Vidyo patch: Rate control for SVC, 1 pass CBR mode.
-Make Rate control work for SVC 1 pass CBR mode.
-Added temporal layering mode.
-Fixed bug in non-rd variance partition.
-Modified/updated the sample encoders (vp9_spatial_svc_encoder, vpx_temporal_svc_encoder).
-Added datarate unittest(s) for 1 pass CBR SVC.
Change-Id: Ie94b1b68a56ea1267b5087c625e5df04def2ee48
diff --git a/examples/vp9_spatial_svc_encoder.c b/examples/vp9_spatial_svc_encoder.c
index a3fd65f..c22a04a 100644
--- a/examples/vp9_spatial_svc_encoder.c
+++ b/examples/vp9_spatial_svc_encoder.c
@@ -14,11 +14,13 @@
* that benefit from a scalable bitstream.
*/
+#include <math.h>
#include <stdarg.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
+
#include "../args.h"
#include "../tools_common.h"
#include "../video_writer.h"
@@ -27,11 +29,18 @@
#include "vpx/vp8cx.h"
#include "vpx/vpx_encoder.h"
#include "../vpxstats.h"
+#define OUTPUT_RC_STATS 1
static const arg_def_t skip_frames_arg =
ARG_DEF("s", "skip-frames", 1, "input frames to skip");
static const arg_def_t frames_arg =
ARG_DEF("f", "frames", 1, "number of frames to encode");
+static const arg_def_t threads_arg =
+ ARG_DEF("th", "threads", 1, "number of threads to use");
+#if OUTPUT_RC_STATS
+static const arg_def_t output_rc_stats_arg =
+ ARG_DEF("rcstat", "output_rc_stats", 1, "output rc stats");
+#endif
static const arg_def_t width_arg = ARG_DEF("w", "width", 1, "source width");
static const arg_def_t height_arg = ARG_DEF("h", "height", 1, "source height");
static const arg_def_t timebase_arg =
@@ -42,6 +51,9 @@
ARG_DEF("sl", "spatial-layers", 1, "number of spatial SVC layers");
static const arg_def_t temporal_layers_arg =
ARG_DEF("tl", "temporal-layers", 1, "number of temporal SVC layers");
+static const arg_def_t temporal_layering_mode_arg =
+ ARG_DEF("tlm", "temporal-layering-mode", 1, "temporal layering scheme."
+ "VP9E_TEMPORAL_LAYERING_MODE");
static const arg_def_t kf_dist_arg =
ARG_DEF("k", "kf-dist", 1, "number of frames between keyframes");
static const arg_def_t scale_factors_arg =
@@ -65,6 +77,8 @@
"generating any outputs");
static const arg_def_t rc_end_usage_arg =
ARG_DEF(NULL, "rc-end-usage", 1, "0 - 3: VBR, CBR, CQ, Q");
+static const arg_def_t speed_arg =
+ ARG_DEF("sp", "speed", 1, "speed configuration");
#if CONFIG_VP9_HIGHBITDEPTH
static const struct arg_enum_list bitdepth_enum[] = {
@@ -85,10 +99,16 @@
&timebase_arg, &bitrate_arg, &skip_frames_arg, &spatial_layers_arg,
&kf_dist_arg, &scale_factors_arg, &passes_arg, &pass_arg,
&fpf_name_arg, &min_q_arg, &max_q_arg, &min_bitrate_arg,
- &max_bitrate_arg, &temporal_layers_arg, &lag_in_frame_arg,
+ &max_bitrate_arg, &temporal_layers_arg, &temporal_layering_mode_arg,
+ &lag_in_frame_arg, &threads_arg,
+#if OUTPUT_RC_STATS
+ &output_rc_stats_arg,
+#endif
+
#if CONFIG_VP9_HIGHBITDEPTH
&bitdepth_arg,
#endif
+ &speed_arg,
&rc_end_usage_arg, NULL
};
@@ -102,6 +122,10 @@
static const uint32_t default_spatial_layers = 5;
static const uint32_t default_temporal_layers = 1;
static const uint32_t default_kf_dist = 100;
+static const uint32_t default_temporal_layering_mode = 0;
+static const uint32_t default_output_rc_stats = 0;
+static const int32_t default_speed = -1; // -1 means use library default.
+static const uint32_t default_threads = 0; // zero means use library default.
typedef struct {
const char *input_filename;
@@ -143,6 +167,12 @@
svc_ctx->log_level = SVC_LOG_DEBUG;
svc_ctx->spatial_layers = default_spatial_layers;
svc_ctx->temporal_layers = default_temporal_layers;
+ svc_ctx->temporal_layering_mode = default_temporal_layering_mode;
+#if OUTPUT_RC_STATS
+ svc_ctx->output_rc_stat = default_output_rc_stats;
+#endif
+ svc_ctx->speed = default_speed;
+ svc_ctx->threads = default_threads;
// start with default encoder configuration
res = vpx_codec_enc_config_default(vpx_codec_vp9_cx(), enc_cfg, 0);
@@ -184,6 +214,20 @@
svc_ctx->spatial_layers = arg_parse_uint(&arg);
} else if (arg_match(&arg, &temporal_layers_arg, argi)) {
svc_ctx->temporal_layers = arg_parse_uint(&arg);
+#if OUTPUT_RC_STATS
+ } else if (arg_match(&arg, &output_rc_stats_arg, argi)) {
+ svc_ctx->output_rc_stat = arg_parse_uint(&arg);
+#endif
+ } else if (arg_match(&arg, &speed_arg, argi)) {
+ svc_ctx->speed = arg_parse_uint(&arg);
+ } else if (arg_match(&arg, &threads_arg, argi)) {
+ svc_ctx->threads = arg_parse_uint(&arg);
+ } else if (arg_match(&arg, &temporal_layering_mode_arg, argi)) {
+ svc_ctx->temporal_layering_mode =
+ enc_cfg->temporal_layering_mode = arg_parse_int(&arg);
+ if (svc_ctx->temporal_layering_mode) {
+ enc_cfg->g_error_resilient = 1;
+ }
} else if (arg_match(&arg, &kf_dist_arg, argi)) {
enc_cfg->kf_min_dist = arg_parse_uint(&arg);
enc_cfg->kf_max_dist = enc_cfg->kf_min_dist;
@@ -316,6 +360,185 @@
enc_cfg->rc_target_bitrate, enc_cfg->kf_max_dist);
}
+#if OUTPUT_RC_STATS
+// For rate control encoding stats.
+struct RateControlStats {
+ // Number of input frames per layer.
+ int layer_input_frames[VPX_MAX_LAYERS];
+ // Total (cumulative) number of encoded frames per layer.
+ int layer_tot_enc_frames[VPX_MAX_LAYERS];
+ // Number of encoded non-key frames per layer.
+ int layer_enc_frames[VPX_MAX_LAYERS];
+ // Framerate per layer (cumulative).
+ double layer_framerate[VPX_MAX_LAYERS];
+ // Target average frame size per layer (per-frame-bandwidth per layer).
+ double layer_pfb[VPX_MAX_LAYERS];
+ // Actual average frame size per layer.
+ double layer_avg_frame_size[VPX_MAX_LAYERS];
+ // Average rate mismatch per layer (|target - actual| / target).
+ double layer_avg_rate_mismatch[VPX_MAX_LAYERS];
+ // Actual encoding bitrate per layer (cumulative).
+ double layer_encoding_bitrate[VPX_MAX_LAYERS];
+ // Average of the short-time encoder actual bitrate.
+ // TODO(marpan): Should we add these short-time stats for each layer?
+ double avg_st_encoding_bitrate;
+ // Variance of the short-time encoder actual bitrate.
+ double variance_st_encoding_bitrate;
+ // Window (number of frames) for computing short-time encoding bitrate.
+ int window_size;
+ // Number of window measurements.
+ int window_count;
+};
+
+// Note: these rate control stats assume only 1 key frame in the
+// sequence (i.e., first frame only).
+static void set_rate_control_stats(struct RateControlStats *rc,
+ vpx_codec_enc_cfg_t *cfg) {
+ unsigned int sl, tl;
+ // Set the layer (cumulative) framerate and the target layer (non-cumulative)
+ // per-frame-bandwidth, for the rate control encoding stats below.
+ const double framerate = cfg->g_timebase.den / cfg->g_timebase.num;
+
+ for (sl = 0; sl < cfg->ss_number_layers; ++sl) {
+ for (tl = 0; tl < cfg->ts_number_layers; ++tl) {
+ const int layer = sl * cfg->ts_number_layers + tl;
+ const int tlayer0 = sl * cfg->ts_number_layers;
+ rc->layer_framerate[layer] =
+ framerate / cfg->ts_rate_decimator[tl];
+ if (tl > 0) {
+ rc->layer_pfb[layer] = 1000.0 *
+ (cfg->layer_target_bitrate[layer] -
+ cfg->layer_target_bitrate[layer - 1]) /
+ (rc->layer_framerate[layer] -
+ rc->layer_framerate[layer - 1]);
+ } else {
+ rc->layer_pfb[tlayer0] = 1000.0 *
+ cfg->layer_target_bitrate[tlayer0] /
+ rc->layer_framerate[tlayer0];
+ }
+ rc->layer_input_frames[layer] = 0;
+ rc->layer_enc_frames[layer] = 0;
+ rc->layer_tot_enc_frames[layer] = 0;
+ rc->layer_encoding_bitrate[layer] = 0.0;
+ rc->layer_avg_frame_size[layer] = 0.0;
+ rc->layer_avg_rate_mismatch[layer] = 0.0;
+ }
+ }
+ rc->window_count = 0;
+ rc->window_size = 15;
+ rc->avg_st_encoding_bitrate = 0.0;
+ rc->variance_st_encoding_bitrate = 0.0;
+}
+
+static void printout_rate_control_summary(struct RateControlStats *rc,
+ vpx_codec_enc_cfg_t *cfg,
+ int frame_cnt) {
+ unsigned int sl, tl;
+ int tot_num_frames = 0;
+ double perc_fluctuation = 0.0;
+ printf("Total number of processed frames: %d\n\n", frame_cnt - 1);
+ printf("Rate control layer stats for sl%d tl%d layer(s):\n\n",
+ cfg->ss_number_layers, cfg->ts_number_layers);
+ for (sl = 0; sl < cfg->ss_number_layers; ++sl) {
+ for (tl = 0; tl < cfg->ts_number_layers; ++tl) {
+ const int layer = sl * cfg->ts_number_layers + tl;
+ const int num_dropped = (tl > 0) ?
+ (rc->layer_input_frames[layer] - rc->layer_enc_frames[layer]) :
+ (rc->layer_input_frames[layer] - rc->layer_enc_frames[layer] - 1);
+ if (!sl)
+ tot_num_frames += rc->layer_input_frames[layer];
+ rc->layer_encoding_bitrate[layer] = 0.001 * rc->layer_framerate[layer] *
+ rc->layer_encoding_bitrate[layer] / tot_num_frames;
+ rc->layer_avg_frame_size[layer] = rc->layer_avg_frame_size[layer] /
+ rc->layer_enc_frames[layer];
+ rc->layer_avg_rate_mismatch[layer] =
+ 100.0 * rc->layer_avg_rate_mismatch[layer] /
+ rc->layer_enc_frames[layer];
+ printf("For layer#: sl%d tl%d \n", sl, tl);
+ printf("Bitrate (target vs actual): %d %f.0 kbps\n",
+ cfg->layer_target_bitrate[layer],
+ rc->layer_encoding_bitrate[layer]);
+ printf("Average frame size (target vs actual): %f %f bits\n",
+ rc->layer_pfb[layer], rc->layer_avg_frame_size[layer]);
+ printf("Average rate_mismatch: %f\n",
+ rc->layer_avg_rate_mismatch[layer]);
+ printf("Number of input frames, encoded (non-key) frames, "
+ "and percent dropped frames: %d %d %f.0 \n",
+ rc->layer_input_frames[layer], rc->layer_enc_frames[layer],
+ 100.0 * num_dropped / rc->layer_input_frames[layer]);
+ printf("\n");
+ }
+ }
+ rc->avg_st_encoding_bitrate = rc->avg_st_encoding_bitrate / rc->window_count;
+ rc->variance_st_encoding_bitrate =
+ rc->variance_st_encoding_bitrate / rc->window_count -
+ (rc->avg_st_encoding_bitrate * rc->avg_st_encoding_bitrate);
+ perc_fluctuation = 100.0 * sqrt(rc->variance_st_encoding_bitrate) /
+ rc->avg_st_encoding_bitrate;
+ printf("Short-time stats, for window of %d frames: \n", rc->window_size);
+ printf("Average, rms-variance, and percent-fluct: %f %f %f \n",
+ rc->avg_st_encoding_bitrate,
+ sqrt(rc->variance_st_encoding_bitrate),
+ perc_fluctuation);
+ if (frame_cnt != tot_num_frames)
+ die("Error: Number of input frames not equal to output encoded frames != "
+ "%d tot_num_frames = %d\n", frame_cnt, tot_num_frames);
+}
+
+vpx_codec_err_t parse_superframe_index(const uint8_t *data,
+ size_t data_sz,
+ uint32_t sizes[8], int *count) {
+ // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+ // it is a super frame index. If the last byte of real video compression
+ // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+ // not the associated matching marker byte at the front of the index we have
+ // an invalid bitstream and need to return an error.
+
+ uint8_t marker;
+
+ marker = *(data + data_sz - 1);
+ *count = 0;
+
+
+ if ((marker & 0xe0) == 0xc0) {
+ const uint32_t frames = (marker & 0x7) + 1;
+ const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+ const size_t index_sz = 2 + mag * frames;
+
+ // This chunk is marked as having a superframe index but doesn't have
+ // enough data for it, thus it's an invalid superframe index.
+ if (data_sz < index_sz)
+ return VPX_CODEC_CORRUPT_FRAME;
+
+ {
+ const uint8_t marker2 = *(data + data_sz - index_sz);
+
+ // This chunk is marked as having a superframe index but doesn't have
+ // the matching marker byte at the front of the index therefore it's an
+ // invalid chunk.
+ if (marker != marker2)
+ return VPX_CODEC_CORRUPT_FRAME;
+ }
+
+ {
+ // Found a valid superframe index.
+ uint32_t i, j;
+ const uint8_t *x = &data[data_sz - index_sz + 1];
+
+ for (i = 0; i < frames; ++i) {
+ uint32_t this_sz = 0;
+
+ for (j = 0; j < mag; ++j)
+ this_sz |= (*x++) << (j * 8);
+ sizes[i] = this_sz;
+ }
+ *count = frames;
+ }
+ }
+ return VPX_CODEC_OK;
+}
+#endif
+
int main(int argc, const char **argv) {
AppInput app_input = {0};
VpxVideoWriter *writer = NULL;
@@ -332,7 +555,15 @@
FILE *infile = NULL;
int end_of_stream = 0;
int frames_received = 0;
-
+#if OUTPUT_RC_STATS
+ VpxVideoWriter *outfile[VPX_TS_MAX_LAYERS] = {NULL};
+ struct RateControlStats rc;
+ vpx_svc_layer_id_t layer_id;
+ int sl, tl;
+ double sum_bitrate = 0.0;
+ double sum_bitrate2 = 0.0;
+ double framerate = 30.0;
+#endif
memset(&svc_ctx, 0, sizeof(svc_ctx));
svc_ctx.log_print = 1;
exec_name = argv[0];
@@ -359,6 +590,13 @@
VPX_CODEC_OK)
die("Failed to initialize encoder\n");
+#if OUTPUT_RC_STATS
+ if (svc_ctx.output_rc_stat) {
+ set_rate_control_stats(&rc, &enc_cfg);
+ framerate = enc_cfg.g_timebase.den / enc_cfg.g_timebase.num;
+ }
+#endif
+
info.codec_fourcc = VP9_FOURCC;
info.time_base.numerator = enc_cfg.g_timebase.num;
info.time_base.denominator = enc_cfg.g_timebase.den;
@@ -370,11 +608,30 @@
if (!writer)
die("Failed to open %s for writing\n", app_input.output_filename);
}
+#if OUTPUT_RC_STATS
+ // For now, just write temporal layer streams.
+ // TODO(wonkap): do spatial by re-writing superframe.
+ if (svc_ctx.output_rc_stat) {
+ for (tl = 0; tl < enc_cfg.ts_number_layers; ++tl) {
+ char file_name[PATH_MAX];
+
+ snprintf(file_name, sizeof(file_name), "%s_t%d.ivf",
+ app_input.output_filename, tl);
+ outfile[tl] = vpx_video_writer_open(file_name, kContainerIVF, &info);
+ if (!outfile[tl])
+ die("Failed to open %s for writing", file_name);
+ }
+ }
+#endif
// skip initial frames
for (i = 0; i < app_input.frames_to_skip; ++i)
vpx_img_read(&raw, infile);
+ if (svc_ctx.speed != -1)
+ vpx_codec_control(&codec, VP8E_SET_CPUUSED, svc_ctx.speed);
+ vpx_codec_control(&codec, VP9E_SET_TILE_COLUMNS, 0);
+
// Encode frames
while (!end_of_stream) {
vpx_codec_iter_t iter = NULL;
@@ -386,7 +643,9 @@
}
res = vpx_svc_encode(&svc_ctx, &codec, (end_of_stream ? NULL : &raw),
- pts, frame_duration, VPX_DL_GOOD_QUALITY);
+ pts, frame_duration, svc_ctx.speed >= 5 ?
+ VPX_DL_REALTIME : VPX_DL_GOOD_QUALITY);
+
printf("%s", vpx_svc_get_message(&svc_ctx));
if (res != VPX_CODEC_OK) {
die_codec(&codec, "Failed to encode frame");
@@ -395,11 +654,90 @@
while ((cx_pkt = vpx_codec_get_cx_data(&codec, &iter)) != NULL) {
switch (cx_pkt->kind) {
case VPX_CODEC_CX_FRAME_PKT: {
- if (cx_pkt->data.frame.sz > 0)
+ if (cx_pkt->data.frame.sz > 0) {
+#if OUTPUT_RC_STATS
+ uint32_t sizes[8];
+ int count = 0;
+#endif
vpx_video_writer_write_frame(writer,
cx_pkt->data.frame.buf,
cx_pkt->data.frame.sz,
cx_pkt->data.frame.pts);
+#if OUTPUT_RC_STATS
+ // TODO(marpan/wonkap): Put this (to line728) in separate function.
+ if (svc_ctx.output_rc_stat) {
+ vpx_codec_control(&codec, VP9E_GET_SVC_LAYER_ID, &layer_id);
+ parse_superframe_index(cx_pkt->data.frame.buf,
+ cx_pkt->data.frame.sz, sizes, &count);
+ for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+ ++rc.layer_input_frames[sl * enc_cfg.ts_number_layers +
+ layer_id.temporal_layer_id];
+ }
+ for (tl = layer_id.temporal_layer_id;
+ tl < enc_cfg.ts_number_layers; ++tl) {
+ vpx_video_writer_write_frame(outfile[tl],
+ cx_pkt->data.frame.buf,
+ cx_pkt->data.frame.sz,
+ cx_pkt->data.frame.pts);
+ }
+
+ for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+ for (tl = layer_id.temporal_layer_id;
+ tl < enc_cfg.ts_number_layers; ++tl) {
+ const int layer = sl * enc_cfg.ts_number_layers + tl;
+ ++rc.layer_tot_enc_frames[layer];
+ rc.layer_encoding_bitrate[layer] += 8.0 * sizes[sl];
+ // Keep count of rate control stats per layer, for non-key
+ // frames.
+ if (tl == layer_id.temporal_layer_id &&
+ !(cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY)) {
+ rc.layer_avg_frame_size[layer] += 8.0 * sizes[sl];
+ rc.layer_avg_rate_mismatch[layer] +=
+ fabs(8.0 * sizes[sl] - rc.layer_pfb[layer]) /
+ rc.layer_pfb[layer];
+ ++rc.layer_enc_frames[layer];
+ }
+ }
+ }
+
+ // Update for short-time encoding bitrate states, for moving
+ // window of size rc->window, shifted by rc->window / 2.
+ // Ignore first window segment, due to key frame.
+ if (frame_cnt > rc.window_size) {
+ tl = layer_id.temporal_layer_id;
+ for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+ sum_bitrate += 0.001 * 8.0 * sizes[sl] * framerate;
+ }
+ if (frame_cnt % rc.window_size == 0) {
+ rc.window_count += 1;
+ rc.avg_st_encoding_bitrate += sum_bitrate / rc.window_size;
+ rc.variance_st_encoding_bitrate +=
+ (sum_bitrate / rc.window_size) *
+ (sum_bitrate / rc.window_size);
+ sum_bitrate = 0.0;
+ }
+ }
+
+ // Second shifted window.
+ if (frame_cnt > rc.window_size + rc.window_size / 2) {
+ tl = layer_id.temporal_layer_id;
+ for (sl = 0; sl < enc_cfg.ss_number_layers; ++sl) {
+ sum_bitrate2 += 0.001 * 8.0 * sizes[sl] * framerate;
+ }
+
+ if (frame_cnt > 2 * rc.window_size &&
+ frame_cnt % rc.window_size == 0) {
+ rc.window_count += 1;
+ rc.avg_st_encoding_bitrate += sum_bitrate2 / rc.window_size;
+ rc.variance_st_encoding_bitrate +=
+ (sum_bitrate2 / rc.window_size) *
+ (sum_bitrate2 / rc.window_size);
+ sum_bitrate2 = 0.0;
+ }
+ }
+ }
+#endif
+ }
printf("SVC frame: %d, kf: %d, size: %d, pts: %d\n", frames_received,
!!(cx_pkt->data.frame.flags & VPX_FRAME_IS_KEY),
@@ -424,25 +762,30 @@
pts += frame_duration;
}
}
-
printf("Processed %d frames\n", frame_cnt);
-
fclose(infile);
+#if OUTPUT_RC_STATS
+ if (svc_ctx.output_rc_stat) {
+ printout_rate_control_summary(&rc, &enc_cfg, frame_cnt);
+ printf("\n");
+ }
+#endif
if (vpx_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec");
-
if (app_input.passes == 2)
stats_close(&app_input.rc_stats, 1);
-
if (writer) {
vpx_video_writer_close(writer);
}
-
+#if OUTPUT_RC_STATS
+ if (svc_ctx.output_rc_stat) {
+ for (tl = 0; tl < enc_cfg.ts_number_layers; ++tl) {
+ vpx_video_writer_close(outfile[tl]);
+ }
+ }
+#endif
vpx_img_free(&raw);
-
// display average size, psnr
printf("%s", vpx_svc_dump_statistics(&svc_ctx));
-
vpx_svc_release(&svc_ctx);
-
return EXIT_SUCCESS;
}
