avcodec/nvenc: surface allocation reduction
This patch aims to reduce the number of input/output surfaces
NVENC allocates per session. Previous default sets allocated surfaces to 32
(unless there is user specified param or lookahead involved). Having large
number of surfaces consumes extra video memory (esp for higher resolution
encoding). The patch changes the surfaces calculation for default, B-frames,
lookahead scenario respectively.
The other change involves surface selection. Previously, if a session
allocates x surfaces, only x-1 surfaces are used (due to combination
of output delay and lock toggle logic). To prevent unused surfaces,
changing surface rotation to using predefined fifo.
Signed-off-by: Timo Rothenpieler <[email protected]>
diff --git a/libavcodec/nvenc.c b/libavcodec/nvenc.c
index cf05455..00766c2 100644
--- a/libavcodec/nvenc.c
+++ b/libavcodec/nvenc.c
@@ -644,16 +644,34 @@
static av_cold int nvenc_recalc_surfaces(AVCodecContext *avctx)
{
NvencContext *ctx = avctx->priv_data;
- int nb_surfaces = 0;
+ // default minimum of 4 surfaces
+ // multiply by 2 for number of NVENCs on gpu (hardcode to 2)
+ // another multiply by 2 to avoid blocking next PBB group
+ int nb_surfaces = FFMAX(4, ctx->encode_config.frameIntervalP * 2 * 2);
+ // lookahead enabled
if (ctx->rc_lookahead > 0) {
- nb_surfaces = ctx->rc_lookahead + ((ctx->encode_config.frameIntervalP > 0) ? ctx->encode_config.frameIntervalP : 0) + 1 + 4;
- if (ctx->nb_surfaces < nb_surfaces) {
+ // +1 is to account for lkd_bound calculation later
+ // +4 is to allow sufficient pipelining with lookahead
+ nb_surfaces = FFMAX(1, FFMAX(nb_surfaces, ctx->rc_lookahead + ctx->encode_config.frameIntervalP + 1 + 4));
+ if (nb_surfaces > ctx->nb_surfaces && ctx->nb_surfaces > 0)
+ {
av_log(avctx, AV_LOG_WARNING,
"Defined rc_lookahead requires more surfaces, "
"increasing used surfaces %d -> %d\n", ctx->nb_surfaces, nb_surfaces);
- ctx->nb_surfaces = nb_surfaces;
}
+ ctx->nb_surfaces = FFMAX(nb_surfaces, ctx->nb_surfaces);
+ } else {
+ if (ctx->encode_config.frameIntervalP > 1 && ctx->nb_surfaces < nb_surfaces && ctx->nb_surfaces > 0)
+ {
+ av_log(avctx, AV_LOG_WARNING,
+ "Defined b-frame requires more surfaces, "
+ "increasing used surfaces %d -> %d\n", ctx->nb_surfaces, nb_surfaces);
+ ctx->nb_surfaces = FFMAX(ctx->nb_surfaces, nb_surfaces);
+ }
+ else if (ctx->nb_surfaces <= 0)
+ ctx->nb_surfaces = nb_surfaces;
+ // otherwise use user specified value
}
ctx->nb_surfaces = FFMAX(1, FFMIN(MAX_REGISTERED_FRAMES, ctx->nb_surfaces));
@@ -1086,6 +1104,7 @@
NvencContext *ctx = avctx->priv_data;
NvencDynLoadFunctions *dl_fn = &ctx->nvenc_dload_funcs;
NV_ENCODE_API_FUNCTION_LIST *p_nvenc = &dl_fn->nvenc_funcs;
+ NvencSurface* tmp_surface = &ctx->surfaces[idx];
NVENCSTATUS nv_status;
NV_ENC_CREATE_BITSTREAM_BUFFER allocOut = { 0 };
@@ -1121,8 +1140,6 @@
ctx->surfaces[idx].height = allocSurf.height;
}
- ctx->surfaces[idx].lockCount = 0;
-
/* 1MB is large enough to hold most output frames.
* NVENC increases this automaticaly if it is not enough. */
allocOut.size = 1024 * 1024;
@@ -1141,6 +1158,8 @@
ctx->surfaces[idx].output_surface = allocOut.bitstreamBuffer;
ctx->surfaces[idx].size = allocOut.size;
+ av_fifo_generic_write(ctx->unused_surface_queue, &tmp_surface, sizeof(tmp_surface), NULL);
+
return 0;
}
@@ -1156,6 +1175,11 @@
ctx->timestamp_list = av_fifo_alloc(ctx->nb_surfaces * sizeof(int64_t));
if (!ctx->timestamp_list)
return AVERROR(ENOMEM);
+
+ ctx->unused_surface_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(NvencSurface*));
+ if (!ctx->unused_surface_queue)
+ return AVERROR(ENOMEM);
+
ctx->output_surface_queue = av_fifo_alloc(ctx->nb_surfaces * sizeof(NvencSurface*));
if (!ctx->output_surface_queue)
return AVERROR(ENOMEM);
@@ -1222,6 +1246,7 @@
av_fifo_freep(&ctx->timestamp_list);
av_fifo_freep(&ctx->output_surface_ready_queue);
av_fifo_freep(&ctx->output_surface_queue);
+ av_fifo_freep(&ctx->unused_surface_queue);
if (ctx->surfaces && avctx->pix_fmt == AV_PIX_FMT_CUDA) {
for (i = 0; i < ctx->nb_surfaces; ++i) {
@@ -1305,16 +1330,14 @@
static NvencSurface *get_free_frame(NvencContext *ctx)
{
- int i;
+ NvencSurface *tmp_surf;
- for (i = 0; i < ctx->nb_surfaces; i++) {
- if (!ctx->surfaces[i].lockCount) {
- ctx->surfaces[i].lockCount = 1;
- return &ctx->surfaces[i];
- }
- }
+ if (!(av_fifo_size(ctx->unused_surface_queue) > 0))
+ // queue empty
+ return NULL;
- return NULL;
+ av_fifo_generic_read(ctx->unused_surface_queue, &tmp_surf, sizeof(tmp_surf), NULL);
+ return tmp_surf;
}
static int nvenc_copy_frame(AVCodecContext *avctx, NvencSurface *nv_surface,
@@ -1712,7 +1735,6 @@
}
if (res) {
- inSurf->lockCount = 0;
return res;
}
@@ -1790,8 +1812,7 @@
if (res)
return res;
- av_assert0(tmpoutsurf->lockCount);
- tmpoutsurf->lockCount--;
+ av_fifo_generic_write(ctx->unused_surface_queue, &tmpoutsurf, sizeof(tmpoutsurf), NULL);
*got_packet = 1;
} else {
