Add support for VDPAU video out, including hardware decoding.
[mplayer/glamo.git] / libmpcodecs / vf_uspp.c
bloba5882ffeca4f20ea4cd28e6667cd91332b4084ba
1 /*
2 * Copyright (C) 2005 Michael Niedermayer <michaelni@gmx.at>
4 * This file is part of MPlayer.
6 * MPlayer 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 * MPlayer 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 MPlayer; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
21 #include <stdio.h>
22 #include <stdlib.h>
23 #include <string.h>
24 #include <inttypes.h>
25 #include <math.h>
26 #include <assert.h>
28 #include "config.h"
30 #include "mp_msg.h"
31 #include "cpudetect.h"
33 #include "libavcodec/avcodec.h"
35 #ifdef HAVE_MALLOC_H
36 #include <malloc.h>
37 #endif
39 #include "img_format.h"
40 #include "mp_image.h"
41 #include "vf.h"
42 #include "libvo/fastmemcpy.h"
44 #define XMIN(a,b) ((a) < (b) ? (a) : (b))
46 #define BLOCK 16
48 //===========================================================================//
49 static const uint8_t __attribute__((aligned(8))) dither[8][8]={
50 { 0*4, 48*4, 12*4, 60*4, 3*4, 51*4, 15*4, 63*4, },
51 { 32*4, 16*4, 44*4, 28*4, 35*4, 19*4, 47*4, 31*4, },
52 { 8*4, 56*4, 4*4, 52*4, 11*4, 59*4, 7*4, 55*4, },
53 { 40*4, 24*4, 36*4, 20*4, 43*4, 27*4, 39*4, 23*4, },
54 { 2*4, 50*4, 14*4, 62*4, 1*4, 49*4, 13*4, 61*4, },
55 { 34*4, 18*4, 46*4, 30*4, 33*4, 17*4, 45*4, 29*4, },
56 { 10*4, 58*4, 6*4, 54*4, 9*4, 57*4, 5*4, 53*4, },
57 { 42*4, 26*4, 38*4, 22*4, 41*4, 25*4, 37*4, 21*4, },
60 static const uint8_t offset[511][2]= {
61 { 0, 0},
62 { 0, 0}, { 8, 8},
63 { 0, 0}, { 4, 4}, {12, 8}, { 8,12},
64 { 0, 0}, {10, 2}, { 4, 4}, {14, 6}, { 8, 8}, { 2,10}, {12,12}, { 6,14},
66 { 0, 0}, {10, 2}, { 4, 4}, {14, 6}, { 8, 8}, { 2,10}, {12,12}, { 6,14},
67 { 5, 1}, {15, 3}, { 9, 5}, { 3, 7}, {13, 9}, { 7,11}, { 1,13}, {11,15},
69 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 5, 1}, {13, 1}, { 5, 9}, {13, 9},
70 { 2, 2}, {10, 2}, { 2,10}, {10,10}, { 7, 3}, {15, 3}, { 7,11}, {15,11},
71 { 4, 4}, {12, 4}, { 4,12}, {12,12}, { 1, 5}, { 9, 5}, { 1,13}, { 9,13},
72 { 6, 6}, {14, 6}, { 6,14}, {14,14}, { 3, 7}, {11, 7}, { 3,15}, {11,15},
74 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 4, 0}, {12, 0}, { 4, 8}, {12, 8},
75 { 1, 1}, { 9, 1}, { 1, 9}, { 9, 9}, { 5, 1}, {13, 1}, { 5, 9}, {13, 9},
76 { 3, 2}, {11, 2}, { 3,10}, {11,10}, { 7, 2}, {15, 2}, { 7,10}, {15,10},
77 { 2, 3}, {10, 3}, { 2,11}, {10,11}, { 6, 3}, {14, 3}, { 6,11}, {14,11},
78 { 0, 4}, { 8, 4}, { 0,12}, { 8,12}, { 4, 4}, {12, 4}, { 4,12}, {12,12},
79 { 1, 5}, { 9, 5}, { 1,13}, { 9,13}, { 5, 5}, {13, 5}, { 5,13}, {13,13},
80 { 3, 6}, {11, 6}, { 3,14}, {11,14}, { 7, 6}, {15, 6}, { 7,14}, {15,14},
81 { 2, 7}, {10, 7}, { 2,15}, {10,15}, { 6, 7}, {14, 7}, { 6,15}, {14,15},
83 { 0, 0}, { 8, 0}, { 0, 8}, { 8, 8}, { 0, 2}, { 8, 2}, { 0,10}, { 8,10},
84 { 0, 4}, { 8, 4}, { 0,12}, { 8,12}, { 0, 6}, { 8, 6}, { 0,14}, { 8,14},
85 { 1, 1}, { 9, 1}, { 1, 9}, { 9, 9}, { 1, 3}, { 9, 3}, { 1,11}, { 9,11},
86 { 1, 5}, { 9, 5}, { 1,13}, { 9,13}, { 1, 7}, { 9, 7}, { 1,15}, { 9,15},
87 { 2, 0}, {10, 0}, { 2, 8}, {10, 8}, { 2, 2}, {10, 2}, { 2,10}, {10,10},
88 { 2, 4}, {10, 4}, { 2,12}, {10,12}, { 2, 6}, {10, 6}, { 2,14}, {10,14},
89 { 3, 1}, {11, 1}, { 3, 9}, {11, 9}, { 3, 3}, {11, 3}, { 3,11}, {11,11},
90 { 3, 5}, {11, 5}, { 3,13}, {11,13}, { 3, 7}, {11, 7}, { 3,15}, {11,15},
91 { 4, 0}, {12, 0}, { 4, 8}, {12, 8}, { 4, 2}, {12, 2}, { 4,10}, {12,10},
92 { 4, 4}, {12, 4}, { 4,12}, {12,12}, { 4, 6}, {12, 6}, { 4,14}, {12,14},
93 { 5, 1}, {13, 1}, { 5, 9}, {13, 9}, { 5, 3}, {13, 3}, { 5,11}, {13,11},
94 { 5, 5}, {13, 5}, { 5,13}, {13,13}, { 5, 7}, {13, 7}, { 5,15}, {13,15},
95 { 6, 0}, {14, 0}, { 6, 8}, {14, 8}, { 6, 2}, {14, 2}, { 6,10}, {14,10},
96 { 6, 4}, {14, 4}, { 6,12}, {14,12}, { 6, 6}, {14, 6}, { 6,14}, {14,14},
97 { 7, 1}, {15, 1}, { 7, 9}, {15, 9}, { 7, 3}, {15, 3}, { 7,11}, {15,11},
98 { 7, 5}, {15, 5}, { 7,13}, {15,13}, { 7, 7}, {15, 7}, { 7,15}, {15,15},
100 { 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},
103 struct vf_priv_s {
104 int log2_count;
105 int qp;
106 int mode;
107 int mpeg2;
108 int temp_stride[3];
109 uint8_t *src[3];
110 int16_t *temp[3];
111 int outbuf_size;
112 uint8_t *outbuf;
113 AVCodecContext *avctx_enc[BLOCK*BLOCK];
114 AVFrame *frame;
115 AVFrame *frame_dec;
118 static void store_slice_c(uint8_t *dst, int16_t *src, int dst_stride, int src_stride, int width, int height, int log2_scale){
119 int y, x;
121 #define STORE(pos) \
122 temp= ((src[x + y*src_stride + pos]<<log2_scale) + d[pos])>>8;\
123 if(temp & 0x100) temp= ~(temp>>31);\
124 dst[x + y*dst_stride + pos]= temp;
126 for(y=0; y<height; y++){
127 const uint8_t *d= dither[y&7];
128 for(x=0; x<width; x+=8){
129 int temp;
130 STORE(0);
131 STORE(1);
132 STORE(2);
133 STORE(3);
134 STORE(4);
135 STORE(5);
136 STORE(6);
137 STORE(7);
142 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){
143 int x, y, i, j;
144 const int count= 1<<p->log2_count;
146 for(i=0; i<3; i++){
147 int is_chroma= !!i;
148 int w= width >>is_chroma;
149 int h= height>>is_chroma;
150 int stride= p->temp_stride[i];
151 int block= BLOCK>>is_chroma;
153 if (!src[i] || !dst[i])
154 continue; // HACK avoid crash for Y8 colourspace
155 for(y=0; y<h; y++){
156 int index= block + block*stride + y*stride;
157 fast_memcpy(p->src[i] + index, src[i] + y*src_stride[i], w);
158 for(x=0; x<block; x++){
159 p->src[i][index - x - 1]= p->src[i][index + x ];
160 p->src[i][index + w + x ]= p->src[i][index + w - x - 1];
163 for(y=0; y<block; y++){
164 fast_memcpy(p->src[i] + ( block-1-y)*stride, p->src[i] + ( y+block )*stride, stride);
165 fast_memcpy(p->src[i] + (h+block +y)*stride, p->src[i] + (h-y+block-1)*stride, stride);
168 p->frame->linesize[i]= stride;
169 memset(p->temp[i], 0, (h+2*block)*stride*sizeof(int16_t));
172 if(p->qp)
173 p->frame->quality= p->qp * FF_QP2LAMBDA;
174 else
175 p->frame->quality= (qp_store[0] * FF_QP2LAMBDA)>>p->mpeg2;
176 // init per MB qscale stuff FIXME
178 for(i=0; i<count; i++){
179 const int x1= offset[i+count-1][0];
180 const int y1= offset[i+count-1][1];
181 int offset, out_size;
182 p->frame->data[0]= p->src[0] + x1 + y1 * p->frame->linesize[0];
183 p->frame->data[1]= p->src[1] + x1/2 + y1/2 * p->frame->linesize[1];
184 p->frame->data[2]= p->src[2] + x1/2 + y1/2 * p->frame->linesize[2];
186 out_size = avcodec_encode_video(p->avctx_enc[i], p->outbuf, p->outbuf_size, p->frame);
187 p->frame_dec = p->avctx_enc[i]->coded_frame;
189 offset= (BLOCK-x1) + (BLOCK-y1)*p->frame_dec->linesize[0];
190 //FIXME optimize
191 for(y=0; y<height; y++){
192 for(x=0; x<width; x++){
193 p->temp[0][ x + y*p->temp_stride[0] ] += p->frame_dec->data[0][ x + y*p->frame_dec->linesize[0] + offset ];
196 offset= (BLOCK/2-x1/2) + (BLOCK/2-y1/2)*p->frame_dec->linesize[1];
197 for(y=0; y<height/2; y++){
198 for(x=0; x<width/2; x++){
199 p->temp[1][ x + y*p->temp_stride[1] ] += p->frame_dec->data[1][ x + y*p->frame_dec->linesize[1] + offset ];
200 p->temp[2][ x + y*p->temp_stride[2] ] += p->frame_dec->data[2][ x + y*p->frame_dec->linesize[2] + offset ];
205 for(j=0; j<3; j++){
206 int is_chroma= !!j;
207 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);
211 static int config(struct vf_instance_s* vf,
212 int width, int height, int d_width, int d_height,
213 unsigned int flags, unsigned int outfmt){
214 int i;
215 AVCodec *enc= avcodec_find_encoder(CODEC_ID_SNOW);
217 for(i=0; i<3; i++){
218 int is_chroma= !!i;
219 int w= ((width + 4*BLOCK-1) & (~(2*BLOCK-1)))>>is_chroma;
220 int h= ((height + 4*BLOCK-1) & (~(2*BLOCK-1)))>>is_chroma;
222 vf->priv->temp_stride[i]= w;
223 vf->priv->temp[i]= malloc(vf->priv->temp_stride[i]*h*sizeof(int16_t));
224 vf->priv->src [i]= malloc(vf->priv->temp_stride[i]*h*sizeof(uint8_t));
226 for(i=0; i< (1<<vf->priv->log2_count); i++){
227 AVCodecContext *avctx_enc;
229 avctx_enc=
230 vf->priv->avctx_enc[i]= avcodec_alloc_context();
231 avctx_enc->width = width + BLOCK;
232 avctx_enc->height = height + BLOCK;
233 avctx_enc->time_base= (AVRational){1,25}; // meaningless
234 avctx_enc->gop_size = 300;
235 avctx_enc->max_b_frames= 0;
236 avctx_enc->pix_fmt = PIX_FMT_YUV420P;
237 avctx_enc->flags = CODEC_FLAG_QSCALE | CODEC_FLAG_LOW_DELAY;
238 avctx_enc->strict_std_compliance = FF_COMPLIANCE_EXPERIMENTAL;
239 avctx_enc->global_quality= 123;
240 avcodec_open(avctx_enc, enc);
241 assert(avctx_enc->codec);
243 vf->priv->frame= avcodec_alloc_frame();
244 vf->priv->frame_dec= avcodec_alloc_frame();
246 vf->priv->outbuf_size= (width + BLOCK)*(height + BLOCK)*10;
247 vf->priv->outbuf= malloc(vf->priv->outbuf_size);
249 return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
252 static void get_image(struct vf_instance_s* vf, mp_image_t *mpi){
253 if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
254 // ok, we can do pp in-place (or pp disabled):
255 vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
256 mpi->type, mpi->flags | MP_IMGFLAG_READABLE, mpi->width, mpi->height);
257 mpi->planes[0]=vf->dmpi->planes[0];
258 mpi->stride[0]=vf->dmpi->stride[0];
259 mpi->width=vf->dmpi->width;
260 if(mpi->flags&MP_IMGFLAG_PLANAR){
261 mpi->planes[1]=vf->dmpi->planes[1];
262 mpi->planes[2]=vf->dmpi->planes[2];
263 mpi->stride[1]=vf->dmpi->stride[1];
264 mpi->stride[2]=vf->dmpi->stride[2];
266 mpi->flags|=MP_IMGFLAG_DIRECT;
269 static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts){
270 mp_image_t *dmpi;
272 if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
273 // no DR, so get a new image! hope we'll get DR buffer:
274 dmpi=vf_get_image(vf->next,mpi->imgfmt,
275 MP_IMGTYPE_TEMP,
276 MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
277 mpi->width,mpi->height);
278 vf_clone_mpi_attributes(dmpi, mpi);
279 }else{
280 dmpi=vf->dmpi;
283 vf->priv->mpeg2= mpi->qscale_type;
284 if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
285 if(mpi->qscale || vf->priv->qp){
286 filter(vf->priv, dmpi->planes, mpi->planes, dmpi->stride, mpi->stride, mpi->w, mpi->h, mpi->qscale, mpi->qstride);
287 }else{
288 memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
289 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]);
290 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]);
294 #if HAVE_MMX
295 if(gCpuCaps.hasMMX) __asm__ volatile ("emms\n\t");
296 #endif
297 #if HAVE_MMX2
298 if(gCpuCaps.hasMMX2) __asm__ volatile ("sfence\n\t");
299 #endif
301 return vf_next_put_image(vf,dmpi, pts);
304 static void uninit(struct vf_instance_s* vf){
305 int i;
306 if(!vf->priv) return;
308 for(i=0; i<3; i++){
309 if(vf->priv->temp[i]) free(vf->priv->temp[i]);
310 vf->priv->temp[i]= NULL;
311 if(vf->priv->src[i]) free(vf->priv->src[i]);
312 vf->priv->src[i]= NULL;
314 for(i=0; i<BLOCK*BLOCK; i++){
315 av_freep(&vf->priv->avctx_enc[i]);
318 free(vf->priv);
319 vf->priv=NULL;
322 //===========================================================================//
323 static int query_format(struct vf_instance_s* vf, unsigned int fmt){
324 switch(fmt){
325 case IMGFMT_YV12:
326 case IMGFMT_I420:
327 case IMGFMT_IYUV:
328 case IMGFMT_Y800:
329 case IMGFMT_Y8:
330 return vf_next_query_format(vf,fmt);
332 return 0;
335 static int control(struct vf_instance_s* vf, int request, void* data){
336 switch(request){
337 case VFCTRL_QUERY_MAX_PP_LEVEL:
338 return 8;
339 case VFCTRL_SET_PP_LEVEL:
340 vf->priv->log2_count= *((unsigned int*)data);
341 //FIXME we have to realloc a few things here
342 return CONTROL_TRUE;
344 return vf_next_control(vf,request,data);
347 static int open(vf_instance_t *vf, char* args){
349 int log2c=-1;
351 vf->config=config;
352 vf->put_image=put_image;
353 vf->get_image=get_image;
354 vf->query_format=query_format;
355 vf->uninit=uninit;
356 vf->control= control;
357 vf->priv=malloc(sizeof(struct vf_priv_s));
358 memset(vf->priv, 0, sizeof(struct vf_priv_s));
360 avcodec_init();
361 avcodec_register_all();
363 vf->priv->log2_count= 4;
365 if (args) sscanf(args, "%d:%d:%d", &log2c, &vf->priv->qp, &vf->priv->mode);
367 if( log2c >=0 && log2c <=8 )
368 vf->priv->log2_count = log2c;
370 if(vf->priv->qp < 0)
371 vf->priv->qp = 0;
373 // #if HAVE_MMX
374 // if(gCpuCaps.hasMMX){
375 // store_slice= store_slice_mmx;
376 // }
377 // #endif
379 return 1;
382 const vf_info_t vf_info_uspp = {
383 "ultra simple/slow postprocess",
384 "uspp",
385 "Michael Niedermayer",
387 open,
388 NULL