FFmpeg
dsp_init.c
Go to the documentation of this file.
1 /*
2  * VVC filters DSP
3  *
4  * Copyright (C) 2024 Zhao Zhili
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #include "libavutil/cpu.h"
24 #include "libavutil/aarch64/cpu.h"
25 #include "libavcodec/aarch64/h26x/dsp.h"
26 #include "libavcodec/vvc/dsp.h"
27 #include "libavcodec/vvc/dec.h"
28 #include "libavcodec/vvc/ctu.h"
29 
30 #define BDOF_BLOCK_SIZE 16
31 #define BDOF_MIN_BLOCK_SIZE 4
32 
33 
34 void ff_alf_classify_sum_neon(int *sum0, int *sum1, int16_t *grad, uint32_t gshift, uint32_t steps);
35 
36 #define BIT_DEPTH 8
37 #include "alf_template.c"
38 #undef BIT_DEPTH
39 
40 #define BIT_DEPTH 10
41 #include "alf_template.c"
42 #undef BIT_DEPTH
43 
44 #define BIT_DEPTH 12
45 #include "alf_template.c"
46 #undef BIT_DEPTH
47 
48 int ff_vvc_sad_neon(const int16_t *src0, const int16_t *src1, int dx, int dy,
49  const int block_w, const int block_h);
50 
51 void ff_vvc_avg_8_neon(uint8_t *dst, ptrdiff_t dst_stride,
52  const int16_t *src0, const int16_t *src1, int width,
53  int height);
54 void ff_vvc_avg_10_neon(uint8_t *dst, ptrdiff_t dst_stride,
55  const int16_t *src0, const int16_t *src1, int width,
56  int height);
57 void ff_vvc_avg_12_neon(uint8_t *dst, ptrdiff_t dst_stride,
58  const int16_t *src0, const int16_t *src1, int width,
59  int height);
60 
61 void ff_vvc_w_avg_8_neon(uint8_t *_dst, ptrdiff_t _dst_stride,
62  const int16_t *src0, const int16_t *src1,
63  int width, int height,
64  uintptr_t w0_w1, uintptr_t offset_shift);
65 void ff_vvc_w_avg_10_neon(uint8_t *_dst, ptrdiff_t _dst_stride,
66  const int16_t *src0, const int16_t *src1,
67  int width, int height,
68  uintptr_t w0_w1, uintptr_t offset_shift);
69 void ff_vvc_w_avg_12_neon(uint8_t *_dst, ptrdiff_t _dst_stride,
70  const int16_t *src0, const int16_t *src1,
71  int width, int height,
72  uintptr_t w0_w1, uintptr_t offset_shift);
73 /* When passing arguments to functions, Apple platforms diverge from the ARM64
74  * standard ABI for functions that require passing arguments on the stack. To
75  * simplify portability in the assembly function interface, use a different
76  * function signature that doesn't require passing arguments on the stack.
77  */
78 #define W_AVG_FUN(bit_depth) \
79 static void vvc_w_avg_ ## bit_depth(uint8_t *dst, ptrdiff_t dst_stride, \
80  const int16_t *src0, const int16_t *src1, int width, int height, \
81  int denom, int w0, int w1, int o0, int o1) \
82 { \
83  int shift = denom + FFMAX(3, 15 - bit_depth); \
84  int offset = ((o0 + o1) * (1 << (bit_depth - 8)) + 1) * (1 << (shift - 1)); \
85  uintptr_t w0_w1 = ((uintptr_t)w0 << 32) | (uint32_t)w1; \
86  uintptr_t offset_shift = ((uintptr_t)offset << 32) | (uint32_t)shift; \
87  ff_vvc_w_avg_ ## bit_depth ## _neon(dst, dst_stride, src0, src1, width, height, w0_w1, offset_shift); \
88 }
89 
90 W_AVG_FUN(8)
91 W_AVG_FUN(10)
92 W_AVG_FUN(12)
93 
94 #define DMVR_FUN(fn, bd) \
95  void ff_vvc_dmvr_ ## fn ## bd ## _neon(int16_t *dst, \
96  const uint8_t *_src, ptrdiff_t _src_stride, int height, \
97  intptr_t mx, intptr_t my, int width);
98 
99 DMVR_FUN(, 8)
100 DMVR_FUN(, 12)
101 DMVR_FUN(h_, 8)
102 DMVR_FUN(h_, 10)
103 DMVR_FUN(h_, 12)
104 DMVR_FUN(v_, 8)
105 DMVR_FUN(hv_, 8)
106 DMVR_FUN(hv_, 10)
107 DMVR_FUN(hv_, 12)
108 
109 #define APPLY_BDOF_FUNC(bd) \
110  void ff_vvc_apply_bdof_ ## bd ## _neon(uint8_t *_dst, ptrdiff_t _dst_stride, \
111  const int16_t *_src0, const int16_t *_src1, \
112  int block_w, int block_h);
113 
114 APPLY_BDOF_FUNC(8)
115 APPLY_BDOF_FUNC(10)
116 APPLY_BDOF_FUNC(12)
117 
118 void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
119 {
120  int cpu_flags = av_get_cpu_flags();
121  if (!have_neon(cpu_flags))
122  return;
123 
124  if (bd == 8) {
125  c->inter.put[0][1][0][0] = ff_vvc_put_pel_pixels4_8_neon;
126  c->inter.put[0][2][0][0] = ff_vvc_put_pel_pixels8_8_neon;
127  c->inter.put[0][3][0][0] = ff_vvc_put_pel_pixels16_8_neon;
128  c->inter.put[0][4][0][0] = ff_vvc_put_pel_pixels32_8_neon;
129  c->inter.put[0][5][0][0] = ff_vvc_put_pel_pixels64_8_neon;
130  c->inter.put[0][6][0][0] = ff_vvc_put_pel_pixels128_8_neon;
131 
132  c->inter.put[0][1][0][1] = ff_vvc_put_qpel_h4_8_neon;
133  c->inter.put[0][2][0][1] = ff_vvc_put_qpel_h8_8_neon;
134  c->inter.put[0][3][0][1] = ff_vvc_put_qpel_h16_8_neon;
135  c->inter.put[0][4][0][1] =
136  c->inter.put[0][5][0][1] =
137  c->inter.put[0][6][0][1] = ff_vvc_put_qpel_h32_8_neon;
138 
139  c->inter.put[0][1][1][0] = ff_vvc_put_qpel_v4_8_neon;
140  c->inter.put[0][2][1][0] =
141  c->inter.put[0][3][1][0] =
142  c->inter.put[0][4][1][0] =
143  c->inter.put[0][5][1][0] =
144  c->inter.put[0][6][1][0] = ff_vvc_put_qpel_v8_8_neon;
145 
146  c->inter.put[0][1][1][1] = ff_vvc_put_qpel_hv4_8_neon;
147  c->inter.put[0][2][1][1] = ff_vvc_put_qpel_hv8_8_neon;
148  c->inter.put[0][3][1][1] = ff_vvc_put_qpel_hv16_8_neon;
149  c->inter.put[0][4][1][1] = ff_vvc_put_qpel_hv32_8_neon;
150  c->inter.put[0][5][1][1] = ff_vvc_put_qpel_hv64_8_neon;
151  c->inter.put[0][6][1][1] = ff_vvc_put_qpel_hv128_8_neon;
152 
153  c->inter.put[1][1][0][0] = ff_vvc_put_pel_pixels4_8_neon;
154  c->inter.put[1][2][0][0] = ff_vvc_put_pel_pixels8_8_neon;
155  c->inter.put[1][3][0][0] = ff_vvc_put_pel_pixels16_8_neon;
156  c->inter.put[1][4][0][0] = ff_vvc_put_pel_pixels32_8_neon;
157  c->inter.put[1][5][0][0] = ff_vvc_put_pel_pixels64_8_neon;
158  c->inter.put[1][6][0][0] = ff_vvc_put_pel_pixels128_8_neon;
159 
160  c->inter.put[1][1][0][1] = ff_vvc_put_epel_h4_8_neon;
161  c->inter.put[1][2][0][1] = ff_vvc_put_epel_h8_8_neon;
162  c->inter.put[1][3][0][1] = ff_vvc_put_epel_h16_8_neon;
163  c->inter.put[1][4][0][1] =
164  c->inter.put[1][5][0][1] =
165  c->inter.put[1][6][0][1] = ff_vvc_put_epel_h32_8_neon;
166 
167  c->inter.put[1][1][1][1] = ff_vvc_put_epel_hv4_8_neon;
168  c->inter.put[1][2][1][1] = ff_vvc_put_epel_hv8_8_neon;
169  c->inter.put[1][3][1][1] = ff_vvc_put_epel_hv16_8_neon;
170  c->inter.put[1][4][1][1] = ff_vvc_put_epel_hv32_8_neon;
171  c->inter.put[1][5][1][1] = ff_vvc_put_epel_hv64_8_neon;
172  c->inter.put[1][6][1][1] = ff_vvc_put_epel_hv128_8_neon;
173 
174  c->inter.put_uni[0][1][0][0] = ff_vvc_put_pel_uni_pixels4_8_neon;
175  c->inter.put_uni[0][2][0][0] = ff_vvc_put_pel_uni_pixels8_8_neon;
176  c->inter.put_uni[0][3][0][0] = ff_vvc_put_pel_uni_pixels16_8_neon;
177  c->inter.put_uni[0][4][0][0] = ff_vvc_put_pel_uni_pixels32_8_neon;
178  c->inter.put_uni[0][5][0][0] = ff_vvc_put_pel_uni_pixels64_8_neon;
179  c->inter.put_uni[0][6][0][0] = ff_vvc_put_pel_uni_pixels128_8_neon;
180 
181  c->inter.put_uni[0][1][0][1] = ff_vvc_put_qpel_uni_h4_8_neon;
182  c->inter.put_uni[0][2][0][1] = ff_vvc_put_qpel_uni_h8_8_neon;
183  c->inter.put_uni[0][3][0][1] = ff_vvc_put_qpel_uni_h16_8_neon;
184  c->inter.put_uni[0][4][0][1] =
185  c->inter.put_uni[0][5][0][1] =
186  c->inter.put_uni[0][6][0][1] = ff_vvc_put_qpel_uni_h32_8_neon;
187 
188  c->inter.put_uni_w[0][1][0][0] = ff_vvc_put_pel_uni_w_pixels4_8_neon;
189  c->inter.put_uni_w[0][2][0][0] = ff_vvc_put_pel_uni_w_pixels8_8_neon;
190  c->inter.put_uni_w[0][3][0][0] = ff_vvc_put_pel_uni_w_pixels16_8_neon;
191  c->inter.put_uni_w[0][4][0][0] = ff_vvc_put_pel_uni_w_pixels32_8_neon;
192  c->inter.put_uni_w[0][5][0][0] = ff_vvc_put_pel_uni_w_pixels64_8_neon;
193  c->inter.put_uni_w[0][6][0][0] = ff_vvc_put_pel_uni_w_pixels128_8_neon;
194 
195  c->inter.avg = ff_vvc_avg_8_neon;
196  c->inter.w_avg = vvc_w_avg_8;
197  c->inter.dmvr[0][0] = ff_vvc_dmvr_8_neon;
198  c->inter.dmvr[0][1] = ff_vvc_dmvr_h_8_neon;
199  c->inter.dmvr[1][0] = ff_vvc_dmvr_v_8_neon;
200  c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_8_neon;
201  c->inter.apply_bdof = ff_vvc_apply_bdof_8_neon;
202 
203  c->sao.band_filter[0] = ff_h26x_sao_band_filter_8x8_8_neon;
204  for (int i = 1; i < FF_ARRAY_ELEMS(c->sao.band_filter); i++)
205  c->sao.band_filter[i] = ff_h26x_sao_band_filter_16x16_8_neon;
206  c->sao.edge_filter[0] = ff_vvc_sao_edge_filter_8x8_8_neon;
207  for (int i = 1; i < FF_ARRAY_ELEMS(c->sao.edge_filter); i++)
208  c->sao.edge_filter[i] = ff_vvc_sao_edge_filter_16x16_8_neon;
209  c->alf.filter[LUMA] = alf_filter_luma_8_neon;
210  c->alf.filter[CHROMA] = alf_filter_chroma_8_neon;
211  c->alf.classify = alf_classify_8_neon;
212 
213  if (have_i8mm(cpu_flags)) {
214  c->inter.put[0][1][0][1] = ff_vvc_put_qpel_h4_8_neon_i8mm;
215  c->inter.put[0][2][0][1] = ff_vvc_put_qpel_h8_8_neon_i8mm;
216  c->inter.put[0][3][0][1] = ff_vvc_put_qpel_h16_8_neon_i8mm;
217  c->inter.put[0][4][0][1] = ff_vvc_put_qpel_h32_8_neon_i8mm;
218  c->inter.put[0][5][0][1] = ff_vvc_put_qpel_h64_8_neon_i8mm;
219  c->inter.put[0][6][0][1] = ff_vvc_put_qpel_h128_8_neon_i8mm;
220 
221  c->inter.put[0][1][1][1] = ff_vvc_put_qpel_hv4_8_neon_i8mm;
222  c->inter.put[0][2][1][1] = ff_vvc_put_qpel_hv8_8_neon_i8mm;
223  c->inter.put[0][3][1][1] = ff_vvc_put_qpel_hv16_8_neon_i8mm;
224  c->inter.put[0][4][1][1] = ff_vvc_put_qpel_hv32_8_neon_i8mm;
225  c->inter.put[0][5][1][1] = ff_vvc_put_qpel_hv64_8_neon_i8mm;
226  c->inter.put[0][6][1][1] = ff_vvc_put_qpel_hv128_8_neon_i8mm;
227 
228  c->inter.put[1][1][0][1] = ff_vvc_put_epel_h4_8_neon_i8mm;
229  c->inter.put[1][2][0][1] = ff_vvc_put_epel_h8_8_neon_i8mm;
230  c->inter.put[1][3][0][1] = ff_vvc_put_epel_h16_8_neon_i8mm;
231  c->inter.put[1][4][0][1] = ff_vvc_put_epel_h32_8_neon_i8mm;
232  c->inter.put[1][5][0][1] = ff_vvc_put_epel_h64_8_neon_i8mm;
233  c->inter.put[1][6][0][1] = ff_vvc_put_epel_h128_8_neon_i8mm;
234 
235  c->inter.put[1][1][1][1] = ff_vvc_put_epel_hv4_8_neon_i8mm;
236  c->inter.put[1][2][1][1] = ff_vvc_put_epel_hv8_8_neon_i8mm;
237  c->inter.put[1][3][1][1] = ff_vvc_put_epel_hv16_8_neon_i8mm;
238  c->inter.put[1][4][1][1] = ff_vvc_put_epel_hv32_8_neon_i8mm;
239  c->inter.put[1][5][1][1] = ff_vvc_put_epel_hv64_8_neon_i8mm;
240  c->inter.put[1][6][1][1] = ff_vvc_put_epel_hv128_8_neon_i8mm;
241  }
242  } else if (bd == 10) {
243  c->inter.avg = ff_vvc_avg_10_neon;
244  c->inter.w_avg = vvc_w_avg_10;
245  c->inter.dmvr[0][1] = ff_vvc_dmvr_h_10_neon;
246  c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_10_neon;
247  c->inter.apply_bdof = ff_vvc_apply_bdof_10_neon;
248 
249  c->alf.filter[LUMA] = alf_filter_luma_10_neon;
250  c->alf.filter[CHROMA] = alf_filter_chroma_10_neon;
251  c->alf.classify = alf_classify_10_neon;
252  } else if (bd == 12) {
253  c->inter.avg = ff_vvc_avg_12_neon;
254  c->inter.w_avg = vvc_w_avg_12;
255  c->inter.dmvr[0][0] = ff_vvc_dmvr_12_neon;
256  c->inter.dmvr[0][1] = ff_vvc_dmvr_h_12_neon;
257  c->inter.dmvr[1][1] = ff_vvc_dmvr_hv_12_neon;
258  c->inter.apply_bdof = ff_vvc_apply_bdof_12_neon;
259 
260  c->alf.filter[LUMA] = alf_filter_luma_12_neon;
261  c->alf.filter[CHROMA] = alf_filter_chroma_12_neon;
262  c->alf.classify = alf_classify_12_neon;
263  }
264 
265  c->inter.sad = ff_vvc_sad_neon;
266 }
_dst
uint8_t * _dst
Definition: dsp.h:56
LUMA
#define LUMA
Definition: filter.c:31
src1
const pixel * src1
Definition: h264pred_template.c:420
ff_vvc_put_qpel_v8_8_neon
void ff_vvc_put_qpel_v8_8_neon(int16_t *dst, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width)
av_get_cpu_flags
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:109
ff_vvc_dsp_init_aarch64
void ff_vvc_dsp_init_aarch64(VVCDSPContext *const c, const int bd)
Definition: dsp_init.c:118
cpu_flags
static atomic_int cpu_flags
Definition: cpu.c:56
dsp.h
W_AVG_FUN
#define W_AVG_FUN(bit_depth)
Definition: dsp_init.c:78
ff_vvc_w_avg_10_neon
void ff_vvc_w_avg_10_neon(uint8_t *_dst, ptrdiff_t _dst_stride, const int16_t *src0, const int16_t *src1, int width, int height, uintptr_t w0_w1, uintptr_t offset_shift)
ff_vvc_put_qpel_v4_8_neon
void ff_vvc_put_qpel_v4_8_neon(int16_t *dst, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width)
FF_ARRAY_ELEMS
#define FF_ARRAY_ELEMS(a)
Definition: sinewin_tablegen.c:29
dsp.h
DMVR_FUN
#define DMVR_FUN(fn, bd)
Definition: dsp_init.c:94
ff_h26x_sao_band_filter_16x16_8_neon
void ff_h26x_sao_band_filter_16x16_8_neon(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, ptrdiff_t stride_src, const int16_t *sao_offset_val, int sao_left_class, int width, int height)
if
if(ret)
Definition: filter_design.txt:179
grad
static double grad(int hash, double x, double y, double z)
Definition: perlin.c:42
ff_vvc_sao_edge_filter_8x8_8_neon
void ff_vvc_sao_edge_filter_8x8_8_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride_dst, const int16_t *sao_offset_val, int eo, int width, int height)
c
Undefined Behavior In the C some operations are like signed integer dereferencing freed accessing outside allocated Undefined Behavior must not occur in a C it is not safe even if the output of undefined operations is unused The unsafety may seem nit picking but Optimizing compilers have in fact optimized code on the assumption that no undefined Behavior occurs Optimizing code based on wrong assumptions can and has in some cases lead to effects beyond the output of computations The signed integer overflow problem in speed critical code Code which is highly optimized and works with signed integers sometimes has the problem that often the output of the computation does not c
Definition: undefined.txt:32
have_i8mm
#define have_i8mm(flags)
Definition: cpu.h:29
height
#define height
Definition: dsp.h:89
dst
uint8_t ptrdiff_t const uint8_t ptrdiff_t int intptr_t intptr_t int int16_t * dst
Definition: dsp.h:87
cpu.h
for
for(k=2;k<=8;++k)
Definition: h264pred_template.c:424
ff_vvc_avg_8_neon
void ff_vvc_avg_8_neon(uint8_t *dst, ptrdiff_t dst_stride, const int16_t *src0, const int16_t *src1, int width, int height)
have_neon
#define have_neon(flags)
Definition: cpu.h:26
ff_h26x_sao_band_filter_8x8_8_neon
void ff_h26x_sao_band_filter_8x8_8_neon(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, ptrdiff_t stride_src, const int16_t *sao_offset_val, int sao_left_class, int width, int height)
ff_vvc_avg_10_neon
void ff_vvc_avg_10_neon(uint8_t *dst, ptrdiff_t dst_stride, const int16_t *src0, const int16_t *src1, int width, int height)
ff_alf_classify_sum_neon
void ff_alf_classify_sum_neon(int *sum0, int *sum1, int16_t *grad, uint32_t gshift, uint32_t steps)
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:256
CHROMA
@ CHROMA
Definition: vf_waveform.c:49
steps
static const int16_t steps[16]
Definition: misc4.c:30
ff_vvc_w_avg_8_neon
void ff_vvc_w_avg_8_neon(uint8_t *_dst, ptrdiff_t _dst_stride, const int16_t *src0, const int16_t *src1, int width, int height, uintptr_t w0_w1, uintptr_t offset_shift)
ff_vvc_w_avg_12_neon
void ff_vvc_w_avg_12_neon(uint8_t *_dst, ptrdiff_t _dst_stride, const int16_t *src0, const int16_t *src1, int width, int height, uintptr_t w0_w1, uintptr_t offset_shift)
src0
const pixel *const src0
Definition: h264pred_template.c:419
ff_vvc_avg_12_neon
void ff_vvc_avg_12_neon(uint8_t *dst, ptrdiff_t dst_stride, const int16_t *src0, const int16_t *src1, int width, int height)
ff_vvc_sao_edge_filter_16x16_8_neon
void ff_vvc_sao_edge_filter_16x16_8_neon(uint8_t *dst, const uint8_t *src, ptrdiff_t stride_dst, const int16_t *sao_offset_val, int eo, int width, int height)
alf_template.c
ff_vvc_sad_neon
int ff_vvc_sad_neon(const int16_t *src0, const int16_t *src1, int dx, int dy, const int block_w, const int block_h)
APPLY_BDOF_FUNC
#define APPLY_BDOF_FUNC(bd)
Definition: dsp_init.c:109
ctu.h
width
#define width
Definition: dsp.h:89
cpu.h
dec.h
VVCDSPContext
Definition: dsp.h:170
Generated on Tue Oct 7 2025 19:21:39 for FFmpeg by   doxygen 1.8.17