Adds an option to specify the number of tile rows and columns, then uses a uniform tiling if possible and otherwise a fixed tiling with equal-sized tiles to fill the frame.
Also adds -tile-columns and -tile-rows options to make tilings with power-of-two numbers of tiles, matching the behaviour of the libvpx/VP9 encoder. --- On 23/09/18 23:55, Mark Thompson wrote: > On 19/09/18 00:15, James Almer wrote: >> On 9/18/2018 7:55 PM, Mark Thompson wrote: >>> On 18/09/18 01:12, James Almer wrote: >>>> On 9/17/2018 8:47 PM, Mark Thompson wrote: >>>>> Adds an option to specify the number of tile rows and columns, then uses >>>>> equal-sized tiles to fill the frame. >>>>> --- >>>>> libavcodec/libaomenc.c | 54 ++++++++++++++++++++++++++++++++++++++++++ >>>>> 1 file changed, 54 insertions(+) >>>>> >>>>> diff --git a/libavcodec/libaomenc.c b/libavcodec/libaomenc.c >>>>> index 6a79d9b873..3ccff0e0fb 100644 >>>>> --- a/libavcodec/libaomenc.c >>>>> +++ b/libavcodec/libaomenc.c >>>>> ... >>>>> @@ -742,6 +795,7 @@ static const AVOption options[] = { >>>>> { "static-thresh", "A change threshold on blocks below which they >>>>> will be skipped by the encoder", OFFSET(static_thresh), AV_OPT_TYPE_INT, >>>>> { .i64 = 0 }, 0, INT_MAX, VE }, >>>>> { "drop-threshold", "Frame drop threshold", offsetof(AOMContext, >>>>> drop_threshold), AV_OPT_TYPE_INT, {.i64 = 0 }, INT_MIN, INT_MAX, VE }, >>>>> { "noise-sensitivity", "Noise sensitivity", >>>>> OFFSET(noise_sensitivity), AV_OPT_TYPE_INT, {.i64 = 0 }, 0, 4, VE}, >>>>> + { "tiles", "Tile rows x columns", OFFSET(tile_cols), >>>>> AV_OPT_TYPE_IMAGE_SIZE, { .str = NULL }, 0, 0, VE }, >>>> >>>> Using separate tile-columns and tile-rows AV_OPT_TYPE_INT options would >>>> be more consistent with the libvpx wrapper, which already has them >>>> called like that and also shares a lot of other option names with the >>>> libaom. >>> >>> The options on libvpx-vp9 are actually log2 of the value, so "-tile-rows 3 >>> -tile-columns 2" gives you 8x4 tiles. (VP9 requires that the number of >>> tiles in each dimension is a power of two, while AV1 lets you set arbitrary >>> sizes.) >>> >>> I don't really mind how this works - I just thought the IMAGE_SIZE method >>> looked nicer. What do you prefer? >> >> I usually prefer consistency in options for similar modules, but the >> equivalent of the VP9 options would be to set the AV1E_SET_TILE_* codec >> control IDs instead of what you're doing here, so your IMAGE_SIZE method >> is fine. >> >> There's for that matter a conflicting patch called "lavc/libaomenc: Add >> -tile-columns/-tile-rows" by Kagami Hiiragi that sets the aforementioned >> codec control IDs with the same option names as the VP9 ones. Maybe both >> patches can be applied, and have the encoder abort if they are used at >> the same time? Otherwise apply yours alone since it allows arbitrary sizes. > > I think having both would make sense - the two different options are > reflecting the different values of uniform_tile_spacing_flag. The explicit > sizing can get the same result with the flag off as it being on, but will > require more bits in every frame header to do so. > > On matching behaviour with libvpx, I don't think tile_cols/rows_log2 in VP9 > and AV1 (with uniform_tile_spacing_flag = 1) actually do have the same effect > - VP9 ensures that you get exactly 2^tile_cols_log2 tile columns, while for > AV1 it's an upper bound and depends on the width. For example, given a width > of 576 (as 9 64x64 superblocks) and tile_cols_log2 = 2, VP9 looks like it > gives you { 2, 2, 2, 3 } while AV1 will give you { 3, 3, 3 } as the tile > column widths. That probably wants to be mentioned in the documentation, and > maybe merits the options not matching precisely. > > So, I'll look into supporting both cases in some consistent way. Here we go; hopefully this covers all of the cases. I went for making the options match precisely for the power-of-two values in VP9 - we can determine what arrangements a uniform tiling will produce and decide from that whether to use one or not. - Mark libavcodec/libaomenc.c | 189 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 189 insertions(+) diff --git a/libavcodec/libaomenc.c b/libavcodec/libaomenc.c index 6a79d9b873..50abc6a5db 100644 --- a/libavcodec/libaomenc.c +++ b/libavcodec/libaomenc.c @@ -34,6 +34,7 @@ #include "libavutil/opt.h" #include "libavutil/pixdesc.h" +#include "av1.h" #include "avcodec.h" #include "internal.h" #include "profiles.h" @@ -68,6 +69,11 @@ typedef struct AOMEncoderContext { int static_thresh; int drop_threshold; int noise_sensitivity; + int tile_cols, tile_rows; + int tile_cols_log2, tile_rows_log2; + + aom_superblock_size_t superblock_size; + int uniform_tiles; } AOMContext; static const char *const ctlidstr[] = { @@ -79,6 +85,9 @@ static const char *const ctlidstr[] = { [AV1E_SET_COLOR_PRIMARIES] = "AV1E_SET_COLOR_PRIMARIES", [AV1E_SET_MATRIX_COEFFICIENTS] = "AV1E_SET_MATRIX_COEFFICIENTS", [AV1E_SET_TRANSFER_CHARACTERISTICS] = "AV1E_SET_TRANSFER_CHARACTERISTICS", + [AV1E_SET_SUPERBLOCK_SIZE] = "AV1E_SET_SUPERBLOCK_SIZE", + [AV1E_SET_TILE_COLUMNS] = "AV1E_SET_TILE_COLUMNS", + [AV1E_SET_TILE_ROWS] = "AV1E_SET_TILE_ROWS", }; static av_cold void log_encoder_error(AVCodecContext *avctx, const char *desc) @@ -143,6 +152,10 @@ static av_cold void dump_enc_cfg(AVCodecContext *avctx, width, "kf_mode:", cfg->kf_mode, width, "kf_min_dist:", cfg->kf_min_dist, width, "kf_max_dist:", cfg->kf_max_dist); + av_log(avctx, level, "tile settings\n" + " %*s%d\n %*s%d\n", + width, "tile_width_count:", cfg->tile_width_count, + width, "tile_height_count:", cfg->tile_height_count); av_log(avctx, level, "\n"); } @@ -284,6 +297,169 @@ static void set_color_range(AVCodecContext *avctx) codecctl_int(avctx, AV1E_SET_COLOR_RANGE, aom_cr); } +static int count_uniform_tiling(int dim, int sb_size, int tiles_log2) +{ + int sb_dim = (dim + sb_size - 1) / sb_size; + int tile_dim = (sb_dim + (1 << tiles_log2) - 1) >> tiles_log2; + av_assert0(tile_dim > 0); + return (sb_dim + tile_dim - 1) / tile_dim; +} + +static int choose_tiling(AVCodecContext *avctx, + struct aom_codec_enc_cfg *enccfg) +{ + AOMContext *ctx = avctx->priv_data; + int sb_128x128_possible, sb_size, sb_width, sb_height; + int uniform_rows, uniform_cols; + int uniform_64x64_possible, uniform_128x128_possible; + int tile_size, rounding, i; + + if (ctx->tile_cols_log2 >= 0) + ctx->tile_cols = 1 << ctx->tile_cols_log2; + if (ctx->tile_rows_log2 >= 0) + ctx->tile_rows = 1 << ctx->tile_rows_log2; + + if (ctx->tile_cols == 0) { + ctx->tile_cols = (avctx->width + AV1_MAX_TILE_WIDTH - 1) / + AV1_MAX_TILE_WIDTH; + if (ctx->tile_cols > 1) { + av_log(avctx, AV_LOG_DEBUG, "Automatically using %d tile " + "columns to fill width.\n", ctx->tile_cols); + } + } + av_assert0(ctx->tile_cols > 0); + if (ctx->tile_rows == 0) { + int max_tile_width = + FFALIGN((FFALIGN(avctx->width, 128) + + ctx->tile_cols - 1) / ctx->tile_cols, 128); + ctx->tile_rows = + (max_tile_width * FFALIGN(avctx->height, 128) + + AV1_MAX_TILE_AREA - 1) / AV1_MAX_TILE_AREA; + if (ctx->tile_rows > 1) { + av_log(avctx, AV_LOG_DEBUG, "Automatically using %d tile " + "rows to fill area.\n", ctx->tile_rows); + } + } + av_assert0(ctx->tile_rows > 0); + + if ((avctx->width + 63) / 64 < ctx->tile_cols || + (avctx->height + 63) / 64 < ctx->tile_rows) { + av_log(avctx, AV_LOG_ERROR, "Invalid tile sizing: frame not " + "large enough to fit specified tile arrangement.\n"); + return AVERROR(EINVAL); + } + if (ctx->tile_cols > AV1_MAX_TILE_COLS || + ctx->tile_rows > AV1_MAX_TILE_ROWS) { + av_log(avctx, AV_LOG_ERROR, "Invalid tile sizing: AV1 does " + "not allow more than %dx%d tiles.\n", + AV1_MAX_TILE_COLS, AV1_MAX_TILE_ROWS); + return AVERROR(EINVAL); + } + if (avctx->width / ctx->tile_cols > AV1_MAX_TILE_WIDTH) { + av_log(avctx, AV_LOG_ERROR, "Invalid tile sizing: AV1 does " + "not allow tiles of width greater than %d.\n", + AV1_MAX_TILE_WIDTH); + return AVERROR(EINVAL); + } + + ctx->superblock_size = AOM_SUPERBLOCK_SIZE_DYNAMIC; + + if (ctx->tile_cols == 1 && ctx->tile_rows == 1) { + av_log(avctx, AV_LOG_DEBUG, "Using a single tile.\n"); + return 0; + } + + sb_128x128_possible = + (avctx->width + 127) / 128 >= ctx->tile_cols && + (avctx->height + 127) / 128 >= ctx->tile_rows; + + ctx->tile_cols_log2 = ctx->tile_cols == 1 ? 0 : + av_log2(ctx->tile_cols - 1) + 1; + ctx->tile_rows_log2 = ctx->tile_rows == 1 ? 0 : + av_log2(ctx->tile_rows - 1) + 1; + + uniform_cols = count_uniform_tiling(avctx->width, + 64, ctx->tile_cols_log2); + uniform_rows = count_uniform_tiling(avctx->height, + 64, ctx->tile_rows_log2); + av_log(avctx, AV_LOG_DEBUG, "Uniform with 64x64 superblocks " + "-> %dx%d tiles.\n", uniform_cols, uniform_rows); + uniform_64x64_possible = uniform_cols == ctx->tile_cols && + uniform_rows == ctx->tile_rows; + + if (sb_128x128_possible) { + uniform_cols = count_uniform_tiling(avctx->width, + 128, ctx->tile_cols_log2); + uniform_rows = count_uniform_tiling(avctx->height, + 128, ctx->tile_rows_log2); + av_log(avctx, AV_LOG_DEBUG, "Uniform with 128x128 superblocks " + "-> %dx%d tiles.\n", uniform_cols, uniform_rows); + uniform_128x128_possible = uniform_cols == ctx->tile_cols && + uniform_rows == ctx->tile_rows; + } else { + av_log(avctx, AV_LOG_DEBUG, "128x128 superblocks not possible.\n"); + uniform_128x128_possible = 0; + } + + ctx->uniform_tiles = 1; + if (uniform_64x64_possible && uniform_128x128_possible) { + av_log(avctx, AV_LOG_DEBUG, "Using uniform tiling with dynamic " + "superblocks (tile_cols_log2 = %d, tile_rows_log2 = %d).\n", + ctx->tile_cols_log2, ctx->tile_rows_log2); + return 0; + } + if (uniform_64x64_possible && !sb_128x128_possible) { + av_log(avctx, AV_LOG_DEBUG, "Using uniform tiling with 64x64 " + "superblocks (tile_cols_log2 = %d, tile_rows_log2 = %d).\n", + ctx->tile_cols_log2, ctx->tile_rows_log2); + ctx->superblock_size = AOM_SUPERBLOCK_SIZE_64X64; + return 0; + } + if (uniform_128x128_possible) { + av_log(avctx, AV_LOG_DEBUG, "Using uniform tiling with 128x128 " + "superblocks (tile_cols_log2 = %d, tile_rows_log2 = %d).\n", + ctx->tile_cols_log2, ctx->tile_rows_log2); + ctx->superblock_size = AOM_SUPERBLOCK_SIZE_128X128; + return 0; + } + ctx->uniform_tiles = 0; + + if (sb_128x128_possible) { + sb_size = 128; + ctx->superblock_size = AOM_SUPERBLOCK_SIZE_128X128; + } else { + sb_size = 64; + ctx->superblock_size = AOM_SUPERBLOCK_SIZE_64X64; + } + av_log(avctx, AV_LOG_DEBUG, "Using fixed tiling with %dx%d " + "superblocks (tile_cols = %d, tile_rows = %d).\n", + sb_size, sb_size, ctx->tile_cols, ctx->tile_rows); + + enccfg->tile_width_count = ctx->tile_cols; + enccfg->tile_height_count = ctx->tile_rows; + + sb_width = (avctx->width + sb_size - 1) / sb_size; + sb_height = (avctx->height + sb_size - 1) / sb_size; + + tile_size = sb_width / ctx->tile_cols; + rounding = sb_width % ctx->tile_cols; + for (i = 0; i < ctx->tile_cols; i++) { + enccfg->tile_widths[i] = tile_size + + (i < rounding / 2 || + i > ctx->tile_cols - 1 - (rounding + 1) / 2); + } + + tile_size = sb_height / ctx->tile_rows; + rounding = sb_height % ctx->tile_rows; + for (i = 0; i < ctx->tile_rows; i++) { + enccfg->tile_heights[i] = tile_size + + (i < rounding / 2 || + i > ctx->tile_rows - 1 - (rounding + 1) / 2); + } + + return 0; +} + static av_cold int aom_init(AVCodecContext *avctx, const struct aom_codec_iface *iface) { @@ -431,6 +607,10 @@ static av_cold int aom_init(AVCodecContext *avctx, enccfg.g_error_resilient = ctx->error_resilient; + res = choose_tiling(avctx, &enccfg); + if (res < 0) + return res; + dump_enc_cfg(avctx, &enccfg); /* Construct Encoder Context */ res = aom_codec_enc_init(&ctx->encoder, iface, &enccfg, flags); @@ -454,6 +634,12 @@ static av_cold int aom_init(AVCodecContext *avctx, codecctl_int(avctx, AV1E_SET_TRANSFER_CHARACTERISTICS, avctx->color_trc); set_color_range(avctx); + codecctl_int(avctx, AV1E_SET_SUPERBLOCK_SIZE, ctx->superblock_size); + if (ctx->uniform_tiles) { + codecctl_int(avctx, AV1E_SET_TILE_COLUMNS, ctx->tile_cols_log2); + codecctl_int(avctx, AV1E_SET_TILE_ROWS, ctx->tile_rows_log2); + } + // provide dummy value to initialize wrapper, values will be updated each _encode() aom_img_wrap(&ctx->rawimg, img_fmt, avctx->width, avctx->height, 1, (unsigned char*)1); @@ -742,6 +928,9 @@ static const AVOption options[] = { { "static-thresh", "A change threshold on blocks below which they will be skipped by the encoder", OFFSET(static_thresh), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, INT_MAX, VE }, { "drop-threshold", "Frame drop threshold", offsetof(AOMContext, drop_threshold), AV_OPT_TYPE_INT, {.i64 = 0 }, INT_MIN, INT_MAX, VE }, { "noise-sensitivity", "Noise sensitivity", OFFSET(noise_sensitivity), AV_OPT_TYPE_INT, {.i64 = 0 }, 0, 4, VE}, + { "tiles", "Tile columns x rows", OFFSET(tile_cols), AV_OPT_TYPE_IMAGE_SIZE, { .str = NULL }, 0, 0, VE }, + { "tile-columns", "Log2 of number of tile columns to use", OFFSET(tile_cols_log2), AV_OPT_TYPE_INT, {.i64 = -1}, -1, 6, VE}, + { "tile-rows", "Log2 of number of tile rows to use", OFFSET(tile_rows_log2), AV_OPT_TYPE_INT, {.i64 = -1}, -1, 6, VE}, { NULL } }; -- 2.18.0 _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel