falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
68854 Commits
32 Branches
283MB
Tree: 9bff052b51
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '9bff052b51'
${ noResults }
FFmpeg/libavfilter/vf_fspp.c

700 lines
25KB
Raw Blame History 

  1. /*

  2.  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>

  3.  * Copyright (C) 2005 Nikolaj Poroshin <porosh3@psu.ru>

  4.  * Copyright (c) 2014 Arwa Arif <arwaarif1994@gmail.com>

  5.  *

  6.  * This file is part of FFmpeg.

  7.  *

  8.  * FFmpeg is free software; you can redistribute it and/or modify

  9.  * it under the terms of the GNU General Public License as published by

  10.  * the Free Software Foundation; either version 2 of the License, or

  11.  * (at your option) any later version.

  12.  *

  13.  * FFmpeg is distributed in the hope that it will be useful,

  14.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  15.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

  16.  * GNU General Public License for more details.

  17.  *

  18.  * You should have received a copy of the GNU General Public License along

  19.  * with FFmpeg; if not, write to the Free Software Foundation, Inc.,

  20.  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.

  21.  */

  22. 

  23. /**

  24.  * @file

  25.  * Fast Simple Post-processing filter

  26.  * This implementation is based on an algorithm described in

  27.  * "Aria Nosratinia Embedded Post-Processing for

  28.  * Enhancement of Compressed Images (1999)"

  29.  * (http://www.utdallas.edu/~aria/papers/vlsisp99.pdf)

  30.  * Further, with splitting (I)DCT into horizontal/vertical passes, one of

  31.  * them can be performed once per block, not per pixel. This allows for much

  32.  * higher speed.

  33.  *

  34.  * Originally written by Michael Niedermayer and Nikolaj for the MPlayer

  35.  * project, and ported by Arwa Arif for FFmpeg.

  36.  */

  37. 

  38. #include "libavutil/avassert.h"

  39. #include "libavutil/imgutils.h"

  40. #include "libavutil/opt.h"

  41. #include "libavutil/pixdesc.h"

  42. #include "internal.h"

  43. #include "libavcodec/avcodec.h" //for reference to FF_QSCALE_TYPE

  44. #include "vf_fspp.h"

  45. 

  46. #define OFFSET(x) offsetof(FSPPContext, x)

  47. #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM

  48. static const AVOption fspp_options[] = {

  49.     { "quality",       "set quality",                          OFFSET(log2_count),    AV_OPT_TYPE_INT, {.i64 = 4},   4, MAX_LEVEL, FLAGS },

  50.     { "qp",            "force a constant quantizer parameter", OFFSET(qp),            AV_OPT_TYPE_INT, {.i64 = 0},   0, 64,        FLAGS },

  51.     { "strength",      "set filter strength",                  OFFSET(strength),      AV_OPT_TYPE_INT, {.i64 = 0}, -15, 32,        FLAGS },

  52.     { "use_bframe_qp", "use B-frames' QP",                     OFFSET(use_bframe_qp), AV_OPT_TYPE_INT, {.i64 = 0},   0, 1,         FLAGS },

  53.     { NULL }

  54. };

  55. 

  56. AVFILTER_DEFINE_CLASS(fspp);

  57. 

  58. DECLARE_ALIGNED(32, static const uint8_t, dither)[8][8] = {

  59.     {  0,  48,  12,  60,   3,  51,  15,  63, },

  60.     { 32,  16,  44,  28,  35,  19,  47,  31, },

  61.     {  8,  56,   4,  52,  11,  59,   7,  55, },

  62.     { 40,  24,  36,  20,  43,  27,  39,  23, },

  63.     {  2,  50,  14,  62,   1,  49,  13,  61, },

  64.     { 34,  18,  46,  30,  33,  17,  45,  29, },

  65.     { 10,  58,   6,  54,   9,  57,   5,  53, },

  66.     { 42,  26,  38,  22,  41,  25,  37,  21, },

  67. };

  68. 

  69. static const short custom_threshold[64] = {

  70. // values (296) can't be too high

  71. // -it causes too big quant dependence

  72. // or maybe overflow(check), which results in some flashing

  73.      71, 296, 295, 237,  71,  40,  38,  19,

  74.     245, 193, 185, 121, 102,  73,  53,  27,

  75.     158, 129, 141, 107,  97,  73,  50,  26,

  76.     102, 116, 109,  98,  82,  66,  45,  23,

  77.      71,  94,  95,  81,  70,  56,  38,  20,

  78.      56,  77,  74,  66,  56,  44,  30,  15,

  79.      38,  53,  50,  45,  38,  30,  21,  11,

  80.      20,  27,  26,  23,  20,  15,  11,   5

  81. };

  82. 

  83. static inline int norm_qscale(int qscale, int type)

  84. {

  85.     switch (type) {

  86.     case FF_QSCALE_TYPE_MPEG1: return qscale;

  87.     case FF_QSCALE_TYPE_MPEG2: return qscale >> 1;

  88.     case FF_QSCALE_TYPE_H264:  return qscale >> 2;

  89.     case FF_QSCALE_TYPE_VP56:  return (63 - qscale + 2) >> 2;

  90.     }

  91.     return qscale;

  92. }

  93. 

  94. //This func reads from 1 slice, 1 and clears 0 & 1

  95. static void store_slice_c(uint8_t *dst, int16_t *src,

  96.                           ptrdiff_t dst_stride, ptrdiff_t src_stride,

  97.                           ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)

  98. {

  99.     int y, x;

  100. #define STORE(pos)                                                             \

  101.     temp = (src[x + pos] + (d[pos] >> log2_scale)) >> (6 - log2_scale);        \

  102.     src[x + pos] = src[x + pos - 8 * src_stride] = 0;                          \

  103.     if (temp & 0x100) temp = ~(temp >> 31);                                    \

  104.     dst[x + pos] = temp;

  105. 

  106.     for (y = 0; y < height; y++) {

  107.         const uint8_t *d = dither[y];

  108.         for (x = 0; x < width; x += 8) {

  109.             int temp;

  110.             STORE(0);

  111.             STORE(1);

  112.             STORE(2);

  113.             STORE(3);

  114.             STORE(4);

  115.             STORE(5);

  116.             STORE(6);

  117.             STORE(7);

  118.         }

  119.         src += src_stride;

  120.         dst += dst_stride;

  121.     }

  122. }

  123. 

  124. //This func reads from 2 slices, 0 & 2  and clears 2-nd

  125. static void store_slice2_c(uint8_t *dst, int16_t *src,

  126.                            ptrdiff_t dst_stride, ptrdiff_t src_stride,

  127.                            ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)

  128. {

  129.     int y, x;

  130. #define STORE2(pos)                                                                                       \

  131.     temp = (src[x + pos] + src[x + pos + 16 * src_stride] + (d[pos] >> log2_scale)) >> (6 - log2_scale);  \

  132.     src[x + pos + 16 * src_stride] = 0;                                                                   \

  133.     if (temp & 0x100) temp = ~(temp >> 31);                                                               \

  134.     dst[x + pos] = temp;

  135. 

  136.     for (y = 0; y < height; y++) {

  137.         const uint8_t *d = dither[y];

  138.         for (x = 0; x < width; x += 8) {

  139.             int temp;

  140.             STORE2(0);

  141.             STORE2(1);

  142.             STORE2(2);

  143.             STORE2(3);

  144.             STORE2(4);

  145.             STORE2(5);

  146.             STORE2(6);

  147.             STORE2(7);

  148.         }

  149.         src += src_stride;

  150.         dst += dst_stride;

  151.     }

  152. }

  153. 

  154. static void mul_thrmat_c(int16_t *thr_adr_noq, int16_t *thr_adr, int q)

  155. {

  156.     int a;

  157.     for (a = 0; a < 64; a++)

  158.         thr_adr[a] = q * thr_adr_noq[a];

  159. }

  160. 

  161. static void filter(FSPPContext *p, uint8_t *dst, uint8_t *src,

  162.                    int dst_stride, int src_stride,

  163.                    int width, int height,

  164.                    uint8_t *qp_store, int qp_stride, int is_luma)

  165. {

  166.     int x, x0, y, es, qy, t;

  167. 

  168.     const int stride = is_luma ? p->temp_stride : (width + 16);

  169.     const int step = 6 - p->log2_count;

  170.     const int qpsh = 4 - p->hsub * !is_luma;

  171.     const int qpsv = 4 - p->vsub * !is_luma;

  172. 

  173.     DECLARE_ALIGNED(32, int32_t, block_align)[4 * 8 * BLOCKSZ + 4 * 8 * BLOCKSZ];

  174.     int16_t *block  = (int16_t *)block_align;

  175.     int16_t *block3 = (int16_t *)(block_align + 4 * 8 * BLOCKSZ);

  176. 

  177.     memset(block3, 0, 4 * 8 * BLOCKSZ);

  178. 

  179.     if (!src || !dst) return;

  180. 

  181.     for (y = 0; y < height; y++) {

  182.         int index = 8 + 8 * stride + y * stride;

  183.         memcpy(p->src + index, src + y * src_stride, width);

  184.         for (x = 0; x < 8; x++) {

  185.             p->src[index         - x - 1] = p->src[index +         x    ];

  186.             p->src[index + width + x    ] = p->src[index + width - x - 1];

  187.         }

  188.     }

  189. 

  190.     for (y = 0; y < 8; y++) {

  191.         memcpy(p->src + (     7 - y    ) * stride, p->src + (     y + 8    ) * stride, stride);

  192.         memcpy(p->src + (height + 8 + y) * stride, p->src + (height - y + 7) * stride, stride);

  193.     }

  194.     //FIXME (try edge emu)

  195. 

  196.     for (y = 8; y < 24; y++)

  197.         memset(p->temp + 8 + y * stride, 0, width * sizeof(int16_t));

  198. 

  199.     for (y = step; y < height + 8; y += step) {    //step= 1,2

  200.         const int y1 = y - 8 + step;                 //l5-7  l4-6;

  201.         qy = y - 4;

  202. 

  203.         if (qy > height - 1) qy = height - 1;

  204.         if (qy < 0) qy = 0;

  205. 

  206.         qy = (qy >> qpsv) * qp_stride;

  207.         p->row_fdct(block, p->src + y * stride + 2 - (y&1), stride, 2);

  208. 

  209.         for (x0 = 0; x0 < width + 8 - 8 * (BLOCKSZ - 1); x0 += 8 * (BLOCKSZ - 1)) {

  210.             p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2 - (y&1), stride, 2 * (BLOCKSZ - 1));

  211. 

  212.             if (p->qp)

  213.                 p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block + 0 * 8, block3 + 0 * 8, 8 * (BLOCKSZ - 1)); //yes, this is a HOTSPOT

  214.             else

  215.                 for (x = 0; x < 8 * (BLOCKSZ - 1); x += 8) {

  216.                     t = x + x0 - 2;                    //correct t=x+x0-2-(y&1), but its the same

  217. 

  218.                     if (t < 0) t = 0;                   //t always < width-2

  219. 

  220.                     t = qp_store[qy + (t >> qpsh)];

  221.                     t = norm_qscale(t, p->qscale_type);

  222. 

  223.                     if (t != p->prev_q) p->prev_q = t, p->mul_thrmat((int16_t *)(&p->threshold_mtx_noq[0]), (int16_t *)(&p->threshold_mtx[0]), t);

  224.                     p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block + x * 8, block3 + x * 8, 8); //yes, this is a HOTSPOT

  225.                 }

  226.             p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 + 2 - (y & 1), stride, 2 * (BLOCKSZ - 1));

  227.             memmove(block,  block  + (BLOCKSZ - 1) * 64, 8 * 8 * sizeof(int16_t)); //cycling

  228.             memmove(block3, block3 + (BLOCKSZ - 1) * 64, 6 * 8 * sizeof(int16_t));

  229.         }

  230. 

  231.         es = width + 8 - x0; //  8, ...

  232.         if (es > 8)

  233.             p->row_fdct(block + 8 * 8, p->src + y * stride + 8 + x0 + 2 - (y & 1), stride, (es - 4) >> 2);

  234. 

  235.         p->column_fidct((int16_t *)(&p->threshold_mtx[0]), block, block3, es&(~1));

  236.         p->row_idct(block3 + 0 * 8, p->temp + (y & 15) * stride + x0 + 2 - (y & 1), stride, es >> 2);

  237. 

  238.         if (!(y1 & 7) && y1) {

  239.             if (y1 & 8)

  240.                 p->store_slice(dst + (y1 - 8) * dst_stride, p->temp + 8 + 8 * stride,

  241.                                dst_stride, stride, width, 8, 5 - p->log2_count);

  242.             else

  243.                 p->store_slice2(dst + (y1 - 8) * dst_stride, p->temp + 8 + 0 * stride,

  244.                                 dst_stride, stride, width, 8, 5 - p->log2_count);

  245.         }

  246.     }

  247. 

  248.     if (y & 7) {  // height % 8 != 0

  249.         if (y & 8)

  250.             p->store_slice(dst + ((y - 8) & ~7) * dst_stride, p->temp + 8 + 8 * stride,

  251.                            dst_stride, stride, width, y&7, 5 - p->log2_count);

  252.         else

  253.             p->store_slice2(dst + ((y - 8) & ~7) * dst_stride, p->temp + 8 + 0 * stride,

  254.                             dst_stride, stride, width, y&7, 5 - p->log2_count);

  255.     }

  256. }

  257. 

  258. static void column_fidct_c(int16_t *thr_adr, int16_t *data, int16_t *output, int cnt)

  259. {

  260.     int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;

  261.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

  262.     int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;

  263.     int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;

  264. 

  265.     int16_t *dataptr;

  266.     int16_t *wsptr;

  267.     int16_t *threshold;

  268.     int ctr;

  269. 

  270.     dataptr = data;

  271.     wsptr = output;

  272. 

  273.     for (; cnt > 0; cnt -= 2) { //start positions

  274.         threshold = (int16_t *)thr_adr;//threshold_mtx

  275.         for (ctr = DCTSIZE; ctr > 0; ctr--) {

  276.             // Process columns from input, add to output.

  277.             tmp0 = dataptr[DCTSIZE * 0] + dataptr[DCTSIZE * 7];

  278.             tmp7 = dataptr[DCTSIZE * 0] - dataptr[DCTSIZE * 7];

  279. 

  280.             tmp1 = dataptr[DCTSIZE * 1] + dataptr[DCTSIZE * 6];

  281.             tmp6 = dataptr[DCTSIZE * 1] - dataptr[DCTSIZE * 6];

  282. 

  283.             tmp2 = dataptr[DCTSIZE * 2] + dataptr[DCTSIZE * 5];

  284.             tmp5 = dataptr[DCTSIZE * 2] - dataptr[DCTSIZE * 5];

  285. 

  286.             tmp3 = dataptr[DCTSIZE * 3] + dataptr[DCTSIZE * 4];

  287.             tmp4 = dataptr[DCTSIZE * 3] - dataptr[DCTSIZE * 4];

  288. 

  289.             // Even part of FDCT

  290. 

  291.             tmp10 = tmp0 + tmp3;

  292.             tmp13 = tmp0 - tmp3;

  293.             tmp11 = tmp1 + tmp2;

  294.             tmp12 = tmp1 - tmp2;

  295. 

  296.             d0 = tmp10 + tmp11;

  297.             d4 = tmp10 - tmp11;

  298. 

  299.             z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);

  300.             d2 = tmp13 + z1;

  301.             d6 = tmp13 - z1;

  302. 

  303.             // Even part of IDCT

  304. 

  305.             THRESHOLD(tmp0, d0, threshold[0 * 8]);

  306.             THRESHOLD(tmp1, d2, threshold[2 * 8]);

  307.             THRESHOLD(tmp2, d4, threshold[4 * 8]);

  308.             THRESHOLD(tmp3, d6, threshold[6 * 8]);

  309.             tmp0 += 2;

  310.             tmp10 = (tmp0 + tmp2) >> 2;

  311.             tmp11 = (tmp0 - tmp2) >> 2;

  312. 

  313.             tmp13 = (tmp1 + tmp3) >>2; //+2 !  (psnr decides)

  314.             tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2

  315. 

  316.             tmp0 = tmp10 + tmp13; //->temps

  317.             tmp3 = tmp10 - tmp13; //->temps

  318.             tmp1 = tmp11 + tmp12; //->temps

  319.             tmp2 = tmp11 - tmp12; //->temps

  320. 

  321.             // Odd part of FDCT

  322. 

  323.             tmp10 = tmp4 + tmp5;

  324.             tmp11 = tmp5 + tmp6;

  325.             tmp12 = tmp6 + tmp7;

  326. 

  327.             z5 = MULTIPLY16H((tmp10 - tmp12) << 2, FIX_0_382683433);

  328.             z2 = MULTIPLY16H(tmp10 << 2, FIX_0_541196100) + z5;

  329.             z4 = MULTIPLY16H(tmp12 << 2, FIX_1_306562965) + z5;

  330.             z3 = MULTIPLY16H(tmp11 << 2, FIX_0_707106781);

  331. 

  332.             z11 = tmp7 + z3;

  333.             z13 = tmp7 - z3;

  334. 

  335.             d5 = z13 + z2;

  336.             d3 = z13 - z2;

  337.             d1 = z11 + z4;

  338.             d7 = z11 - z4;

  339. 

  340.             // Odd part of IDCT

  341. 

  342.             THRESHOLD(tmp4, d1, threshold[1 * 8]);

  343.             THRESHOLD(tmp5, d3, threshold[3 * 8]);

  344.             THRESHOLD(tmp6, d5, threshold[5 * 8]);

  345.             THRESHOLD(tmp7, d7, threshold[7 * 8]);

  346. 

  347.             //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0

  348.             z13 = tmp6 + tmp5;

  349.             z10 = (tmp6 - tmp5) << 1;

  350.             z11 = tmp4 + tmp7;

  351.             z12 = (tmp4 - tmp7) << 1;

  352. 

  353.             tmp7  = (z11 + z13) >> 2; //+2 !

  354.             tmp11 = MULTIPLY16H((z11 - z13) << 1, FIX_1_414213562);

  355.             z5    = MULTIPLY16H(z10 + z12,        FIX_1_847759065);

  356.             tmp10 = MULTIPLY16H(z12,              FIX_1_082392200) - z5;

  357.             tmp12 = MULTIPLY16H(z10,              FIX_2_613125930) + z5; // - !!

  358. 

  359.             tmp6 = tmp12 - tmp7;

  360.             tmp5 = tmp11 - tmp6;

  361.             tmp4 = tmp10 + tmp5;

  362. 

  363.             wsptr[DCTSIZE * 0] +=  (tmp0 + tmp7);

  364.             wsptr[DCTSIZE * 1] +=  (tmp1 + tmp6);

  365.             wsptr[DCTSIZE * 2] +=  (tmp2 + tmp5);

  366.             wsptr[DCTSIZE * 3] +=  (tmp3 - tmp4);

  367.             wsptr[DCTSIZE * 4] +=  (tmp3 + tmp4);

  368.             wsptr[DCTSIZE * 5] +=  (tmp2 - tmp5);

  369.             wsptr[DCTSIZE * 6]  =  (tmp1 - tmp6);

  370.             wsptr[DCTSIZE * 7]  =  (tmp0 - tmp7);

  371.             //

  372.             dataptr++; //next column

  373.             wsptr++;

  374.             threshold++;

  375.         }

  376.         dataptr += 8; //skip each second start pos

  377.         wsptr   += 8;

  378.     }

  379. }

  380. 

  381. static void row_idct_c(int16_t *workspace, int16_t *output_adr, ptrdiff_t output_stride, int cnt)

  382. {

  383.     int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;

  384.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

  385.     int_simd16_t z5, z10, z11, z12, z13;

  386.     int16_t *outptr;

  387.     int16_t *wsptr;

  388. 

  389.     cnt *= 4;

  390.     wsptr = workspace;

  391.     outptr = output_adr;

  392.     for (; cnt > 0; cnt--) {

  393.         // Even part

  394.         //Simd version reads 4x4 block and transposes it

  395.         tmp10 = wsptr[2] +  wsptr[3];

  396.         tmp11 = wsptr[2] -  wsptr[3];

  397. 

  398.         tmp13 = wsptr[0] +  wsptr[1];

  399.         tmp12 = (MULTIPLY16H(wsptr[0] - wsptr[1], FIX_1_414213562_A) << 2) - tmp13;//this shift order to avoid overflow

  400. 

  401.         tmp0 = tmp10 + tmp13; //->temps

  402.         tmp3 = tmp10 - tmp13; //->temps

  403.         tmp1 = tmp11 + tmp12;

  404.         tmp2 = tmp11 - tmp12;

  405. 

  406.         // Odd part

  407.         //Also transpose, with previous:

  408.         // ---- ----      ||||

  409.         // ---- ---- idct ||||

  410.         // ---- ---- ---> ||||

  411.         // ---- ----      ||||

  412.         z13 = wsptr[4] + wsptr[5];

  413.         z10 = wsptr[4] - wsptr[5];

  414.         z11 = wsptr[6] + wsptr[7];

  415.         z12 = wsptr[6] - wsptr[7];

  416. 

  417.         tmp7 = z11 + z13;

  418.         tmp11 = MULTIPLY16H(z11 - z13, FIX_1_414213562);

  419. 

  420.         z5 =    MULTIPLY16H(z10 + z12, FIX_1_847759065);

  421.         tmp10 = MULTIPLY16H(z12,       FIX_1_082392200) - z5;

  422.         tmp12 = MULTIPLY16H(z10,       FIX_2_613125930) + z5; // - FIX_

  423. 

  424.         tmp6 = (tmp12 << 3) - tmp7;

  425.         tmp5 = (tmp11 << 3) - tmp6;

  426.         tmp4 = (tmp10 << 3) + tmp5;

  427. 

  428.         // Final output stage: descale and write column

  429.         outptr[0 * output_stride] += DESCALE(tmp0 + tmp7, 3);

  430.         outptr[1 * output_stride] += DESCALE(tmp1 + tmp6, 3);

  431.         outptr[2 * output_stride] += DESCALE(tmp2 + tmp5, 3);

  432.         outptr[3 * output_stride] += DESCALE(tmp3 - tmp4, 3);

  433.         outptr[4 * output_stride] += DESCALE(tmp3 + tmp4, 3);

  434.         outptr[5 * output_stride] += DESCALE(tmp2 - tmp5, 3);

  435.         outptr[6 * output_stride] += DESCALE(tmp1 - tmp6, 3); //no += ?

  436.         outptr[7 * output_stride] += DESCALE(tmp0 - tmp7, 3); //no += ?

  437.         outptr++;

  438. 

  439.         wsptr += DCTSIZE;       // advance pointer to next row

  440.     }

  441. }

  442. 

  443. static void row_fdct_c(int16_t *data, const uint8_t *pixels, ptrdiff_t line_size, int cnt)

  444. {

  445.     int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;

  446.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

  447.     int_simd16_t z1, z2, z3, z4, z5, z11, z13;

  448.     int16_t *dataptr;

  449. 

  450.     cnt *= 4;

  451.     // Pass 1: process rows.

  452. 

  453.     dataptr = data;

  454.     for (; cnt > 0; cnt--) {

  455.         tmp0 = pixels[line_size * 0] + pixels[line_size * 7];

  456.         tmp7 = pixels[line_size * 0] - pixels[line_size * 7];

  457.         tmp1 = pixels[line_size * 1] + pixels[line_size * 6];

  458.         tmp6 = pixels[line_size * 1] - pixels[line_size * 6];

  459.         tmp2 = pixels[line_size * 2] + pixels[line_size * 5];

  460.         tmp5 = pixels[line_size * 2] - pixels[line_size * 5];

  461.         tmp3 = pixels[line_size * 3] + pixels[line_size * 4];

  462.         tmp4 = pixels[line_size * 3] - pixels[line_size * 4];

  463. 

  464.         // Even part

  465. 

  466.         tmp10 = tmp0 + tmp3;

  467.         tmp13 = tmp0 - tmp3;

  468.         tmp11 = tmp1 + tmp2;

  469.         tmp12 = tmp1 - tmp2;

  470.         //Even columns are written first, this leads to different order of columns

  471.         //in column_fidct(), but they are processed independently, so all ok.

  472.         //Later in the row_idct() columns readed at the same order.

  473.         dataptr[2] = tmp10 + tmp11;

  474.         dataptr[3] = tmp10 - tmp11;

  475. 

  476.         z1 = MULTIPLY16H((tmp12 + tmp13) << 2, FIX_0_707106781);

  477.         dataptr[0] = tmp13 + z1;

  478.         dataptr[1] = tmp13 - z1;

  479. 

  480.         // Odd part

  481. 

  482.         tmp10 = (tmp4 + tmp5) << 2;

  483.         tmp11 = (tmp5 + tmp6) << 2;

  484.         tmp12 = (tmp6 + tmp7) << 2;

  485. 

  486.         z5 = MULTIPLY16H(tmp10 - tmp12, FIX_0_382683433);

  487.         z2 = MULTIPLY16H(tmp10,         FIX_0_541196100) + z5;

  488.         z4 = MULTIPLY16H(tmp12,         FIX_1_306562965) + z5;

  489.         z3 = MULTIPLY16H(tmp11,         FIX_0_707106781);

  490. 

  491.         z11 = tmp7 + z3;

  492.         z13 = tmp7 - z3;

  493. 

  494.         dataptr[4] = z13 + z2;

  495.         dataptr[5] = z13 - z2;

  496.         dataptr[6] = z11 + z4;

  497.         dataptr[7] = z11 - z4;

  498. 

  499.         pixels++;               // advance pointer to next column

  500.         dataptr += DCTSIZE;

  501.     }

  502. }

  503. 

  504. static int query_formats(AVFilterContext *ctx)

  505. {

  506.     static const enum PixelFormat pix_fmts[] = {

  507.         AV_PIX_FMT_YUV444P,  AV_PIX_FMT_YUV422P,

  508.         AV_PIX_FMT_YUV420P,  AV_PIX_FMT_YUV411P,

  509.         AV_PIX_FMT_YUV410P,  AV_PIX_FMT_YUV440P,

  510.         AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ422P,

  511.         AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ440P,

  512.         AV_PIX_FMT_GBRP, AV_PIX_FMT_GRAY8,

  513.         AV_PIX_FMT_NONE

  514.     };

  515.     ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));

  516.     return 0;

  517. }

  518. 

  519. static int config_input(AVFilterLink *inlink)

  520. {

  521.     AVFilterContext *ctx = inlink->dst;

  522.     FSPPContext *fspp = ctx->priv;

  523.     const int h = FFALIGN(inlink->h + 16, 16);

  524.     const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(inlink->format);

  525. 

  526.     fspp->hsub = desc->log2_chroma_w;

  527.     fspp->vsub = desc->log2_chroma_h;

  528. 

  529.     fspp->temp_stride = FFALIGN(inlink->w + 16, 16);

  530.     fspp->temp = av_malloc_array(fspp->temp_stride, h * sizeof(*fspp->temp));

  531.     fspp->src  = av_malloc_array(fspp->temp_stride, h * sizeof(*fspp->src));

  532. 

  533.     if (!fspp->temp || !fspp->src)

  534.         return AVERROR(ENOMEM);

  535. 

  536.     if (!fspp->use_bframe_qp && !fspp->qp) {

  537.         fspp->non_b_qp_alloc_size = FF_CEIL_RSHIFT(inlink->w, 4) * FF_CEIL_RSHIFT(inlink->h, 4);

  538.         fspp->non_b_qp_table = av_calloc(fspp->non_b_qp_alloc_size, sizeof(*fspp->non_b_qp_table));

  539.         if (!fspp->non_b_qp_table)

  540.             return AVERROR(ENOMEM);

  541.     }

  542. 

  543.     fspp->store_slice  = store_slice_c;

  544.     fspp->store_slice2 = store_slice2_c;

  545.     fspp->mul_thrmat   = mul_thrmat_c;

  546.     fspp->column_fidct = column_fidct_c;

  547.     fspp->row_idct     = row_idct_c;

  548.     fspp->row_fdct     = row_fdct_c;

  549. 

  550.     if (ARCH_X86)

  551.         ff_fspp_init_x86(fspp);

  552. 

  553.     return 0;

  554. }

  555. 

  556. static int filter_frame(AVFilterLink *inlink, AVFrame *in)

  557. {

  558.     AVFilterContext *ctx = inlink->dst;

  559.     FSPPContext *fspp = ctx->priv;

  560.     AVFilterLink *outlink = ctx->outputs[0];

  561.     AVFrame *out = in;

  562. 

  563.     int qp_stride = 0;

  564.     uint8_t *qp_table = NULL;

  565.     int i, bias;

  566.     int custom_threshold_m[64];

  567. 

  568.     bias = (1 << 4) + fspp->strength;

  569. 

  570.     for (i = 0; i < 64; i++) //FIXME: tune custom_threshold[] and remove this !

  571.         custom_threshold_m[i] = (int)(custom_threshold[i] * (bias / 71.0) + 0.5);

  572. 

  573.     for (i = 0; i < 8; i++) {

  574.         fspp->threshold_mtx_noq[2 * i] = (uint64_t)custom_threshold_m[i * 8 + 2]

  575.                                       |(((uint64_t)custom_threshold_m[i * 8 + 6]) << 16)

  576.                                       |(((uint64_t)custom_threshold_m[i * 8 + 0]) << 32)

  577.                                       |(((uint64_t)custom_threshold_m[i * 8 + 4]) << 48);

  578. 

  579.         fspp->threshold_mtx_noq[2 * i + 1] = (uint64_t)custom_threshold_m[i * 8 + 5]

  580.                                           |(((uint64_t)custom_threshold_m[i * 8 + 3]) << 16)

  581.                                           |(((uint64_t)custom_threshold_m[i * 8 + 1]) << 32)

  582.                                           |(((uint64_t)custom_threshold_m[i * 8 + 7]) << 48);

  583.     }

  584. 

  585.     if (fspp->qp)

  586.         fspp->prev_q = fspp->qp, fspp->mul_thrmat((int16_t *)(&fspp->threshold_mtx_noq[0]), (int16_t *)(&fspp->threshold_mtx[0]), fspp->qp);

  587. 

  588.     /* if we are not in a constant user quantizer mode and we don't want to use

  589.      * the quantizers from the B-frames (B-frames often have a higher QP), we

  590.      * need to save the qp table from the last non B-frame; this is what the

  591.      * following code block does */

  592.     if (!fspp->qp) {

  593.         qp_table = av_frame_get_qp_table(in, &qp_stride, &fspp->qscale_type);

  594. 

  595.         if (qp_table && !fspp->use_bframe_qp && in->pict_type != AV_PICTURE_TYPE_B) {

  596.             int w, h;

  597. 

  598.             /* if the qp stride is not set, it means the QP are only defined on

  599.              * a line basis */

  600.            if (!qp_stride) {

  601.                 w = FF_CEIL_RSHIFT(inlink->w, 4);

  602.                 h = 1;

  603.             } else {

  604.                 w = qp_stride;

  605.                 h = FF_CEIL_RSHIFT(inlink->h, 4);

  606.             }

  607.             if (w * h > fspp->non_b_qp_alloc_size) {

  608.                 int ret = av_reallocp_array(&fspp->non_b_qp_table, w, h);

  609.                 if (ret < 0) {

  610.                     fspp->non_b_qp_alloc_size = 0;

  611.                     return ret;

  612.                 }

  613.                 fspp->non_b_qp_alloc_size = w * h;

  614.             }

  615. 

  616.             av_assert0(w * h <= fspp->non_b_qp_alloc_size);

  617.             memcpy(fspp->non_b_qp_table, qp_table, w * h);

  618.         }

  619.     }

  620. 

  621.     if (fspp->log2_count && !ctx->is_disabled) {

  622.         if (!fspp->use_bframe_qp && fspp->non_b_qp_table)

  623.             qp_table = fspp->non_b_qp_table;

  624. 

  625.         if (qp_table || fspp->qp) {

  626.             const int cw = FF_CEIL_RSHIFT(inlink->w, fspp->hsub);

  627.             const int ch = FF_CEIL_RSHIFT(inlink->h, fspp->vsub);

  628. 

  629.             /* get a new frame if in-place is not possible or if the dimensions

  630.              * are not multiple of 8 */

  631.             if (!av_frame_is_writable(in) || (inlink->w & 7) || (inlink->h & 7)) {

  632.                 const int aligned_w = FFALIGN(inlink->w, 8);

  633.                 const int aligned_h = FFALIGN(inlink->h, 8);

  634. 

  635.                 out = ff_get_video_buffer(outlink, aligned_w, aligned_h);

  636.                 if (!out) {

  637.                     av_frame_free(&in);

  638.                     return AVERROR(ENOMEM);

  639.                 }

  640.                 av_frame_copy_props(out, in);

  641.             }

  642. 

  643.             filter(fspp, out->data[0], in->data[0], out->linesize[0], in->linesize[0],

  644.                    inlink->w, inlink->h, qp_table, qp_stride, 1);

  645.             filter(fspp, out->data[1], in->data[1], out->linesize[1], in->linesize[1],

  646.                    cw,        ch,        qp_table, qp_stride, 0);

  647.             filter(fspp, out->data[2], in->data[2], out->linesize[2], in->linesize[2],

  648.                    cw,        ch,        qp_table, qp_stride, 0);

  649.             emms_c();

  650.         }

  651.     }

  652. 

  653.     if (in != out) {

  654.         if (in->data[3])

  655.             av_image_copy_plane(out->data[3], out->linesize[3],

  656.                                 in ->data[3], in ->linesize[3],

  657.                                 inlink->w, inlink->h);

  658.         av_frame_free(&in);

  659.     }

  660.     return ff_filter_frame(outlink, out);

  661. }

  662. 

  663. static av_cold void uninit(AVFilterContext *ctx)

  664. {

  665.     FSPPContext *fspp = ctx->priv;

  666.     av_freep(&fspp->temp);

  667.     av_freep(&fspp->src);

  668.     av_freep(&fspp->non_b_qp_table);

  669. }

  670. 

  671. static const AVFilterPad fspp_inputs[] = {

  672.     {

  673.         .name         = "default",

  674.         .type         = AVMEDIA_TYPE_VIDEO,

  675.         .config_props = config_input,

  676.         .filter_frame = filter_frame,

  677.     },

  678.     { NULL }

  679. };

  680. 

  681. static const AVFilterPad fspp_outputs[] = {

  682.     {

  683.         .name = "default",

  684.         .type = AVMEDIA_TYPE_VIDEO,

  685.     },

  686.     { NULL }

  687. };

  688. 

  689. AVFilter ff_vf_fspp = {

  690.     .name            = "fspp",

  691.     .description     = NULL_IF_CONFIG_SMALL("Apply Fast Simple Post-processing filter."),

  692.     .priv_size       = sizeof(FSPPContext),

  693.     .uninit          = uninit,

  694.     .query_formats   = query_formats,

  695.     .inputs          = fspp_inputs,

  696.     .outputs         = fspp_outputs,

  697.     .priv_class      = &fspp_class,

  698.     .flags           = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,

  699. };