Improve perf of opj_t1_dec_sigpass_mqc_vsc() and opj_t1_dec_refpass_mqc_vsc() with loop unrolling

This commit is contained in:
Even Rouault 2016-05-23 13:45:15 +02:00
parent 8371491a99
commit 107eb31531

View File

@ -660,27 +660,48 @@ static void opj_t1_dec_sigpass_mqc_vsc(
{ {
OPJ_INT32 one, half, oneplushalf, vsc; OPJ_INT32 one, half, oneplushalf, vsc;
OPJ_UINT32 i, j, k; OPJ_UINT32 i, j, k;
opj_colflag_t *colflags1 = &t1->colflags[t1->flags_stride + 1]; OPJ_INT32 *data1 = t1->data;
opj_flag_t *flags1 = &t1->flags[1];
opj_colflag_t *colflags1 = &t1->colflags[t1->flags_stride + 1];
one = 1 << bpno; one = 1 << bpno;
half = one >> 1; half = one >> 1;
oneplushalf = one | half; oneplushalf = one | half;
for (k = 0; k < t1->h; k += 4) { for (k = 0; k < (t1->h & ~3); k += 4) {
for (i = 0; i < t1->w; ++i) { for (i = 0; i < t1->w; ++i) {
OPJ_INT32 *data2 = data1 + i;
opj_flag_t *flags2 = flags1 + i;
opj_colflag_t *colflags2 = colflags1 + i; opj_colflag_t *colflags2 = colflags1 + i;
for (j = k; j < k + 4 && j < t1->h; ++j) { flags2 += t1->flags_stride;
vsc = (j == k + 3 || j == t1->h - 1) ? 1 : 0; opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 0);
opj_t1_dec_sigpass_step_mqc_vsc( data2 += t1->w;
t1, flags2 += t1->flags_stride;
&t1->flags[((j+1) * t1->flags_stride) + i + 1], opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 1);
colflags2, data2 += t1->w;
&t1->data[(j * t1->w) + i], flags2 += t1->flags_stride;
oneplushalf, opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 2);
vsc, data2 += t1->w;
j - k); flags2 += t1->flags_stride;
} opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 1, 3);
data2 += t1->w;
} }
data1 += t1->w << 2;
flags1 += t1->flags_stride << 2;
colflags1 += t1->flags_stride; colflags1 += t1->flags_stride;
} }
for (i = 0; i < t1->w; ++i) {
opj_colflag_t *colflags2 = colflags1 + i;
for (j = k; j < t1->h; ++j) {
vsc = (j == t1->h - 1) ? 1 : 0;
opj_t1_dec_sigpass_step_mqc_vsc(
t1,
&t1->flags[((j+1) * t1->flags_stride) + i + 1],
colflags2,
&t1->data[(j * t1->w) + i],
oneplushalf,
vsc,
j - k);
}
}
} /* VSC and BYPASS by Antonin */ } /* VSC and BYPASS by Antonin */
@ -921,27 +942,48 @@ static void opj_t1_dec_refpass_mqc_vsc(
OPJ_INT32 one, poshalf, neghalf; OPJ_INT32 one, poshalf, neghalf;
OPJ_UINT32 i, j, k; OPJ_UINT32 i, j, k;
OPJ_INT32 vsc; OPJ_INT32 vsc;
OPJ_INT32 *data1 = t1->data;
opj_flag_t *flags1 = &t1->flags[1];
opj_colflag_t *colflags1 = &t1->colflags[t1->flags_stride + 1]; opj_colflag_t *colflags1 = &t1->colflags[t1->flags_stride + 1];
one = 1 << bpno; one = 1 << bpno;
poshalf = one >> 1; poshalf = one >> 1;
neghalf = bpno > 0 ? -poshalf : -1; neghalf = bpno > 0 ? -poshalf : -1;
for (k = 0; k < t1->h; k += 4) { for (k = 0; k < (t1->h & ~3); k += 4) {
for (i = 0; i < t1->w; ++i) { for (i = 0; i < t1->w; ++i) {
OPJ_INT32 *data2 = data1 + i;
opj_flag_t *flags2 = flags1 + i;
opj_colflag_t *colflags2 = colflags1 + i; opj_colflag_t *colflags2 = colflags1 + i;
for (j = k; j < k + 4 && j < t1->h; ++j) { flags2 += t1->flags_stride;
vsc = ((j == k + 3 || j == t1->h - 1)) ? 1 : 0; opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 0);
opj_t1_dec_refpass_step_mqc_vsc( data2 += t1->w;
t1, flags2 += t1->flags_stride;
&t1->flags[((j+1) * t1->flags_stride) + i + 1], opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 1);
colflags2, data2 += t1->w;
&t1->data[(j * t1->w) + i], flags2 += t1->flags_stride;
poshalf, opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 2);
neghalf, data2 += t1->w;
vsc, j - k); flags2 += t1->flags_stride;
} opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 1, 3);
data2 += t1->w;
} }
data1 += t1->w << 2;
flags1 += t1->flags_stride << 2;
colflags1 += t1->flags_stride; colflags1 += t1->flags_stride;
} }
for (i = 0; i < t1->w; ++i) {
opj_colflag_t *colflags2 = colflags1 + i;
for (j = k; j < t1->h; ++j) {
vsc = (j == t1->h - 1) ? 1 : 0;
opj_t1_dec_refpass_step_mqc_vsc(
t1,
&t1->flags[((j+1) * t1->flags_stride) + i + 1],
colflags2,
&t1->data[(j * t1->w) + i],
poshalf, neghalf,
vsc,
j - k);
}
}
} /* VSC and BYPASS by Antonin */ } /* VSC and BYPASS by Antonin */