mlib_status mlib_ImageConvKernelConvert(mlib_s32 *ikernel,
mlib_s32 *iscale,
const mlib_d64 *fkernel,
mlib_s32 m,
mlib_s32 n,
mlib_type type)
{
mlib_d64 sum_pos, sum_neg, sum, norm, max, f;
mlib_s32 isum_pos, isum_neg, isum, test;
mlib_s32 i, scale, scale1, chk_flag;
if (ikernel == NULL || iscale == NULL || fkernel == NULL || m < 1 || n < 1) {
return MLIB_FAILURE;
}
if ((type == MLIB_BYTE) || (type == MLIB_SHORT) || (type == MLIB_USHORT)) {
if (type != MLIB_SHORT) { /* MLIB_BYTE, MLIB_USHORT */
sum_pos = 0;
sum_neg = 0;
for (i = 0; i < m * n; i++) {
if (fkernel[i] > 0)
sum_pos += fkernel[i];
else
sum_neg -= fkernel[i];
}
sum = (sum_pos > sum_neg) ? sum_pos : sum_neg;
scale = mlib_ilogb(sum);
scale++;
scale = 31 - scale;
}
else { /* MLIB_SHORT */
sum = 0;
max = 0;
for (i = 0; i < m * n; i++) {
f = mlib_fabs(fkernel[i]);
sum += f;
max = (max > f) ? max : f;
}
scale1 = mlib_ilogb(max) + 1;
scale = mlib_ilogb(sum);
scale = (scale > scale1) ? scale : scale1;
scale++;
scale = 32 - scale;
}
if (scale <= 16)
return MLIB_FAILURE;
if (scale > 31)
scale = 31;
*iscale = scale;
chk_flag = mlib_ImageConvVersion(m, n, scale, type);
if (!chk_flag) {
norm = (1u << scale);
for (i = 0; i < m * n; i++) {
CLAMP_S32(ikernel[i], fkernel[i] * norm);
}
return MLIB_SUCCESS;
}
/* try to round coefficients */
#ifdef __sparc
scale1 = 16; /* shift of coefficients is 16 */
#else
if (chk_flag == 3)
scale1 = 16; /* MMX */
else
scale1 = (type == MLIB_BYTE) ? 8 : 16;
#endif /* __sparc */
norm = (1u << (scale - scale1));
for (i = 0; i < m * n; i++) {
if (fkernel[i] > 0)
ikernel[i] = (mlib_s32) (fkernel[i] * norm + 0.5);
else
ikernel[i] = (mlib_s32) (fkernel[i] * norm - 0.5);
}
isum_pos = 0;
isum_neg = 0;
test = 0;
for (i = 0; i < m * n; i++) {
if (ikernel[i] > 0)
isum_pos += ikernel[i];
else
isum_neg -= ikernel[i];
}
if (type == MLIB_BYTE || type == MLIB_USHORT) {
isum = (isum_pos > isum_neg) ? isum_pos : isum_neg;
if (isum >= (1 << (31 - scale1)))
test = 1;
}
else {
isum = isum_pos + isum_neg;
if (isum >= (1 << (32 - scale1)))
test = 1;
for (i = 0; i < m * n; i++) {
if (abs(ikernel[i]) >= (1 << (31 - scale1)))
test = 1;
}
}
if (test == 1) { /* rounding according scale1 cause overflow, truncate instead of round */
for (i = 0; i < m * n; i++)
ikernel[i] = (mlib_s32) (fkernel[i] * norm) << scale1;
}
else { /* rounding is Ok */
for (i = 0; i < m * n; i++)
ikernel[i] = ikernel[i] << scale1;
}
return MLIB_SUCCESS;
}
else if ((type == MLIB_INT) || (type == MLIB_BIT)) {
max = 0;
for (i = 0; i < m * n; i++) {
f = mlib_fabs(fkernel[i]);
max = (max > f) ? max : f;
}
scale = mlib_ilogb(max);
if (scale > 29)
return MLIB_FAILURE;
if (scale < -100)
scale = -100;
*iscale = 29 - scale;
scale = 29 - scale;
norm = 1.0;
while (scale > 30) {
norm *= (1 << 30);
scale -= 30;
}
norm *= (1 << scale);
for (i = 0; i < m * n; i++) {
if (fkernel[i] > 0) {
CLAMP_S32(ikernel[i], fkernel[i] * norm + 0.5);
}
else {
CLAMP_S32(ikernel[i], fkernel[i] * norm - 0.5);
}
}
return MLIB_SUCCESS;
}
else {
return MLIB_FAILURE;
}
}
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);
}