Packing bitstream on-the-fly with delayed context updates
Produce the token partitions on-the-fly, while processing each MB.
Context is updated at the beginning of each frame based on the
previoud frame's counters. Optimally encoder outputs partitions in
separate buffers. For frame based output, partitions are concatenated
internally.
Limitations:
- enabled just in combination with realtime-only mode
- number of encoding threads has to be equal or less than the
number of token partitions. For this reason, by default the encoder
will do 8 token partitions.
- vpxenc supports partition output (-P) just in combination with
IVF output format (--ivf)
Performance:
- Realtime encoder can be up to 13% faster (ARM) depending on the number
of threads and bitrate settings. Constant gain over the 5-16 speed
range.
- Token buffer reduced from one frame to 8 MBs
Quality:
- quality is affected by the delayed context updates. This again
dependents on input material, speed and bitrate settings. For VC
style input the loss seen is up to 0.2dB. If error-resilient=2
mode is used than the effect of this change is negligible.
Example:
./configure --enable-realtime-only --enable-onthefly-bitpacking
./vpxenc --rt --end-usage=1 --fps=30000/1000 -w 640 -h 480
--target-bitrate=1000 --token-parts=3 --static-thresh=2000
--ivf -P -t 4 -o strm.ivf tanya_640x480.yuv
Change-Id: I127295cb85b835fc287e1c0201a67e378d025d76
diff --git a/configure b/configure
index 4ac97b1..cf3eb32 100755
--- a/configure
+++ b/configure
@@ -39,6 +39,7 @@
${toggle_multithread} multithreaded encoding and decoding
${toggle_spatial_resampling} spatial sampling (scaling) support
${toggle_realtime_only} enable this option while building for real-time encoding
+ ${toggle_onthefly_bitpacking} enable on-the-fly bitpacking in real-time encoding
${toggle_error_concealment} enable this option to get a decoder which is able to conceal losses
${toggle_runtime_cpu_detect} runtime cpu detection
${toggle_shared} shared library support
@@ -253,6 +254,7 @@
static_msvcrt
spatial_resampling
realtime_only
+ onthefly_bitpacking
error_concealment
shared
static
@@ -297,6 +299,7 @@
mem_tracker
spatial_resampling
realtime_only
+ onthefly_bitpacking
error_concealment
shared
static
diff --git a/vp8/encoder/bitstream.c b/vp8/encoder/bitstream.c
index f26b4e4..af00f7c 100644
--- a/vp8/encoder/bitstream.c
+++ b/vp8/encoder/bitstream.c
@@ -24,6 +24,7 @@
#include "bitstream.h"
#include "defaultcoefcounts.h"
+#include "vp8/common/common.h"
const int vp8cx_base_skip_false_prob[128] =
{
@@ -159,7 +160,7 @@
);
}
-static void pack_tokens_c(vp8_writer *w, const TOKENEXTRA *p, int xcount)
+void vp8_pack_tokens_c(vp8_writer *w, const TOKENEXTRA *p, int xcount)
{
const TOKENEXTRA *const stop = p + xcount;
unsigned int split;
@@ -398,7 +399,7 @@
const TOKENEXTRA *stop = cpi->tplist[mb_row].stop;
int tokens = stop - p;
- pack_tokens_c(w, p, tokens);
+ vp8_pack_tokens_c(w, p, tokens);
}
vp8_stop_encode(w);
@@ -417,7 +418,7 @@
const TOKENEXTRA *stop = cpi->tplist[mb_row].stop;
int tokens = stop - p;
- pack_tokens_c(w, p, tokens);
+ vp8_pack_tokens_c(w, p, tokens);
}
}
@@ -783,6 +784,7 @@
}
}
+#if 0
/* This function is used for debugging probability trees. */
static void print_prob_tree(vp8_prob
coef_probs[BLOCK_TYPES][COEF_BANDS][PREV_COEF_CONTEXTS][ENTROPY_NODES])
@@ -814,6 +816,7 @@
fprintf(f, "}\n");
fclose(f);
}
+#endif
static void sum_probs_over_prev_coef_context(
const unsigned int probs[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS],
@@ -943,7 +946,6 @@
int t = 0; /* token/prob index */
-
vp8_tree_probs_from_distribution(
MAX_ENTROPY_TOKENS, vp8_coef_encodings, vp8_coef_tree,
cpi->frame_coef_probs [i][j][k],
@@ -1048,10 +1050,33 @@
return savings;
}
-static void update_coef_probs(VP8_COMP *cpi)
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+int vp8_update_coef_context(VP8_COMP *cpi)
+{
+ int savings = 0;
+
+
+ if (cpi->common.frame_type == KEY_FRAME)
+ {
+ /* Reset to default counts/probabilities at key frames */
+ vp8_copy(cpi->coef_counts, default_coef_counts);
+ }
+
+ if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+ savings += independent_coef_context_savings(cpi);
+ else
+ savings += default_coef_context_savings(cpi);
+
+ return savings;
+}
+#endif
+
+void vp8_update_coef_probs(VP8_COMP *cpi)
{
int i = 0;
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
vp8_writer *const w = cpi->bc;
+#endif
int savings = 0;
vp8_clear_system_state(); //__asm emms;
@@ -1131,7 +1156,11 @@
cpi->common.frame_type == KEY_FRAME && newp != *Pold)
u = 1;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ cpi->update_probs[i][j][k][t] = u;
+#else
vp8_write(w, u, upd);
+#endif
#ifdef ENTROPY_STATS
@@ -1143,7 +1172,9 @@
/* send/use new probability */
*Pold = newp;
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
vp8_write_literal(w, newp, 8);
+#endif
savings += s;
@@ -1172,6 +1203,50 @@
while (++i < BLOCK_TYPES);
}
+
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+static void pack_coef_probs(VP8_COMP *cpi)
+{
+ int i = 0;
+ vp8_writer *const w = cpi->bc;
+
+ do
+ {
+ int j = 0;
+
+ do
+ {
+ int k = 0;
+
+ do
+ {
+ int t = 0; /* token/prob index */
+
+ do
+ {
+ const vp8_prob newp = cpi->common.fc.coef_probs [i][j][k][t];
+ const vp8_prob upd = vp8_coef_update_probs [i][j][k][t];
+
+ const char u = cpi->update_probs[i][j][k][t] ;
+
+ vp8_write(w, u, upd);
+
+ if (u)
+ {
+ /* send/use new probability */
+ vp8_write_literal(w, newp, 8);
+ }
+ }
+ while (++t < ENTROPY_NODES);
+ }
+ while (++k < PREV_COEF_CONTEXTS);
+ }
+ while (++j < COEF_BANDS);
+ }
+ while (++i < BLOCK_TYPES);
+}
+#endif
+
#ifdef PACKET_TESTING
FILE *vpxlogc = 0;
#endif
@@ -1434,6 +1509,7 @@
vp8_write_bit(bc, pc->ref_frame_sign_bias[ALTREF_FRAME]);
}
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
{
if (pc->frame_type == KEY_FRAME)
@@ -1441,6 +1517,7 @@
else
pc->refresh_entropy_probs = 0;
}
+#endif
vp8_write_bit(bc, pc->refresh_entropy_probs);
@@ -1458,13 +1535,17 @@
vp8_clear_system_state(); //__asm emms;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ pack_coef_probs(cpi);
+#else
if (pc->refresh_entropy_probs == 0)
{
// save a copy for later refresh
vpx_memcpy(&cpi->common.lfc, &cpi->common.fc, sizeof(cpi->common.fc));
}
- update_coef_probs(cpi);
+ vp8_update_coef_probs(cpi);
+#endif
#ifdef ENTROPY_STATS
active_section = 2;
@@ -1512,6 +1593,45 @@
cpi->partition_sz[0] = *size;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ {
+ const int num_part = (1 << pc->multi_token_partition);
+ unsigned char * dp = cpi->partition_d[0] + cpi->partition_sz[0];
+
+ if (num_part > 1)
+ {
+ /* write token part sizes (all but last) if more than 1 */
+ validate_buffer(dp, 3 * (num_part - 1), cpi->partition_d_end[0],
+ &pc->error);
+
+ cpi->partition_sz[0] += 3*(num_part-1);
+
+ for(i = 1; i < num_part; i++)
+ {
+ write_partition_size(dp, cpi->partition_sz[i]);
+ dp += 3;
+ }
+ }
+
+ if (!cpi->output_partition)
+ {
+ /* concatenate partition buffers */
+ for(i = 0; i < num_part; i++)
+ {
+ vpx_memmove(dp, cpi->partition_d[i+1], cpi->partition_sz[i+1]);
+ cpi->partition_d[i+1] = dp;
+ dp += cpi->partition_sz[i+1];
+ }
+ }
+
+ /* update total size */
+ *size = 0;
+ for(i = 0; i < num_part+1; i++)
+ {
+ *size += cpi->partition_sz[i];
+ }
+ }
+#else
if (pc->multi_token_partition != ONE_PARTITION)
{
int num_part = 1 << pc->multi_token_partition;
@@ -1561,6 +1681,7 @@
*size += cpi->bc[1].pos;
cpi->partition_sz[1] = cpi->bc[1].pos;
}
+#endif
}
#ifdef ENTROPY_STATS
diff --git a/vp8/encoder/bitstream.h b/vp8/encoder/bitstream.h
index 736b9d4..e0bbdc4 100644
--- a/vp8/encoder/bitstream.h
+++ b/vp8/encoder/bitstream.h
@@ -35,7 +35,10 @@
# define pack_mb_row_tokens(a,b) \
vp8cx_pack_mb_row_tokens_armv5(a,b,vp8_coef_encodings,vp8_extra_bits,vp8_coef_tree)
#else
-# define pack_tokens(a,b,c) pack_tokens_c(a,b,c)
+
+void vp8_pack_tokens_c(vp8_writer *w, const TOKENEXTRA *p, int xcount);
+
+# define pack_tokens(a,b,c) vp8_pack_tokens_c(a,b,c)
# define pack_tokens_into_partitions(a,b,c,d) pack_tokens_into_partitions_c(a,b,c,d)
# define pack_mb_row_tokens(a,b) pack_mb_row_tokens_c(a,b)
#endif
diff --git a/vp8/encoder/encodeframe.c b/vp8/encoder/encodeframe.c
index 9899565..21757f8 100644
--- a/vp8/encoder/encodeframe.c
+++ b/vp8/encoder/encodeframe.c
@@ -28,6 +28,9 @@
#include <limits.h>
#include "vp8/common/invtrans.h"
#include "vpx_ports/vpx_timer.h"
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+#include "bitstream.h"
+#endif
extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t) ;
extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
@@ -373,10 +376,17 @@
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
int map_index = (mb_row * cpi->common.mb_cols);
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ const int num_part = (1 << cm->multi_token_partition);
+ TOKENEXTRA * tp_start = cpi->tok;
+ vp8_writer *w;
+#endif
+
#if CONFIG_MULTITHREAD
const int nsync = cpi->mt_sync_range;
- const int rightmost_col = cm->mb_cols - 1;
+ const int rightmost_col = cm->mb_cols + nsync;
volatile const int *last_row_current_mb_col;
+ volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
@@ -384,6 +394,13 @@
last_row_current_mb_col = &rightmost_col;
#endif
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ if(num_part > 1)
+ w= &cpi->bc[1 + (mb_row % num_part)];
+ else
+ w = &cpi->bc[1];
+#endif
+
// reset above block coeffs
xd->above_context = cm->above_context;
@@ -411,6 +428,10 @@
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
{
+
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ *tp = cpi->tok;
+#endif
// Distance of Mb to the left & right edges, specified in
// 1/8th pel units as they are always compared to values
// that are in 1/8th pel units
@@ -435,12 +456,13 @@
vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
#if CONFIG_MULTITHREAD
- if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
+ if (cpi->b_multi_threaded != 0)
{
+ *current_mb_col = mb_col - 1; // set previous MB done
+
if ((mb_col & (nsync - 1)) == 0)
{
- while (mb_col > (*last_row_current_mb_col - nsync)
- && (*last_row_current_mb_col) != (cm->mb_cols - 1))
+ while (mb_col > (*last_row_current_mb_col - nsync))
{
x86_pause_hint();
thread_sleep(0);
@@ -495,13 +517,13 @@
#endif
- // Count of last ref frame 0,0 useage
+ // Count of last ref frame 0,0 usage
if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
cpi->inter_zz_count ++;
// Special case code for cyclic refresh
// If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
- // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
+ // during vp8cx_encode_inter_macroblock()) back into the global segmentation map
if ((cpi->current_layer == 0) &&
(cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled))
{
@@ -525,7 +547,14 @@
cpi->tplist[mb_row].stop = *tp;
- // Increment pointer into gf useage flags structure.
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ /* pack tokens for this MB */
+ {
+ int tok_count = *tp - tp_start;
+ pack_tokens(w, tp_start, tok_count);
+ }
+#endif
+ // Increment pointer into gf usage flags structure.
x->gf_active_ptr++;
// Increment the activity mask pointers.
@@ -539,39 +568,29 @@
recon_yoffset += 16;
recon_uvoffset += 8;
- // Keep track of segment useage
+ // Keep track of segment usage
segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
// skip to next mb
xd->mode_info_context++;
x->partition_info++;
-
xd->above_context++;
-#if CONFIG_MULTITHREAD
- if (cpi->b_multi_threaded != 0)
- {
- cpi->mt_current_mb_col[mb_row] = mb_col;
- }
-#endif
}
//extend the recon for intra prediction
- vp8_extend_mb_row(
- &cm->yv12_fb[dst_fb_idx],
- xd->dst.y_buffer + 16,
- xd->dst.u_buffer + 8,
- xd->dst.v_buffer + 8);
+ vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
+ xd->dst.y_buffer + 16,
+ xd->dst.u_buffer + 8,
+ xd->dst.v_buffer + 8);
+
+#if CONFIG_MULTITHREAD
+ if (cpi->b_multi_threaded != 0)
+ *current_mb_col = rightmost_col;
+#endif
// this is to account for the border
xd->mode_info_context++;
x->partition_info++;
-
-#if CONFIG_MULTITHREAD
- if ((cpi->b_multi_threaded != 0) && (mb_row == cm->mb_rows - 1))
- {
- sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
- }
-#endif
}
void init_encode_frame_mb_context(VP8_COMP *cpi)
@@ -599,7 +618,7 @@
if (cm->frame_type == KEY_FRAME)
vp8_init_mbmode_probs(cm);
- // Copy data over into macro block data sturctures.
+ // Copy data over into macro block data structures.
x->src = * cpi->Source;
xd->pre = cm->yv12_fb[cm->lst_fb_idx];
xd->dst = cm->yv12_fb[cm->new_fb_idx];
@@ -656,10 +675,13 @@
MACROBLOCK *const x = & cpi->mb;
VP8_COMMON *const cm = & cpi->common;
MACROBLOCKD *const xd = & x->e_mbd;
-
TOKENEXTRA *tp = cpi->tok;
int segment_counts[MAX_MB_SEGMENTS];
int totalrate;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ BOOL_CODER * bc = &cpi->bc[1]; // bc[0] is for control partition
+ const int num_part = (1 << cm->multi_token_partition);
+#endif
vpx_memset(segment_counts, 0, sizeof(segment_counts));
totalrate = 0;
@@ -694,6 +716,7 @@
cpi->prediction_error = 0;
cpi->intra_error = 0;
cpi->skip_true_count = 0;
+ cpi->tok_count = 0;
#if 0
// Experimental code
@@ -704,6 +727,7 @@
xd->mode_info_context = cm->mi;
vp8_zero(cpi->MVcount);
+
vp8_zero(cpi->coef_counts);
vp8cx_frame_init_quantizer(cpi);
@@ -722,9 +746,22 @@
build_activity_map(cpi);
}
- // re-initencode frame context.
+ // re-init encode frame context.
init_encode_frame_mb_context(cpi);
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ {
+ int i;
+ for(i = 0; i < num_part; i++)
+ {
+ vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
+ cpi->partition_d_end[i + 1]);
+ bc[i].error = &cm->error;
+ }
+ }
+
+#endif
+
{
struct vpx_usec_timer emr_timer;
vpx_usec_timer_start(&emr_timer);
@@ -748,7 +785,11 @@
{
vp8_zero(cm->left_context)
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ tp = cpi->tok;
+#else
tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
+#endif
encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
@@ -761,12 +802,14 @@
x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
+ if(mb_row == cm->mb_rows - 1)
+ {
+ sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
+ }
}
sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
- cpi->tok_count = 0;
-
for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
{
cpi->tok_count += cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start;
@@ -799,9 +842,12 @@
// for each macroblock row in image
for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
{
-
vp8_zero(cm->left_context)
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ tp = cpi->tok;
+#endif
+
encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
// adjust to the next row of mbs
@@ -811,16 +857,25 @@
}
cpi->tok_count = tp - cpi->tok;
-
}
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ {
+ int i;
+ for(i = 0; i < num_part; i++)
+ {
+ vp8_stop_encode(&bc[i]);
+ cpi->partition_sz[i+1] = bc[i].pos;
+ }
+ }
+#endif
+
vpx_usec_timer_mark(&emr_timer);
cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
-
}
- // Work out the segment probabilites if segmentation is enabled
+ // Work out the segment probabilities if segmentation is enabled
if (xd->segmentation_enabled)
{
int tot_count;
@@ -908,20 +963,16 @@
}
#endif
- // Adjust the projected reference frame useage probability numbers to reflect
- // what we have just seen. This may be usefull when we make multiple itterations
+#if ! CONFIG_REALTIME_ONLY
+ // Adjust the projected reference frame usage probability numbers to reflect
+ // what we have just seen. This may be useful when we make multiple iterations
// of the recode loop rather than continuing to use values from the previous frame.
if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
(!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
{
vp8_convert_rfct_to_prob(cpi);
}
-
-#if 0
- // Keep record of the total distortion this time around for future use
- cpi->last_frame_distortion = cpi->frame_distortion;
#endif
-
}
void vp8_setup_block_ptrs(MACROBLOCK *x)
{
diff --git a/vp8/encoder/ethreading.c b/vp8/encoder/ethreading.c
index 2874e78..b549a7d 100644
--- a/vp8/encoder/ethreading.c
+++ b/vp8/encoder/ethreading.c
@@ -13,6 +13,8 @@
#include "vp8/common/common.h"
#include "vp8/common/extend.h"
+#include "bitstream.h"
+
#if CONFIG_MULTITHREAD
extern int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
@@ -74,6 +76,10 @@
MACROBLOCK *x = &mbri->mb;
MACROBLOCKD *xd = &x->e_mbd;
TOKENEXTRA *tp ;
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ TOKENEXTRA *tp_start = cpi->tok + (1 + ithread) * (16 * 24);
+ const int num_part = (1 << cm->multi_token_partition);
+#endif
int *segment_counts = mbri->segment_counts;
int *totalrate = &mbri->totalrate;
@@ -91,9 +97,15 @@
int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
int map_index = (mb_row * cm->mb_cols);
- volatile int *last_row_current_mb_col;
+ volatile const int *last_row_current_mb_col;
+ volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ vp8_writer *w = &cpi->bc[1 + (mb_row % num_part)];
+#else
tp = cpi->tok + (mb_row * (cm->mb_cols * 16 * 24));
+ cpi->tplist[mb_row].start = tp;
+#endif
last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
@@ -107,25 +119,27 @@
recon_yoffset = (mb_row * recon_y_stride * 16);
recon_uvoffset = (mb_row * recon_uv_stride * 8);
- cpi->tplist[mb_row].start = tp;
-
- //printf("Thread mb_row = %d\n", mb_row);
-
// Set the mb activity pointer to the start of the row.
x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
// for each macroblock col in image
for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
{
+ *current_mb_col = mb_col - 1;
+
if ((mb_col & (nsync - 1)) == 0)
{
- while (mb_col > (*last_row_current_mb_col - nsync) && *last_row_current_mb_col != cm->mb_cols - 1)
+ while (mb_col > (*last_row_current_mb_col - nsync))
{
x86_pause_hint();
thread_sleep(0);
}
}
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ tp = tp_start;
+#endif
+
// Distance of Mb to the various image edges.
// These specified to 8th pel as they are always compared to values that are in 1/8th pel units
xd->mb_to_left_edge = -((mb_col * 16) << 3);
@@ -154,7 +168,7 @@
vp8_activity_masking(cpi, x);
// Is segmentation enabled
- // MB level adjutment to quantizer
+ // MB level adjustment to quantizer
if (xd->segmentation_enabled)
{
// Code to set segment id in xd->mbmi.segment_id for current MB (with range checking)
@@ -196,13 +210,13 @@
#endif
- // Count of last ref frame 0,0 useage
+ // Count of last ref frame 0,0 usage
if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
cpi->inter_zz_count++;
// Special case code for cyclic refresh
// If cyclic update enabled then copy xd->mbmi.segment_id; (which may have been updated based on mode
- // during vp8cx_encode_inter_macroblock()) back into the global sgmentation map
+ // during vp8cx_encode_inter_macroblock()) back into the global segmentation map
if (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)
{
const MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
@@ -223,9 +237,17 @@
}
}
- cpi->tplist[mb_row].stop = tp;
- // Increment pointer into gf useage flags structure.
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ /* pack tokens for this MB */
+ {
+ int tok_count = tp - tp_start;
+ pack_tokens(w, tp_start, tok_count);
+ }
+#else
+ cpi->tplist[mb_row].stop = tp;
+#endif
+ // Increment pointer into gf usage flags structure.
x->gf_active_ptr++;
// Increment the activity mask pointers.
@@ -239,23 +261,21 @@
recon_yoffset += 16;
recon_uvoffset += 8;
- // Keep track of segment useage
+ // Keep track of segment usage
segment_counts[xd->mode_info_context->mbmi.segment_id]++;
// skip to next mb
xd->mode_info_context++;
x->partition_info++;
xd->above_context++;
-
- cpi->mt_current_mb_col[mb_row] = mb_col;
}
- //extend the recon for intra prediction
- vp8_extend_mb_row(
- &cm->yv12_fb[dst_fb_idx],
- xd->dst.y_buffer + 16,
- xd->dst.u_buffer + 8,
- xd->dst.v_buffer + 8);
+ vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
+ xd->dst.y_buffer + 16,
+ xd->dst.u_buffer + 8,
+ xd->dst.v_buffer + 8);
+
+ *current_mb_col = mb_col + nsync;
// this is to account for the border
xd->mode_info_context++;
@@ -271,7 +291,6 @@
if (mb_row == cm->mb_rows - 1)
{
- //SetEvent(cpi->h_event_main);
sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
}
}
diff --git a/vp8/encoder/onyx_if.c b/vp8/encoder/onyx_if.c
index 6645442..5ad51e8 100644
--- a/vp8/encoder/onyx_if.c
+++ b/vp8/encoder/onyx_if.c
@@ -42,6 +42,11 @@
#include <stdio.h>
#include <limits.h>
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+extern int vp8_update_coef_context(VP8_COMP *cpi);
+extern void vp8_update_coef_probs(VP8_COMP *cpi);
+#endif
+
extern void vp8cx_pick_filter_level_fast(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi);
extern void vp8cx_set_alt_lf_level(VP8_COMP *cpi, int filt_val);
extern void vp8cx_pick_filter_level(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi);
@@ -1106,8 +1111,11 @@
vpx_free(cpi->tok);
{
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ unsigned int tokens = 8 * 24 * 16; /* one MB for each thread */
+#else
unsigned int tokens = cm->mb_rows * cm->mb_cols * 24 * 16;
-
+#endif
CHECK_MEM_ERROR(cpi->tok, vpx_calloc(tokens, sizeof(*cpi->tok)));
}
@@ -1514,6 +1522,10 @@
cm->refresh_last_frame = 1;
cm->refresh_entropy_probs = 1;
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ cpi->oxcf.token_partitions = 3;
+#endif
+
if (cpi->oxcf.token_partitions >= 0 && cpi->oxcf.token_partitions <= 3)
cm->multi_token_partition =
(TOKEN_PARTITION) cpi->oxcf.token_partitions;
@@ -3725,12 +3737,40 @@
}
#endif
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ {
+ if(cpi->oxcf.error_resilient_mode)
+ cm->refresh_entropy_probs = 0;
+
+ if (cpi->oxcf.error_resilient_mode & VPX_ERROR_RESILIENT_PARTITIONS)
+ {
+ if (cm->frame_type == KEY_FRAME)
+ cm->refresh_entropy_probs = 1;
+ }
+
+ if (cm->refresh_entropy_probs == 0)
+ {
+ // save a copy for later refresh
+ vpx_memcpy(&cm->lfc, &cm->fc, sizeof(cm->fc));
+ }
+
+ vp8_update_coef_context(cpi);
+
+ vp8_update_coef_probs(cpi);
+
+ // transform / motion compensation build reconstruction frame
+ // +pack coef partitions
+ vp8_encode_frame(cpi);
+
+ /* cpi->projected_frame_size is not needed for RT mode */
+ }
+#else
// transform / motion compensation build reconstruction frame
vp8_encode_frame(cpi);
cpi->projected_frame_size -= vp8_estimate_entropy_savings(cpi);
cpi->projected_frame_size = (cpi->projected_frame_size > 0) ? cpi->projected_frame_size : 0;
-
+#endif
vp8_clear_system_state(); //__asm emms;
// Test to see if the stats generated for this frame indicate that we should have coded a key frame
@@ -4093,10 +4133,12 @@
update_reference_frames(cm);
+#if !(CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
if (cpi->oxcf.error_resilient_mode)
{
cm->refresh_entropy_probs = 0;
}
+#endif
#if CONFIG_MULTITHREAD
/* wait that filter_level is picked so that we can continue with stream packing */
@@ -4819,6 +4861,29 @@
vpx_usec_timer_start(&ticktimer);
}
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ {
+ int i;
+ const int num_part = (1 << cm->multi_token_partition);
+ /* the available bytes in dest */
+ const unsigned long dest_size = dest_end - dest;
+ const int tok_part_buff_size = (dest_size * 9) / (10 * num_part);
+
+ unsigned char *dp = dest;
+
+ cpi->partition_d[0] = dp;
+ dp += dest_size/10; /* reserve 1/10 for control partition */
+ cpi->partition_d_end[0] = dp;
+
+ for(i = 0; i < num_part; i++)
+ {
+ cpi->partition_d[i + 1] = dp;
+ dp += tok_part_buff_size;
+ cpi->partition_d_end[i + 1] = dp;
+ }
+ }
+#endif
+
// start with a 0 size frame
*size = 0;
diff --git a/vp8/encoder/onyx_int.h b/vp8/encoder/onyx_int.h
index dcefbac..e17d8ea 100644
--- a/vp8/encoder/onyx_int.h
+++ b/vp8/encoder/onyx_int.h
@@ -443,9 +443,12 @@
unsigned int MVcount [2] [MVvals]; /* (row,col) MV cts this frame */
unsigned int coef_counts [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS]; /* for this frame */
+
//DECLARE_ALIGNED(16, int, coef_counts_backup [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS]); //not used any more
//save vp8_tree_probs_from_distribution result for each frame to avoid repeat calculation
vp8_prob frame_coef_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+ char update_probs [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES];
+
unsigned int frame_branch_ct [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES][2];
int gfu_boost;
@@ -543,6 +546,8 @@
TOKENLIST *tplist;
unsigned int partition_sz[MAX_PARTITIONS];
+ unsigned char *partition_d[MAX_PARTITIONS];
+ unsigned char *partition_d_end[MAX_PARTITIONS];
// end of multithread data
diff --git a/vp8/vp8_cx_iface.c b/vp8/vp8_cx_iface.c
index 42da7be..fa78ec3 100644
--- a/vp8/vp8_cx_iface.c
+++ b/vp8/vp8_cx_iface.c
@@ -66,7 +66,11 @@
0, /* noise_sensitivity */
0, /* Sharpness */
0, /* static_thresh */
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ VP8_EIGHT_TOKENPARTITION,
+#else
VP8_ONE_TOKENPARTITION, /* token_partitions */
+#endif
0, /* arnr_max_frames */
3, /* arnr_strength */
3, /* arnr_type*/
@@ -241,6 +245,11 @@
RANGE_CHECK_HI(cfg, ts_layer_id[i], cfg->ts_number_layers-1);
}
+#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
+ if(cfg->g_threads > (1 << vp8_cfg->token_partitions))
+ ERROR("g_threads cannot be bigger than number of token partitions");
+#endif
+
return VPX_CODEC_OK;
}
@@ -919,16 +928,28 @@
for (i = 0; i < num_partitions; ++i)
{
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ pkt.data.frame.buf = cpi->partition_d[i];
+#else
pkt.data.frame.buf = cx_data;
+ cx_data += cpi->partition_sz[i];
+ cx_data_sz -= cpi->partition_sz[i];
+#endif
pkt.data.frame.sz = cpi->partition_sz[i];
pkt.data.frame.partition_id = i;
/* don't set the fragment bit for the last partition */
if (i == (num_partitions - 1))
pkt.data.frame.flags &= ~VPX_FRAME_IS_FRAGMENT;
vpx_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
- cx_data += cpi->partition_sz[i];
- cx_data_sz -= cpi->partition_sz[i];
}
+#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
+ /* In lagged mode the encoder can buffer multiple frames.
+ * We don't want this in partitioned output because
+ * partitions are spread all over the output buffer.
+ * So, force an exit!
+ */
+ cx_data_sz -= ctx->cx_data_sz / 2;
+#endif
}
else
{
diff --git a/vpxenc.c b/vpxenc.c
index 7d6758a..5475b44 100644
--- a/vpxenc.c
+++ b/vpxenc.c
@@ -489,6 +489,13 @@
if(fwrite(header, 1, 12, outfile));
}
+static void write_ivf_frame_size(FILE *outfile, size_t size)
+{
+ char header[4];
+ mem_put_le32(header, size);
+ fwrite(header, 1, 4, outfile);
+}
+
typedef off_t EbmlLoc;
@@ -945,7 +952,6 @@
#include "args.h"
-
static const arg_def_t debugmode = ARG_DEF("D", "debug", 0,
"Debug mode (makes output deterministic)");
static const arg_def_t outputfile = ARG_DEF("o", "output", 1,
@@ -980,6 +986,8 @@
"Stream frame rate (rate/scale)");
static const arg_def_t use_ivf = ARG_DEF(NULL, "ivf", 0,
"Output IVF (default is WebM)");
+static const arg_def_t out_part = ARG_DEF("P", "output-partitions", 0,
+ "Makes encoder output partitions. Requires IVF output!");
static const arg_def_t q_hist_n = ARG_DEF(NULL, "q-hist", 1,
"Show quantizer histogram (n-buckets)");
static const arg_def_t rate_hist_n = ARG_DEF(NULL, "rate-hist", 1,
@@ -989,7 +997,7 @@
&debugmode,
&outputfile, &codecarg, &passes, &pass_arg, &fpf_name, &limit, &deadline,
&best_dl, &good_dl, &rt_dl,
- &verbosearg, &psnrarg, &use_ivf, &q_hist_n, &rate_hist_n,
+ &verbosearg, &psnrarg, &use_ivf, &out_part, &q_hist_n, &rate_hist_n,
NULL
};
@@ -1492,6 +1500,7 @@
int show_psnr;
int have_framerate;
struct vpx_rational framerate;
+ int out_part;
int debug;
int show_q_hist_buckets;
int show_rate_hist_buckets;
@@ -1603,6 +1612,8 @@
global->framerate = arg_parse_rational(&arg);
global->have_framerate = 1;
}
+ else if (arg_match(&arg,&out_part, argi))
+ global->out_part = 1;
else if (arg_match(&arg, &debugmode, argi))
global->debug = 1;
else if (arg_match(&arg, &q_hist_n, argi))
@@ -2078,11 +2089,14 @@
struct global_config *global)
{
int i;
+ int flags = 0;
+
+ flags |= global->show_psnr ? VPX_CODEC_USE_PSNR : 0;
+ flags |= global->out_part ? VPX_CODEC_USE_OUTPUT_PARTITION : 0;
/* Construct Encoder Context */
vpx_codec_enc_init(&stream->encoder, global->codec->iface,
- &stream->config.cfg,
- global->show_psnr ? VPX_CODEC_USE_PSNR : 0);
+ &stream->config.cfg, flags);
ctx_exit_on_error(&stream->encoder, "Failed to initialize encoder");
/* Note that we bypass the vpx_codec_control wrapper macro because
@@ -2151,12 +2165,18 @@
while ((pkt = vpx_codec_get_cx_data(&stream->encoder, &iter)))
{
+ static size_t fsize = 0;
+ static off_t ivf_header_pos = 0;
+
*got_data = 1;
switch (pkt->kind)
{
case VPX_CODEC_CX_FRAME_PKT:
- stream->frames_out++;
+ if (!(pkt->data.frame.flags & VPX_FRAME_IS_FRAGMENT))
+ {
+ stream->frames_out++;
+ }
fprintf(stderr, " %6luF",
(unsigned long)pkt->data.frame.sz);
@@ -2172,9 +2192,28 @@
}
else
{
- write_ivf_frame_header(stream->file, pkt);
- if(fwrite(pkt->data.frame.buf, 1,
- pkt->data.frame.sz, stream->file));
+ if (pkt->data.frame.partition_id <= 0)
+ {
+ ivf_header_pos = ftello(stream->file);
+ fsize = pkt->data.frame.sz;
+
+ write_ivf_frame_header(stream->file, pkt);
+ }
+ else
+ {
+ fsize += pkt->data.frame.sz;
+
+ if (!(pkt->data.frame.flags & VPX_FRAME_IS_FRAGMENT))
+ {
+ off_t currpos = ftello(stream->file);
+ fseeko(stream->file, ivf_header_pos, SEEK_SET);
+ write_ivf_frame_size(stream->file, fsize);
+ fseeko(stream->file, currpos, SEEK_SET);
+ }
+ }
+
+ fwrite(pkt->data.frame.buf, 1,
+ pkt->data.frame.sz, stream->file);
}
stream->nbytes += pkt->data.raw.sz;
break;
