd = buffs[n] + wid;
buffo = (mlib_s32*)(buffd + wid);
buffi = buffo + (wid &~ 1);
chan1 = nchannel;
chan2 = chan1 + chan1;
wid -= (m - 1);
hgt -= (n - 1);
adr_dst += dn*dll + dm*nchannel;
for (c = 0; c < nchannel; c++) {
if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
sl = adr_src + c;
dl = adr_dst + c;
for (l = 0; l < n; l++) {
FTYPE *buff = buffs[l];
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i < wid + (m - 1); i++) {
buff[i] = (FTYPE)sl[i*chan1];
}
sl += sll;
}
buff_ind = 0;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i < wid; i++) buffd[i] = 0.0;
for (j = 0; j < hgt; j++) {
FTYPE **buffc = buffs + buff_ind;
FTYPE *buffn = buffc[n];
FTYPE *pk = k;
for (l = 0; l < n; l++) {
FTYPE *buff_l = buffc[l];
for (off = 0; off < m;) {
FTYPE *buff = buff_l + off;
kw = m - off;
if (kw > 2*MAX_KER) kw = MAX_KER; else
if (kw > MAX_KER) kw = kw/2;
off += kw;
sp = sl;
dp = dl;
p2 = buff[0]; p3 = buff[1]; p4 = buff[2];
p5 = buff[3]; p6 = buff[4]; p7 = buff[5];
k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
k4 = pk[4]; k5 = pk[5]; k6 = pk[6];
pk += kw;
if (kw == 7) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
p6 = buff[i + 6]; p7 = buff[i + 7];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6; p5 = p7;
p6 = buff[i + 6]; p7 = buff[i + 7];
LOAD_BUFF(buffi);
dd.d64 = *(FTYPE *)(buffi + i);
buffn[i ] = (FTYPE)dd.i32s.i0;
buffn[i + 1] = (FTYPE)dd.i32s.i1;
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + p6*k6 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + p7*k6 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 6) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
p5 = buff[i + 5]; p6 = buff[i + 6];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5; p4 = p6;
p5 = buff[i + 5]; p6 = buff[i + 6];
buffn[i ] = (FTYPE)sp[0];
buffn[i + 1] = (FTYPE)sp[chan1];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + p5*k5 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + p6*k5 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 5) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5;
p4 = buff[i + 4]; p5 = buff[i + 5];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4; p3 = p5;
p4 = buff[i + 4]; p5 = buff[i + 5];
buffn[i ] = (FTYPE)sp[0];
buffn[i + 1] = (FTYPE)sp[chan1];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + p4*k4 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + p5*k4 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 4) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = buff[i + 3]; p4 = buff[i + 4];
buffd[i ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = buff[i + 3]; p4 = buff[i + 4];
buffn[i ] = (FTYPE)sp[0];
buffn[i + 1] = (FTYPE)sp[chan1];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
} else if (kw == 3) {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3;
p2 = buff[i + 2]; p3 = buff[i + 3];
buffd[i ] += p0*k0 + p1*k1 + p2*k2;
buffd[i + 1] += p1*k0 + p2*k1 + p3*k2;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2; p1 = p3;
p2 = buff[i + 2]; p3 = buff[i + 3];
buffn[i ] = (FTYPE)sp[0];
buffn[i + 1] = (FTYPE)sp[chan1];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
} else /*if (kw == 2)*/ {
if (l < (n - 1) || off < m) {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2;
p1 = buff[i + 1]; p2 = buff[i + 2];
buffd[i ] += p0*k0 + p1*k1;
buffd[i + 1] += p1*k0 + p2*k1;
}
} else {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (i = 0; i <= (wid - 2); i += 2) {
p0 = p2;
p1 = buff[i + 1]; p2 = buff[i + 2];
buffn[i ] = (FTYPE)sp[0];
buffn[i + 1] = (FTYPE)sp[chan1];
d0 = D2I(p0*k0 + p1*k1 + buffd[i ]);
d1 = D2I(p1*k0 + p2*k1 + buffd[i + 1]);
dp[0 ] = FROM_S32(d0);
dp[chan1] = FROM_S32(d1);
buffd[i ] = 0.0;
buffd[i + 1] = 0.0;
sp += chan2;
dp += chan2;
}
}
}
}
}
/* last pixels */
for (; i < wid; i++) {
FTYPE *pk = k, s = 0;
mlib_s32 x, d0;
for (l = 0; l < n; l++) {
FTYPE *buff = buffc[l] + i;
for (x = 0; x < m; x++) s += buff[x] * (*pk++);
}
d0 = D2I(s);
dp[0] = FROM_S32(d0);
buffn[i] = (FTYPE)sp[0];
sp += chan1;
dp += chan1;
}
for (l = 0; l < (m - 1); l++) buffn[wid + l] = sp[l*chan1];
/* next line */
sl += sll;
dl += dll;
buff_ind++;
if (buff_ind >= n + 1) buff_ind = 0;
}
}
FREE_AND_RETURN_STATUS;
}
static mlib_status mlib_ImageConv1xN(mlib_image *dst,
const mlib_image *src,
const mlib_d64 *k,
mlib_s32 n,
mlib_s32 dn,
mlib_s32 cmask)
{
FTYPE buff[BUFF_SIZE];
mlib_s32 off, kh;
mlib_s32 d0, d1;
const FTYPE *pk;
FTYPE k0, k1, k2, k3;
FTYPE p0, p1, p2, p3, p4;
DEF_VARS(DTYPE);
DTYPE *sl_c, *dl_c, *sl0;
mlib_s32 l, hsize, max_hsize;
GET_SRC_DST_PARAMETERS(DTYPE);
hgt -= (n - 1);
adr_dst += dn*dll;
max_hsize = (CACHE_SIZE/sizeof(DTYPE))/sll;
if (!max_hsize) max_hsize = 1;
if (max_hsize > BUFF_SIZE) {
pbuff = mlib_malloc(sizeof(FTYPE)*max_hsize);
}
chan1 = nchannel;
sl_c = adr_src;
dl_c = adr_dst;
for (l = 0; l < hgt; l += hsize) {
hsize = hgt - l;
if (hsize > max_hsize) hsize = max_hsize;
for (c = 0; c < nchannel; c++) {
if (!(cmask & (1 << (chan1 - 1 - c)))) continue;
sl = sl_c + c;
dl = dl_c + c;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j < hsize; j++) pbuff[j] = 0.0;
for (i = 0; i < wid; i++) {
sl0 = sl;
for (off = 0; off < (n - 4); off += 4) {
pk = k + off;
sp = sl0;
k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
sp += 3*sll;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j < hsize; j += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = sp[0];
p4 = sp[sll];
pbuff[j ] += p0*k0 + p1*k1 + p2*k2 + p3*k3;
pbuff[j + 1] += p1*k0 + p2*k1 + p3*k2 + p4*k3;
sp += 2*sll;
}
sl0 += 4*sll;
}
pk = k + off;
sp = sl0;
k0 = pk[0]; k1 = pk[1]; k2 = pk[2]; k3 = pk[3];
p2 = sp[0]; p3 = sp[sll]; p4 = sp[2*sll];
dp = dl;
kh = n - off;
if (kh == 4) {
sp += 3*sll;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j <= (hsize - 2); j += 2) {
p0 = p2; p1 = p3; p2 = p4;
p3 = sp[0];
p4 = sp[sll];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + p4*k3 + pbuff[j + 1]);
dp[0 ] = FROM_S32(d0);
dp[dll] = FROM_S32(d1);
pbuff[j] = 0;
pbuff[j + 1] = 0;
sp += 2*sll;
dp += 2*dll;
}
if (j < hsize) {
p0 = p2; p1 = p3; p2 = p4;
p3 = sp[0];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + p3*k3 + pbuff[j]);
pbuff[j] = 0;
dp[0] = FROM_S32(d0);
}
} else if (kh == 3) {
sp += 2*sll;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j <= (hsize - 2); j += 2) {
p0 = p2; p1 = p3;
p2 = sp[0];
p3 = sp[sll];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
d1 = D2I(p1*k0 + p2*k1 + p3*k2 + pbuff[j + 1]);
dp[0 ] = FROM_S32(d0);
dp[dll] = FROM_S32(d1);
pbuff[j] = 0;
pbuff[j + 1] = 0;
sp += 2*sll;
dp += 2*dll;
}
if (j < hsize) {
p0 = p2; p1 = p3;
p2 = sp[0];
d0 = D2I(p0*k0 + p1*k1 + p2*k2 + pbuff[j]);
pbuff[j] = 0;
dp[0] = FROM_S32(d0);
}
} else if (kh == 2) {
sp += sll;
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j <= (hsize - 2); j += 2) {
p0 = p2;
p1 = sp[0];
p2 = sp[sll];
d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
d1 = D2I(p1*k0 + p2*k1 + pbuff[j + 1]);
dp[0 ] = FROM_S32(d0);
dp[dll] = FROM_S32(d1);
pbuff[j] = 0;
pbuff[j + 1] = 0;
sp += 2*sll;
dp += 2*dll;
}
if (j < hsize) {
p0 = p2;
p1 = sp[0];
d0 = D2I(p0*k0 + p1*k1 + pbuff[j]);
pbuff[j] = 0;
dp[0] = FROM_S32(d0);
}
} else /* if (kh == 1) */ {
#ifdef __SUNPRO_C
#pragma pipeloop(0)
#endif /* __SUNPRO_C */
for (j = 0; j < hsize; j++) {
p0 = sp[0];
d0 = D2I(p0*k0 + pbuff[j]);
dp[0] = FROM_S32(d0);
pbuff[j] = 0;
sp += sll;
dp += dll;
}
}
sl += chan1;
dl += chan1;
}
}
sl_c += max_hsize*sll;
dl_c += max_hsize*dll;
}
if (pbuff != buff) 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_v_conv2x2_u16nw_mask(
mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 scalef_expon,
mlib_s32 cmask)
{
/* pointers to dst row */
mlib_u16 *da, *d_a;
/* pointers to src, dst data */
mlib_u16 *adr_dst, *adr_src, *dend;
/* pointers to src rows */
mlib_u16 *sa, *sa1, *sa2, *sa_2;
/* pointers to rows in interm. src buf */
mlib_u16 *buff_src, *sbuf1, *sbuf2, *prow;
mlib_u16 *s_buf1;
/* mlib_d64 pointers to rows in interm. src buf */
mlib_d64 *s1, *s2;
/* src, dst and interm. buf. strides */
mlib_s32 dlb, slb, buf_slb;
mlib_s32 dh, dw;
mlib_d64 out0, out1, tmp0, tmp1, tmp2, tmp3;
/* data */
mlib_d64 d1, d2, d_1, d_2;
/* shifted data */
mlib_d64 d21, d22;
/* coefficients */
mlib_f32 k1, k2, k3, k4;
int gsr_scale, i, j, nchannel, chan, testchan;
mlib_u16 t1, t2, t3, t4, t5, t6, t7, t8;
type_mlib_d64 str;
mlib_d64 ker_off, mask8000 = vis_to_double_dup(0x80008000);
nchannel = mlib_ImageGetChannels(src);
GET_SRC_DST_PARAMETERS();
LOAD_KERNEL_INTO_FLOAT();
gsr_scale = 32 - scalef_expon;
vis_write_gsr((gsr_scale << 3) + 2);
/* buf_slb - 8-byte aligned */
buf_slb = (2 * dw + 26) & (~7);
/* alloc. interm. src buffer */
buff_src =
(mlib_u16 *)__mlib_malloc(2 * buf_slb * sizeof (mlib_u8) + 8);
if (buff_src == NULL)
return (MLIB_FAILURE);
buf_slb >>= 1;
sbuf1 = (mlib_u16 *)((mlib_addr)(buff_src + 8) & (~7));
sbuf2 = sbuf1 + buf_slb;
dw -= 1;
/* edge - no write */
dh -= 1;
testchan = 1;
for (chan = nchannel - 1; chan >= 0; chan--) {
if ((cmask & testchan) == 0) {
testchan <<= 1;
continue;
}
testchan <<= 1;
sa = adr_src + chan;
sa1 = sa + slb;
sa_2 = sa2 = sa1 + slb;
d_a = adr_dst + chan;
/* load interm. src buff */
for (i = 0, j = 0; j < (dw + 1); i += nchannel, j++) {
sbuf1[j] = sa1[i];
sbuf2[j] = sa[i];
}
for (j = 0; j < dh - 1; j++) {
da = d_a;
prow = sbuf1;
sbuf1 = sbuf2;
sbuf2 = prow;
s1 = (mlib_d64 *)sbuf1;
s2 = (mlib_d64 *)sbuf2;
dend = da + (dw - 1) * nchannel;
s_buf1 = sbuf1;
d1 = *s1;
d2 = *s2;
d1 = vis_fxor(d1, mask8000);
d2 = vis_fxor(d2, mask8000);
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_1 = vis_fxor(d_1, mask8000);
d_2 = vis_fxor(d_2, mask8000);
CONV_16_BEGIN(d1, k1);
CONV_16(d2, k3);
d21 = vis_faligndata(d1, d_1);
d22 = vis_faligndata(d2, d_2);
CONV_16(d21, k2);
CONV_16(d22, k4);
str.value =
vis_fxor(vis_fpackfix_pair(out0, out1), mask8000);
d1 = d_1;
d2 = d_2;
s1++;
s2++;
/*
* in each iteration store result from prev. iterat.
* and load data for processing next row
*/
#pragma pipeloop(0)
for (i = 0; i < dw - 4; i += 4) {
t1 = *sa_2;
sa_2 += nchannel;
t2 = *sa_2;
sa_2 += nchannel;
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_1 = vis_fxor(d_1, mask8000);
d_2 = vis_fxor(d_2, mask8000);
CONV_16_BEGIN(d1, k1);
t3 = *sa_2;
sa_2 += nchannel;
t4 = *sa_2;
sa_2 += nchannel;
CONV_16(d2, k3);
t5 = str.forshort.ushort0;
t6 = str.forshort.ushort1;
d21 = vis_faligndata(d1, d_1);
t7 = str.forshort.ushort2;
d22 = vis_faligndata(d2, d_2);
t8 = str.forshort.ushort3;
CONV_16(d21, k2);
(*s_buf1++) = t1;
(*s_buf1++) = t2;
CONV_16(d22, k4);
(*s_buf1++) = t3;
(*s_buf1++) = t4;
*da = t5;
da += nchannel;
str.value =
vis_fxor(vis_fpackfix_pair(out0, out1),
mask8000);
*da = t6;
da += nchannel;
d1 = d_1;
d2 = d_2;
*da = t7;
da += nchannel;
s1++;
s2++;
*da = t8;
da += nchannel;
}
for (; i < dw + 1; i++) {
(*s_buf1++) = *sa_2;
sa_2 += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort0;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort1;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort2;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort3;
}
sa_2 = sa2 = sa2 + slb;
d_a += dlb;
}
/* process last row - no need to load data */
da = d_a;
prow = sbuf1;
sbuf1 = sbuf2;
sbuf2 = prow;
s1 = (mlib_d64 *)sbuf1;
s2 = (mlib_d64 *)sbuf2;
dend = da + (dw - 1) * nchannel;
d1 = *s1;
d2 = *s2;
d1 = vis_fxor(d1, mask8000);
d2 = vis_fxor(d2, mask8000);
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_1 = vis_fxor(d_1, mask8000);
d_2 = vis_fxor(d_2, mask8000);
CONV_16_BEGIN(d1, k1);
CONV_16(d2, k3);
d21 = vis_faligndata(d1, d_1);
d22 = vis_faligndata(d2, d_2);
CONV_16(d21, k2);
CONV_16(d22, k4);
d1 = d_1;
d2 = d_2;
s1++;
s2++;
#pragma pipeloop(0)
for (i = 4; i < dw; i += 4) {
str.value =
vis_fxor(vis_fpackfix_pair(out0, out1), mask8000);
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_1 = vis_fxor(d_1, mask8000);
d_2 = vis_fxor(d_2, mask8000);
CONV_16_BEGIN(d1, k1);
t5 = str.forshort.ushort0;
CONV_16(d2, k3);
d21 = vis_faligndata(d1, d_1);
t6 = str.forshort.ushort1;
d22 = vis_faligndata(d2, d_2);
CONV_16(d21, k2);
t7 = str.forshort.ushort2;
CONV_16(d22, k4);
t8 = str.forshort.ushort3;
*da = t5;
da += nchannel;
*da = t6;
da += nchannel;
*da = t7;
da += nchannel;
d1 = d_1;
d2 = d_2;
*da = t8;
da += nchannel;
s1++;
s2++;
}
str.value = vis_fxor(vis_fpackfix_pair(out0, out1), mask8000);
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort0;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort1;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort2;
da += nchannel;
}
if ((mlib_addr)da <= (mlib_addr)dend) {
*da = str.forshort.ushort3;
}
}
__mlib_free(buff_src);
return (MLIB_SUCCESS);
}
mlib_status
mlib_v_conv2x2_u16nw_4(
mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 scalef_expon)
{
/* pointers to dst row */
mlib_u16 *da, *d_a;
/* pointers to src, dst data */
mlib_u16 *adr_dst, *adr_src, *dend;
/* pointers to src rows */
mlib_u16 *sa, *sa1;
/* pointers to rows in interm. src buf */
mlib_d64 *buff_src, *sbuf1, *sbuf2, *prow;
/* pointer to row in interm. dst buf */
mlib_d64 *dbuf;
/* mlib_d64 pointers to rows in interm. src buf */
mlib_d64 *s1, *s2;
/* mlib_d64 pointer to row in interm. dst buf */
mlib_d64 *ddst;
/* data */
mlib_d64 d1, d2, d_1, d_2;
mlib_f32 k1, k2, k3, k4;
/* src, dst and interm. buf. strides */
mlib_s32 dlb, slb, buf_slb;
mlib_s32 dh, dw;
mlib_d64 out0, out1, tmp0, tmp1, tmp2, tmp3;
mlib_d64 *dsa, *dp;
mlib_d64 sd0, sd1;
mlib_s32 emask;
int gsr_scale, i, j;
mlib_d64 ker_off, mask8000 = vis_to_double_dup(0x80008000);
GET_SRC_DST_PARAMETERS();
LOAD_KERNEL_INTO_FLOAT();
gsr_scale = 32 - scalef_expon;
vis_write_gsr((gsr_scale << 3));
buf_slb = (8 * dw + 16) >> 3;
PREPARE_INTERM_BUFFERS();
dw -= 1;
dw *= 4;
dh -= 1;
sa = adr_src;
sa1 = sa + slb;
d_a = adr_dst;
/* load interm. src buff */
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(sbuf2, sa, 4);
#pragma pipeloop(0)
for (j = 0; j < dh; j++) {
LOOP_INI();
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(sbuf2, sa1, 4);
d1 = *s1;
d2 = *s2;
d1 = vis_fxor(d1, mask8000);
d2 = vis_fxor(d2, mask8000);
#pragma pipeloop(0)
for (i = 0; i < dw; i += 4) {
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_1 = vis_fxor(d_1, mask8000);
d_2 = vis_fxor(d_2, mask8000);
CONV_16_BEGIN(d1, k1);
CONV_16(d2, k3);
CONV_16(d_1, k2);
CONV_16(d_2, k4);
(*ddst++) =
vis_fxor(vis_fpackfix_pair(out0, out1), mask8000);
d1 = d_1;
d2 = d_2;
s1++;
s2++;
}
PREPARE_TO_COPY_INTERM_BUF_TO_DST();
#pragma pipeloop(0)
COPY_INTERM_BUF_TO_DST();
COPY_TAIL();
sa1 = sa1 + slb;
d_a += dlb;
}
__mlib_free(buff_src);
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);
}
mlib_status
mlib_v_conv3x3_8nw_4(
mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 scalef_expon,
mlib_s32 cmask)
{
/* pointers to dst row */
mlib_u8 *da, *d_a;
/* pointers to src, dst data */
mlib_u8 *adr_dst, *adr_src, *dend;
/* pointers to src rows */
mlib_u8 *sa, *sa1, *sa2;
/* pointers to rows in interm. src buf */
mlib_d64 *buff_src, *sbuf1, *sbuf2, *prow;
/* pointers to rows in interm. src buf */
mlib_d64 *sbuf3;
/* pointer to row in interm. dst buf */
mlib_d64 *dbuf;
/* mlib_d64 pointers to rows in interm. src buf */
mlib_d64 *s1, *s2, *s3;
/* mlib_d64 pointer to row in interm. dst buf */
mlib_d64 *ddst;
/* data */
mlib_d64 d1, d2, d_1, d_2, d21, d22;
/* data */
mlib_d64 d3, d_3, d23;
mlib_f32 k1k2, k3k4, k5k6, k7k8, k9k9;
/* src, dst and interm. buf. strides */
mlib_s32 dlb, slb, buf_slb;
mlib_s32 dh, dw;
mlib_d64 out0, out1;
mlib_d64 tmp0, tmp1, rnd;
mlib_d64 *dsa, *dp;
mlib_d64 sd0, sd1, sd00;
mlib_s32 emask, cmask1;
mlib_s32 rval, gsr_scale, i, j;
gsr_scale = 31 - scalef_expon;
vis_write_gsr((gsr_scale << 3));
rval = mlib_round_8[gsr_scale];
rnd = vis_freg_pair(vis_to_float(rval), vis_to_float(rval));
cmask = ((cmask & 0xf) << 4) + (cmask & 0xf);
cmask = (cmask << 8) + (cmask);
GET_SRC_DST_PARAMETERS();
LOAD_KERNEL_INTO_FLOAT();
buf_slb = (4 * dw + 24) >> 3;
PREPARE_INTERM_BUFFERS();
dw -= 2;
dw *= 4;
dh -= 2;
sa = adr_src;
sa1 = sa + slb;
sa2 = sa1 + slb;
d_a = adr_dst + dlb + 4;
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf2, sa);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(8);
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf3, sa1);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(8);
#pragma pipeloop(0)
for (j = 0; j < dh; j++) {
LOOP_INI();
PREPARE_TO_LOAD_LINE(sbuf3, sa2);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(8);
vis_alignaddr(s1, 4);
d1 = *s1;
d2 = *s2;
d3 = *s3;
#pragma pipeloop(0)
for (i = 0; i < dw; i += 8) {
d_1 = *(s1 + 1);
d_2 = *(s2 + 1);
d_3 = *(s3 + 1);
out0 = out1 = rnd;
CONV_AU(d1, k1k2);
CONV_AL(d2, k3k4);
CONV_AU(d3, k7k8);
d21 = vis_faligndata(d1, d_1);
d22 = vis_faligndata(d2, d_2);
d23 = vis_faligndata(d3, d_3);
CONV_AL(d21, k1k2);
CONV_AU(d22, k5k6);
CONV_AL(d23, k7k8);
CONV_AU(d_1, k3k4);
CONV_AL(d_2, k5k6);
CONV_AU(d_3, k9k9);
(*ddst++) = vis_fpack16_pair(out0, out1);
d1 = d_1;
d2 = d_2;
d3 = d_3;
s1++;
s2++;
s3++;
}
ddst = dbuf;
/* prepare the destination addresses */
dp = (mlib_d64 *)((mlib_addr)da & (~7));
i = (mlib_addr)dp - (mlib_addr)da;
cmask1 = cmask >> (-i);
ddst = vis_alignaddr(ddst, i);
/* generate edge mask for the start point */
emask = vis_edge8(da, dend);
sd1 = ddst[0];
if (emask != 0xff) {
sd0 = sd1;
sd1 = ddst[1];
sd0 = vis_faligndata(sd0, sd1);
vis_pst_8(sd0, dp++, emask & cmask1);
ddst++;
i += 8;
}
#pragma pipeloop(0)
for (; i <= (dw - 8); i += 8) {
sd0 = sd1;
sd1 = ddst[1];
sd00 = vis_faligndata(sd0, sd1);
vis_pst_8(sd00, dp++, cmask1);
ddst++;
}
if (i < dw) {
sd0 = vis_faligndata(sd1, ddst[1]);
emask = vis_edge8(dp, dend);
vis_pst_8(sd0, dp, emask & cmask1);
}
sa2 = sa2 + slb;
d_a += dlb;
}
__mlib_free(buff_src);
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_c_conv2x2ext_u8(
mlib_image *dst,
const mlib_image *src,
mlib_s32 dx_l,
mlib_s32 dx_r,
mlib_s32 dy_t,
mlib_s32 dy_b,
const mlib_s32 *kern,
mlib_s32 scalef_expon,
mlib_s32 cmask)
{
mlib_d64 buff_arr[4 * BUFF_LINE];
mlib_s32 *pbuff = (mlib_s32 *)buff_arr, *buff0, *buff1, *buff2, *buffT;
DTYPE *adr_src, *sl, *sp, *sl1;
DTYPE *adr_dst, *dl, *dp;
mlib_d64 k0, k1, k2, k3, scalef = 1.0;
mlib_d64 p00, p01, p02, p10, p11, p12;
mlib_s32 wid, hgt, sll, dll, wid1;
mlib_s32 nchannel, chan1, chan2;
mlib_s32 i, j, c, swid;
LOAD_KERNEL_INTO_DOUBLE();
GET_SRC_DST_PARAMETERS(DTYPE);
vis_write_gsr(23 << 3);
swid = wid + D_KER;
wid1 = (swid + 1) & ~1;
if (wid1 > BUFF_LINE) {
pbuff = __mlib_malloc(4 * sizeof (mlib_s32) * wid1);
if (pbuff == NULL)
return (MLIB_FAILURE);
}
buff0 = pbuff;
buff1 = buff0 + wid1;
buff2 = buff1 + wid1;
chan1 = nchannel;
chan2 = chan1 + chan1;
swid -= dx_r;
for (c = 0; c < nchannel; c++) {
if (!(cmask & (1 << (nchannel - 1 - c))))
continue;
sl = adr_src + c;
dl = adr_dst + c;
if ((hgt - dy_b) > 0)
sl1 = sl + sll;
else
sl1 = sl;
#pragma pipeloop(0)
for (i = 0; i < swid; i++) {
buff0[i - 1] = (mlib_s32)sl[i * chan1];
buff1[i - 1] = (mlib_s32)sl1[i * chan1];
}
if (dx_r != 0) {
buff0[swid - 1] = buff0[swid - 2];
buff1[swid - 1] = buff1[swid - 2];
}
if ((hgt - dy_b) > 1)
sl = sl1 + sll;
else
sl = sl1;
for (j = 0; j < hgt; j++) {
sp = sl;
dp = dl;
buff2[-1] = (mlib_s32)sp[0];
sp += chan1;
p02 = buff0[-1];
p12 = buff1[-1];
#pragma pipeloop(0)
for (i = 0; i <= (wid - 2); i += 2) {
d64_2x32 sd0, sd1;
d64_2x32 dd0, dd1;
p00 = p02;
p10 = p12;
sd0.d64 = *(TYPE_64BIT *) (buff0 + i);
sd1.d64 = *(TYPE_64BIT *) (buff1 + i);
p01 = (mlib_d64)sd0.i32s.i0;
p02 = (mlib_d64)sd0.i32s.i1;
p11 = (mlib_d64)sd1.i32s.i0;
p12 = (mlib_d64)sd1.i32s.i1;
LOAD_BUFF(buff2);
dd0.i32s.i0 =
CLAMP_S32(p00 * k0 + p01 * k1 + p10 * k2 +
p11 * k3);
dd0.i32s.i1 =
CLAMP_S32(p01 * k0 + p02 * k1 + p11 * k2 +
p12 * k3);
dd1.d64 = vis_fpack32(dd1.d64, dd0.d64);
STORE2(dd1.i32s.i0, dd1.i32s.i1);
sp += chan2;
dp += chan2;
}
for (; i < wid; i++) {
d64_2x32 dd0, dd1;
p00 = buff0[i - 1];
p10 = buff1[i - 1];
p01 = buff0[i];
p11 = buff1[i];
buff2[i] = (mlib_s32)sp[0];
dd0.i32s.i1 =
CLAMP_S32(p00 * k0 + p01 * k1 + p10 * k2 +
p11 * k3);
dd1.d64 = vis_fpack32(dd1.d64, dd0.d64);
dp[0] = dd1.i32s.i1;
sp += chan1;
dp += chan1;
}
if (dx_r != 0)
buff2[swid - 1] = buff2[swid - 2];
if (j < hgt - dy_b - 2)
sl += sll;
dl += dll;
buffT = buff0;
buff0 = buff1;
buff1 = buff2;
buff2 = buffT;
}
}
if (pbuff != (mlib_s32 *)buff_arr)
__mlib_free(pbuff);
return (MLIB_SUCCESS);
}
mlib_status
mlib_v_conv5x5_8nw_4(
mlib_image *dst,
const mlib_image *src,
const mlib_s32 *kernel,
mlib_s32 scalef_expon)
{
/* pointers to dst row */
mlib_u8 *da, *d_a;
/* pointers to src, dst data */
mlib_u8 *adr_dst, *adr_src, *dend;
/* pointers to src rows */
mlib_u8 *sa, *sa1, *sa2, *sa3, *sa4;
/* pointers to rows in interm. src buf */
mlib_d64 *buff_src, *sbuf1, *sbuf2, *prow;
/* pointers to rows in interm. src buf */
mlib_d64 *sbuf3, *sbuf4, *sbuf5;
/* pointer to row in interm. dst buf */
mlib_d64 *dbuf, *dbuf1;
/* mlib_d64 pointers to rows in interm. src buf */
mlib_d64 *s1, *s2, *s3, *s4, *s5;
/* mlib_d64 pointer to row in interm. dst buf */
mlib_d64 *ddst;
/* data */
mlib_d64 d1, d2, d3, d4, d5;
/* data */
mlib_d64 d11, d12, d13, d14, d15;
/* data */
mlib_d64 d21, d22, d23, d24, d25;
/* data */
mlib_d64 dt_1, dt_2, dt_3, dt_4, dt_5;
mlib_f32 k1k2, k3k4, k5k6, k7k8;
mlib_f32 k9k10, k11k12, k13k14, k15k16;
mlib_f32 k17k18, k19k20, k21k22, k23k24, k25;
/* src, dst and interm. buf. strides */
mlib_s32 dlb, slb, buf_slb;
mlib_s32 dh, dw;
mlib_d64 out0, out1;
mlib_d64 tmp0, tmp1, rnd;
mlib_d64 *dsa, *dp;
mlib_d64 sd0, sd1;
mlib_s32 emask;
mlib_s32 rval, gsr_scale, i, j;
gsr_scale = 31 - scalef_expon;
vis_write_gsr((gsr_scale << 3));
rval = mlib_round_8[gsr_scale];
rnd = vis_freg_pair(vis_to_float(rval), vis_to_float(rval));
GET_SRC_DST_PARAMETERS();
LOAD_KERNEL_INTO_FLOAT();
buf_slb = (4 * dw + 24) >> 3;
PREPARE_INTERM_BUFFERS();
dw -= 4;
dw *= 4;
dh -= 4;
sa = adr_src;
sa1 = sa + slb;
sa2 = sa1 + slb;
sa3 = sa2 + slb;
sa4 = sa3 + slb;
d_a = adr_dst + 2 * dlb + 8;
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf2, sa);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(16);
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf3, sa1);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(16);
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf4, sa2);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(16);
/* load interm. src buff */
PREPARE_TO_LOAD_LINE(sbuf5, sa3);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER(16);
#pragma pipeloop(0)
for (j = 0; j < dh; j++) {
LOOP_INI();
PREPARE_TO_LOAD_LINE(sbuf5, sa4);
#pragma pipeloop(0)
LOAD_LINE_INTO_BUFFER_NF(16);
vis_alignaddr(s1, 4);
dbuf1 = dbuf;
d1 = *s1;
d2 = *s2;
d3 = *s3;
d11 = *(s1 + 1);
d12 = *(s2 + 1);
d13 = *(s3 + 1);
#pragma pipeloop(0)
for (i = 0; i < dw; i += 8) {
d21 = *(s1 + 2);
d22 = *(s2 + 2);
d23 = *(s3 + 2);
out0 = out1 = rnd;
CONV_AU(d1, k1k2);
CONV_AL(d2, k5k6);
CONV_AU(d3, k11k12);
dt_1 = vis_faligndata(d1, d11);
dt_2 = vis_faligndata(d2, d12);
dt_3 = vis_faligndata(d3, d13);
CONV_AL(dt_1, k1k2);
CONV_AU(dt_2, k7k8);
CONV_AL(dt_3, k11k12);
CONV_AU(d11, k3k4);
CONV_AL(d12, k7k8);
CONV_AU(d13, k13k14);
dt_1 = vis_faligndata(d11, d21);
dt_2 = vis_faligndata(d12, d22);
dt_3 = vis_faligndata(d13, d23);
CONV_AL(dt_1, k3k4);
CONV_AU(dt_2, k9k10);
CONV_AL(dt_3, k13k14);
CONV_AU(d21, k5k6);
CONV_AL(d22, k9k10);
CONV_AU(d23, k15k16);
dbuf1[0] = out0;
dbuf1[1] = out1;
dbuf1 += 2;
d1 = d11;
d2 = d12;
d3 = d13;
d11 = d21;
d12 = d22;
d13 = d23;
s1++;
s2++;
s3++;
}
dbuf1 = dbuf;
d4 = *s4;
d5 = *s5;
d14 = *(s4 + 1);
d15 = *(s5 + 1);
#pragma pipeloop(0)
for (i = 0; i < dw; i += 8) {
d24 = *(s4 + 2);
d25 = *(s5 + 2);
out0 = dbuf1[0];
out1 = dbuf1[1];
CONV_AL(d4, k15k16);
CONV_AU(d5, k21k22);
dt_4 = vis_faligndata(d4, d14);
dt_5 = vis_faligndata(d5, d15);
CONV_AU(dt_4, k17k18);
CONV_AL(dt_5, k21k22);
CONV_AL(d14, k17k18);
CONV_AU(d15, k23k24);
dt_4 = vis_faligndata(d14, d24);
dt_5 = vis_faligndata(d15, d25);
CONV_AU(dt_4, k19k20);
CONV_AL(dt_5, k23k24);
CONV_AL(d24, k19k20);
CONV_AU(d25, k25);
dbuf1 += 2;
(*ddst++) = vis_fpack16_pair(out0, out1);
d4 = d14;
d5 = d15;
d14 = d24;
d15 = d25;
s4++;
s5++;
}
PREPARE_TO_COPY_INTERM_BUF_TO_DST();
#pragma pipeloop(0)
COPY_INTERM_BUF_TO_DST();
COPY_TAIL();
sa4 = sa4 + slb;
d_a += dlb;
}
__mlib_free(buff_src);
return (MLIB_SUCCESS);
}
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);
}