2 * Copyright (c) 2016 Martin Storsjo
4 * This file is part of FFmpeg.
6 * FFmpeg 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.
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
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License along
17 * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
23 #include "libavcodec/avcodec.h"
24 #include "libavcodec/h264dsp.h"
25 #include "libavcodec/h264data.h"
26 #include "libavutil/common.h"
27 #include "libavutil/internal.h"
28 #include "libavutil/intreadwrite.h"
30 static const uint32_t pixel_mask[3] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff };
32 #define SIZEOF_PIXEL ((bit_depth + 7) / 8)
33 #define SIZEOF_COEF (2 * ((bit_depth + 7) / 8))
34 #define PIXEL_STRIDE 16
36 #define randomize_buffers() \
39 uint32_t mask = pixel_mask[bit_depth - 8]; \
40 for (y = 0; y < sz; y++) { \
41 for (x = 0; x < PIXEL_STRIDE; x += 4) { \
42 AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask); \
43 AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask); \
45 for (x = 0; x < sz; x++) { \
46 if (bit_depth == 8) { \
47 coef[y * sz + x] = src[y * PIXEL_STRIDE + x] - \
48 dst[y * PIXEL_STRIDE + x]; \
50 ((int32_t *)coef)[y * sz + x] = \
51 ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] - \
52 ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x]; \
58 #define dct4x4_impl(size, dctcoef) \
59 static void dct4x4_##size(dctcoef *coef) \
63 for (i = 0; i < 4; i++) { \
64 const int z0 = coef[i*4 + 0] + coef[i*4 + 3]; \
65 const int z1 = coef[i*4 + 1] + coef[i*4 + 2]; \
66 const int z2 = coef[i*4 + 0] - coef[i*4 + 3]; \
67 const int z3 = coef[i*4 + 1] - coef[i*4 + 2]; \
68 tmp[i + 4*0] = z0 + z1; \
69 tmp[i + 4*1] = 2*z2 + z3; \
70 tmp[i + 4*2] = z0 - z1; \
71 tmp[i + 4*3] = z2 - 2*z3; \
73 for (i = 0; i < 4; i++) { \
74 const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3]; \
75 const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2]; \
76 const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3]; \
77 const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2]; \
78 coef[i*4 + 0] = z0 + z1; \
79 coef[i*4 + 1] = 2*z2 + z3; \
80 coef[i*4 + 2] = z0 - z1; \
81 coef[i*4 + 3] = z2 - 2*z3; \
83 for (y = 0; y < 4; y++) { \
84 for (x = 0; x < 4; x++) { \
85 static const int scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
86 const int idx = (y & 1) + (x & 1); \
87 coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15; \
92 #define DCT8_1D(src, srcstride, dst, dststride) do { \
93 const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7]; \
94 const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7]; \
95 const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6]; \
96 const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6]; \
97 const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5]; \
98 const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5]; \
99 const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4]; \
100 const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4]; \
101 const int b0 = a0 + a6; \
102 const int b1 = a2 + a4; \
103 const int b2 = a0 - a6; \
104 const int b3 = a2 - a4; \
105 const int b4 = a3 + a5 + (a1 + (a1 >> 1)); \
106 const int b5 = a1 - a7 - (a5 + (a5 >> 1)); \
107 const int b6 = a1 + a7 - (a3 + (a3 >> 1)); \
108 const int b7 = a3 - a5 + (a7 + (a7 >> 1)); \
109 (dst)[dststride * 0] = b0 + b1; \
110 (dst)[dststride * 1] = b4 + (b7 >> 2); \
111 (dst)[dststride * 2] = b2 + (b3 >> 1); \
112 (dst)[dststride * 3] = b5 + (b6 >> 2); \
113 (dst)[dststride * 4] = b0 - b1; \
114 (dst)[dststride * 5] = b6 - (b5 >> 2); \
115 (dst)[dststride * 6] = (b2 >> 1) - b3; \
116 (dst)[dststride * 7] = (b4 >> 2) - b7; \
119 #define dct8x8_impl(size, dctcoef) \
120 static void dct8x8_##size(dctcoef *coef) \
124 for (i = 0; i < 8; i++) \
125 DCT8_1D(coef + i, 8, tmp + i, 8); \
127 for (i = 0; i < 8; i++) \
128 DCT8_1D(tmp + 8*i, 1, coef + i, 8); \
130 for (y = 0; y < 8; y++) { \
131 for (x = 0; x < 8; x++) { \
132 static const int scale[] = { \
133 13107 * 20, 11428 * 18, 20972 * 32, \
134 12222 * 19, 16777 * 25, 15481 * 24, \
136 static const int idxmap[] = { \
142 const int idx = idxmap[(y & 3) * 4 + (x & 3)]; \
143 coef[y*8 + x] = ((int64_t)coef[y*8 + x] * \
144 scale[idx] + (1 << 17)) >> 18; \
149 dct4x4_impl(16, int16_t)
150 dct4x4_impl(32, int32_t)
152 dct8x8_impl(16, int16_t)
153 dct8x8_impl(32, int32_t)
155 static void dct4x4(int16_t *coef, int bit_depth)
160 dct4x4_32((int32_t *) coef);
163 static void dct8x8(int16_t *coef, int bit_depth)
165 if (bit_depth == 8) {
168 dct8x8_32((int32_t *) coef);
173 static void check_idct(void)
175 LOCAL_ALIGNED_16(uint8_t, src, [8 * 8 * 2]);
176 LOCAL_ALIGNED_16(uint8_t, dst, [8 * 8 * 2]);
177 LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]);
178 LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]);
179 LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
180 LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
181 LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
183 int bit_depth, sz, align, dc;
184 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
186 for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
187 ff_h264dsp_init(&h, bit_depth, 1);
188 for (sz = 4; sz <= 8; sz += 4) {
192 dct4x4(coef, bit_depth);
194 dct8x8(coef, bit_depth);
196 for (dc = 0; dc <= 1; dc++) {
197 void (*idct)(uint8_t *, int16_t *, int) = NULL;
198 switch ((sz << 1) | dc) {
199 case (4 << 1) | 0: idct = h.h264_idct_add; break;
200 case (4 << 1) | 1: idct = h.h264_idct_dc_add; break;
201 case (8 << 1) | 0: idct = h.h264_idct8_add; break;
202 case (8 << 1) | 1: idct = h.h264_idct8_dc_add; break;
204 if (check_func(idct, "h264_idct%d_add%s_%dbpp", sz, dc ? "_dc" : "", bit_depth)) {
205 for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) {
206 uint8_t *dst1 = dst1_base + align;
208 memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
209 memcpy(subcoef0, coef, SIZEOF_COEF);
211 memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
213 memcpy(dst0, dst, sz * PIXEL_STRIDE);
214 memcpy(dst1, dst, sz * PIXEL_STRIDE);
215 memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
216 call_ref(dst0, subcoef0, PIXEL_STRIDE);
217 call_new(dst1, subcoef1, PIXEL_STRIDE);
218 if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
219 memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
221 bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
229 static void check_idct_multiple(void)
231 LOCAL_ALIGNED_16(uint8_t, dst_full, [16 * 16 * 2]);
232 LOCAL_ALIGNED_16(int16_t, coef_full, [16 * 16 * 2]);
233 LOCAL_ALIGNED_16(uint8_t, dst0, [16 * 16 * 2]);
234 LOCAL_ALIGNED_16(uint8_t, dst1, [16 * 16 * 2]);
235 LOCAL_ALIGNED_16(int16_t, coef0, [16 * 16 * 2]);
236 LOCAL_ALIGNED_16(int16_t, coef1, [16 * 16 * 2]);
237 LOCAL_ALIGNED_16(uint8_t, nnzc, [15 * 8]);
239 int bit_depth, i, y, func;
240 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]);
242 for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
243 ff_h264dsp_init(&h, bit_depth, 1);
244 for (func = 0; func < 3; func++) {
245 void (*idct)(uint8_t *, const int *, int16_t *, int, const uint8_t[]) = NULL;
247 int sz = 4, intra = 0;
248 int block_offset[16] = { 0 };
251 idct = h.h264_idct_add16;
252 name = "h264_idct_add16";
255 idct = h.h264_idct_add16intra;
256 name = "h264_idct_add16intra";
260 idct = h.h264_idct8_add4;
261 name = "h264_idct8_add4";
265 memset(nnzc, 0, 15 * 8);
266 memset(coef_full, 0, 16 * 16 * SIZEOF_COEF);
267 for (i = 0; i < 16 * 16; i += sz * sz) {
268 uint8_t src[8 * 8 * 2];
269 uint8_t dst[8 * 8 * 2];
270 int16_t coef[8 * 8 * 2];
272 int block_y = (index / 16) * sz;
273 int block_x = index % 16;
274 int offset = (block_y * 16 + block_x) * SIZEOF_PIXEL;
279 dct4x4(coef, bit_depth);
281 dct8x8(coef, bit_depth);
283 for (y = 0; y < sz; y++)
284 memcpy(&dst_full[offset + y * 16 * SIZEOF_PIXEL],
285 &dst[PIXEL_STRIDE * y], sz * SIZEOF_PIXEL);
289 memcpy(&coef_full[i * SIZEOF_COEF/sizeof(coef[0])],
290 coef, nnz * SIZEOF_COEF);
292 if (intra && nnz == 1)
295 nnzc[scan8[i / 16]] = nnz;
296 block_offset[i / 16] = offset;
299 if (check_func(idct, "%s_%dbpp", name, bit_depth)) {
300 memcpy(coef0, coef_full, 16 * 16 * SIZEOF_COEF);
301 memcpy(coef1, coef_full, 16 * 16 * SIZEOF_COEF);
302 memcpy(dst0, dst_full, 16 * 16 * SIZEOF_PIXEL);
303 memcpy(dst1, dst_full, 16 * 16 * SIZEOF_PIXEL);
304 call_ref(dst0, block_offset, coef0, 16 * SIZEOF_PIXEL, nnzc);
305 call_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
306 if (memcmp(dst0, dst1, 16 * 16 * SIZEOF_PIXEL) ||
307 memcmp(coef0, coef1, 16 * 16 * SIZEOF_COEF))
309 bench_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
315 void checkasm_check_h264dsp(void)
318 check_idct_multiple();