FFmpeg coverage

Directory: ../../../ffmpeg/
File: src/libavcodec/x86/h264dsp_init.c
Date: 2026-04-24 19:58:39
Exec Total Coverage
Lines: 112 112 100.0%
Functions: 1 1 100.0%
Branches: 42 42 100.0%
Line Branch Exec Source
1 /*
2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
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/cpu.h"
23 #include "libavutil/x86/asm.h"
24 #include "libavutil/x86/cpu.h"
25 #include "libavcodec/h264dsp.h"
26
27 /***********************************/
28 /* IDCT */
29 #define IDCT_ADD_FUNC(NUM, DEPTH, OPT) \
30 void ff_h264_idct ## NUM ## _add_ ## DEPTH ## _ ## OPT(uint8_t *dst, \
31 int16_t *block, \
32 int stride);
33
34 IDCT_ADD_FUNC(, 8, sse2)
35 IDCT_ADD_FUNC(, 8, avx)
36 IDCT_ADD_FUNC(, 10, sse2)
37 IDCT_ADD_FUNC(_dc, 8, sse2)
38 IDCT_ADD_FUNC(_dc, 8, avx)
39 IDCT_ADD_FUNC(_dc, 10, mmxext)
40 IDCT_ADD_FUNC(8_dc, 8, mmxext)
41 IDCT_ADD_FUNC(8_dc, 10, sse2)
42 IDCT_ADD_FUNC(8, 8, sse2)
43 IDCT_ADD_FUNC(8, 10, sse2)
44 IDCT_ADD_FUNC(, 10, avx)
45 IDCT_ADD_FUNC(8_dc, 10, avx)
46 IDCT_ADD_FUNC(8, 10, avx)
47
48
49 #define IDCT_ADD_REP_FUNC(NUM, REP, DEPTH, OPT) \
50 void ff_h264_idct ## NUM ## _add ## REP ## _ ## DEPTH ## _ ## OPT \
51 (uint8_t *dst, const int *block_offset, \
52 int16_t *block, int stride, const uint8_t nnzc[5 * 8]);
53
54 IDCT_ADD_REP_FUNC(8, 4, 8, sse2)
55 IDCT_ADD_REP_FUNC(8, 4, 10, sse2)
56 IDCT_ADD_REP_FUNC(8, 4, 10, avx)
57 IDCT_ADD_REP_FUNC(, 16, 8, sse2)
58 IDCT_ADD_REP_FUNC(, 16, 10, sse2)
59 IDCT_ADD_REP_FUNC(, 16intra, 8, sse2)
60 IDCT_ADD_REP_FUNC(, 16intra, 10, sse2)
61 IDCT_ADD_REP_FUNC(, 16, 10, avx)
62 IDCT_ADD_REP_FUNC(, 16intra, 10, avx)
63
64
65 #define IDCT_ADD_REP_FUNC2(NUM, REP, DEPTH, OPT) \
66 void ff_h264_idct ## NUM ## _add ## REP ## _ ## DEPTH ## _ ## OPT \
67 (uint8_t **dst, const int *block_offset, \
68 int16_t *block, int stride, const uint8_t nnzc[15 * 8]);
69
70 IDCT_ADD_REP_FUNC2(, 8, 8, sse2)
71 IDCT_ADD_REP_FUNC2(, 8, 10, sse2)
72 IDCT_ADD_REP_FUNC2(, 8, 10, avx)
73
74 IDCT_ADD_REP_FUNC2(, 8_422, 8, mmx)
75
76 IDCT_ADD_REP_FUNC2(, 8_422, 10, sse2)
77 IDCT_ADD_REP_FUNC2(, 8_422, 10, avx)
78
79 void ff_h264_luma_dc_dequant_idct_sse2(int16_t *output, int16_t *input, int qmul);
80
81 /***********************************/
82 /* deblocking */
83
84 void ff_h264_loop_filter_strength_mmxext(int16_t bS[2][4][4], uint8_t nnz[40],
85 int8_t ref[2][40],
86 int16_t mv[2][40][2],
87 int bidir, int edges, int step,
88 int mask_mv0, int mask_mv1, int field);
89
90 #define LF_FUNC(DIR, TYPE, DEPTH, OPT) \
91 void ff_deblock_ ## DIR ## _ ## TYPE ## _ ## DEPTH ## _ ## OPT(uint8_t *pix, \
92 ptrdiff_t stride, \
93 int alpha, \
94 int beta, \
95 int8_t *tc0);
96 #define LF_IFUNC(DIR, TYPE, DEPTH, OPT) \
97 void ff_deblock_ ## DIR ## _ ## TYPE ## _ ## DEPTH ## _ ## OPT(uint8_t *pix, \
98 ptrdiff_t stride, \
99 int alpha, \
100 int beta);
101
102 #define LF_FUNCS(type, depth) \
103 LF_FUNC(h, luma, depth, sse2) \
104 LF_IFUNC(h, luma_intra, depth, sse2) \
105 LF_FUNC(v, luma, depth, sse2) \
106 LF_IFUNC(v, luma_intra, depth, sse2) \
107 LF_FUNC(h, chroma, depth, sse2) \
108 LF_IFUNC(h, chroma_intra, depth, sse2) \
109 LF_FUNC(h, chroma422, depth, sse2) \
110 LF_IFUNC(h, chroma422_intra, depth, sse2) \
111 LF_FUNC(v, chroma, depth, sse2) \
112 LF_IFUNC(v, chroma_intra, depth, sse2) \
113 LF_FUNC(h, luma, depth, avx) \
114 LF_IFUNC(h, luma_intra, depth, avx) \
115 LF_FUNC(v, luma, depth, avx) \
116 LF_IFUNC(v, luma_intra, depth, avx) \
117 LF_FUNC(h, chroma, depth, avx) \
118 LF_IFUNC(h, chroma_intra, depth, avx) \
119 LF_FUNC(h, chroma422, depth, avx) \
120 LF_IFUNC(h, chroma422_intra, depth, avx) \
121 LF_FUNC(v, chroma, depth, avx) \
122 LF_IFUNC(v, chroma_intra, depth, avx)
123
124 LF_FUNC(h, luma_mbaff, 8, sse2)
125 LF_FUNC(h, luma_mbaff, 8, avx)
126
127 LF_FUNCS(uint8_t, 8)
128 LF_FUNCS(uint16_t, 10)
129
130 /***********************************/
131 /* weighted prediction */
132
133 #define H264_WEIGHT(W, OPT) \
134 void ff_h264_weight_ ## W ## _ ## OPT(uint8_t *dst, ptrdiff_t stride, \
135 int height, int log2_denom, \
136 int weight, int offset);
137
138 #define H264_BIWEIGHT(W, OPT) \
139 void ff_h264_biweight_ ## W ## _ ## OPT(uint8_t *dst, uint8_t *src, \
140 ptrdiff_t stride, int height, \
141 int log2_denom, int weightd, \
142 int weights, int offset);
143
144 #define H264_BIWEIGHT_MMX(W) \
145 H264_WEIGHT(W, mmxext) \
146 H264_BIWEIGHT(W, mmxext)
147
148 #define H264_BIWEIGHT_SSE(W) \
149 H264_WEIGHT(W, sse2) \
150 H264_BIWEIGHT(W, sse2) \
151 H264_BIWEIGHT(W, ssse3)
152
153 H264_BIWEIGHT_SSE(16)
154 H264_BIWEIGHT_SSE(8)
155 H264_BIWEIGHT_MMX(4)
156
157 #define H264_WEIGHT_10(W, DEPTH, OPT) \
158 void ff_h264_weight_ ## W ## _ ## DEPTH ## _ ## OPT(uint8_t *dst, \
159 ptrdiff_t stride, \
160 int height, \
161 int log2_denom, \
162 int weight, \
163 int offset);
164
165 #define H264_BIWEIGHT_10(W, DEPTH, OPT) \
166 void ff_h264_biweight_ ## W ## _ ## DEPTH ## _ ## OPT(uint8_t *dst, \
167 uint8_t *src, \
168 ptrdiff_t stride, \
169 int height, \
170 int log2_denom, \
171 int weightd, \
172 int weights, \
173 int offset);
174
175 #define H264_BIWEIGHT_10_SSE(W, DEPTH) \
176 H264_WEIGHT_10(W, DEPTH, sse2) \
177 H264_WEIGHT_10(W, DEPTH, sse4) \
178 H264_BIWEIGHT_10(W, DEPTH, sse2) \
179 H264_BIWEIGHT_10(W, DEPTH, sse4)
180
181 H264_BIWEIGHT_10_SSE(16, 10)
182 H264_BIWEIGHT_10_SSE(8, 10)
183 H264_BIWEIGHT_10_SSE(4, 10)
184
185 1458 av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth,
186 const int chroma_format_idc)
187 {
188 1458 int cpu_flags = av_get_cpu_flags();
189
190
4/4
✓ Branch 0 taken 503 times.
✓ Branch 1 taken 955 times.
✓ Branch 2 taken 424 times.
✓ Branch 3 taken 79 times.
 1458 if (EXTERNAL_MMXEXT(cpu_flags) && chroma_format_idc <= 1)
191 424 c->loop_filter_strength = ff_h264_loop_filter_strength_mmxext;
192
193
2/2
✓ Branch 0 taken 1150 times.
✓ Branch 1 taken 308 times.
 1458 if (bit_depth == 8) {
194
2/2
✓ Branch 0 taken 282 times.
✓ Branch 1 taken 868 times.
 1150 if (EXTERNAL_MMX(cpu_flags)) {
195
2/2
✓ Branch 0 taken 33 times.
✓ Branch 1 taken 249 times.
 282 if (chroma_format_idc <= 1) {
196 } else {
197 33 c->idct_add8 = ff_h264_idct_add8_422_8_mmx;
198 }
199 }
200
2/2
✓ Branch 0 taken 275 times.
✓ Branch 1 taken 875 times.
 1150 if (EXTERNAL_MMXEXT(cpu_flags)) {
201 275 c->idct8_dc_add = ff_h264_idct8_dc_add_8_mmxext;
202
203 275 c->weight_pixels_tab[2] = ff_h264_weight_4_mmxext;
204
205 275 c->biweight_pixels_tab[2] = ff_h264_biweight_4_mmxext;
206 }
207
2/2
✓ Branch 0 taken 261 times.
✓ Branch 1 taken 889 times.
 1150 if (EXTERNAL_SSE2(cpu_flags)) {
208 261 c->idct8_add = ff_h264_idct8_add_8_sse2;
209
210 261 c->idct_add16 = ff_h264_idct_add16_8_sse2;
211 261 c->idct8_add4 = ff_h264_idct8_add4_8_sse2;
212
2/2
✓ Branch 0 taken 234 times.
✓ Branch 1 taken 27 times.
 261 if (chroma_format_idc <= 1)
213 234 c->idct_add8 = ff_h264_idct_add8_8_sse2;
214 261 c->idct_add16intra = ff_h264_idct_add16intra_8_sse2;
215 261 c->luma_dc_dequant_idct = ff_h264_luma_dc_dequant_idct_sse2;
216
217 261 c->weight_pixels_tab[0] = ff_h264_weight_16_sse2;
218 261 c->weight_pixels_tab[1] = ff_h264_weight_8_sse2;
219
220 261 c->biweight_pixels_tab[0] = ff_h264_biweight_16_sse2;
221 261 c->biweight_pixels_tab[1] = ff_h264_biweight_8_sse2;
222
223 261 c->v_loop_filter_luma = ff_deblock_v_luma_8_sse2;
224 261 c->h_loop_filter_luma = ff_deblock_h_luma_8_sse2;
225 261 c->v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_sse2;
226 261 c->h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_sse2;
227
228 #if ARCH_X86_64
229 261 c->h_loop_filter_luma_mbaff = ff_deblock_h_luma_mbaff_8_sse2;
230 #endif
231
232 261 c->v_loop_filter_chroma = ff_deblock_v_chroma_8_sse2;
233 261 c->v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_8_sse2;
234
2/2
✓ Branch 0 taken 234 times.
✓ Branch 1 taken 27 times.
 261 if (chroma_format_idc <= 1) {
235 234 c->h_loop_filter_chroma = ff_deblock_h_chroma_8_sse2;
236 234 c->h_loop_filter_chroma_intra = ff_deblock_h_chroma_intra_8_sse2;
237 } else {
238 27 c->h_loop_filter_chroma = ff_deblock_h_chroma422_8_sse2;
239 27 c->h_loop_filter_chroma_intra = ff_deblock_h_chroma422_intra_8_sse2;
240 }
241
242 261 c->idct_add = ff_h264_idct_add_8_sse2;
243 261 c->idct_dc_add = ff_h264_idct_dc_add_8_sse2;
244 }
245
2/2
✓ Branch 0 taken 247 times.
✓ Branch 1 taken 903 times.
 1150 if (EXTERNAL_SSSE3(cpu_flags)) {
246 247 c->biweight_pixels_tab[0] = ff_h264_biweight_16_ssse3;
247 247 c->biweight_pixels_tab[1] = ff_h264_biweight_8_ssse3;
248 }
249
2/2
✓ Branch 0 taken 212 times.
✓ Branch 1 taken 938 times.
 1150 if (EXTERNAL_AVX(cpu_flags)) {
250 212 c->v_loop_filter_luma = ff_deblock_v_luma_8_avx;
251 212 c->h_loop_filter_luma = ff_deblock_h_luma_8_avx;
252 212 c->v_loop_filter_luma_intra = ff_deblock_v_luma_intra_8_avx;
253 212 c->h_loop_filter_luma_intra = ff_deblock_h_luma_intra_8_avx;
254 #if ARCH_X86_64
255 212 c->h_loop_filter_luma_mbaff = ff_deblock_h_luma_mbaff_8_avx;
256 #endif
257
258 212 c->v_loop_filter_chroma = ff_deblock_v_chroma_8_avx;
259 212 c->v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_8_avx;
260
2/2
✓ Branch 0 taken 199 times.
✓ Branch 1 taken 13 times.
 212 if (chroma_format_idc <= 1) {
261 199 c->h_loop_filter_chroma = ff_deblock_h_chroma_8_avx;
262 199 c->h_loop_filter_chroma_intra = ff_deblock_h_chroma_intra_8_avx;
263 } else {
264 13 c->h_loop_filter_chroma = ff_deblock_h_chroma422_8_avx;
265 13 c->h_loop_filter_chroma_intra = ff_deblock_h_chroma422_intra_8_avx;
266 }
267
268 212 c->idct_add = ff_h264_idct_add_8_avx;
269 212 c->idct_dc_add = ff_h264_idct_dc_add_8_avx;
270 }
271
2/2
✓ Branch 0 taken 149 times.
✓ Branch 1 taken 159 times.
 308 } else if (bit_depth == 10) {
272
2/2
✓ Branch 0 taken 96 times.
✓ Branch 1 taken 53 times.
 149 if (EXTERNAL_MMXEXT(cpu_flags)) {
273 96 c->idct_dc_add = ff_h264_idct_dc_add_10_mmxext;
274 }
275
2/2
✓ Branch 0 taken 82 times.
✓ Branch 1 taken 67 times.
 149 if (EXTERNAL_SSE2(cpu_flags)) {
276 82 c->idct_add = ff_h264_idct_add_10_sse2;
277 82 c->idct8_dc_add = ff_h264_idct8_dc_add_10_sse2;
278
279 82 c->idct_add16 = ff_h264_idct_add16_10_sse2;
280
2/2
✓ Branch 0 taken 62 times.
✓ Branch 1 taken 20 times.
 82 if (chroma_format_idc <= 1) {
281 62 c->idct_add8 = ff_h264_idct_add8_10_sse2;
282 } else {
283 20 c->idct_add8 = ff_h264_idct_add8_422_10_sse2;
284 }
285 82 c->idct_add16intra = ff_h264_idct_add16intra_10_sse2;
286 #if HAVE_ALIGNED_STACK
287 82 c->idct8_add = ff_h264_idct8_add_10_sse2;
288 82 c->idct8_add4 = ff_h264_idct8_add4_10_sse2;
289 #endif /* HAVE_ALIGNED_STACK */
290
291 82 c->weight_pixels_tab[0] = ff_h264_weight_16_10_sse2;
292 82 c->weight_pixels_tab[1] = ff_h264_weight_8_10_sse2;
293 82 c->weight_pixels_tab[2] = ff_h264_weight_4_10_sse2;
294
295 82 c->biweight_pixels_tab[0] = ff_h264_biweight_16_10_sse2;
296 82 c->biweight_pixels_tab[1] = ff_h264_biweight_8_10_sse2;
297 82 c->biweight_pixels_tab[2] = ff_h264_biweight_4_10_sse2;
298
299 82 c->v_loop_filter_chroma = ff_deblock_v_chroma_10_sse2;
300 82 c->v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_10_sse2;
301
2/2
✓ Branch 0 taken 62 times.
✓ Branch 1 taken 20 times.
 82 if (chroma_format_idc <= 1) {
302 62 c->h_loop_filter_chroma = ff_deblock_h_chroma_10_sse2;
303 } else {
304 20 c->h_loop_filter_chroma = ff_deblock_h_chroma422_10_sse2;
305 }
306 82 c->v_loop_filter_luma = ff_deblock_v_luma_10_sse2;
307 82 c->h_loop_filter_luma = ff_deblock_h_luma_10_sse2;
308 82 c->v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_sse2;
309 82 c->h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_sse2;
310 }
311
2/2
✓ Branch 0 taken 61 times.
✓ Branch 1 taken 88 times.
 149 if (EXTERNAL_SSE4(cpu_flags)) {
312 61 c->weight_pixels_tab[0] = ff_h264_weight_16_10_sse4;
313 61 c->weight_pixels_tab[1] = ff_h264_weight_8_10_sse4;
314 61 c->weight_pixels_tab[2] = ff_h264_weight_4_10_sse4;
315
316 61 c->biweight_pixels_tab[0] = ff_h264_biweight_16_10_sse4;
317 61 c->biweight_pixels_tab[1] = ff_h264_biweight_8_10_sse4;
318 61 c->biweight_pixels_tab[2] = ff_h264_biweight_4_10_sse4;
319 }
320
2/2
✓ Branch 0 taken 33 times.
✓ Branch 1 taken 116 times.
 149 if (EXTERNAL_AVX(cpu_flags)) {
321 33 c->idct_dc_add =
322 33 c->idct_add = ff_h264_idct_add_10_avx;
323 33 c->idct8_dc_add = ff_h264_idct8_dc_add_10_avx;
324
325 33 c->idct_add16 = ff_h264_idct_add16_10_avx;
326
2/2
✓ Branch 0 taken 27 times.
✓ Branch 1 taken 6 times.
 33 if (chroma_format_idc <= 1) {
327 27 c->idct_add8 = ff_h264_idct_add8_10_avx;
328 } else {
329 6 c->idct_add8 = ff_h264_idct_add8_422_10_avx;
330 }
331 33 c->idct_add16intra = ff_h264_idct_add16intra_10_avx;
332 #if HAVE_ALIGNED_STACK
333 33 c->idct8_add = ff_h264_idct8_add_10_avx;
334 33 c->idct8_add4 = ff_h264_idct8_add4_10_avx;
335 #endif /* HAVE_ALIGNED_STACK */
336
337 33 c->v_loop_filter_chroma = ff_deblock_v_chroma_10_avx;
338 33 c->v_loop_filter_chroma_intra = ff_deblock_v_chroma_intra_10_avx;
339
2/2
✓ Branch 0 taken 27 times.
✓ Branch 1 taken 6 times.
 33 if (chroma_format_idc <= 1) {
340 27 c->h_loop_filter_chroma = ff_deblock_h_chroma_10_avx;
341 } else {
342 6 c->h_loop_filter_chroma = ff_deblock_h_chroma422_10_avx;
343 }
344 33 c->v_loop_filter_luma = ff_deblock_v_luma_10_avx;
345 33 c->h_loop_filter_luma = ff_deblock_h_luma_10_avx;
346 33 c->v_loop_filter_luma_intra = ff_deblock_v_luma_intra_10_avx;
347 33 c->h_loop_filter_luma_intra = ff_deblock_h_luma_intra_10_avx;
348 }
349 }
350 1458 }
351