FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/tests/checkasm/h264dsp.c
Date: 2024-11-20 23:03:26
Exec Total Coverage
Lines: 178 180 98.9%
Functions: 11 11 100.0%
Branches: 243 391 62.1%
Line Branch Exec Source
1 /*
2 * Copyright (c) 2016 Martin Storsjo
3 *
4 * This file is part of FFmpeg.
5 *
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.
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
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 FFmpeg; 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/h264dsp.h"
24 #include "libavcodec/h264data.h"
25 #include "libavcodec/h264_parse.h"
26 #include "libavutil/common.h"
27 #include "libavutil/intreadwrite.h"
28 #include "libavutil/mem_internal.h"
29
30 static const uint32_t pixel_mask[5] = { 0xffffffff, 0x01ff01ff, 0x03ff03ff, 0x0fff0fff, 0x3fff3fff };
31 static const uint32_t pixel_mask_lf[3] = { 0xff0fff0f, 0x01ff000f, 0x03ff000f };
32
33 #define SIZEOF_PIXEL ((bit_depth + 7) / 8)
34 #define SIZEOF_COEF (2 * ((bit_depth + 7) / 8))
35 #define PIXEL_STRIDE 16
36
37 #define randomize_buffers(idx) \
38 do { \
39 int x, y; \
40 uint32_t mask = pixel_mask[(idx)]; \
41 for (y = 0; y < sz; y++) { \
42 for (x = 0; x < PIXEL_STRIDE; x += 4) { \
43 AV_WN32A(src + y * PIXEL_STRIDE + x, rnd() & mask); \
44 AV_WN32A(dst + y * PIXEL_STRIDE + x, rnd() & mask); \
45 } \
46 for (x = 0; x < sz; x++) { \
47 if (bit_depth == 8) { \
48 coef[y * sz + x] = src[y * PIXEL_STRIDE + x] - \
49 dst[y * PIXEL_STRIDE + x]; \
50 } else { \
51 ((int32_t *)coef)[y * sz + x] = \
52 ((uint16_t *)src)[y * (PIXEL_STRIDE/2) + x] - \
53 ((uint16_t *)dst)[y * (PIXEL_STRIDE/2) + x]; \
54 } \
55 } \
56 } \
57 } while (0)
58
59 #define dct4x4_impl(size, dctcoef) \
60 static void dct4x4_##size(dctcoef *coef) \
61 { \
62 int i, y, x; \
63 dctcoef tmp[16]; \
64 for (i = 0; i < 4; i++) { \
65 const int z0 = coef[i*4 + 0] + coef[i*4 + 3]; \
66 const int z1 = coef[i*4 + 1] + coef[i*4 + 2]; \
67 const int z2 = coef[i*4 + 0] - coef[i*4 + 3]; \
68 const int z3 = coef[i*4 + 1] - coef[i*4 + 2]; \
69 tmp[i + 4*0] = z0 + z1; \
70 tmp[i + 4*1] = 2*z2 + z3; \
71 tmp[i + 4*2] = z0 - z1; \
72 tmp[i + 4*3] = z2 - 2*z3; \
73 } \
74 for (i = 0; i < 4; i++) { \
75 const int z0 = tmp[i*4 + 0] + tmp[i*4 + 3]; \
76 const int z1 = tmp[i*4 + 1] + tmp[i*4 + 2]; \
77 const int z2 = tmp[i*4 + 0] - tmp[i*4 + 3]; \
78 const int z3 = tmp[i*4 + 1] - tmp[i*4 + 2]; \
79 coef[i*4 + 0] = z0 + z1; \
80 coef[i*4 + 1] = 2*z2 + z3; \
81 coef[i*4 + 2] = z0 - z1; \
82 coef[i*4 + 3] = z2 - 2*z3; \
83 } \
84 for (y = 0; y < 4; y++) { \
85 for (x = 0; x < 4; x++) { \
86 const int64_t scale[] = { 13107 * 10, 8066 * 13, 5243 * 16 }; \
87 const int idx = (y & 1) + (x & 1); \
88 coef[y*4 + x] = (coef[y*4 + x] * scale[idx] + (1 << 14)) >> 15; \
89 } \
90 } \
91 }
92
93 #define DCT8_1D(src, srcstride, dst, dststride) do { \
94 const int a0 = (src)[srcstride * 0] + (src)[srcstride * 7]; \
95 const int a1 = (src)[srcstride * 0] - (src)[srcstride * 7]; \
96 const int a2 = (src)[srcstride * 1] + (src)[srcstride * 6]; \
97 const int a3 = (src)[srcstride * 1] - (src)[srcstride * 6]; \
98 const int a4 = (src)[srcstride * 2] + (src)[srcstride * 5]; \
99 const int a5 = (src)[srcstride * 2] - (src)[srcstride * 5]; \
100 const int a6 = (src)[srcstride * 3] + (src)[srcstride * 4]; \
101 const int a7 = (src)[srcstride * 3] - (src)[srcstride * 4]; \
102 const int b0 = a0 + a6; \
103 const int b1 = a2 + a4; \
104 const int b2 = a0 - a6; \
105 const int b3 = a2 - a4; \
106 const int b4 = a3 + a5 + (a1 + (a1 >> 1)); \
107 const int b5 = a1 - a7 - (a5 + (a5 >> 1)); \
108 const int b6 = a1 + a7 - (a3 + (a3 >> 1)); \
109 const int b7 = a3 - a5 + (a7 + (a7 >> 1)); \
110 (dst)[dststride * 0] = b0 + b1; \
111 (dst)[dststride * 1] = b4 + (b7 >> 2); \
112 (dst)[dststride * 2] = b2 + (b3 >> 1); \
113 (dst)[dststride * 3] = b5 + (b6 >> 2); \
114 (dst)[dststride * 4] = b0 - b1; \
115 (dst)[dststride * 5] = b6 - (b5 >> 2); \
116 (dst)[dststride * 6] = (b2 >> 1) - b3; \
117 (dst)[dststride * 7] = (b4 >> 2) - b7; \
118 } while (0)
119
120 #define dct8x8_impl(size, dctcoef) \
121 static void dct8x8_##size(dctcoef *coef) \
122 { \
123 int i, x, y; \
124 dctcoef tmp[64]; \
125 for (i = 0; i < 8; i++) \
126 DCT8_1D(coef + i, 8, tmp + i, 8); \
127 \
128 for (i = 0; i < 8; i++) \
129 DCT8_1D(tmp + 8*i, 1, coef + i, 8); \
130 \
131 for (y = 0; y < 8; y++) { \
132 for (x = 0; x < 8; x++) { \
133 static const int scale[] = { \
134 13107 * 20, 11428 * 18, 20972 * 32, \
135 12222 * 19, 16777 * 25, 15481 * 24, \
136 }; \
137 static const int idxmap[] = { \
138 0, 3, 4, 3, \
139 3, 1, 5, 1, \
140 4, 5, 2, 5, \
141 3, 1, 5, 1, \
142 }; \
143 const int idx = idxmap[(y & 3) * 4 + (x & 3)]; \
144 coef[y*8 + x] = ((int64_t)coef[y*8 + x] * \
145 scale[idx] + (1 << 17)) >> 18; \
146 } \
147 } \
148 }
149
150
8/8
✓ Branch 0 taken 1820 times.
✓ Branch 1 taken 455 times.
✓ Branch 2 taken 1820 times.
✓ Branch 3 taken 455 times.
✓ Branch 4 taken 7280 times.
✓ Branch 5 taken 1820 times.
✓ Branch 6 taken 1820 times.
✓ Branch 7 taken 455 times.
 13195 dct4x4_impl(16, int16_t)
151
8/8
✓ Branch 0 taken 3952 times.
✓ Branch 1 taken 988 times.
✓ Branch 2 taken 3952 times.
✓ Branch 3 taken 988 times.
✓ Branch 4 taken 15808 times.
✓ Branch 5 taken 3952 times.
✓ Branch 6 taken 3952 times.
✓ Branch 7 taken 988 times.
 28652 dct4x4_impl(32, int32_t)
152
153
8/8
✓ Branch 0 taken 728 times.
✓ Branch 1 taken 91 times.
✓ Branch 2 taken 728 times.
✓ Branch 3 taken 91 times.
✓ Branch 4 taken 5824 times.
✓ Branch 5 taken 728 times.
✓ Branch 6 taken 728 times.
✓ Branch 7 taken 91 times.
 8099 dct8x8_impl(16, int16_t)
154
8/8
✓ Branch 0 taken 2080 times.
✓ Branch 1 taken 260 times.
✓ Branch 2 taken 2080 times.
✓ Branch 3 taken 260 times.
✓ Branch 4 taken 16640 times.
✓ Branch 5 taken 2080 times.
✓ Branch 6 taken 2080 times.
✓ Branch 7 taken 260 times.
 23140 dct8x8_impl(32, int32_t)
155
156 1443 static void dct4x4(int16_t *coef, int bit_depth)
157 {
158
2/2
✓ Branch 0 taken 455 times.
✓ Branch 1 taken 988 times.
 1443 if (bit_depth == 8)
159 455 dct4x4_16(coef);
160 else
161 988 dct4x4_32((int32_t *) coef);
162 1443 }
163
164 351 static void dct8x8(int16_t *coef, int bit_depth)
165 {
166
2/2
✓ Branch 0 taken 91 times.
✓ Branch 1 taken 260 times.
 351 if (bit_depth == 8) {
167 91 dct8x8_16(coef);
168 } else {
169 260 dct8x8_32((int32_t *) coef);
170 }
171 351 }
172
173
174 13 static void check_idct(void)
175 {
176 static const int depths[5] = { 8, 9, 10, 12, 14 };
177 13 LOCAL_ALIGNED_16(uint8_t, src, [8 * 8 * 2]);
178 13 LOCAL_ALIGNED_16(uint8_t, dst, [8 * 8 * 2]);
179 13 LOCAL_ALIGNED_16(uint8_t, dst0, [8 * 8 * 2]);
180 13 LOCAL_ALIGNED_16(uint8_t, dst1_base, [8 * 8 * 2 + 32]);
181 13 LOCAL_ALIGNED_16(int16_t, coef, [8 * 8 * 2]);
182 13 LOCAL_ALIGNED_16(int16_t, subcoef0, [8 * 8 * 2]);
183 13 LOCAL_ALIGNED_16(int16_t, subcoef1, [8 * 8 * 2]);
184 H264DSPContext h;
185 int bit_depth, sz, align, dc, i;
186
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
 13 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *dst, int16_t *block, int stride);
187
188
2/2
✓ Branch 0 taken 65 times.
✓ Branch 1 taken 13 times.
 78 for (i = 0; i < FF_ARRAY_ELEMS(depths); i++) {
189 65 bit_depth = depths[i];
190 65 ff_h264dsp_init(&h, bit_depth, 1);
191
192
2/2
✓ Branch 0 taken 195 times.
✓ Branch 1 taken 65 times.
 260 for (dc = 0; dc <= 2; dc++) {
193
2/2
✓ Branch 0 taken 390 times.
✓ Branch 1 taken 195 times.
 585 for (sz = 4; sz <= 8; sz += 4) {
194 390 void (*idct)(uint8_t *, int16_t *, int) = NULL;
195 390 const char fmts[3][28] = {
196 "h264_idct%d_add_%dbpp", "h264_idct%d_dc_add_%dbpp",
197 "h264_add_pixels%d_%dbpp",
198 };
199
200
8/8
✓ Branch 2 taken 9360 times.
✓ Branch 3 taken 2340 times.
✓ Branch 4 taken 3120 times.
✓ Branch 5 taken 12480 times.
✓ Branch 6 taken 15600 times.
✓ Branch 7 taken 2340 times.
✓ Branch 8 taken 2340 times.
✓ Branch 9 taken 390 times.
 27690 randomize_buffers(i);
201
202
2/2
✓ Branch 0 taken 195 times.
✓ Branch 1 taken 195 times.
 390 if (sz == 4)
203 195 dct4x4(coef, bit_depth);
204 else
205 195 dct8x8(coef, bit_depth);
206
207
6/7
✓ Branch 0 taken 65 times.
✓ Branch 1 taken 65 times.
✓ Branch 2 taken 65 times.
✓ Branch 3 taken 65 times.
✓ Branch 4 taken 65 times.
✓ Branch 5 taken 65 times.
✗ Branch 6 not taken.
 390 switch ((sz << 2) | dc) {
208 65 case (4 << 2) | 0: idct = h.h264_idct_add; break;
209 65 case (4 << 2) | 1: idct = h.h264_idct_dc_add; break;
210 65 case (4 << 2) | 2: idct = h.h264_add_pixels4_clear; break;
211 65 case (8 << 2) | 0: idct = h.h264_idct8_add; break;
212 65 case (8 << 2) | 1: idct = h.h264_idct8_dc_add; break;
213 65 case (8 << 2) | 2: idct = h.h264_add_pixels8_clear; break;
214 }
215
216
2/2
✓ Branch 3 taken 44 times.
✓ Branch 4 taken 346 times.
 390 if (check_func(idct, fmts[dc], sz, bit_depth)) {
217
2/2
✓ Branch 0 taken 86 times.
✓ Branch 1 taken 44 times.
 130 for (align = 0; align < 16; align += sz * SIZEOF_PIXEL) {
218 86 uint8_t *dst1 = dst1_base + align;
219
2/2
✓ Branch 0 taken 52 times.
✓ Branch 1 taken 34 times.
 86 if (dc) {
220 52 memset(subcoef0, 0, sz * sz * SIZEOF_COEF);
221 52 memcpy(subcoef0, coef, SIZEOF_COEF);
222 } else {
223 34 memcpy(subcoef0, coef, sz * sz * SIZEOF_COEF);
224 }
225 86 memcpy(dst0, dst, sz * PIXEL_STRIDE);
226 86 memcpy(dst1, dst, sz * PIXEL_STRIDE);
227 86 memcpy(subcoef1, subcoef0, sz * sz * SIZEOF_COEF);
228 86 call_ref(dst0, subcoef0, PIXEL_STRIDE);
229 86 call_new(dst1, subcoef1, PIXEL_STRIDE);
230
1/2
✓ Branch 0 taken 86 times.
✗ Branch 1 not taken.
 86 if (memcmp(dst0, dst1, sz * PIXEL_STRIDE) ||
231
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 86 times.
 86 memcmp(subcoef0, subcoef1, sz * sz * SIZEOF_COEF))
232 ✗ fail();
233
1/8
✗ Branch 1 not taken.
✓ Branch 2 taken 86 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
 86 bench_new(dst1, subcoef1, sz * SIZEOF_PIXEL);
234 }
235 }
236 }
237 }
238 }
239 13 }
240
241 13 static void check_idct_multiple(void)
242 {
243 13 LOCAL_ALIGNED_16(uint8_t, dst_full, [16 * 16 * 2]);
244 13 LOCAL_ALIGNED_16(int16_t, coef_full, [16 * 16 * 2]);
245 13 LOCAL_ALIGNED_16(uint8_t, dst0, [16 * 16 * 2]);
246 13 LOCAL_ALIGNED_16(uint8_t, dst1, [16 * 16 * 2]);
247 13 LOCAL_ALIGNED_16(int16_t, coef0, [16 * 16 * 2]);
248 13 LOCAL_ALIGNED_16(int16_t, coef1, [16 * 16 * 2]);
249 13 LOCAL_ALIGNED_16(uint8_t, nnzc, [15 * 8]);
250 H264DSPContext h;
251 int bit_depth, i, y, func;
252
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
 13 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]);
253
254
2/2
✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
 52 for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
255 39 ff_h264dsp_init(&h, bit_depth, 1);
256
2/2
✓ Branch 0 taken 117 times.
✓ Branch 1 taken 39 times.
 156 for (func = 0; func < 3; func++) {
257 117 void (*idct)(uint8_t *, const int *, int16_t *, int, const uint8_t[]) = NULL;
258 const char *name;
259 117 int sz = 4, intra = 0;
260 117 int block_offset[16] = { 0 };
261
3/4
✓ Branch 0 taken 39 times.
✓ Branch 1 taken 39 times.
✓ Branch 2 taken 39 times.
✗ Branch 3 not taken.
 117 switch (func) {
262 39 case 0:
263 39 idct = h.h264_idct_add16;
264 39 name = "h264_idct_add16";
265 39 break;
266 39 case 1:
267 39 idct = h.h264_idct_add16intra;
268 39 name = "h264_idct_add16intra";
269 39 intra = 1;
270 39 break;
271 39 case 2:
272 39 idct = h.h264_idct8_add4;
273 39 name = "h264_idct8_add4";
274 39 sz = 8;
275 39 break;
276 }
277 117 memset(nnzc, 0, 15 * 8);
278 117 memset(coef_full, 0, 16 * 16 * SIZEOF_COEF);
279
2/2
✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 117 times.
 1521 for (i = 0; i < 16 * 16; i += sz * sz) {
280 uint8_t src[8 * 8 * 2];
281 uint8_t dst[8 * 8 * 2];
282 int16_t coef[8 * 8 * 2];
283 1404 int index = i / sz;
284 1404 int block_y = (index / 16) * sz;
285 1404 int block_x = index % 16;
286 1404 int offset = (block_y * 16 + block_x) * SIZEOF_PIXEL;
287 1404 int nnz = rnd() % 3;
288
289
8/8
✓ Branch 2 taken 24960 times.
✓ Branch 3 taken 6240 times.
✓ Branch 4 taken 9984 times.
✓ Branch 5 taken 19968 times.
✓ Branch 6 taken 29952 times.
✓ Branch 7 taken 6240 times.
✓ Branch 8 taken 6240 times.
✓ Branch 9 taken 1404 times.
 62556 randomize_buffers(bit_depth - 8);
290
2/2
✓ Branch 0 taken 1248 times.
✓ Branch 1 taken 156 times.
 1404 if (sz == 4)
291 1248 dct4x4(coef, bit_depth);
292 else
293 156 dct8x8(coef, bit_depth);
294
295
2/2
✓ Branch 0 taken 6240 times.
✓ Branch 1 taken 1404 times.
 7644 for (y = 0; y < sz; y++)
296 6240 memcpy(&dst_full[offset + y * 16 * SIZEOF_PIXEL],
297 6240 &dst[PIXEL_STRIDE * y], sz * SIZEOF_PIXEL);
298
299
2/2
✓ Branch 0 taken 486 times.
✓ Branch 1 taken 918 times.
 1404 if (nnz > 1)
300 486 nnz = sz * sz;
301 1404 memcpy(&coef_full[i * SIZEOF_COEF/sizeof(coef[0])],
302 1404 coef, nnz * SIZEOF_COEF);
303
304
4/4
✓ Branch 0 taken 624 times.
✓ Branch 1 taken 780 times.
✓ Branch 2 taken 189 times.
✓ Branch 3 taken 435 times.
 1404 if (intra && nnz == 1)
305 189 nnz = 0;
306
307 1404 nnzc[scan8[i / 16]] = nnz;
308 1404 block_offset[i / 16] = offset;
309 }
310
311
2/2
✓ Branch 3 taken 18 times.
✓ Branch 4 taken 99 times.
 117 if (check_func(idct, "%s_%dbpp", name, bit_depth)) {
312 18 memcpy(coef0, coef_full, 16 * 16 * SIZEOF_COEF);
313 18 memcpy(coef1, coef_full, 16 * 16 * SIZEOF_COEF);
314 18 memcpy(dst0, dst_full, 16 * 16 * SIZEOF_PIXEL);
315 18 memcpy(dst1, dst_full, 16 * 16 * SIZEOF_PIXEL);
316 18 call_ref(dst0, block_offset, coef0, 16 * SIZEOF_PIXEL, nnzc);
317 18 call_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
318
1/2
✓ Branch 0 taken 18 times.
✗ Branch 1 not taken.
 18 if (memcmp(dst0, dst1, 16 * 16 * SIZEOF_PIXEL) ||
319
1/2
✗ Branch 0 not taken.
✓ Branch 1 taken 18 times.
 18 memcmp(coef0, coef1, 16 * 16 * SIZEOF_COEF))
320 ✗ fail();
321
1/8
✗ Branch 1 not taken.
✓ Branch 2 taken 18 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
 18 bench_new(dst1, block_offset, coef1, 16 * SIZEOF_PIXEL, nnzc);
322 }
323 }
324 }
325 13 }
326
327
328 13 static void check_loop_filter(void)
329 {
330 13 LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
331 13 LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
332 13 LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
333 H264DSPContext h;
334 int bit_depth;
335 int alphas[36], betas[36];
336 int8_t tc0[36][4];
337
338
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
 13 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
339 int alpha, int beta, int8_t *tc0);
340
341
2/2
✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
 52 for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
342 int i, j, a, c;
343 39 uint32_t mask = pixel_mask_lf[bit_depth - 8];
344 39 ff_h264dsp_init(&h, bit_depth, 1);
345
2/2
✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 39 times.
 1443 for (i = 35, a = 255, c = 250; i >= 0; i--) {
346 1404 alphas[i] = a << (bit_depth - 8);
347 1404 betas[i] = (i + 1) / 2 << (bit_depth - 8);
348 1404 tc0[i][0] = tc0[i][3] = (c + 6) / 10;
349 1404 tc0[i][1] = (c + 7) / 15;
350 1404 tc0[i][2] = (c + 9) / 20;
351 1404 a = a*9/10;
352 1404 c = c*9/10;
353 }
354
355 #define CHECK_LOOP_FILTER(name, align, idc) \
356 do { \
357 if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) { \
358 for (j = 0; j < 36; j++) { \
359 intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \
360 for (i = 0; i < 1024; i+=4) { \
361 AV_WN32A(dst + i, rnd() & mask); \
362 } \
363 memcpy(dst0, dst, 32 * 16 * 2); \
364 memcpy(dst1, dst, 32 * 16 * 2); \
365 \
366 call_ref(dst0 + off, 32, alphas[j], betas[j], tc0[j]); \
367 call_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]); \
368 if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) { \
369 fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d " \
370 "tc0:{%d,%d,%d,%d}\n", j, alphas[j], betas[j], \
371 tc0[j][0], tc0[j][1], tc0[j][2], tc0[j][3]); \
372 fail(); \
373 } \
374 bench_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]);\
375 } \
376 } \
377 } while (0)
378
379
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_v_loop_filter_luma, 1,);
380
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_h_loop_filter_luma, 0,);
381
8/16
✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
 46299 CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff, 0,);
382
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_v_loop_filter_chroma, 1,);
383
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0,);
384
8/16
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
 27795 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0,);
385
386 39 ff_h264dsp_init(&h, bit_depth, 2);
387
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0, 422);
388
8/16
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
 27795 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0, 422);
389 #undef CHECK_LOOP_FILTER
390 }
391 13 }
392
393 13 static void check_loop_filter_intra(void)
394 {
395 13 LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
396 13 LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
397 13 LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
398 H264DSPContext h;
399 int bit_depth;
400 int alphas[36], betas[36];
401
402
2/2
✓ Branch 1 taken 12 times.
✓ Branch 2 taken 1 times.
 13 declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, ptrdiff_t stride,
403 int alpha, int beta);
404
405
2/2
✓ Branch 0 taken 39 times.
✓ Branch 1 taken 13 times.
 52 for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
406 int i, j, a;
407 39 uint32_t mask = pixel_mask_lf[bit_depth - 8];
408 39 ff_h264dsp_init(&h, bit_depth, 1);
409
2/2
✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 39 times.
 1443 for (i = 35, a = 255; i >= 0; i--) {
410 1404 alphas[i] = a << (bit_depth - 8);
411 1404 betas[i] = (i + 1) / 2 << (bit_depth - 8);
412 1404 a = a*9/10;
413 }
414
415 #define CHECK_LOOP_FILTER(name, align, idc) \
416 do { \
417 if (check_func(h.name, #name #idc "_%dbpp", bit_depth)) { \
418 for (j = 0; j < 36; j++) { \
419 intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \
420 for (i = 0; i < 1024; i+=4) { \
421 AV_WN32A(dst + i, rnd() & mask); \
422 } \
423 memcpy(dst0, dst, 32 * 16 * 2); \
424 memcpy(dst1, dst, 32 * 16 * 2); \
425 \
426 call_ref(dst0 + off, 32, alphas[j], betas[j]); \
427 call_new(dst1 + off, 32, alphas[j], betas[j]); \
428 if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) { \
429 fprintf(stderr, #name #idc ": j:%d, alpha:%d beta:%d\n", \
430 j, alphas[j], betas[j]); \
431 fail(); \
432 } \
433 bench_new(dst1 + off, 32, alphas[j], betas[j]); \
434 } \
435 } \
436 } while (0)
437
438
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_v_loop_filter_luma_intra, 1,);
439
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_h_loop_filter_luma_intra, 0,);
440
8/16
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
 27795 CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff_intra, 0,);
441
8/16
✓ Branch 3 taken 7 times.
✓ Branch 4 taken 32 times.
✓ Branch 6 taken 64512 times.
✓ Branch 7 taken 252 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 252 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 252 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 252 times.
✓ Branch 40 taken 7 times.
 64803 CHECK_LOOP_FILTER(h264_v_loop_filter_chroma_intra, 1,);
442
8/16
✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
 46299 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0,);
443
8/16
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
 27795 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0,);
444
445 39 ff_h264dsp_init(&h, bit_depth, 2);
446
8/16
✓ Branch 3 taken 5 times.
✓ Branch 4 taken 34 times.
✓ Branch 6 taken 46080 times.
✓ Branch 7 taken 180 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 180 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 180 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 180 times.
✓ Branch 40 taken 5 times.
 46299 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_intra, 0, 422);
447
8/16
✓ Branch 3 taken 3 times.
✓ Branch 4 taken 36 times.
✓ Branch 6 taken 27648 times.
✓ Branch 7 taken 108 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 108 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 108 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✗ Branch 33 not taken.
✗ Branch 34 not taken.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 39 taken 108 times.
✓ Branch 40 taken 3 times.
 27795 CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff_intra, 0, 422);
448 #undef CHECK_LOOP_FILTER
449 }
450 13 }
451
452 13 void checkasm_check_h264dsp(void)
453 {
454 13 check_idct();
455 13 check_idct_multiple();
456 13 report("idct");
457
458 13 check_loop_filter();
459 13 report("loop_filter");
460
461 13 check_loop_filter_intra();
462 13 report("loop_filter_intra");
463 13 }
464