From: "Tomas Härdin" <tjoppen@acc.umu.se> To: FFmpeg development discussions and patches <ffmpeg-devel@ffmpeg.org> Subject: [FFmpeg-devel] [PATCH 01/13] lavc/jpeg2000dec: Finer granularity threading Date: Tue, 14 Jun 2022 16:39:00 +0200 Message-ID: <10ec51ef44325c2de6d5de7b994a9b6c8eb5e3a2.camel@acc.umu.se> (raw) [-- Attachment #1: Type: text/plain, Size: 106 bytes --] Patch 12 in this series is optional since it's just me getting the speed up on a specific machine /Tomas [-- Attachment #2: 0001-lavc-jpeg2000dec-Finer-granularity-threading.patch --] [-- Type: text/x-patch, Size: 11603 bytes --] From 115aa26c343419e81c1b5ba0bfdb1615cbec27e9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Tomas=20H=C3=A4rdin?= <git@haerdin.se> Date: Fri, 10 Jun 2022 14:10:02 +0200 Subject: [PATCH 01/13] lavc/jpeg2000dec: Finer granularity threading Decoding and dequant is now threaded on codeblock level. IDWT is threaded on component level. MCT and write_frame() remain threaded on tile level. This brings lossless 4K J2K with -lowres 2 -thread_type slice -threads 96 on an AMD EPYC 7R32 from 4.8 fps (177% CPU) to 31 fps (1284% CPU). --- libavcodec/jpeg2000dec.c | 196 ++++++++++++++++++++++++++++----------- 1 file changed, 142 insertions(+), 54 deletions(-) diff --git a/libavcodec/jpeg2000dec.c b/libavcodec/jpeg2000dec.c index 92966b11f5..d9754fc50e 100644 --- a/libavcodec/jpeg2000dec.c +++ b/libavcodec/jpeg2000dec.c @@ -92,6 +92,15 @@ typedef struct Jpeg2000Tile { int coord[2][2]; // border coordinates {{x0, x1}, {y0, y1}} } Jpeg2000Tile; +typedef struct Jpeg2000IdwtThread { + int cb_start, cb_end; +} Jpeg2000IdwtThread; + +typedef struct Jpeg2000CodeblockThread { + int tileno, compno, reslevelno, bandno, precno, cblkno; + int coded; +} Jpeg2000CodeblockThread; + typedef struct Jpeg2000DecoderContext { AVClass *class; AVCodecContext *avctx; @@ -136,6 +145,11 @@ typedef struct Jpeg2000DecoderContext { /*options parameters*/ int reduction_factor; + + Jpeg2000IdwtThread *idwt; + unsigned int idwt_size; + Jpeg2000CodeblockThread *cb; + unsigned int cb_size; } Jpeg2000DecoderContext; /* get_bits functions for JPEG2000 packet bitstream @@ -1937,54 +1951,33 @@ static inline void roi_scale_cblk(Jpeg2000Cblk *cblk, } } -static inline void tile_codeblocks(Jpeg2000DecoderContext *s, Jpeg2000Tile *tile) +static int jpeg2000_decode_cb(AVCodecContext *avctx, void *td, + int jobnr, int threadnr) { Jpeg2000T1Context t1; - - int compno, reslevelno, bandno; - - /* Loop on tile components */ - for (compno = 0; compno < s->ncomponents; compno++) { - Jpeg2000Component *comp = tile->comp + compno; - Jpeg2000CodingStyle *codsty = tile->codsty + compno; - int coded = 0; - - t1.stride = (1<<codsty->log2_cblk_width) + 2; - - /* Loop on resolution levels */ - for (reslevelno = 0; reslevelno < codsty->nreslevels2decode; reslevelno++) { - Jpeg2000ResLevel *rlevel = comp->reslevel + reslevelno; - /* Loop on bands */ - for (bandno = 0; bandno < rlevel->nbands; bandno++) { - int nb_precincts, precno; - Jpeg2000Band *band = rlevel->band + bandno; - int cblkno = 0, bandpos; - - bandpos = bandno + (reslevelno > 0); - - if (band->coord[0][0] == band->coord[0][1] || - band->coord[1][0] == band->coord[1][1]) - continue; - - nb_precincts = rlevel->num_precincts_x * rlevel->num_precincts_y; - /* Loop on precincts */ - for (precno = 0; precno < nb_precincts; precno++) { - Jpeg2000Prec *prec = band->prec + precno; - - /* Loop on codeblocks */ - for (cblkno = 0; - cblkno < prec->nb_codeblocks_width * prec->nb_codeblocks_height; - cblkno++) { - int x, y; - Jpeg2000Cblk *cblk = prec->cblk + cblkno; - int ret = decode_cblk(s, codsty, &t1, cblk, + Jpeg2000DecoderContext *s = avctx->priv_data; + Jpeg2000CodeblockThread *cb = s->cb + jobnr; + Jpeg2000Tile *tile = s->tile + cb->tileno; + Jpeg2000Component *comp = tile->comp + cb->compno; + Jpeg2000CodingStyle *codsty = tile->codsty + cb->compno; + Jpeg2000ResLevel *rlevel = comp->reslevel + cb->reslevelno; + Jpeg2000Band *band = rlevel->band + cb->bandno; + Jpeg2000Prec *prec = band->prec + cb->precno; + Jpeg2000Cblk *cblk = prec->cblk + cb->cblkno; + int ret, x, y, bandpos = cb->bandno + (cb->reslevelno > 0); + + t1.stride = (1<<codsty->log2_cblk_width) + 2; + cb->coded = 0; + + ret = decode_cblk(s, codsty, &t1, cblk, cblk->coord[0][1] - cblk->coord[0][0], cblk->coord[1][1] - cblk->coord[1][0], bandpos, comp->roi_shift); if (ret) - coded = 1; + cb->coded = 1; else - continue; + return 0; + x = cblk->coord[0][0] - band->coord[0][0]; y = cblk->coord[1][0] - band->coord[1][0]; @@ -1996,16 +1989,28 @@ static inline void tile_codeblocks(Jpeg2000DecoderContext *s, Jpeg2000Tile *tile dequantization_int_97(x, y, cblk, comp, &t1, band); else dequantization_int(x, y, cblk, comp, &t1, band); - } /* end cblk */ - } /*end prec */ - } /* end band */ - } /* end reslevel */ - /* inverse DWT */ - if (coded) + return 0; +} + +static int jpeg2000_idwt(AVCodecContext *avctx, void *td, + int jobnr, int threadnr) +{ + Jpeg2000DecoderContext *s = avctx->priv_data; + Jpeg2000IdwtThread *idwt = s->idwt + jobnr; + Jpeg2000Tile *tile = s->tile + jobnr / s->ncomponents; + int compno = jobnr % s->ncomponents; + Jpeg2000Component *comp = tile->comp + compno; + Jpeg2000CodingStyle *codsty = tile->codsty + compno; + + for (int i = idwt->cb_start; i < idwt->cb_end; i++) { + if (s->cb[i].coded) { ff_dwt_decode(&comp->dwt, codsty->transform == FF_DWT97 ? (void*)comp->f_data : (void*)comp->i_data); + break; + } + } - } /*end comp */ + return 0; } #define WRITE_FRAME(D, PIXEL) \ @@ -2075,15 +2080,13 @@ WRITE_FRAME(16, uint16_t) #undef WRITE_FRAME -static int jpeg2000_decode_tile(AVCodecContext *avctx, void *td, - int jobnr, int threadnr) +static int jpeg2000_mct_write_frame(AVCodecContext *avctx, void *td, + int jobnr, int threadnr) { Jpeg2000DecoderContext *s = avctx->priv_data; AVFrame *picture = td; Jpeg2000Tile *tile = s->tile + jobnr; - tile_codeblocks(s, tile); - /* inverse MCT transformation */ if (tile->codsty[0].mct) mct_decode(s, tile); @@ -2473,11 +2476,80 @@ static av_cold int jpeg2000_decode_init(AVCodecContext *avctx) return 0; } +static int jpeg2000_setup_cbs(Jpeg2000DecoderContext *s, int *cbs_out) +{ + if (s->numXtiles * s->numYtiles > INT_MAX/sizeof(*s->idwt)/s->ncomponents) + return AVERROR(ENOMEM); + + av_fast_malloc(&s->idwt, &s->idwt_size, s->numXtiles * s->numYtiles * s->ncomponents * sizeof(*s->idwt)); + if (!s->idwt) + return AVERROR(ENOMEM); + + for (int pass = 0; pass < 2; pass++) { + int cbs = 0; + for (int tileno = 0; tileno < s->numXtiles * s->numYtiles; tileno++) { + for (int compno = 0; compno < s->ncomponents; compno++) { + Jpeg2000Tile *tile = s->tile + tileno; + Jpeg2000Component *comp = tile->comp + compno; + Jpeg2000CodingStyle *codsty = tile->codsty + compno; + Jpeg2000IdwtThread *idwt = s->idwt + compno + tileno * s->ncomponents; + + idwt->cb_start = cbs; + + for (int reslevelno = 0; reslevelno < codsty->nreslevels2decode; reslevelno++) { + Jpeg2000ResLevel *rlevel = comp->reslevel + reslevelno; + for (int bandno = 0; bandno < rlevel->nbands; bandno++) { + int nb_precincts = rlevel->num_precincts_x * rlevel->num_precincts_y; + Jpeg2000Band *band = rlevel->band + bandno; + + if (band->coord[0][0] == band->coord[0][1] || + band->coord[1][0] == band->coord[1][1]) + continue; + + for (int precno = 0; precno < nb_precincts; precno++) { + Jpeg2000Prec *prec = band->prec + precno; + int prec_cbs = prec->nb_codeblocks_width * prec->nb_codeblocks_height; + + if (cbs > INT_MAX - prec_cbs) + return AVERROR(ENOMEM); + + for (int cblkno = 0; cblkno < prec_cbs; cblkno++, cbs++) { + if (pass == 1) { + Jpeg2000CodeblockThread *cb = s->cb + cbs; + cb->tileno = tileno; + cb->compno = compno; + cb->reslevelno = reslevelno; + cb->bandno = bandno; + cb->precno = precno; + cb->cblkno = cblkno; + } + } + } + } + } + + idwt->cb_end = cbs; + } + } + + if (pass == 0) { + if (cbs > INT_MAX/sizeof(*s->cb)) + return AVERROR(ENOMEM); + av_fast_malloc(&s->cb, &s->cb_size, cbs*sizeof(*s->cb)); + if (!s->cb) + return AVERROR(ENOMEM); + } + + *cbs_out = cbs; + } + return 0; +} + static int jpeg2000_decode_frame(AVCodecContext *avctx, AVFrame *picture, int *got_frame, AVPacket *avpkt) { Jpeg2000DecoderContext *s = avctx->priv_data; - int ret; + int ret, cbs; s->avctx = avctx; bytestream2_init(&s->g, avpkt->data, avpkt->size); @@ -2535,7 +2607,12 @@ static int jpeg2000_decode_frame(AVCodecContext *avctx, AVFrame *picture, } } - avctx->execute2(avctx, jpeg2000_decode_tile, picture, NULL, s->numXtiles * s->numYtiles); + if ((ret = jpeg2000_setup_cbs(s, &cbs))) + goto end; + + avctx->execute2(avctx, jpeg2000_decode_cb, NULL, NULL, cbs); + avctx->execute2(avctx, jpeg2000_idwt, NULL, NULL, s->numXtiles * s->numYtiles * s->ncomponents); + avctx->execute2(avctx, jpeg2000_mct_write_frame, picture, NULL, s->numXtiles * s->numYtiles); jpeg2000_dec_cleanup(s); @@ -2554,6 +2631,16 @@ end: return ret; } +static av_cold int jpeg2000_decode_close(AVCodecContext *avctx) +{ + Jpeg2000DecoderContext *s = avctx->priv_data; + + av_freep(&s->idwt); + av_freep(&s->cb); + + return 0; +} + #define OFFSET(x) offsetof(Jpeg2000DecoderContext, x) #define VD AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_DECODING_PARAM @@ -2579,6 +2666,7 @@ const FFCodec ff_jpeg2000_decoder = { .priv_data_size = sizeof(Jpeg2000DecoderContext), .init = jpeg2000_decode_init, FF_CODEC_DECODE_CB(jpeg2000_decode_frame), + .close = jpeg2000_decode_close, .p.priv_class = &jpeg2000_class, .p.max_lowres = 5, .p.profiles = NULL_IF_CONFIG_SMALL(ff_jpeg2000_profiles), -- 2.30.2 [-- Attachment #3: Type: text/plain, Size: 251 bytes --] _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
next reply other threads:[~2022-06-14 14:39 UTC|newest] Thread overview: 31+ messages / expand[flat|nested] mbox.gz Atom feed top 2022-06-14 14:39 Tomas Härdin [this message] 2022-06-14 14:39 ` [FFmpeg-devel] [PATCH 02/13] lavc/jpeg2000dec: Reindent Tomas Härdin 2022-06-14 14:40 ` [FFmpeg-devel] [PATCH 03/13] lavc/jpeg2000dwt: Implement sliced transforms Tomas Härdin 2022-06-14 14:40 ` [FFmpeg-devel] [PATCH 04/13] lavc/jpeg2000dec: Implement IDWT slicing Tomas Härdin 2022-06-14 14:41 ` [FFmpeg-devel] [PATCH 05/13] lavc/jpeg2000dec: Thread init_tile() Tomas Härdin 2022-06-14 21:11 ` Michael Niedermayer 2022-06-15 13:11 ` Tomas Härdin 2022-06-15 21:05 ` Michael Niedermayer 2022-06-16 9:28 ` Tomas Härdin 2022-06-14 14:42 ` [FFmpeg-devel] [PATCH 06/13] lavu/mem: Add ff_fast_recalloc() Tomas Härdin 2022-06-14 20:26 ` Michael Niedermayer 2022-06-15 9:59 ` Tomas Härdin 2022-06-15 12:15 ` James Almer 2022-06-16 12:44 ` Tomas Härdin 2022-06-18 14:57 ` Anton Khirnov 2022-06-21 8:04 ` Tomas Härdin 2022-06-14 14:42 ` [FFmpeg-devel] [PATCH 07/13] lavc/jpeg2000*: Use ff_fast_recalloc() to eliminate lots of allocations Tomas Härdin 2022-06-14 15:23 ` Andreas Rheinhardt 2022-06-15 10:03 ` Tomas Härdin 2022-06-14 14:43 ` [FFmpeg-devel] [PATCH 08/13] lavc/jpeg2000: Switch Jpeg2000TgtNode to int32_t parent Tomas Härdin 2022-06-14 14:43 ` [FFmpeg-devel] [PATCH 09/13] lavc/jpeg2000: Speed up ff_jpeg2000_tag_tree_init() using stereotypes for sizes <= 4x4 Tomas Härdin 2022-06-18 15:00 ` Anton Khirnov 2022-06-21 7:57 ` Tomas Härdin 2022-06-14 14:43 ` [FFmpeg-devel] [PATCH 10/13] lavc/jpeg2000: Reindent Tomas Härdin 2022-06-14 14:44 ` [FFmpeg-devel] [PATCH 11/13] lavc/jpeg2000: Minimize calls to av_codec_is_encoder() Tomas Härdin 2022-06-14 15:04 ` Andreas Rheinhardt 2022-06-15 10:20 ` Tomas Härdin 2022-06-14 14:44 ` [FFmpeg-devel] [PATCH 12/13] lavc/jpeg2000dec: Use coarser slicing for initial reslevels Tomas Härdin 2022-06-14 14:47 ` [FFmpeg-devel] [PATCH 13/13] lavc/jpeg2000dec: Component-level threading of write_frame() Tomas Härdin 2022-06-18 14:50 ` [FFmpeg-devel] [PATCH 01/13] lavc/jpeg2000dec: Finer granularity threading Anton Khirnov 2022-06-24 8:19 ` Tomas Härdin
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=10ec51ef44325c2de6d5de7b994a9b6c8eb5e3a2.camel@acc.umu.se \ --to=tjoppen@acc.umu.se \ --cc=ffmpeg-devel@ffmpeg.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: link
Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel This inbox may be cloned and mirrored by anyone: git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git # If you have public-inbox 1.1+ installed, you may # initialize and index your mirror using the following commands: public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \ ffmpegdev@gitmailbox.com public-inbox-index ffmpegdev Example config snippet for mirrors. AGPL code for this site: git clone https://public-inbox.org/public-inbox.git