Commit 7cb7588b authored by Marco's avatar Marco Committed by Gerrit Code Review

Merge "Various updates to vp8."

parents 33e61df7 af898b56
......@@ -8,292 +8,730 @@
* be found in the AUTHORS file in the root of the source tree.
*/
/*
* This is an example demonstrating multi-resolution encoding in VP8.
* High-resolution input video is down-sampled to lower-resolutions. The
* encoder then encodes the video and outputs multiple bitstreams with
* different resolutions.
*
* This test also allows for settings temporal layers for each spatial layer.
* Different number of temporal layers per spatial stream may be used.
* Currently up to 3 temporal layers per spatial stream (encoder) are supported
* in this test.
*/
// This is an example demonstrating multi-resolution encoding in VP8.
// High-resolution input video is down-sampled to lower-resolutions. The
// encoder then encodes the video and outputs multiple bitstreams with
// different resolutions.
//
// Configure with --enable-multi-res-encoding flag to enable this example.
#include "./vpx_config.h"
#include <stdio.h>
#include <stdlib.h>
#include <stdarg.h>
#include <string.h>
#include <math.h>
#include <assert.h>
#include <sys/time.h>
#if USE_POSIX_MMAP
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <unistd.h>
#endif
#include "vpx_ports/vpx_timer.h"
#define VPX_CODEC_DISABLE_COMPAT 1
#include "vpx/vpx_encoder.h"
#include "vpx/vp8cx.h"
#include "vpx_ports/mem_ops.h"
#include "./tools_common.h"
#define interface (vpx_codec_vp8_cx())
#define fourcc 0x30385056
void usage_exit() {
exit(EXIT_FAILURE);
}
/*
* The input video frame is downsampled several times to generate a multi-level
* hierarchical structure. NUM_ENCODERS is defined as the number of encoding
* levels required. For example, if the size of input video is 1280x720,
* NUM_ENCODERS is 3, and down-sampling factor is 2, the encoder outputs 3
* bitstreams with resolution of 1280x720(level 0), 640x360(level 1), and
* 320x180(level 2) respectively.
*/
/* Number of encoders (spatial resolutions) used in this test. */
#define NUM_ENCODERS 3
/* Maximum number of temporal layers allowed for this test. */
#define MAX_NUM_TEMPORAL_LAYERS 3
/* This example uses the scaler function in libyuv. */
#include "third_party/libyuv/include/libyuv/basic_types.h"
#include "third_party/libyuv/include/libyuv/scale.h"
#include "third_party/libyuv/include/libyuv/cpu_id.h"
#include "vpx/vpx_encoder.h"
#include "vpx/vp8cx.h"
int (*read_frame_p)(FILE *f, vpx_image_t *img);
#include "./tools_common.h"
#include "./video_writer.h"
static int read_frame(FILE *f, vpx_image_t *img) {
size_t nbytes, to_read;
int res = 1;
// The input video frame is downsampled several times to generate a
// multi-level hierarchical structure. kNumEncoders is defined as the number
// of encoding levels required. For example, if the size of input video is
// 1280x720, kNumEncoders is 3, and down-sampling factor is 2, the encoder
// outputs 3 bitstreams with resolution of 1280x720(level 0),
// 640x360(level 1), and 320x180(level 2) respectively.
#define kNumEncoders 3
to_read = img->w*img->h*3/2;
nbytes = fread(img->planes[0], 1, to_read, f);
if(nbytes != to_read) {
res = 0;
if(nbytes > 0)
printf("Warning: Read partial frame. Check your width & height!\n");
}
return res;
}
static const char *exec_name;
static int read_frame_by_row(FILE *f, vpx_image_t *img) {
size_t nbytes, to_read;
int res = 1;
int plane;
void usage_exit() {
fprintf(stderr,
"Usage: %s <width> <height> <infile> <outfile(s)> <output psnr?>\n",
exec_name);
exit(EXIT_FAILURE);
for (plane = 0; plane < 3; plane++)
{
unsigned char *ptr;
int w = (plane ? (1 + img->d_w) / 2 : img->d_w);
int h = (plane ? (1 + img->d_h) / 2 : img->d_h);
int r;
/* Determine the correct plane based on the image format. The for-loop
* always counts in Y,U,V order, but this may not match the order of
* the data on disk.
*/
switch (plane)
{
case 1:
ptr = img->planes[img->fmt==VPX_IMG_FMT_YV12? VPX_PLANE_V : VPX_PLANE_U];
break;
case 2:
ptr = img->planes[img->fmt==VPX_IMG_FMT_YV12?VPX_PLANE_U : VPX_PLANE_V];
break;
default:
ptr = img->planes[plane];
}
for (r = 0; r < h; r++)
{
to_read = w;
nbytes = fread(ptr, 1, to_read, f);
if(nbytes != to_read) {
res = 0;
if(nbytes > 0)
printf("Warning: Read partial frame. Check your width & height!\n");
break;
}
ptr += img->stride[plane];
}
if (!res)
break;
}
return res;
}
int main(int argc, char *argv[]) {
int frame_cnt = 0;
FILE *infile = NULL;
VpxVideoWriter *writers[kNumEncoders];
vpx_codec_ctx_t codec[kNumEncoders];
vpx_codec_enc_cfg_t cfg[kNumEncoders];
vpx_image_t raw[kNumEncoders];
const VpxInterface *const encoder = get_vpx_encoder_by_name("vp8");
// Currently, only realtime mode is supported in multi-resolution encoding.
const int arg_deadline = VPX_DL_REALTIME;
int i;
int width = 0;
int height = 0;
int frame_avail = 0;
int got_data = 0;
// Set show_psnr to 1/0 to show/not show PSNR. Choose show_psnr=0 if you
// don't need to know PSNR, which will skip PSNR calculation and save
// encoding time.
int show_psnr = 0;
uint64_t psnr_sse_total[kNumEncoders] = {0};
uint64_t psnr_samples_total[kNumEncoders] = {0};
double psnr_totals[kNumEncoders][4] = {{0, 0}};
int psnr_count[kNumEncoders] = {0};
// Set the required target bitrates for each resolution level.
// If target bitrate for highest-resolution level is set to 0,
// (i.e. target_bitrate[0]=0), we skip encoding at that level.
unsigned int target_bitrate[kNumEncoders] = {1000, 500, 100};
// Enter the frame rate of the input video.
const int framerate = 30;
// Set down-sampling factor for each resolution level.
// dsf[0] controls down sampling from level 0 to level 1;
// dsf[1] controls down sampling from level 1 to level 2;
// dsf[2] is not used.
vpx_rational_t dsf[kNumEncoders] = {{2, 1}, {2, 1}, {1, 1}};
exec_name = argv[0];
if (!encoder)
die("Unsupported codec.");
// exe_name, input width, input height, input file,
// output file 1, output file 2, output file 3, psnr on/off
if (argc != (5 + kNumEncoders))
die("Invalid number of input options.");
printf("Using %s\n", vpx_codec_iface_name(encoder->codec_interface()));
width = strtol(argv[1], NULL, 0);
height = strtol(argv[2], NULL, 0);
if (width < 16 || width % 2 || height < 16 || height % 2)
die("Invalid resolution: %ldx%ld", width, height);
// Open input video file for encoding
if (!(infile = fopen(argv[3], "rb")))
die("Failed to open %s for reading", argv[3]);
show_psnr = strtol(argv[kNumEncoders + 4], NULL, 0);
// Populate default encoder configuration
for (i = 0; i < kNumEncoders; ++i) {
vpx_codec_err_t res =
vpx_codec_enc_config_default(encoder->codec_interface(), &cfg[i], 0);
if (res != VPX_CODEC_OK) {
printf("Failed to get config: %s\n", vpx_codec_err_to_string(res));
return EXIT_FAILURE;
static void write_ivf_file_header(FILE *outfile,
const vpx_codec_enc_cfg_t *cfg,
int frame_cnt) {
char header[32];
if(cfg->g_pass != VPX_RC_ONE_PASS && cfg->g_pass != VPX_RC_LAST_PASS)
return;
header[0] = 'D';
header[1] = 'K';
header[2] = 'I';
header[3] = 'F';
mem_put_le16(header+4, 0); /* version */
mem_put_le16(header+6, 32); /* headersize */
mem_put_le32(header+8, fourcc); /* headersize */
mem_put_le16(header+12, cfg->g_w); /* width */
mem_put_le16(header+14, cfg->g_h); /* height */
mem_put_le32(header+16, cfg->g_timebase.den); /* rate */
mem_put_le32(header+20, cfg->g_timebase.num); /* scale */
mem_put_le32(header+24, frame_cnt); /* length */
mem_put_le32(header+28, 0); /* unused */
(void) fwrite(header, 1, 32, outfile);
}
static void write_ivf_frame_header(FILE *outfile,
const vpx_codec_cx_pkt_t *pkt)
{
char header[12];
vpx_codec_pts_t pts;
if(pkt->kind != VPX_CODEC_CX_FRAME_PKT)
return;
pts = pkt->data.frame.pts;
mem_put_le32(header, pkt->data.frame.sz);
mem_put_le32(header+4, pts&0xFFFFFFFF);
mem_put_le32(header+8, pts >> 32);
(void) fwrite(header, 1, 12, outfile);
}
/* Temporal scaling parameters */
/* This sets all the temporal layer parameters given |num_temporal_layers|,
* including the target bit allocation across temporal layers. Bit allocation
* parameters will be passed in as user parameters in another version.
*/
static void set_temporal_layer_pattern(int num_temporal_layers,
vpx_codec_enc_cfg_t *cfg,
int bitrate,
int *layer_flags)
{
assert(num_temporal_layers <= MAX_NUM_TEMPORAL_LAYERS);
switch (num_temporal_layers)
{
case 1:
{
/* 1-layer */
cfg->ts_number_layers = 1;
cfg->ts_periodicity = 1;
cfg->ts_rate_decimator[0] = 1;
cfg->ts_layer_id[0] = 0;
cfg->ts_target_bitrate[0] = bitrate;
// Update L only.
layer_flags[0] = VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF;
break;
}
case 2:
{
/* 2-layers, with sync point at first frame of layer 1. */
cfg->ts_number_layers = 2;
cfg->ts_periodicity = 2;
cfg->ts_rate_decimator[0] = 2;
cfg->ts_rate_decimator[1] = 1;
cfg->ts_layer_id[0] = 0;
cfg->ts_layer_id[1] = 1;
// Use 60/40 bit allocation as example.
cfg->ts_target_bitrate[0] = 0.6f * bitrate;
cfg->ts_target_bitrate[1] = bitrate;
/* 0=L, 1=GF */
// ARF is used as predictor for all frames, and is only updated on
// key frame. Sync point every 8 frames.
// Layer 0: predict from L and ARF, update L and G.
layer_flags[0] = VP8_EFLAG_NO_REF_GF |
VP8_EFLAG_NO_UPD_ARF;
// Layer 1: sync point: predict from L and ARF, and update G.
layer_flags[1] = VP8_EFLAG_NO_REF_GF |
VP8_EFLAG_NO_UPD_LAST |
VP8_EFLAG_NO_UPD_ARF;
// Layer 0, predict from L and ARF, update L.
layer_flags[2] = VP8_EFLAG_NO_REF_GF |
VP8_EFLAG_NO_UPD_GF |
VP8_EFLAG_NO_UPD_ARF;
// Layer 1: predict from L, G and ARF, and update G.
layer_flags[3] = VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_UPD_LAST |
VP8_EFLAG_NO_UPD_ENTROPY;
// Layer 0
layer_flags[4] = layer_flags[2];
// Layer 1
layer_flags[5] = layer_flags[3];
// Layer 0
layer_flags[6] = layer_flags[4];
// Layer 1
layer_flags[7] = layer_flags[5];
break;
}
}
// Update the default configuration according to needs of the application.
// Highest-resolution encoder settings
cfg[0].g_w = width;
cfg[0].g_h = height;
cfg[0].g_threads = 1;
cfg[0].rc_dropframe_thresh = 30;
cfg[0].rc_end_usage = VPX_CBR;
cfg[0].rc_resize_allowed = 0;
cfg[0].rc_min_quantizer = 4;
cfg[0].rc_max_quantizer = 56;
cfg[0].rc_undershoot_pct = 98;
cfg[0].rc_overshoot_pct = 100;
cfg[0].rc_buf_initial_sz = 500;
cfg[0].rc_buf_optimal_sz = 600;
cfg[0].rc_buf_sz = 1000;
cfg[0].g_error_resilient = 1;
cfg[0].g_lag_in_frames = 0;
cfg[0].kf_mode = VPX_KF_AUTO; // VPX_KF_DISABLED
cfg[0].kf_min_dist = 3000;
cfg[0].kf_max_dist = 3000;
cfg[0].rc_target_bitrate = target_bitrate[0];
cfg[0].g_timebase.num = 1;
cfg[0].g_timebase.den = framerate;
// Other-resolution encoder settings
for (i = 1; i < kNumEncoders; ++i) {
cfg[i] = cfg[0];
cfg[i].g_threads = 1;
cfg[i].rc_target_bitrate = target_bitrate[i];
// Note: Width & height of other-resolution encoders are calculated
// from the highest-resolution encoder's size and the corresponding
// down_sampling_factor.
case 3:
default:
{
unsigned int iw = cfg[i - 1].g_w * dsf[i - 1].den + dsf[i - 1].num - 1;
unsigned int ih = cfg[i - 1].g_h * dsf[i - 1].den + dsf[i - 1].num - 1;
cfg[i].g_w = iw / dsf[i - 1].num;
cfg[i].g_h = ih / dsf[i - 1].num;
// 3-layers structure where ARF is used as predictor for all frames,
// and is only updated on key frame.
// Sync points for layer 1 and 2 every 8 frames.
cfg->ts_number_layers = 3;
cfg->ts_periodicity = 4;
cfg->ts_rate_decimator[0] = 4;
cfg->ts_rate_decimator[1] = 2;
cfg->ts_rate_decimator[2] = 1;
cfg->ts_layer_id[0] = 0;
cfg->ts_layer_id[1] = 2;
cfg->ts_layer_id[2] = 1;
cfg->ts_layer_id[3] = 2;
// Use 40/20/40 bit allocation as example.
cfg->ts_target_bitrate[0] = 0.4f * bitrate;
cfg->ts_target_bitrate[1] = 0.6f * bitrate;
cfg->ts_target_bitrate[2] = bitrate;
/* 0=L, 1=GF, 2=ARF */
// Layer 0: predict from L and ARF; update L and G.
layer_flags[0] = VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_REF_GF;
// Layer 2: sync point: predict from L and ARF; update none.
layer_flags[1] = VP8_EFLAG_NO_REF_GF |
VP8_EFLAG_NO_UPD_GF |
VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_UPD_LAST |
VP8_EFLAG_NO_UPD_ENTROPY;
// Layer 1: sync point: predict from L and ARF; update G.
layer_flags[2] = VP8_EFLAG_NO_REF_GF |
VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_UPD_LAST;
// Layer 2: predict from L, G, ARF; update none.
layer_flags[3] = VP8_EFLAG_NO_UPD_GF |
VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_UPD_LAST |
VP8_EFLAG_NO_UPD_ENTROPY;
// Layer 0: predict from L and ARF; update L.
layer_flags[4] = VP8_EFLAG_NO_UPD_GF |
VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_REF_GF;
// Layer 2: predict from L, G, ARF; update none.
layer_flags[5] = layer_flags[3];
// Layer 1: predict from L, G, ARF; update G.
layer_flags[6] = VP8_EFLAG_NO_UPD_ARF |
VP8_EFLAG_NO_UPD_LAST;
// Layer 2: predict from L, G, ARF; update none.
layer_flags[7] = layer_flags[3];
break;
}
}
}
/* The periodicity of the pattern given the number of temporal layers. */
static int periodicity_to_num_layers[MAX_NUM_TEMPORAL_LAYERS] = {1, 8, 8};
int main(int argc, char **argv)
{
FILE *infile, *outfile[NUM_ENCODERS];
FILE *downsampled_input[NUM_ENCODERS - 1];
char filename[50];
vpx_codec_ctx_t codec[NUM_ENCODERS];
vpx_codec_enc_cfg_t cfg[NUM_ENCODERS];
int frame_cnt = 0;
vpx_image_t raw[NUM_ENCODERS];
vpx_codec_err_t res[NUM_ENCODERS];
int i;
long width;
long height;
int length_frame;
int frame_avail;
int got_data;
int flags = 0;
int layer_id = 0;
int layer_flags[VPX_TS_MAX_PERIODICITY * NUM_ENCODERS]
= {0};
int flag_periodicity;
/*Currently, only realtime mode is supported in multi-resolution encoding.*/
int arg_deadline = VPX_DL_REALTIME;
/* Set show_psnr to 1/0 to show/not show PSNR. Choose show_psnr=0 if you
don't need to know PSNR, which will skip PSNR calculation and save
encoding time. */
int show_psnr = 0;
int key_frame_insert = 0;
uint64_t psnr_sse_total[NUM_ENCODERS] = {0};
uint64_t psnr_samples_total[NUM_ENCODERS] = {0};
double psnr_totals[NUM_ENCODERS][4] = {{0,0}};
int psnr_count[NUM_ENCODERS] = {0};
double cx_time = 0;
struct timeval tv1, tv2, difftv;
/* Set the required target bitrates for each resolution level.
* If target bitrate for highest-resolution level is set to 0,
* (i.e. target_bitrate[0]=0), we skip encoding at that level.
*/
unsigned int target_bitrate[NUM_ENCODERS]={1000, 500, 100};
/* Enter the frame rate of the input video */
int framerate = 30;
/* Set down-sampling factor for each resolution level.
dsf[0] controls down sampling from level 0 to level 1;
dsf[1] controls down sampling from level 1 to level 2;
dsf[2] is not used. */
vpx_rational_t dsf[NUM_ENCODERS] = {{2, 1}, {2, 1}, {1, 1}};
/* Set the number of temporal layers for each encoder/resolution level,
* starting from highest resoln down to lowest resoln. */
unsigned int num_temporal_layers[NUM_ENCODERS] = {3, 3, 3};
if(argc!= (7 + 3 * NUM_ENCODERS))
die("Usage: %s <width> <height> <frame_rate> <infile> <outfile(s)> "
"<rate_encoder(s)> <temporal_layer(s)> <key_frame_insert> <output psnr?> \n",
argv[0]);
printf("Using %s\n",vpx_codec_iface_name(interface));
width = strtol(argv[1], NULL, 0);
height = strtol(argv[2], NULL, 0);
framerate = strtol(argv[3], NULL, 0);
if(width < 16 || width%2 || height <16 || height%2)
die("Invalid resolution: %ldx%ld", width, height);
/* Open input video file for encoding */
if(!(infile = fopen(argv[4], "rb")))
die("Failed to open %s for reading", argv[4]);
/* Open output file for each encoder to output bitstreams */
for (i=0; i< NUM_ENCODERS; i++)
{
if(!target_bitrate[i])
{
outfile[i] = NULL;
continue;
}
// Make width & height to be multiplier of 2.
if ((cfg[i].g_w) % 2)
cfg[i].g_w++;
if ((cfg[i].g_h) % 2)
cfg[i].g_h++;
}
// Open output file for each encoder to output bitstreams
for (i = 0; i < kNumEncoders; ++i) {
VpxVideoInfo info = {
encoder->fourcc,
cfg[i].g_w,
cfg[i].g_h,
{cfg[i].g_timebase.num, cfg[i].g_timebase.den}
};
if (!(writers[i] = vpx_video_writer_open(argv[i+4], kContainerIVF, &info)))
die("Failed to open %s for writing", argv[i+4]);
}
// Allocate image for each encoder
for (i = 0; i < kNumEncoders; ++i)
if (!vpx_img_alloc(&raw[i], VPX_IMG_FMT_I420, cfg[i].g_w, cfg[i].g_h, 32))
die("Failed to allocate image", cfg[i].g_w, cfg[i].g_h);
// Initialize multi-encoder
if (vpx_codec_enc_init_multi(&codec[0], encoder->codec_interface(), &cfg[0],
kNumEncoders,
show_psnr ? VPX_CODEC_USE_PSNR : 0, &dsf[0]))
die_codec(&codec[0], "Failed to initialize encoder");
// The extra encoding configuration parameters can be set as follows.
for (i = 0; i < kNumEncoders; i++) {
// Set encoding speed
if (vpx_codec_control(&codec[i], VP8E_SET_CPUUSED, -6))
die_codec(&codec[i], "Failed to set cpu_used");
// Set static threshold.
if (vpx_codec_control(&codec[i], VP8E_SET_STATIC_THRESHOLD, 1))
die_codec(&codec[i], "Failed to set static threshold");
// Set NOISE_SENSITIVITY to do TEMPORAL_DENOISING
// Enable denoising for the highest-resolution encoder.
if (vpx_codec_control(&codec[0], VP8E_SET_NOISE_SENSITIVITY, i == 0))
die_codec(&codec[0], "Failed to set noise_sensitivity");
}
frame_avail = 1;
got_data = 0;
while (frame_avail || got_data) {
vpx_codec_iter_t iter[kNumEncoders] = {NULL};
const vpx_codec_cx_pkt_t *pkt[kNumEncoders];
frame_avail = vpx_img_read(&raw[0], infile);
if (frame_avail) {
for (i = 1; i < kNumEncoders; ++i) {
vpx_image_t *const prev = &raw[i - 1];
// Scale the image down a number of times by downsampling factor
// FilterMode 1 or 2 give better psnr than FilterMode 0.
I420Scale(prev->planes[VPX_PLANE_Y], prev->stride[VPX_PLANE_Y],
prev->planes[VPX_PLANE_U], prev->stride[VPX_PLANE_U],
prev->planes[VPX_PLANE_V], prev->stride[VPX_PLANE_V],
prev->d_w, prev->d_h,
raw[i].planes[VPX_PLANE_Y], raw[i].stride[VPX_PLANE_Y],
raw[i].planes[VPX_PLANE_U], raw[i].stride[VPX_PLANE_U],
raw[i].planes[VPX_PLANE_V], raw[i].stride[VPX_PLANE_V],
raw[i].d_w, raw[i].d_h, 1);
}
if(!(outfile[i] = fopen(argv[i+5], "wb")))
die("Failed to open %s for writing", argv[i+4]);
}
// Encode frame.
if (vpx_codec_encode(&codec[0], frame_avail? &raw[0] : NULL,
frame_cnt, 1, 0, arg_deadline)) {
die_codec(&codec[0], "Failed to encode frame");
// Bitrates per spatial layer: overwrite default rates above.
for (i=0; i< NUM_ENCODERS; i++)
{
target_bitrate[i] = strtol(argv[NUM_ENCODERS + 5 + i], NULL, 0);
}