git.sesse.net Git - ffmpeg/blob - libavcodec/arm/hevcdsp_init_neon.c

   1 /*
   2  * Copyright (c) 2014 Seppo Tomperi <seppo.tomperi@vtt.fi>
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/attributes.h"
  22 #include "libavutil/arm/cpu.h"
  23 #include "libavcodec/hevcdsp.h"
  24 #include "hevcdsp_arm.h"
  25
  26 void ff_hevc_v_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
  27 void ff_hevc_h_loop_filter_luma_neon(uint8_t *_pix, ptrdiff_t _stride, int _beta, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
  28 void ff_hevc_v_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
  29 void ff_hevc_h_loop_filter_chroma_neon(uint8_t *_pix, ptrdiff_t _stride, int *_tc, uint8_t *_no_p, uint8_t *_no_q);
  30 void ff_hevc_add_residual_4x4_8_neon(uint8_t *_dst, int16_t *coeffs,
  31                                      ptrdiff_t stride);
  32 void ff_hevc_add_residual_4x4_10_neon(uint8_t *_dst, int16_t *coeffs,
  33                                       ptrdiff_t stride);
  34 void ff_hevc_add_residual_8x8_8_neon(uint8_t *_dst, int16_t *coeffs,
  35                                      ptrdiff_t stride);
  36 void ff_hevc_add_residual_8x8_10_neon(uint8_t *_dst, int16_t *coeffs,
  37                                       ptrdiff_t stride);
  38 void ff_hevc_add_residual_16x16_8_neon(uint8_t *_dst, int16_t *coeffs,
  39                                        ptrdiff_t stride);
  40 void ff_hevc_add_residual_16x16_10_neon(uint8_t *_dst, int16_t *coeffs,
  41                                         ptrdiff_t stride);
  42 void ff_hevc_add_residual_32x32_8_neon(uint8_t *_dst, int16_t *coeffs,
  43                                        ptrdiff_t stride);
  44 void ff_hevc_add_residual_32x32_10_neon(uint8_t *_dst, int16_t *coeffs,
  45                                         ptrdiff_t stride);
  46 void ff_hevc_idct_4x4_dc_8_neon(int16_t *coeffs);
  47 void ff_hevc_idct_8x8_dc_8_neon(int16_t *coeffs);
  48 void ff_hevc_idct_16x16_dc_8_neon(int16_t *coeffs);
  49 void ff_hevc_idct_32x32_dc_8_neon(int16_t *coeffs);
  50 void ff_hevc_idct_4x4_dc_10_neon(int16_t *coeffs);
  51 void ff_hevc_idct_8x8_dc_10_neon(int16_t *coeffs);
  52 void ff_hevc_idct_16x16_dc_10_neon(int16_t *coeffs);
  53 void ff_hevc_idct_32x32_dc_10_neon(int16_t *coeffs);
  54 void ff_hevc_idct_4x4_8_neon(int16_t *coeffs, int col_limit);
  55 void ff_hevc_idct_8x8_8_neon(int16_t *coeffs, int col_limit);
  56 void ff_hevc_idct_16x16_8_neon(int16_t *coeffs, int col_limit);
  57 void ff_hevc_idct_32x32_8_neon(int16_t *coeffs, int col_limit);
  58 void ff_hevc_idct_4x4_10_neon(int16_t *coeffs, int col_limit);
  59 void ff_hevc_idct_8x8_10_neon(int16_t *coeffs, int col_limit);
  60 void ff_hevc_idct_16x16_10_neon(int16_t *coeffs, int col_limit);
  61 void ff_hevc_idct_32x32_10_neon(int16_t *coeffs, int col_limit);
  62 void ff_hevc_transform_luma_4x4_neon_8(int16_t *coeffs);
  63
  64 #define PUT_PIXELS(name) \
  65     void name(int16_t *dst, uint8_t *src, \
  66                                 ptrdiff_t srcstride, int height, \
  67                                 intptr_t mx, intptr_t my, int width)
  68 PUT_PIXELS(ff_hevc_put_pixels_w2_neon_8);
  69 PUT_PIXELS(ff_hevc_put_pixels_w4_neon_8);
  70 PUT_PIXELS(ff_hevc_put_pixels_w6_neon_8);
  71 PUT_PIXELS(ff_hevc_put_pixels_w8_neon_8);
  72 PUT_PIXELS(ff_hevc_put_pixels_w12_neon_8);
  73 PUT_PIXELS(ff_hevc_put_pixels_w16_neon_8);
  74 PUT_PIXELS(ff_hevc_put_pixels_w24_neon_8);
  75 PUT_PIXELS(ff_hevc_put_pixels_w32_neon_8);
  76 PUT_PIXELS(ff_hevc_put_pixels_w48_neon_8);
  77 PUT_PIXELS(ff_hevc_put_pixels_w64_neon_8);
  78 #undef PUT_PIXELS
  79
  80 static void (*put_hevc_qpel_neon[4][4])(int16_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride,
  81                                    int height, int width);
  82 static void (*put_hevc_qpel_uw_neon[4][4])(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride,
  83                                    int width, int height, int16_t* src2, ptrdiff_t src2stride);
  84 void ff_hevc_put_qpel_neon_wrapper(int16_t *dst, uint8_t *src, ptrdiff_t srcstride,
  85                                    int height, intptr_t mx, intptr_t my, int width);
  86 void ff_hevc_put_qpel_uni_neon_wrapper(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride,
  87                                    int height, intptr_t mx, intptr_t my, int width);
  88 void ff_hevc_put_qpel_bi_neon_wrapper(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride,
  89                                        int16_t *src2,
  90                                        int height, intptr_t mx, intptr_t my, int width);
  91 #define QPEL_FUNC(name) \
  92     void name(int16_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride, \
  93                                    int height, int width)
  94
  95 QPEL_FUNC(ff_hevc_put_qpel_v1_neon_8);
  96 QPEL_FUNC(ff_hevc_put_qpel_v2_neon_8);
  97 QPEL_FUNC(ff_hevc_put_qpel_v3_neon_8);
  98 QPEL_FUNC(ff_hevc_put_qpel_h1_neon_8);
  99 QPEL_FUNC(ff_hevc_put_qpel_h2_neon_8);
 100 QPEL_FUNC(ff_hevc_put_qpel_h3_neon_8);
 101 QPEL_FUNC(ff_hevc_put_qpel_h1v1_neon_8);
 102 QPEL_FUNC(ff_hevc_put_qpel_h1v2_neon_8);
 103 QPEL_FUNC(ff_hevc_put_qpel_h1v3_neon_8);
 104 QPEL_FUNC(ff_hevc_put_qpel_h2v1_neon_8);
 105 QPEL_FUNC(ff_hevc_put_qpel_h2v2_neon_8);
 106 QPEL_FUNC(ff_hevc_put_qpel_h2v3_neon_8);
 107 QPEL_FUNC(ff_hevc_put_qpel_h3v1_neon_8);
 108 QPEL_FUNC(ff_hevc_put_qpel_h3v2_neon_8);
 109 QPEL_FUNC(ff_hevc_put_qpel_h3v3_neon_8);
 110 #undef QPEL_FUNC
 111
 112 #define QPEL_FUNC_UW_PIX(name) \
 113     void name(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, \
 114                                    int height, intptr_t mx, intptr_t my, int width);
 115 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w4_neon_8);
 116 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w8_neon_8);
 117 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w16_neon_8);
 118 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w24_neon_8);
 119 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w32_neon_8);
 120 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w48_neon_8);
 121 QPEL_FUNC_UW_PIX(ff_hevc_put_qpel_uw_pixels_w64_neon_8);
 122 #undef QPEL_FUNC_UW_PIX
 123
 124 #define QPEL_FUNC_UW(name) \
 125     void name(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, \
 126                                    int width, int height, int16_t* src2, ptrdiff_t src2stride);
 127 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_pixels_neon_8);
 128 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_v1_neon_8);
 129 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_v2_neon_8);
 130 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_v3_neon_8);
 131 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h1_neon_8);
 132 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h2_neon_8);
 133 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3_neon_8);
 134 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h1v1_neon_8);
 135 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h1v2_neon_8);
 136 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h1v3_neon_8);
 137 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h2v1_neon_8);
 138 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h2v2_neon_8);
 139 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h2v3_neon_8);
 140 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v1_neon_8);
 141 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v2_neon_8);
 142 QPEL_FUNC_UW(ff_hevc_put_qpel_uw_h3v3_neon_8);
 143 #undef QPEL_FUNC_UW
 144
 145 void ff_hevc_put_qpel_neon_wrapper(int16_t *dst, uint8_t *src, ptrdiff_t srcstride,
 146                                    int height, intptr_t mx, intptr_t my, int width) {
 147
 148     put_hevc_qpel_neon[my][mx](dst, MAX_PB_SIZE, src, srcstride, height, width);
 149 }
 150
 151 void ff_hevc_put_qpel_uni_neon_wrapper(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride,
 152                                    int height, intptr_t mx, intptr_t my, int width) {
 153
 154     put_hevc_qpel_uw_neon[my][mx](dst, dststride, src, srcstride, width, height, NULL, 0);
 155 }
 156
 157 void ff_hevc_put_qpel_bi_neon_wrapper(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride,
 158                                        int16_t *src2,
 159                                        int height, intptr_t mx, intptr_t my, int width) {
 160     put_hevc_qpel_uw_neon[my][mx](dst, dststride, src, srcstride, width, height, src2, MAX_PB_SIZE);
 161 }
 162
 163 av_cold void ff_hevc_dsp_init_neon(HEVCDSPContext *c, const int bit_depth)
 164 {
 165     if (bit_depth == 8) {
 166         int x;
 167         c->hevc_v_loop_filter_luma     = ff_hevc_v_loop_filter_luma_neon;
 168         c->hevc_h_loop_filter_luma     = ff_hevc_h_loop_filter_luma_neon;
 169         c->hevc_v_loop_filter_chroma   = ff_hevc_v_loop_filter_chroma_neon;
 170         c->hevc_h_loop_filter_chroma   = ff_hevc_h_loop_filter_chroma_neon;
 171         c->add_residual[0]             = ff_hevc_add_residual_4x4_8_neon;
 172         c->add_residual[1]             = ff_hevc_add_residual_8x8_8_neon;
 173         c->add_residual[2]             = ff_hevc_add_residual_16x16_8_neon;
 174         c->add_residual[3]             = ff_hevc_add_residual_32x32_8_neon;
 175         c->idct_dc[0]                  = ff_hevc_idct_4x4_dc_8_neon;
 176         c->idct_dc[1]                  = ff_hevc_idct_8x8_dc_8_neon;
 177         c->idct_dc[2]                  = ff_hevc_idct_16x16_dc_8_neon;
 178         c->idct_dc[3]                  = ff_hevc_idct_32x32_dc_8_neon;
 179         c->idct[0]                     = ff_hevc_idct_4x4_8_neon;
 180         c->idct[1]                     = ff_hevc_idct_8x8_8_neon;
 181         c->idct[2]                     = ff_hevc_idct_16x16_8_neon;
 182         c->idct[3]                     = ff_hevc_idct_32x32_8_neon;
 183         c->transform_4x4_luma          = ff_hevc_transform_luma_4x4_neon_8;
 184         put_hevc_qpel_neon[1][0]       = ff_hevc_put_qpel_v1_neon_8;
 185         put_hevc_qpel_neon[2][0]       = ff_hevc_put_qpel_v2_neon_8;
 186         put_hevc_qpel_neon[3][0]       = ff_hevc_put_qpel_v3_neon_8;
 187         put_hevc_qpel_neon[0][1]       = ff_hevc_put_qpel_h1_neon_8;
 188         put_hevc_qpel_neon[0][2]       = ff_hevc_put_qpel_h2_neon_8;
 189         put_hevc_qpel_neon[0][3]       = ff_hevc_put_qpel_h3_neon_8;
 190         put_hevc_qpel_neon[1][1]       = ff_hevc_put_qpel_h1v1_neon_8;
 191         put_hevc_qpel_neon[1][2]       = ff_hevc_put_qpel_h2v1_neon_8;
 192         put_hevc_qpel_neon[1][3]       = ff_hevc_put_qpel_h3v1_neon_8;
 193         put_hevc_qpel_neon[2][1]       = ff_hevc_put_qpel_h1v2_neon_8;
 194         put_hevc_qpel_neon[2][2]       = ff_hevc_put_qpel_h2v2_neon_8;
 195         put_hevc_qpel_neon[2][3]       = ff_hevc_put_qpel_h3v2_neon_8;
 196         put_hevc_qpel_neon[3][1]       = ff_hevc_put_qpel_h1v3_neon_8;
 197         put_hevc_qpel_neon[3][2]       = ff_hevc_put_qpel_h2v3_neon_8;
 198         put_hevc_qpel_neon[3][3]       = ff_hevc_put_qpel_h3v3_neon_8;
 199         put_hevc_qpel_uw_neon[1][0]      = ff_hevc_put_qpel_uw_v1_neon_8;
 200         put_hevc_qpel_uw_neon[2][0]      = ff_hevc_put_qpel_uw_v2_neon_8;
 201         put_hevc_qpel_uw_neon[3][0]      = ff_hevc_put_qpel_uw_v3_neon_8;
 202         put_hevc_qpel_uw_neon[0][1]      = ff_hevc_put_qpel_uw_h1_neon_8;
 203         put_hevc_qpel_uw_neon[0][2]      = ff_hevc_put_qpel_uw_h2_neon_8;
 204         put_hevc_qpel_uw_neon[0][3]      = ff_hevc_put_qpel_uw_h3_neon_8;
 205         put_hevc_qpel_uw_neon[1][1]      = ff_hevc_put_qpel_uw_h1v1_neon_8;
 206         put_hevc_qpel_uw_neon[1][2]      = ff_hevc_put_qpel_uw_h2v1_neon_8;
 207         put_hevc_qpel_uw_neon[1][3]      = ff_hevc_put_qpel_uw_h3v1_neon_8;
 208         put_hevc_qpel_uw_neon[2][1]      = ff_hevc_put_qpel_uw_h1v2_neon_8;
 209         put_hevc_qpel_uw_neon[2][2]      = ff_hevc_put_qpel_uw_h2v2_neon_8;
 210         put_hevc_qpel_uw_neon[2][3]      = ff_hevc_put_qpel_uw_h3v2_neon_8;
 211         put_hevc_qpel_uw_neon[3][1]      = ff_hevc_put_qpel_uw_h1v3_neon_8;
 212         put_hevc_qpel_uw_neon[3][2]      = ff_hevc_put_qpel_uw_h2v3_neon_8;
 213         put_hevc_qpel_uw_neon[3][3]      = ff_hevc_put_qpel_uw_h3v3_neon_8;
 214         for (x = 0; x < 10; x++) {
 215             c->put_hevc_qpel[x][1][0]         = ff_hevc_put_qpel_neon_wrapper;
 216             c->put_hevc_qpel[x][0][1]         = ff_hevc_put_qpel_neon_wrapper;
 217             c->put_hevc_qpel[x][1][1]         = ff_hevc_put_qpel_neon_wrapper;
 218             c->put_hevc_qpel_uni[x][1][0]     = ff_hevc_put_qpel_uni_neon_wrapper;
 219             c->put_hevc_qpel_uni[x][0][1]     = ff_hevc_put_qpel_uni_neon_wrapper;
 220             c->put_hevc_qpel_uni[x][1][1]     = ff_hevc_put_qpel_uni_neon_wrapper;
 221             c->put_hevc_qpel_bi[x][1][0]      = ff_hevc_put_qpel_bi_neon_wrapper;
 222             c->put_hevc_qpel_bi[x][0][1]      = ff_hevc_put_qpel_bi_neon_wrapper;
 223             c->put_hevc_qpel_bi[x][1][1]      = ff_hevc_put_qpel_bi_neon_wrapper;
 224         }
 225         c->put_hevc_qpel[0][0][0]  = ff_hevc_put_pixels_w2_neon_8;
 226         c->put_hevc_qpel[1][0][0]  = ff_hevc_put_pixels_w4_neon_8;
 227         c->put_hevc_qpel[2][0][0]  = ff_hevc_put_pixels_w6_neon_8;
 228         c->put_hevc_qpel[3][0][0]  = ff_hevc_put_pixels_w8_neon_8;
 229         c->put_hevc_qpel[4][0][0]  = ff_hevc_put_pixels_w12_neon_8;
 230         c->put_hevc_qpel[5][0][0]  = ff_hevc_put_pixels_w16_neon_8;
 231         c->put_hevc_qpel[6][0][0]  = ff_hevc_put_pixels_w24_neon_8;
 232         c->put_hevc_qpel[7][0][0]  = ff_hevc_put_pixels_w32_neon_8;
 233         c->put_hevc_qpel[8][0][0]  = ff_hevc_put_pixels_w48_neon_8;
 234         c->put_hevc_qpel[9][0][0]  = ff_hevc_put_pixels_w64_neon_8;
 235
 236         c->put_hevc_qpel_uni[1][0][0]  = ff_hevc_put_qpel_uw_pixels_w4_neon_8;
 237         c->put_hevc_qpel_uni[3][0][0]  = ff_hevc_put_qpel_uw_pixels_w8_neon_8;
 238         c->put_hevc_qpel_uni[5][0][0]  = ff_hevc_put_qpel_uw_pixels_w16_neon_8;
 239         c->put_hevc_qpel_uni[6][0][0]  = ff_hevc_put_qpel_uw_pixels_w24_neon_8;
 240         c->put_hevc_qpel_uni[7][0][0]  = ff_hevc_put_qpel_uw_pixels_w32_neon_8;
 241         c->put_hevc_qpel_uni[8][0][0]  = ff_hevc_put_qpel_uw_pixels_w48_neon_8;
 242         c->put_hevc_qpel_uni[9][0][0]  = ff_hevc_put_qpel_uw_pixels_w64_neon_8;
 243     }
 244
 245     if (bit_depth == 10) {
 246         c->add_residual[0] = ff_hevc_add_residual_4x4_10_neon;
 247         c->add_residual[1] = ff_hevc_add_residual_8x8_10_neon;
 248         c->add_residual[2] = ff_hevc_add_residual_16x16_10_neon;
 249         c->add_residual[3] = ff_hevc_add_residual_32x32_10_neon;
 250
 251         c->idct_dc[0] = ff_hevc_idct_4x4_dc_10_neon;
 252         c->idct_dc[1] = ff_hevc_idct_8x8_dc_10_neon;
 253         c->idct_dc[2] = ff_hevc_idct_16x16_dc_10_neon;
 254         c->idct_dc[3] = ff_hevc_idct_32x32_dc_10_neon;
 255
 256         c->idct[0] = ff_hevc_idct_4x4_10_neon;
 257         c->idct[1] = ff_hevc_idct_8x8_10_neon;
 258         c->idct[2] = ff_hevc_idct_16x16_10_neon;
 259         c->idct[3] = ff_hevc_idct_32x32_10_neon;
 260     }
 261 }