diff options
| author | Even Rouault <even.rouault@spatialys.com> | 2016-05-23 13:45:15 +0200 |
|---|---|---|
| committer | Even Rouault <even.rouault@spatialys.com> | 2016-05-23 13:45:15 +0200 |
| commit | 107eb31531ca688e2799406e69e9383efc13448f (patch) | |
| tree | ea630b3364734ee4d1bf9652897aa97d4309d335 /src | |
| parent | 8371491a9968a31ce16d6ce37b775ef3c7d090c8 (diff) | |
Improve perf of opj_t1_dec_sigpass_mqc_vsc() and opj_t1_dec_refpass_mqc_vsc() with loop unrolling
Diffstat (limited to 'src')
| -rw-r--r-- | src/lib/openjp2/t1.c | 92 |
1 files changed, 67 insertions, 25 deletions
diff --git a/src/lib/openjp2/t1.c b/src/lib/openjp2/t1.c index 0023ad74..277261d7 100644 --- a/src/lib/openjp2/t1.c +++ b/src/lib/openjp2/t1.c @@ -660,27 +660,48 @@ static void opj_t1_dec_sigpass_mqc_vsc( { OPJ_INT32 one, half, oneplushalf, vsc; 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; half = one >> 1; 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) { + OPJ_INT32 *data2 = data1 + i; + opj_flag_t *flags2 = flags1 + i; opj_colflag_t *colflags2 = colflags1 + i; - for (j = k; j < k + 4 && j < t1->h; ++j) { - vsc = (j == k + 3 || 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); - } + flags2 += t1->flags_stride; + opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 0); + data2 += t1->w; + flags2 += t1->flags_stride; + opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 1); + data2 += t1->w; + flags2 += t1->flags_stride; + opj_t1_dec_sigpass_step_mqc_vsc(t1, flags2, colflags2, data2, oneplushalf, 0, 2); + data2 += t1->w; + 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; } + 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 */ @@ -921,27 +942,48 @@ static void opj_t1_dec_refpass_mqc_vsc( OPJ_INT32 one, poshalf, neghalf; OPJ_UINT32 i, j, k; 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]; one = 1 << bpno; poshalf = one >> 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) { + OPJ_INT32 *data2 = data1 + i; + opj_flag_t *flags2 = flags1 + i; opj_colflag_t *colflags2 = colflags1 + i; - for (j = k; j < k + 4 && j < t1->h; ++j) { - vsc = ((j == k + 3 || 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); - } + flags2 += t1->flags_stride; + opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 0); + data2 += t1->w; + flags2 += t1->flags_stride; + opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 1); + data2 += t1->w; + flags2 += t1->flags_stride; + opj_t1_dec_refpass_step_mqc_vsc(t1, flags2, colflags2, data2, poshalf, neghalf, 0, 2); + data2 += t1->w; + 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; } + 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 */ |