libs: encoder: h265: Add ensure_tile to calculate tiles.

We need consider tiles and slices together, separate tiles uniformly
and then assign slices uniformly to each tiles.

Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer-vaapi/-/merge_requests/294>
This commit is contained in:
He Junyan 2020-03-05 17:56:51 +08:00
parent 32db615685
commit cefd1a665f

View file

@ -119,6 +119,13 @@ struct _GstVaapiEncoderH265
gboolean low_delay_b; gboolean low_delay_b;
guint32 num_tile_cols; guint32 num_tile_cols;
guint32 num_tile_rows; guint32 num_tile_rows;
/* CTUs start address used in stream pack */
guint32 *tile_slice_address;
/* CTUs in this slice */
guint32 *tile_slice_ctu_num;
/* map the tile_slice_address to CTU start address in picture,
which is used by VA API. */
guint32 *tile_slice_address_map;
/* maximum required size of the decoded picture buffer */ /* maximum required size of the decoded picture buffer */
guint32 max_dec_pic_buffering; guint32 max_dec_pic_buffering;
@ -1774,6 +1781,11 @@ fill_sequence (GstVaapiEncoderH265 * encoder, GstVaapiEncSequence * sequence)
return TRUE; return TRUE;
} }
/* CTUs in each tile column */
static guint32 tile_ctu_cols[GST_VAAPI_H265_MAX_COL_TILES];
/* CTUs in each tile row */
static guint32 tile_ctu_rows[GST_VAAPI_H265_MAX_ROW_TILES];
/* Fills in VA picture parameter buffer */ /* Fills in VA picture parameter buffer */
static gboolean static gboolean
fill_picture (GstVaapiEncoderH265 * encoder, GstVaapiEncPicture * picture, fill_picture (GstVaapiEncoderH265 * encoder, GstVaapiEncPicture * picture,
@ -2374,6 +2386,299 @@ reset_properties (GstVaapiEncoderH265 * encoder)
reorder_pool->frame_index = 0; reorder_pool->frame_index = 0;
} }
static void
reset_tile (GstVaapiEncoderH265 * encoder)
{
memset (tile_ctu_cols, 0, sizeof (tile_ctu_cols));
memset (tile_ctu_rows, 0, sizeof (tile_ctu_rows));
if (encoder->tile_slice_address)
g_free (encoder->tile_slice_address);
encoder->tile_slice_address = NULL;
if (encoder->tile_slice_ctu_num)
g_free (encoder->tile_slice_ctu_num);
encoder->tile_slice_ctu_num = NULL;
if (encoder->tile_slice_address_map)
g_free (encoder->tile_slice_address_map);
encoder->tile_slice_address_map = NULL;
}
static void
recalculate_slices_num_by_tile (GstVaapiEncoderH265 * encoder)
{
GstVaapiDisplay *const display = GST_VAAPI_ENCODER_DISPLAY (encoder);
/* If driver has the requirement that the slice should not span tiles,
we need to increase slice number if needed. */
if (gst_vaapi_display_has_driver_quirks (display,
GST_VAAPI_DRIVER_QUIRK_HEVC_ENC_SLICE_NOT_SPAN_TILE)) {
if (encoder->num_slices < encoder->num_tile_cols * encoder->num_tile_rows) {
/* encoder->num_slices > 1 means user set it */
if (encoder->num_slices > 1)
GST_WARNING ("user set num-slices to %d, which is smaller than tile"
" num %d. We should make slice not span tiles, just set the"
" num-slices to tile num here.",
encoder->num_slices,
encoder->num_tile_cols * encoder->num_tile_rows);
else
GST_INFO ("set default slice num to %d, the same as the tile num.",
encoder->num_tile_cols * encoder->num_tile_rows);
encoder->num_slices = encoder->num_tile_cols * encoder->num_tile_rows;
}
}
}
static GstVaapiEncoderStatus
calculate_slices_start_address (GstVaapiEncoderH265 * encoder)
{
GstVaapiDisplay *const display = GST_VAAPI_ENCODER_DISPLAY (encoder);
guint32 ctu_per_slice;
guint32 left_slices;
gint32 i, j, k;
/* If driver has the requirement that the slice should not span tiles,
firstly we should scatter slices uniformly into each tile, bigger
tile gets more slices. Then we should assign CTUs within one tile
uniformly to each slice in that tile. */
if (gst_vaapi_display_has_driver_quirks (display,
GST_VAAPI_DRIVER_QUIRK_HEVC_ENC_SLICE_NOT_SPAN_TILE)) {
guint32 *slices_per_tile = g_malloc (encoder->num_tile_cols *
encoder->num_tile_rows * sizeof (guint32));
if (!slices_per_tile)
return GST_VAAPI_ENCODER_STATUS_ERROR_ALLOCATION_FAILED;
ctu_per_slice = (encoder->ctu_width * encoder->ctu_height +
encoder->num_slices - 1) / encoder->num_slices;
g_assert (ctu_per_slice > 0);
left_slices = encoder->num_slices;
for (i = 0; i < encoder->num_tile_cols * encoder->num_tile_rows; i++) {
slices_per_tile[i] = 1;
left_slices--;
}
while (left_slices) {
/* Find the biggest CTUs/slices, and assign more. */
gfloat largest = 0.0f;
k = -1;
for (i = 0; i < encoder->num_tile_cols * encoder->num_tile_rows; i++) {
gfloat f;
f = ((gfloat) (tile_ctu_cols[i % encoder->num_tile_cols] *
tile_ctu_rows[i / encoder->num_tile_cols])) /
(gfloat) slices_per_tile[i];
g_assert (f >= 1.0f);
if (f > largest) {
k = i;
largest = f;
}
}
g_assert (k >= 0);
slices_per_tile[k]++;
left_slices--;
}
/* Assign CTUs in one tile uniformly to each slice. Note: the slice start
address is CTB address in tile scan(see spec 6.5), that is, we accumulate
all CTUs in tile0, then tile1, and tile2..., not from the picture's
perspective. */
encoder->tile_slice_address[0] = 0;
k = 1;
for (i = 0; i < encoder->num_tile_rows; i++) {
for (j = 0; j < encoder->num_tile_cols; j++) {
guint32 s_num = slices_per_tile[i * encoder->num_tile_cols + j];
guint32 one_tile_ctus = tile_ctu_cols[j] * tile_ctu_rows[i];
guint32 s;
GST_LOG ("Tile(row %d col %d), has CTU in col %d,"
" CTU in row is %d, total CTU %d, assigned %d slices", i, j,
tile_ctu_cols[j], tile_ctu_rows[i], one_tile_ctus, s_num);
g_assert (s_num > 0);
for (s = 0; s < s_num; s++) {
encoder->tile_slice_address[k] =
encoder->tile_slice_address[k - 1] + ((s +
1) * one_tile_ctus) / s_num - (s * one_tile_ctus) / s_num;
encoder->tile_slice_ctu_num[k - 1] =
encoder->tile_slice_address[k] - encoder->tile_slice_address[k -
1];
k++;
}
}
}
g_assert (k == encoder->num_slices + 1);
/* Calculate the last one */
encoder->tile_slice_ctu_num[encoder->num_slices - 1] =
encoder->ctu_width * encoder->ctu_height -
encoder->tile_slice_address[encoder->num_slices - 1];
g_free (slices_per_tile);
}
/* The easy way, just assign CTUs to each slice uniformly */
else {
ctu_per_slice = (encoder->ctu_width * encoder->ctu_height +
encoder->num_slices - 1) / encoder->num_slices;
g_assert (ctu_per_slice > 0);
for (i = 0; i < encoder->num_slices - 1; i++)
encoder->tile_slice_ctu_num[i] = ctu_per_slice;
encoder->tile_slice_ctu_num[encoder->num_slices - 1] =
encoder->ctu_width * encoder->ctu_height -
(encoder->num_slices - 1) * ctu_per_slice;
encoder->tile_slice_address[0] = 0;
for (i = 1; i <= encoder->num_slices; i++)
encoder->tile_slice_address[i] = encoder->tile_slice_address[i - 1] +
encoder->tile_slice_ctu_num[i - 1];
}
return GST_VAAPI_ENCODER_STATUS_SUCCESS;
}
static GstVaapiEncoderStatus
ensure_tile (GstVaapiEncoderH265 * encoder)
{
gint32 i, j, k;
guint32 ctu_tile_width_accu[GST_VAAPI_H265_MAX_COL_TILES + 1];
guint32 ctu_tile_height_accu[GST_VAAPI_H265_MAX_ROW_TILES + 1];
guint32 num_slices;
GstVaapiEncoderStatus ret;
reset_tile (encoder);
if (!h265_is_tile_enabled (encoder))
return GST_VAAPI_ENCODER_STATUS_SUCCESS;
if (!gst_vaapi_encoder_ensure_tile_support (GST_VAAPI_ENCODER (encoder),
encoder->profile, encoder->entrypoint)) {
GST_ERROR ("The profile:%s, entrypoint:%d does not support tile.",
gst_vaapi_utils_h265_get_profile_string (encoder->profile),
encoder->entrypoint);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
if (encoder->num_tile_cols >
gst_vaapi_utils_h265_get_level_limits (encoder->level)->MaxTileColumns) {
GST_ERROR ("num_tile_cols:%d exceeds MaxTileColumns:%d",
encoder->num_tile_cols,
gst_vaapi_utils_h265_get_level_limits (encoder->level)->MaxTileColumns);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
if (encoder->num_tile_rows >
gst_vaapi_utils_h265_get_level_limits (encoder->level)->MaxTileRows) {
GST_ERROR ("num_tile_rows:%d exceeds MaxTileRows:%d",
encoder->num_tile_rows,
gst_vaapi_utils_h265_get_level_limits (encoder->level)->MaxTileRows);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
if (encoder->ctu_width < encoder->num_tile_cols) {
GST_WARNING
("Only %d CTUs in width, not enough to split into %d tile columns",
encoder->ctu_width, encoder->num_tile_cols);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
if (encoder->ctu_height < encoder->num_tile_rows) {
GST_WARNING
("Only %d CTUs in height, not enough to split into %d tile rows",
encoder->ctu_height, encoder->num_tile_rows);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
recalculate_slices_num_by_tile (encoder);
/* ensure not exceed max supported slices */
num_slices = encoder->num_slices;
gst_vaapi_encoder_ensure_num_slices (GST_VAAPI_ENCODER_CAST (encoder),
encoder->profile, encoder->entrypoint,
(encoder->ctu_width * encoder->ctu_height + 1) / 2, &num_slices);
if (num_slices != encoder->num_slices) {
GST_ERROR ("The tile setting need at least %d slices, but the max"
" slice number is just %d", encoder->num_slices, num_slices);
return GST_VAAPI_ENCODER_STATUS_ERROR_UNKNOWN;
}
encoder->tile_slice_address =
/* Add one as sentinel, hold val to calculate ctu_num */
g_malloc ((encoder->num_slices + 1) * sizeof (guint32));
if (!encoder->tile_slice_address)
return GST_VAAPI_ENCODER_STATUS_ERROR_ALLOCATION_FAILED;
encoder->tile_slice_ctu_num =
g_malloc (encoder->num_slices * sizeof (guint32));
if (!encoder->tile_slice_ctu_num)
return GST_VAAPI_ENCODER_STATUS_ERROR_ALLOCATION_FAILED;
encoder->tile_slice_address_map =
g_malloc (encoder->ctu_width * encoder->ctu_height * sizeof (guint32));
if (!encoder->tile_slice_address_map)
return GST_VAAPI_ENCODER_STATUS_ERROR_ALLOCATION_FAILED;
/* firstly uniformly separate CTUs into tiles, as the spec 6.5.1 define */
for (i = 0; i < encoder->num_tile_cols; i++)
tile_ctu_cols[i] =
((i + 1) * encoder->ctu_width) / encoder->num_tile_cols -
(i * encoder->ctu_width) / encoder->num_tile_cols;
for (i = 0; i < encoder->num_tile_rows; i++)
tile_ctu_rows[i] =
((i + 1) * encoder->ctu_height) / encoder->num_tile_rows -
(i * encoder->ctu_height) / encoder->num_tile_rows;
ret = calculate_slices_start_address (encoder);
if (ret != GST_VAAPI_ENCODER_STATUS_SUCCESS)
return ret;
/* Build the map to specifying the conversion between a CTB address in CTB
raster scan of a picture and a CTB address in tile scan(see spec 6.5.1
for details). */
ctu_tile_width_accu[0] = 0;
for (i = 1; i <= encoder->num_tile_cols; i++)
ctu_tile_width_accu[i] = ctu_tile_width_accu[i - 1] + tile_ctu_cols[i - 1];
ctu_tile_height_accu[0] = 0;
for (i = 1; i <= encoder->num_tile_rows; i++)
ctu_tile_height_accu[i] =
ctu_tile_height_accu[i - 1] + tile_ctu_rows[i - 1];
for (k = 0; k < encoder->ctu_width * encoder->ctu_height; k++) {
/* The ctu coordinate in the picture. */
guint32 x = k % encoder->ctu_width;
guint32 y = k / encoder->ctu_width;
/* The ctu coordinate in the tile mode. */
guint32 tile_x = 0;
guint32 tile_y = 0;
/* The index of the CTU in the tile mode. */
guint32 tso = 0;
for (i = 0; i < encoder->num_tile_cols; i++)
if (x >= ctu_tile_width_accu[i])
tile_x = i;
g_assert (tile_x <= encoder->num_tile_cols - 1);
for (j = 0; j < encoder->num_tile_rows; j++)
if (y >= ctu_tile_height_accu[j])
tile_y = j;
g_assert (tile_y <= encoder->num_tile_rows - 1);
/* add all ctus in the tiles the same line before us */
for (i = 0; i < tile_x; i++)
tso += tile_ctu_rows[tile_y] * tile_ctu_cols[i];
/* add all ctus in the tiles above us */
for (j = 0; j < tile_y; j++)
tso += encoder->ctu_width * tile_ctu_rows[j];
/* add the ctus inside the same tile before us */
tso += (y - ctu_tile_height_accu[tile_y]) * tile_ctu_cols[tile_x]
+ x - ctu_tile_width_accu[tile_x];
g_assert (tso < encoder->ctu_width * encoder->ctu_height);
encoder->tile_slice_address_map[tso] = k;
}
return GST_VAAPI_ENCODER_STATUS_SUCCESS;
}
static GstVaapiEncoderStatus static GstVaapiEncoderStatus
gst_vaapi_encoder_h265_encode (GstVaapiEncoder * base_encoder, gst_vaapi_encoder_h265_encode (GstVaapiEncoder * base_encoder,
GstVaapiEncPicture * picture, GstVaapiCodedBufferProxy * codedbuf) GstVaapiEncPicture * picture, GstVaapiCodedBufferProxy * codedbuf)
@ -2837,6 +3142,9 @@ gst_vaapi_encoder_h265_reconfigure (GstVaapiEncoder * base_encoder)
} }
reset_properties (encoder); reset_properties (encoder);
status = ensure_tile (encoder);
if (status != GST_VAAPI_ENCODER_STATUS_SUCCESS)
return status;
ensure_control_rate_params (encoder); ensure_control_rate_params (encoder);
return set_context_info (base_encoder); return set_context_info (base_encoder);
} }
@ -2907,6 +3215,8 @@ gst_vaapi_encoder_h265_finalize (GObject * object)
} }
g_queue_clear (&reorder_pool->reorder_frame_list); g_queue_clear (&reorder_pool->reorder_frame_list);
reset_tile (encoder);
G_OBJECT_CLASS (gst_vaapi_encoder_h265_parent_class)->finalize (object); G_OBJECT_CLASS (gst_vaapi_encoder_h265_parent_class)->finalize (object);
} }