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.
100686 Commits
32 Branches
824MB
Tree: b51eae1abe
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'b51eae1abe'
${ noResults }
FFmpeg/libavfilter/vf_fspp.c

682 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/mem_internal.h"

  41. #include "libavutil/opt.h"

  42. #include "libavutil/pixdesc.h"

  43. #include "internal.h"

  44. #include "qp_table.h"

  45. #include "vf_fspp.h"

  46. 

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

  48. #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM

  49. static const AVOption fspp_options[] = {

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

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

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

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

  54.     { NULL }

  55. };

  56. 

  57. AVFILTER_DEFINE_CLASS(fspp);

  58. 

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

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

  61.     { 32,  16,  44,  28,  35,  19,  47,  31, },

  62.     {  8,  56,   4,  52,  11,  59,   7,  55, },

  63.     { 40,  24,  36,  20,  43,  27,  39,  23, },

  64.     {  2,  50,  14,  62,   1,  49,  13,  61, },

  65.     { 34,  18,  46,  30,  33,  17,  45,  29, },

  66.     { 10,  58,   6,  54,   9,  57,   5,  53, },

  67.     { 42,  26,  38,  22,  41,  25,  37,  21, },

  68. };

  69. 

  70. static const short custom_threshold[64] = {

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

  72. // -it causes too big quant dependence

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

  74.      71, 296, 295, 237,  71,  40,  38,  19,

  75.     245, 193, 185, 121, 102,  73,  53,  27,

  76.     158, 129, 141, 107,  97,  73,  50,  26,

  77.     102, 116, 109,  98,  82,  66,  45,  23,

  78.      71,  94,  95,  81,  70,  56,  38,  20,

  79.      56,  77,  74,  66,  56,  44,  30,  15,

  80.      38,  53,  50,  45,  38,  30,  21,  11,

  81.      20,  27,  26,  23,  20,  15,  11,   5

  82. };

  83. 

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

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

  86.                           ptrdiff_t dst_stride, ptrdiff_t src_stride,

  87.                           ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)

  88. {

  89.     int y, x;

  90. #define STORE(pos)                                                             \

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

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

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

  94.     dst[x + pos] = temp;

  95. 

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

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

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

  99.             int temp;

  100.             STORE(0);

  101.             STORE(1);

  102.             STORE(2);

  103.             STORE(3);

  104.             STORE(4);

  105.             STORE(5);

  106.             STORE(6);

  107.             STORE(7);

  108.         }

  109.         src += src_stride;

  110.         dst += dst_stride;

  111.     }

  112. }

  113. 

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

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

  116.                            ptrdiff_t dst_stride, ptrdiff_t src_stride,

  117.                            ptrdiff_t width, ptrdiff_t height, ptrdiff_t log2_scale)

  118. {

  119.     int y, x;

  120. #define STORE2(pos)                                                                                       \

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

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

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

  124.     dst[x + pos] = temp;

  125. 

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

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

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

  129.             int temp;

  130.             STORE2(0);

  131.             STORE2(1);

  132.             STORE2(2);

  133.             STORE2(3);

  134.             STORE2(4);

  135.             STORE2(5);

  136.             STORE2(6);

  137.             STORE2(7);

  138.         }

  139.         src += src_stride;

  140.         dst += dst_stride;

  141.     }

  142. }

  143. 

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

  145. {

  146.     int a;

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

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

  149. }

  150. 

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

  152.                    int dst_stride, int src_stride,

  153.                    int width, int height,

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

  155. {

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

  157. 

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

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

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

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

  162. 

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

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

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

  166. 

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

  168. 

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

  170. 

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

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

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

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

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

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

  177.         }

  178.     }

  179. 

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

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

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

  183.     }

  184.     //FIXME (try edge emu)

  185. 

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

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

  188. 

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

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

  191.         qy = y - 4;

  192. 

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

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

  195. 

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

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

  198. 

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

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

  201. 

  202.             if (p->qp)

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

  204.             else

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

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

  207. 

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

  209. 

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

  211.                     t = ff_norm_qscale(t, p->qscale_type);

  212. 

  213.                     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);

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

  215.                 }

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

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

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

  219.         }

  220. 

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

  222.         if (es > 8)

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

  224. 

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

  226.         if (es > 3)

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

  228. 

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

  230.             if (y1 & 8)

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

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

  233.             else

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

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

  236.         }

  237.     }

  238. 

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

  240.         if (y & 8)

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

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

  243.         else

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

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

  246.     }

  247. }

  248. 

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

  250. {

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

  252.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

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

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

  255. 

  256.     int16_t *dataptr;

  257.     int16_t *wsptr;

  258.     int16_t *threshold;

  259.     int ctr;

  260. 

  261.     dataptr = data;

  262.     wsptr = output;

  263. 

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

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

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

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

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

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

  270. 

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

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

  273. 

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

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

  276. 

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

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

  279. 

  280.             // Even part of FDCT

  281. 

  282.             tmp10 = tmp0 + tmp3;

  283.             tmp13 = tmp0 - tmp3;

  284.             tmp11 = tmp1 + tmp2;

  285.             tmp12 = tmp1 - tmp2;

  286. 

  287.             d0 = tmp10 + tmp11;

  288.             d4 = tmp10 - tmp11;

  289. 

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

  291.             d2 = tmp13 + z1;

  292.             d6 = tmp13 - z1;

  293. 

  294.             // Even part of IDCT

  295. 

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

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

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

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

  300.             tmp0 += 2;

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

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

  303. 

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

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

  306. 

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

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

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

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

  311. 

  312.             // Odd part of FDCT

  313. 

  314.             tmp10 = tmp4 + tmp5;

  315.             tmp11 = tmp5 + tmp6;

  316.             tmp12 = tmp6 + tmp7;

  317. 

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

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

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

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

  322. 

  323.             z11 = tmp7 + z3;

  324.             z13 = tmp7 - z3;

  325. 

  326.             d5 = z13 + z2;

  327.             d3 = z13 - z2;

  328.             d1 = z11 + z4;

  329.             d7 = z11 - z4;

  330. 

  331.             // Odd part of IDCT

  332. 

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

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

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

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

  337. 

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

  339.             z13 = tmp6 + tmp5;

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

  341.             z11 = tmp4 + tmp7;

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

  343. 

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

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

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

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

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

  349. 

  350.             tmp6 = tmp12 - tmp7;

  351.             tmp5 = tmp11 - tmp6;

  352.             tmp4 = tmp10 + tmp5;

  353. 

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

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

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

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

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

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

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

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

  362.             //

  363.             dataptr++; //next column

  364.             wsptr++;

  365.             threshold++;

  366.         }

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

  368.         wsptr   += 8;

  369.     }

  370. }

  371. 

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

  373. {

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

  375.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

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

  377.     int16_t *outptr;

  378.     int16_t *wsptr;

  379. 

  380.     cnt *= 4;

  381.     wsptr = workspace;

  382.     outptr = output_adr;

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

  384.         // Even part

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

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

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

  388. 

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

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

  391. 

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

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

  394.         tmp1 = tmp11 + tmp12;

  395.         tmp2 = tmp11 - tmp12;

  396. 

  397.         // Odd part

  398.         //Also transpose, with previous:

  399.         // ---- ----      ||||

  400.         // ---- ---- idct ||||

  401.         // ---- ---- ---> ||||

  402.         // ---- ----      ||||

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

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

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

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

  407. 

  408.         tmp7 = z11 + z13;

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

  410. 

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

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

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

  414. 

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

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

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

  418. 

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

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

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

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

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

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

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

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

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

  428.         outptr++;

  429. 

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

  431.     }

  432. }

  433. 

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

  435. {

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

  437.     int_simd16_t tmp10, tmp11, tmp12, tmp13;

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

  439.     int16_t *dataptr;

  440. 

  441.     cnt *= 4;

  442.     // Pass 1: process rows.

  443. 

  444.     dataptr = data;

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

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

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

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

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

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

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

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

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

  454. 

  455.         // Even part

  456. 

  457.         tmp10 = tmp0 + tmp3;

  458.         tmp13 = tmp0 - tmp3;

  459.         tmp11 = tmp1 + tmp2;

  460.         tmp12 = tmp1 - tmp2;

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

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

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

  464.         dataptr[2] = tmp10 + tmp11;

  465.         dataptr[3] = tmp10 - tmp11;

  466. 

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

  468.         dataptr[0] = tmp13 + z1;

  469.         dataptr[1] = tmp13 - z1;

  470. 

  471.         // Odd part

  472. 

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

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

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

  476. 

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

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

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

  480.         z3 = MULTIPLY16H(tmp11,         FIX_0_707106781);

  481. 

  482.         z11 = tmp7 + z3;

  483.         z13 = tmp7 - z3;

  484. 

  485.         dataptr[4] = z13 + z2;

  486.         dataptr[5] = z13 - z2;

  487.         dataptr[6] = z11 + z4;

  488.         dataptr[7] = z11 - z4;

  489. 

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

  491.         dataptr += DCTSIZE;

  492.     }

  493. }

  494. 

  495. static int query_formats(AVFilterContext *ctx)

  496. {

  497.     static const enum AVPixelFormat pix_fmts[] = {

  498.         AV_PIX_FMT_YUV444P,  AV_PIX_FMT_YUV422P,

  499.         AV_PIX_FMT_YUV420P,  AV_PIX_FMT_YUV411P,

  500.         AV_PIX_FMT_YUV410P,  AV_PIX_FMT_YUV440P,

  501.         AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ422P,

  502.         AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ440P,

  503.         AV_PIX_FMT_GBRP, AV_PIX_FMT_GRAY8,

  504.         AV_PIX_FMT_NONE

  505.     };

  506. 

  507.     AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);

  508.     if (!fmts_list)

  509.         return AVERROR(ENOMEM);

  510.     return ff_set_common_formats(ctx, fmts_list);

  511. }

  512. 

  513. static int config_input(AVFilterLink *inlink)

  514. {

  515.     AVFilterContext *ctx = inlink->dst;

  516.     FSPPContext *fspp = ctx->priv;

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

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

  519. 

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

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

  522. 

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

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

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

  526. 

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

  528.         return AVERROR(ENOMEM);

  529. 

  530.     fspp->store_slice  = store_slice_c;

  531.     fspp->store_slice2 = store_slice2_c;

  532.     fspp->mul_thrmat   = mul_thrmat_c;

  533.     fspp->column_fidct = column_fidct_c;

  534.     fspp->row_idct     = row_idct_c;

  535.     fspp->row_fdct     = row_fdct_c;

  536. 

  537.     if (ARCH_X86)

  538.         ff_fspp_init_x86(fspp);

  539. 

  540.     return 0;

  541. }

  542. 

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

  544. {

  545.     AVFilterContext *ctx = inlink->dst;

  546.     FSPPContext *fspp = ctx->priv;

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

  548.     AVFrame *out = in;

  549. 

  550.     int qp_stride = 0;

  551.     int8_t *qp_table = NULL;

  552.     int i, bias;

  553.     int ret = 0;

  554.     int custom_threshold_m[64];

  555. 

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

  557. 

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

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

  560. 

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

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

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

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

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

  566. 

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

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

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

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

  571.     }

  572. 

  573.     if (fspp->qp)

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

  575. 

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

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

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

  579.      * following code block does */

  580.     if (!fspp->qp && (fspp->use_bframe_qp || in->pict_type != AV_PICTURE_TYPE_B)) {

  581.         ret = ff_qp_table_extract(in, &qp_table, &qp_stride, NULL, &fspp->qscale_type);

  582.         if (ret < 0) {

  583.             av_frame_free(&in);

  584.             return ret;

  585.         }

  586. 

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

  588.             av_freep(&fspp->non_b_qp_table);

  589.             fspp->non_b_qp_table  = qp_table;

  590.             fspp->non_b_qp_stride = qp_stride;

  591.         }

  592.     }

  593. 

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

  595.         if (!fspp->use_bframe_qp && fspp->non_b_qp_table) {

  596.             qp_table = fspp->non_b_qp_table;

  597.             qp_stride = fspp->non_b_qp_stride;

  598.         }

  599. 

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

  601.             const int cw = AV_CEIL_RSHIFT(inlink->w, fspp->hsub);

  602.             const int ch = AV_CEIL_RSHIFT(inlink->h, fspp->vsub);

  603. 

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

  605.              * are not multiple of 8 */

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

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

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

  609. 

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

  611.                 if (!out) {

  612.                     av_frame_free(&in);

  613.                     ret = AVERROR(ENOMEM);

  614.                     goto finish;

  615.                 }

  616.                 av_frame_copy_props(out, in);

  617.                 out->width = in->width;

  618.                 out->height = in->height;

  619.             }

  620. 

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

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

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

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

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

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

  627.             emms_c();

  628.         }

  629.     }

  630. 

  631.     if (in != out) {

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

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

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

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

  636.         av_frame_free(&in);

  637.     }

  638.     ret = ff_filter_frame(outlink, out);

  639. finish:

  640.     if (qp_table != fspp->non_b_qp_table)

  641.         av_freep(&qp_table);

  642.     return ret;

  643. }

  644. 

  645. static av_cold void uninit(AVFilterContext *ctx)

  646. {

  647.     FSPPContext *fspp = ctx->priv;

  648.     av_freep(&fspp->temp);

  649.     av_freep(&fspp->src);

  650.     av_freep(&fspp->non_b_qp_table);

  651. }

  652. 

  653. static const AVFilterPad fspp_inputs[] = {

  654.     {

  655.         .name         = "default",

  656.         .type         = AVMEDIA_TYPE_VIDEO,

  657.         .config_props = config_input,

  658.         .filter_frame = filter_frame,

  659.     },

  660.     { NULL }

  661. };

  662. 

  663. static const AVFilterPad fspp_outputs[] = {

  664.     {

  665.         .name = "default",

  666.         .type = AVMEDIA_TYPE_VIDEO,

  667.     },

  668.     { NULL }

  669. };

  670. 

  671. AVFilter ff_vf_fspp = {

  672.     .name            = "fspp",

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

  674.     .priv_size       = sizeof(FSPPContext),

  675.     .uninit          = uninit,

  676.     .query_formats   = query_formats,

  677.     .inputs          = fspp_inputs,

  678.     .outputs         = fspp_outputs,

  679.     .priv_class      = &fspp_class,

  680.     .flags           = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,

  681. };