From 1105cb797c67d05bf4666e2e33140debcfec12a7 Mon Sep 17 00:00:00 2001 From: Andreas Rheinhardt Date: Thu, 22 May 2025 15:57:13 +0200 Subject: [PATCH 1/3] avcodec/asvenc: Don't waste bits encoding non-visible part Up until now, the encoder replicated all the border pixels for incomplete 16x16 macroblocks. In case the available width or height is <= 8, some of the luma blocks of the MB do not correspond to actual input, so that we should encode them using the least amount of bits. Zeroing the block coefficients (as this commit does) achieves this, replicating the pixels and performing an FDCT does not. This commit also removes the frame copying code for insufficiently aligned dimensions. The vsynth3-asv[12] FATE tests use a 34x34 input file and are therefore affected by this. As the ref updates show, the size and checksum of the encoded changes, yet the decoded output stays the same. Signed-off-by: Andreas Rheinhardt --- libavcodec/asvenc.c | 131 +++++++++++++++++++++------------- tests/ref/vsynth/vsynth3-asv1 | 4 +- tests/ref/vsynth/vsynth3-asv2 | 4 +- 3 files changed, 84 insertions(+), 55 deletions(-) diff --git a/libavcodec/asvenc.c b/libavcodec/asvenc.c index 52666ee547..a53dc7c670 100644 --- a/libavcodec/asvenc.c +++ b/libavcodec/asvenc.c @@ -26,6 +26,7 @@ #include "config_components.h" #include "libavutil/attributes.h" +#include "libavutil/intreadwrite.h" #include "libavutil/mem.h" #include "libavutil/mem_internal.h" @@ -228,55 +229,65 @@ static inline void dct_get(ASVEncContext *a, const AVFrame *frame, } } -static int encode_frame(AVCodecContext *avctx, AVPacket *pkt, - const AVFrame *pict, int *got_packet) +static void handle_partial_mb(ASVEncContext *a, const uint8_t *const data[3], + const int linesizes[3], + int valid_width, int valid_height) { - ASVEncContext *const a = avctx->priv_data; - const ASVCommonContext *const c = &a->c; - int size, ret; - - if (pict->width % 16 || pict->height % 16) { - AVFrame *clone = av_frame_alloc(); - int i; - - if (!clone) - return AVERROR(ENOMEM); - clone->format = pict->format; - clone->width = FFALIGN(pict->width, 16); - clone->height = FFALIGN(pict->height, 16); - ret = av_frame_get_buffer(clone, 0); - if (ret < 0) { - av_frame_free(&clone); - return ret; + const int nb_blocks = a->c.avctx->flags & AV_CODEC_FLAG_GRAY ? 4 : 6; + static const struct Descriptor { + uint8_t x_offset, y_offset; + uint8_t component, subsampling; + } block_descriptor[] = { + { 0, 0, 0, 0 }, { 8, 0, 0, 0 }, { 0, 8, 0, 0 }, { 8, 8, 0, 0 }, + { 0, 0, 1, 1 }, { 0, 0, 2, 1 }, + }; + + for (int i = 0; i < nb_blocks; ++i) { + const struct Descriptor *const desc = block_descriptor + i; + int width_avail = AV_CEIL_RSHIFT(valid_width, desc->subsampling) - desc->x_offset; + int height_avail = AV_CEIL_RSHIFT(valid_height, desc->subsampling) - desc->y_offset; + + if (width_avail <= 0 || height_avail <= 0) { + // This block is outside of the visible part; don't replicate pixels, + // just zero the block, so that only the dc value will be coded. + memset(a->block[i], 0, sizeof(a->block[i])); + continue; } - - ret = av_frame_copy(clone, pict); - if (ret < 0) { - av_frame_free(&clone); - return ret; + width_avail = FFMIN(width_avail, 8); + height_avail = FFMIN(height_avail, 8); + + ptrdiff_t linesize = linesizes[desc->component]; + const uint8_t *src = data[desc->component] + desc->y_offset * linesize + desc->x_offset; + int16_t *block = a->block[i]; + + for (int h = 0;; block += 8, src += linesize) { + int16_t last; + for (int w = 0; w < width_avail; ++w) + last = block[w] = src[w]; + for (int w = width_avail; w < 8; ++w) + block[w] = last; + if (++h == height_avail) + break; } - - for (i = 0; i<3; i++) { - int x, y; - int w = AV_CEIL_RSHIFT(pict->width, !!i); - int h = AV_CEIL_RSHIFT(pict->height, !!i); - int w2 = AV_CEIL_RSHIFT(clone->width, !!i); - int h2 = AV_CEIL_RSHIFT(clone->height, !!i); - for (y=0; ydata[i][x + y*clone->linesize[i]] = - clone->data[i][w - 1 + y*clone->linesize[i]]; - for (y=h; ydata[i][x + y*clone->linesize[i]] = - clone->data[i][x + (h-1)*clone->linesize[i]]; + const int16_t *const last_row = block; + for (int h = height_avail; h < 8; ++h) { + block += 8; + AV_COPY128(block, last_row); } - ret = encode_frame(avctx, pkt, clone, got_packet); - av_frame_free(&clone); - return ret; + a->fdsp.fdct(a->block[i]); } + encode_mb(a, a->block); +} + +static int encode_frame(AVCodecContext *avctx, AVPacket *pkt, + const AVFrame *pict, int *got_packet) +{ + ASVEncContext *const a = avctx->priv_data; + const ASVCommonContext *const c = &a->c; + int size, ret; + ret = ff_alloc_packet(avctx, pkt, c->mb_height * c->mb_width * MAX_MB_SIZE + 3); if (ret < 0) return ret; @@ -290,19 +301,37 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt, } } - if (c->mb_width2 != c->mb_width) { - int mb_x = c->mb_width2; + if (avctx->width & 15) { + const uint8_t *src[3] = { + pict->data[0] + c->mb_width2 * 16, + pict->data[1] + c->mb_width2 * 8, + pict->data[2] + c->mb_width2 * 8, + }; + int available_width = avctx->width & 15; + for (int mb_y = 0; mb_y < c->mb_height2; mb_y++) { - dct_get(a, pict, mb_x, mb_y); - encode_mb(a, a->block); + handle_partial_mb(a, src, pict->linesize, available_width, 16); + src[0] += 16 * pict->linesize[0]; + src[1] += 8 * pict->linesize[1]; + src[2] += 8 * pict->linesize[2]; } } - if (c->mb_height2 != c->mb_height) { - int mb_y = c->mb_height2; - for (int mb_x = 0; mb_x < c->mb_width; mb_x++) { - dct_get(a, pict, mb_x, mb_y); - encode_mb(a, a->block); + if (avctx->height & 15) { + const uint8_t *src[3] = { + pict->data[0] + c->mb_height2 * 16 * pict->linesize[0], + pict->data[1] + c->mb_height2 * 8 * pict->linesize[1], + pict->data[2] + c->mb_height2 * 8 * pict->linesize[2], + }; + int available_height = avctx->height & 15; + + for (int remaining = avctx->width;; remaining -= 16) { + handle_partial_mb(a, src, pict->linesize, remaining, available_height); + if (remaining <= 16) + break; + src[0] += 16; + src[1] += 8; + src[2] += 8; } } diff --git a/tests/ref/vsynth/vsynth3-asv1 b/tests/ref/vsynth/vsynth3-asv1 index 0abbf787ec..af1dc644b0 100644 --- a/tests/ref/vsynth/vsynth3-asv1 +++ b/tests/ref/vsynth/vsynth3-asv1 @@ -1,4 +1,4 @@ -81eeea0d0e6219b2f381cf2100e9a12f *tests/data/fate/vsynth3-asv1.avi -34704 tests/data/fate/vsynth3-asv1.avi +69ae6df10440e68c53bee4e713851199 *tests/data/fate/vsynth3-asv1.avi +31524 tests/data/fate/vsynth3-asv1.avi 3c8636e22a96267451684f42d7a6f608 *tests/data/fate/vsynth3-asv1.out.rawvideo stddev: 13.16 PSNR: 25.74 MAXDIFF: 112 bytes: 86700/ 86700 diff --git a/tests/ref/vsynth/vsynth3-asv2 b/tests/ref/vsynth/vsynth3-asv2 index 90b8a47f34..9fa9822c0b 100644 --- a/tests/ref/vsynth/vsynth3-asv2 +++ b/tests/ref/vsynth/vsynth3-asv2 @@ -1,4 +1,4 @@ -8402fb1112fb8119c019154a472b5cd0 *tests/data/fate/vsynth3-asv2.avi -36208 tests/data/fate/vsynth3-asv2.avi +63000eaedeb60bede8baeb090f02881a *tests/data/fate/vsynth3-asv2.avi +33696 tests/data/fate/vsynth3-asv2.avi 5469c0735b7c9279e5e8e3439fc6acab *tests/data/fate/vsynth3-asv2.out.rawvideo stddev: 9.07 PSNR: 28.97 MAXDIFF: 51 bytes: 86700/ 86700 -- 2.45.2