git.sesse.net Git - ffmpeg/blob - tests/checkasm/h264dsp.c

   1 /*
   2  * Copyright (c) 2016 Martin Storsjo
   3  *
   4  * This file is part of Libav.
   5  *
   6  * Libav is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * Libav 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
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along
  17  * with Libav; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 #include <string.h>
  22 #include "checkasm.h"
  23 #include "libavcodec/avcodec.h"
  24 #include "libavcodec/h264dsp.h"
  25 #include "libavutil/common.h"
  26 #include "libavutil/internal.h"
  27 #include "libavutil/intreadwrite.h"
  28
  29 static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
  30
  31 #define SIZEOF_PIXEL ((bit_depth + 7) / 8)
  32 #define SIZEOF_COEF  (2 * ((bit_depth + 7) / 8))
  33 #define PIXEL_STRIDE 16
  34
  35 #define randomize_buffers()                                                  \
  36     do {                                                                     \
  37         int x, y;                                                            \
  38         uint32_t mask = pixel_mask[bit_depth - 8];                           \
  39         for (y = 0; y < sz; y++) {                                           \
  40             for (x = 0; x < PIXEL_STRIDE; x += 4) {                          \
  41                 AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask);          \
  42                 AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask);          \
  43             }                                                                \
  44             for (x = 0; x < sz; x++) {                                       \
  45                 if (bit_depth == 8) {                                        \
  46                     coef[y * sz + x] = src[y * PIXEL_STRIDE + x] -           \
  47                                        dst[y * PIXEL_STRIDE + x];            \
  48                 } else {                                                     \
  49                     ((int32_t *)coef)[y * sz + x] =                          \
  50                         ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] -        \
  51                         ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x];         \
  52                 }                                                            \
  53             }                                                                \
  54         }                                                                    \
  55     } while (0)
  56
  57 #define dct4x4_impl(size, dctcoef)                                           \
  58 static void dct4x4_##size(dctcoef *coef)                                     \
  59 {                                                                            \
  60     int i, y, x;                                                             \
  61     dctcoef tmp[16];                                                         \
  62     for (i = 0; i < 4; i++) {                                                \
  63         const int z0 = coef[i*4 + 0] + coef[i*4 + 3];                        \
  64         const int z1 = coef[i*4 + 1] + coef[i*4 + 2];                        \
  65         const int z2 = coef[i*4 + 0] - coef[i*4 + 3];                        \
  66         const int z3 = coef[i*4 + 1] - coef[i*4 + 2];                        \
  67         tmp[i + 4*0] =   z0 +   z1;                                          \
  68         tmp[i + 4*1] = 2*z2 +   z3;                                          \
  69         tmp[i + 4*2] =   z0 -   z1;                                          \
  70         tmp[i + 4*3] =   z2 - 2*z3;                                          \
  71     }                                                                        \
  72     for (i = 0; i < 4; i++) {                                                \
  73         const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3];                          \
  74         const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2];                          \
  75         const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3];                          \
  76         const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2];                          \
  77         coef[i*4 + 0] =   z0 +   z1;                                         \
  78         coef[i*4 + 1] = 2*z2 +   z3;                                         \
  79         coef[i*4 + 2] =   z0 -   z1;                                         \
  80         coef[i*4 + 3] =   z2 - 2*z3;                                         \
  81     }                                                                        \
  82     for (y = 0; y < 4; y++) {                                                \
  83         for (x = 0; x < 4; x++) {                                            \
  84             static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
  85             const int idx = (y & 1) + (x & 1);                               \
  86             coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15;  \
  87         }                                                                    \
  88     }                                                                        \
  89 }
  90
  91 #define DCT8_1D(src, srcstride, dst, dststride) do {                         \
  92     const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7];              \
  93     const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7];              \
  94     const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6];              \
  95     const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6];              \
  96     const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5];              \
  97     const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5];              \
  98     const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4];              \
  99     const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4];              \
 100     const int b0 = a0 + a6;                                                  \
 101     const int b1 = a2 + a4;                                                  \
 102     const int b2 = a0 - a6;                                                  \
 103     const int b3 = a2 - a4;                                                  \
 104     const int b4 = a3 + a5 + (a1 + (a1 >> 1));                               \
 105     const int b5 = a1 - a7 - (a5 + (a5 >> 1));                               \
 106     const int b6 = a1 + a7 - (a3 + (a3 >> 1));                               \
 107     const int b7 = a3 - a5 + (a7 + (a7 >> 1));                               \
 108     (dst)[dststride * 0] =  b0 +  b1;                                        \
 109     (dst)[dststride * 1] =  b4 + (b7 >> 2);                                  \
 110     (dst)[dststride * 2] =  b2 + (b3 >> 1);                                  \
 111     (dst)[dststride * 3] =  b5 + (b6 >> 2);                                  \
 112     (dst)[dststride * 4] =  b0  - b1;                                        \
 113     (dst)[dststride * 5] =  b6 - (b5 >> 2);                                  \
 114     (dst)[dststride * 6] = (b2 >> 1) - b3;                                   \
 115     (dst)[dststride * 7] = (b4 >> 2) - b7;                                   \
 116 } while (0)
 117
 118 #define dct8x8_impl(size, dctcoef)                                           \
 119 static void dct8x8_##size(dctcoef *coef)                                     \
 120 {                                                                            \
 121     int i, x, y;                                                             \
 122     dctcoef tmp[64];                                                         \
 123     for (i = 0; i < 8; i++)                                                  \
 124         DCT8_1D(coef + i, 8, tmp + i, 8);                                    \
 125                                                                              \
 126     for (i = 0; i < 8; i++)                                                  \
 127         DCT8_1D(tmp + 8*i, 1, coef + i, 8);                                  \
 128                                                                              \
 129     for (y = 0; y < 8; y++) {                                                \
 130         for (x = 0; x < 8; x++) {                                            \
 131             static const int scale[] = {                                     \
 132                 13107 * 20, 11428 * 18, 20972 * 32,                          \
 133                 12222 * 19, 16777 * 25, 15481 * 24,                          \
 134             };                                                               \
 135             static const int idxmap[] = {                                    \
 136                 0, 3, 4, 3,                                                  \
 137                 3, 1, 5, 1,                                                  \
 138                 4, 5, 2, 5,                                                  \
 139                 3, 1, 5, 1,                                                  \
 140             };                                                               \
 141             const int idx = idxmap[(y & 3) * 4 + (x & 3)];                   \
 142             coef[y*8 + x] = ((int64_t)coef[y*8 + x] *                        \
 143                              scale[idx] + (1 << 17)) >> 18;                  \
 144         }                                                                    \
 145     }                                                                        \
 146 }
 147
 148 dct4x4_impl(16, int16_t)
 149 dct4x4_impl(32, int32_t)
 150
 151 dct8x8_impl(16, int16_t)
 152 dct8x8_impl(32, int32_t)
 153
 154 static void dct4x4(int16_t *coef, int bit_depth)
 155 {
 156     if (bit_depth == 8)
 157         dct4x4_16(coef);
 158     else
 159         dct4x4_32((int32_t *) coef);
 160 }
 161
 162 static void dct8x8(int16_t *coef, int bit_depth)
 163 {
 164     if (bit_depth == 8) {
 165         dct8x8_16(coef);
 166     } else {
 167         dct8x8_32((int32_t *) coef);
 168     }
 169 }
 170
 171
 172 static void check_idct(void)
 173 {
 174     LOCAL_ALIGNED_16(uint8_t, src,  [8 * 8 * 2]);
 175     LOCAL_ALIGNED_16(uint8_t, dst,  [8 * 8 * 2]);
 176     LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]);
 177     LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]);
 178     LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
 179     LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
 180     LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
 181     H264DSPContext h;
 182     int bit_depth, sz, align, dc;
 183     declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
 184
 185     for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
 186         ff_h264dsp_init(&h, bit_depth, 1);
 187         for (sz = 4; sz <= 8; sz += 4) {
 188             randomize_buffers();
 189
 190             if (sz == 4)
 191                 dct4x4(coef, bit_depth);
 192             else
 193                 dct8x8(coef, bit_depth);
 194
 195             for (dc = 0; dc <= 1; dc++) {
 196                 void (*idct)(uint8_t *, int16_t *, int) = NULL;
 197                 switch ((sz << 1) | dc) {
 198                 case (4 << 1) | 0: idct = h.h264_idct_add; break;
 199                 case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
 200                 case (8 << 1) | 0: idct = h.h264_idct8_add; break;
 201                 case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
 202                 }
 203                 if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
 204                     for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) {
 205                         uint8_t *dst1 = dst1_base + align;
 206                         if (dc) {
 207                             memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
 208                             memcpy(subcoef0, coef, SIZEOF_COEF);
 209                         } else {
 210                             memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
 211                         }
 212                         memcpy(dst0, dst, sz * PIXEL_STRIDE);
 213                         memcpy(dst1, dst, sz * PIXEL_STRIDE);
 214                         memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
 215                         call_ref(dst0, subcoef0, PIXEL_STRIDE);
 216                         call_new(dst1, subcoef1, PIXEL_STRIDE);
 217                         if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
 218                             memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
 219                             fail();
 220                         bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
 221                     }
 222                 }
 223             }
 224         }
 225     }
 226     report("idct");
 227 }
 228
 229 void checkasm_check_h264dsp(void)
 230 {
 231     check_idct();
 232 }