2 Copyright (C) 2005 Michael Niedermayer <michaelni@gmx.at>
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
29 #include "cpudetect.h"
31 #include "libavcodec/avcodec.h"
37 #include "img_format.h"
40 #include "libvo/fastmemcpy.h"
42 #define XMIN(a,b) ((a) < (b) ? (a) : (b))
46 //===========================================================================//
47 static const uint8_t __attribute__((aligned(8))) dither
[8][8]={
48 { 0*4, 48*4, 12*4, 60*4, 3*4, 51*4, 15*4, 63*4, },
49 { 32*4, 16*4, 44*4, 28*4, 35*4, 19*4, 47*4, 31*4, },
50 { 8*4, 56*4, 4*4, 52*4, 11*4, 59*4, 7*4, 55*4, },
51 { 40*4, 24*4, 36*4, 20*4, 43*4, 27*4, 39*4, 23*4, },
52 { 2*4, 50*4, 14*4, 62*4, 1*4, 49*4, 13*4, 61*4, },
53 { 34*4, 18*4, 46*4, 30*4, 33*4, 17*4, 45*4, 29*4, },
54 { 10*4, 58*4, 6*4, 54*4, 9*4, 57*4, 5*4, 53*4, },
55 { 42*4, 26*4, 38*4, 22*4, 41*4, 25*4, 37*4, 21*4, },
58 static const uint8_t offset
[511][2]= {
61 { 0, 0}, { 4, 4}, {12, 8}, { 8,12},
62 { 0, 0}, {10, 2}, { 4, 4}, {14, 6}, { 8, 8}, { 2,10}, {12,12}, { 6,14},
64 { 0, 0}, {10, 2}, { 4, 4}, {14, 6}, { 8, 8}, { 2,10}, {12,12}, { 6,14},
65 { 5, 1}, {15, 3}, { 9, 5}, { 3, 7}, {13, 9}, { 7,11}, { 1,13}, {11,15},
67 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 5, 1}, {13, 1}, { 5, 9}, {13, 9},
68 { 2, 2}, {10, 2}, { 2,10}, {10,10}, { 7, 3}, {15, 3}, { 7,11}, {15,11},
69 { 4, 4}, {12, 4}, { 4,12}, {12,12}, { 1, 5}, { 9, 5}, { 1,13}, { 9,13},
70 { 6, 6}, {14, 6}, { 6,14}, {14,14}, { 3, 7}, {11, 7}, { 3,15}, {11,15},
72 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 4, 0}, {12, 0}, { 4, 8}, {12, 8},
73 { 1, 1}, { 9, 1}, { 1, 9}, { 9, 9}, { 5, 1}, {13, 1}, { 5, 9}, {13, 9},
74 { 3, 2}, {11, 2}, { 3,10}, {11,10}, { 7, 2}, {15, 2}, { 7,10}, {15,10},
75 { 2, 3}, {10, 3}, { 2,11}, {10,11}, { 6, 3}, {14, 3}, { 6,11}, {14,11},
76 { 0, 4}, { 8, 4}, { 0,12}, { 8,12}, { 4, 4}, {12, 4}, { 4,12}, {12,12},
77 { 1, 5}, { 9, 5}, { 1,13}, { 9,13}, { 5, 5}, {13, 5}, { 5,13}, {13,13},
78 { 3, 6}, {11, 6}, { 3,14}, {11,14}, { 7, 6}, {15, 6}, { 7,14}, {15,14},
79 { 2, 7}, {10, 7}, { 2,15}, {10,15}, { 6, 7}, {14, 7}, { 6,15}, {14,15},
81 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 0, 2}, { 8, 2}, { 0,10}, { 8,10},
82 { 0, 4}, { 8, 4}, { 0,12}, { 8,12}, { 0, 6}, { 8, 6}, { 0,14}, { 8,14},
83 { 1, 1}, { 9, 1}, { 1, 9}, { 9, 9}, { 1, 3}, { 9, 3}, { 1,11}, { 9,11},
84 { 1, 5}, { 9, 5}, { 1,13}, { 9,13}, { 1, 7}, { 9, 7}, { 1,15}, { 9,15},
85 { 2, 0}, {10, 0}, { 2, 8}, {10, 8}, { 2, 2}, {10, 2}, { 2,10}, {10,10},
86 { 2, 4}, {10, 4}, { 2,12}, {10,12}, { 2, 6}, {10, 6}, { 2,14}, {10,14},
87 { 3, 1}, {11, 1}, { 3, 9}, {11, 9}, { 3, 3}, {11, 3}, { 3,11}, {11,11},
88 { 3, 5}, {11, 5}, { 3,13}, {11,13}, { 3, 7}, {11, 7}, { 3,15}, {11,15},
89 { 4, 0}, {12, 0}, { 4, 8}, {12, 8}, { 4, 2}, {12, 2}, { 4,10}, {12,10},
90 { 4, 4}, {12, 4}, { 4,12}, {12,12}, { 4, 6}, {12, 6}, { 4,14}, {12,14},
91 { 5, 1}, {13, 1}, { 5, 9}, {13, 9}, { 5, 3}, {13, 3}, { 5,11}, {13,11},
92 { 5, 5}, {13, 5}, { 5,13}, {13,13}, { 5, 7}, {13, 7}, { 5,15}, {13,15},
93 { 6, 0}, {14, 0}, { 6, 8}, {14, 8}, { 6, 2}, {14, 2}, { 6,10}, {14,10},
94 { 6, 4}, {14, 4}, { 6,12}, {14,12}, { 6, 6}, {14, 6}, { 6,14}, {14,14},
95 { 7, 1}, {15, 1}, { 7, 9}, {15, 9}, { 7, 3}, {15, 3}, { 7,11}, {15,11},
96 { 7, 5}, {15, 5}, { 7,13}, {15,13}, { 7, 7}, {15, 7}, { 7,15}, {15,15},
98 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 4, 4}, {12, 4}, { 4,12}, {12,12}, { 0, 4}, { 8, 4}, { 0,12}, { 8,12}, { 4, 0}, {12, 0}, { 4, 8}, {12, 8}, { 2, 2}, {10, 2}, { 2,10}, {10,10}, { 6, 6}, {14, 6}, { 6,14}, {14,14}, { 2, 6}, {10, 6}, { 2,14}, {10,14}, { 6, 2}, {14, 2}, { 6,10}, {14,10}, { 0, 2}, { 8, 2}, { 0,10}, { 8,10}, { 4, 6}, {12, 6}, { 4,14}, {12,14}, { 0, 6}, { 8, 6}, { 0,14}, { 8,14}, { 4, 2}, {12, 2}, { 4,10}, {12,10}, { 2, 0}, {10, 0}, { 2, 8}, {10, 8}, { 6, 4}, {14, 4}, { 6,12}, {14,12}, { 2, 4}, {10, 4}, { 2,12}, {10,12}, { 6, 0}, {14, 0}, { 6, 8}, {14, 8}, { 1, 1}, { 9, 1}, { 1, 9}, { 9, 9}, { 5, 5}, {13, 5}, { 5,13}, {13,13}, { 1, 5}, { 9, 5}, { 1,13}, { 9,13}, { 5, 1}, {13, 1}, { 5, 9}, {13, 9}, { 3, 3}, {11, 3}, { 3,11}, {11,11}, { 7, 7}, {15, 7}, { 7,15}, {15,15}, { 3, 7}, {11, 7}, { 3,15}, {11,15}, { 7, 3}, {15, 3}, { 7,11}, {15,11}, { 1, 3}, { 9, 3}, { 1,11}, { 9,11}, { 5, 7}, {13, 7}, { 5,15}, {13,15}, { 1, 7}, { 9, 7}, { 1,15}, { 9,15}, { 5, 3}, {13, 3}, { 5,11}, {13,11}, { 3, 1}, {11, 1}, { 3, 9}, {11, 9}, { 7, 5}, {15, 5}, { 7,13}, {15,13}, { 3, 5}, {11, 5}, { 3,13}, {11,13}, { 7, 1}, {15, 1}, { 7, 9}, {15, 9}, { 0, 1}, { 8, 1}, { 0, 9}, { 8, 9}, { 4, 5}, {12, 5}, { 4,13}, {12,13}, { 0, 5}, { 8, 5}, { 0,13}, { 8,13}, { 4, 1}, {12, 1}, { 4, 9}, {12, 9}, { 2, 3}, {10, 3}, { 2,11}, {10,11}, { 6, 7}, {14, 7}, { 6,15}, {14,15}, { 2, 7}, {10, 7}, { 2,15}, {10,15}, { 6, 3}, {14, 3}, { 6,11}, {14,11}, { 0, 3}, { 8, 3}, { 0,11}, { 8,11}, { 4, 7}, {12, 7}, { 4,15}, {12,15}, { 0, 7}, { 8, 7}, { 0,15}, { 8,15}, { 4, 3}, {12, 3}, { 4,11}, {12,11}, { 2, 1}, {10, 1}, { 2, 9}, {10, 9}, { 6, 5}, {14, 5}, { 6,13}, {14,13}, { 2, 5}, {10, 5}, { 2,13}, {10,13}, { 6, 1}, {14, 1}, { 6, 9}, {14, 9}, { 1, 0}, { 9, 0}, { 1, 8}, { 9, 8}, { 5, 4}, {13, 4}, { 5,12}, {13,12}, { 1, 4}, { 9, 4}, { 1,12}, { 9,12}, { 5, 0}, {13, 0}, { 5, 8}, {13, 8}, { 3, 2}, {11, 2}, { 3,10}, {11,10}, { 7, 6}, {15, 6}, { 7,14}, {15,14}, { 3, 6}, {11, 6}, { 3,14}, {11,14}, { 7, 2}, {15, 2}, { 7,10}, {15,10}, { 1, 2}, { 9, 2}, { 1,10}, { 9,10}, { 5, 6}, {13, 6}, { 5,14}, {13,14}, { 1, 6}, { 9, 6}, { 1,14}, { 9,14}, { 5, 2}, {13, 2}, { 5,10}, {13,10}, { 3, 0}, {11, 0}, { 3, 8}, {11, 8}, { 7, 4}, {15, 4}, { 7,12}, {15,12}, { 3, 4}, {11, 4}, { 3,12}, {11,12}, { 7, 0}, {15, 0}, { 7, 8}, {15, 8},
111 AVCodecContext
*avctx_enc
[BLOCK
*BLOCK
];
116 static void store_slice_c(uint8_t *dst
, int16_t *src
, int dst_stride
, int src_stride
, int width
, int height
, int log2_scale
){
120 temp= ((src[x + y*src_stride + pos]<<log2_scale) + d[pos])>>8;\
121 if(temp & 0x100) temp= ~(temp>>31);\
122 dst[x + y*dst_stride + pos]= temp;
124 for(y
=0; y
<height
; y
++){
125 const uint8_t *d
= dither
[y
&7];
126 for(x
=0; x
<width
; x
+=8){
140 static void filter(struct vf_priv_s
*p
, uint8_t *dst
[3], uint8_t *src
[3], int dst_stride
[3], int src_stride
[3], int width
, int height
, uint8_t *qp_store
, int qp_stride
){
142 const int count
= 1<<p
->log2_count
;
146 int w
= width
>>is_chroma
;
147 int h
= height
>>is_chroma
;
148 int stride
= p
->temp_stride
[i
];
149 int block
= BLOCK
>>is_chroma
;
151 if (!src
[i
] || !dst
[i
])
152 continue; // HACK avoid crash for Y8 colourspace
154 int index
= block
+ block
*stride
+ y
*stride
;
155 fast_memcpy(p
->src
[i
] + index
, src
[i
] + y
*src_stride
[i
], w
);
156 for(x
=0; x
<block
; x
++){
157 p
->src
[i
][index
- x
- 1]= p
->src
[i
][index
+ x
];
158 p
->src
[i
][index
+ w
+ x
]= p
->src
[i
][index
+ w
- x
- 1];
161 for(y
=0; y
<block
; y
++){
162 fast_memcpy(p
->src
[i
] + ( block
-1-y
)*stride
, p
->src
[i
] + ( y
+block
)*stride
, stride
);
163 fast_memcpy(p
->src
[i
] + (h
+block
+y
)*stride
, p
->src
[i
] + (h
-y
+block
-1)*stride
, stride
);
166 p
->frame
->linesize
[i
]= stride
;
167 memset(p
->temp
[i
], 0, (h
+2*block
)*stride
*sizeof(int16_t));
171 p
->frame
->quality
= p
->qp
* FF_QP2LAMBDA
;
173 p
->frame
->quality
= (qp_store
[0] * FF_QP2LAMBDA
)>>p
->mpeg2
;
174 // init per MB qscale stuff FIXME
176 for(i
=0; i
<count
; i
++){
177 const int x1
= offset
[i
+count
-1][0];
178 const int y1
= offset
[i
+count
-1][1];
179 int offset
, out_size
;
180 p
->frame
->data
[0]= p
->src
[0] + x1
+ y1
* p
->frame
->linesize
[0];
181 p
->frame
->data
[1]= p
->src
[1] + x1
/2 + y1
/2 * p
->frame
->linesize
[1];
182 p
->frame
->data
[2]= p
->src
[2] + x1
/2 + y1
/2 * p
->frame
->linesize
[2];
184 out_size
= avcodec_encode_video(p
->avctx_enc
[i
], p
->outbuf
, p
->outbuf_size
, p
->frame
);
185 p
->frame_dec
= p
->avctx_enc
[i
]->coded_frame
;
187 offset
= (BLOCK
-x1
) + (BLOCK
-y1
)*p
->frame_dec
->linesize
[0];
189 for(y
=0; y
<height
; y
++){
190 for(x
=0; x
<width
; x
++){
191 p
->temp
[0][ x
+ y
*p
->temp_stride
[0] ] += p
->frame_dec
->data
[0][ x
+ y
*p
->frame_dec
->linesize
[0] + offset
];
194 offset
= (BLOCK
/2-x1
/2) + (BLOCK
/2-y1
/2)*p
->frame_dec
->linesize
[1];
195 for(y
=0; y
<height
/2; y
++){
196 for(x
=0; x
<width
/2; x
++){
197 p
->temp
[1][ x
+ y
*p
->temp_stride
[1] ] += p
->frame_dec
->data
[1][ x
+ y
*p
->frame_dec
->linesize
[1] + offset
];
198 p
->temp
[2][ x
+ y
*p
->temp_stride
[2] ] += p
->frame_dec
->data
[2][ x
+ y
*p
->frame_dec
->linesize
[2] + offset
];
205 store_slice_c(dst
[j
], p
->temp
[j
], dst_stride
[j
], p
->temp_stride
[j
], width
>>is_chroma
, height
>>is_chroma
, 8-p
->log2_count
);
209 static int config(struct vf_instance_s
* vf
,
210 int width
, int height
, int d_width
, int d_height
,
211 unsigned int flags
, unsigned int outfmt
){
213 AVCodec
*enc
= avcodec_find_encoder(CODEC_ID_SNOW
);
217 int w
= ((width
+ 4*BLOCK
-1) & (~(2*BLOCK
-1)))>>is_chroma
;
218 int h
= ((height
+ 4*BLOCK
-1) & (~(2*BLOCK
-1)))>>is_chroma
;
220 vf
->priv
->temp_stride
[i
]= w
;
221 vf
->priv
->temp
[i
]= malloc(vf
->priv
->temp_stride
[i
]*h
*sizeof(int16_t));
222 vf
->priv
->src
[i
]= malloc(vf
->priv
->temp_stride
[i
]*h
*sizeof(uint8_t));
224 for(i
=0; i
< (1<<vf
->priv
->log2_count
); i
++){
225 AVCodecContext
*avctx_enc
;
228 vf
->priv
->avctx_enc
[i
]= avcodec_alloc_context();
229 avctx_enc
->width
= width
+ BLOCK
;
230 avctx_enc
->height
= height
+ BLOCK
;
231 avctx_enc
->time_base
= (AVRational
){1,25}; // meaningless
232 avctx_enc
->gop_size
= 300;
233 avctx_enc
->max_b_frames
= 0;
234 avctx_enc
->pix_fmt
= PIX_FMT_YUV420P
;
235 avctx_enc
->flags
= CODEC_FLAG_QSCALE
| CODEC_FLAG_LOW_DELAY
;
236 avctx_enc
->strict_std_compliance
= FF_COMPLIANCE_EXPERIMENTAL
;
237 avctx_enc
->global_quality
= 123;
238 avcodec_open(avctx_enc
, enc
);
239 assert(avctx_enc
->codec
);
241 vf
->priv
->frame
= avcodec_alloc_frame();
242 vf
->priv
->frame_dec
= avcodec_alloc_frame();
244 vf
->priv
->outbuf_size
= (width
+ BLOCK
)*(height
+ BLOCK
)*10;
245 vf
->priv
->outbuf
= malloc(vf
->priv
->outbuf_size
);
247 return vf_next_config(vf
,width
,height
,d_width
,d_height
,flags
,outfmt
);
250 static void get_image(struct vf_instance_s
* vf
, mp_image_t
*mpi
){
251 if(mpi
->flags
&MP_IMGFLAG_PRESERVE
) return; // don't change
252 // ok, we can do pp in-place (or pp disabled):
253 vf
->dmpi
=vf_get_image(vf
->next
,mpi
->imgfmt
,
254 mpi
->type
, mpi
->flags
| MP_IMGFLAG_READABLE
, mpi
->width
, mpi
->height
);
255 mpi
->planes
[0]=vf
->dmpi
->planes
[0];
256 mpi
->stride
[0]=vf
->dmpi
->stride
[0];
257 mpi
->width
=vf
->dmpi
->width
;
258 if(mpi
->flags
&MP_IMGFLAG_PLANAR
){
259 mpi
->planes
[1]=vf
->dmpi
->planes
[1];
260 mpi
->planes
[2]=vf
->dmpi
->planes
[2];
261 mpi
->stride
[1]=vf
->dmpi
->stride
[1];
262 mpi
->stride
[2]=vf
->dmpi
->stride
[2];
264 mpi
->flags
|=MP_IMGFLAG_DIRECT
;
267 static int put_image(struct vf_instance_s
* vf
, mp_image_t
*mpi
, double pts
){
270 if(!(mpi
->flags
&MP_IMGFLAG_DIRECT
)){
271 // no DR, so get a new image! hope we'll get DR buffer:
272 dmpi
=vf_get_image(vf
->next
,mpi
->imgfmt
,
274 MP_IMGFLAG_ACCEPT_STRIDE
|MP_IMGFLAG_PREFER_ALIGNED_STRIDE
,
275 mpi
->width
,mpi
->height
);
276 vf_clone_mpi_attributes(dmpi
, mpi
);
281 vf
->priv
->mpeg2
= mpi
->qscale_type
;
282 if(vf
->priv
->log2_count
|| !(mpi
->flags
&MP_IMGFLAG_DIRECT
)){
283 if(mpi
->qscale
|| vf
->priv
->qp
){
284 filter(vf
->priv
, dmpi
->planes
, mpi
->planes
, dmpi
->stride
, mpi
->stride
, mpi
->w
, mpi
->h
, mpi
->qscale
, mpi
->qstride
);
286 memcpy_pic(dmpi
->planes
[0], mpi
->planes
[0], mpi
->w
, mpi
->h
, dmpi
->stride
[0], mpi
->stride
[0]);
287 memcpy_pic(dmpi
->planes
[1], mpi
->planes
[1], mpi
->w
>>mpi
->chroma_x_shift
, mpi
->h
>>mpi
->chroma_y_shift
, dmpi
->stride
[1], mpi
->stride
[1]);
288 memcpy_pic(dmpi
->planes
[2], mpi
->planes
[2], mpi
->w
>>mpi
->chroma_x_shift
, mpi
->h
>>mpi
->chroma_y_shift
, dmpi
->stride
[2], mpi
->stride
[2]);
293 if(gCpuCaps
.hasMMX
) asm volatile ("emms\n\t");
296 if(gCpuCaps
.hasMMX2
) asm volatile ("sfence\n\t");
299 return vf_next_put_image(vf
,dmpi
, pts
);
302 static void uninit(struct vf_instance_s
* vf
){
304 if(!vf
->priv
) return;
307 if(vf
->priv
->temp
[i
]) free(vf
->priv
->temp
[i
]);
308 vf
->priv
->temp
[i
]= NULL
;
309 if(vf
->priv
->src
[i
]) free(vf
->priv
->src
[i
]);
310 vf
->priv
->src
[i
]= NULL
;
312 for(i
=0; i
<BLOCK
*BLOCK
; i
++){
313 av_freep(&vf
->priv
->avctx_enc
[i
]);
320 //===========================================================================//
321 static int query_format(struct vf_instance_s
* vf
, unsigned int fmt
){
328 return vf_next_query_format(vf
,fmt
);
333 static int control(struct vf_instance_s
* vf
, int request
, void* data
){
335 case VFCTRL_QUERY_MAX_PP_LEVEL
:
337 case VFCTRL_SET_PP_LEVEL
:
338 vf
->priv
->log2_count
= *((unsigned int*)data
);
339 //FIXME we have to realloc a few things here
342 return vf_next_control(vf
,request
,data
);
345 static int open(vf_instance_t
*vf
, char* args
){
350 vf
->put_image
=put_image
;
351 vf
->get_image
=get_image
;
352 vf
->query_format
=query_format
;
354 vf
->control
= control
;
355 vf
->priv
=malloc(sizeof(struct vf_priv_s
));
356 memset(vf
->priv
, 0, sizeof(struct vf_priv_s
));
359 avcodec_register_all();
361 vf
->priv
->log2_count
= 4;
363 if (args
) sscanf(args
, "%d:%d:%d", &log2c
, &vf
->priv
->qp
, &vf
->priv
->mode
);
365 if( log2c
>=0 && log2c
<=8 )
366 vf
->priv
->log2_count
= log2c
;
372 // if(gCpuCaps.hasMMX){
373 // store_slice= store_slice_mmx;
380 const vf_info_t vf_info_uspp
= {
381 "ultra simple/slow postprocess",
383 "Michael Niedermayer",