[FFmpeg-devel] libaomenc: Add support for tiles

Submitted by Mark Thompson on Oct. 2, 2018, 11:50 p.m.

Details

Message ID 20181002235017.19636-1-sw@jkqxz.net
State New
Headers show

Commit Message

Mark Thompson Oct. 2, 2018, 11:50 p.m.
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(+)

Patch hide | download patch | download mbox

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 }
 };