|
|
|
|
/*
|
|
|
|
|
* VC-1 and WMV3 parser
|
|
|
|
|
* Copyright (c) 2006-2007 Konstantin Shishkov
|
|
|
|
|
* Partly based on vc9.c (c) 2005 Anonymous, Alex Beregszaszi, Michael Niedermayer
|
|
|
|
|
*
|
|
|
|
|
* This file is part of FFmpeg.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg is free software; you can redistribute it and/or
|
|
|
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
|
|
|
* License as published by the Free Software Foundation; either
|
|
|
|
|
* version 2.1 of the License, or (at your option) any later version.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg is distributed in the hope that it will be useful,
|
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
|
* Lesser General Public License for more details.
|
|
|
|
|
*
|
|
|
|
|
* You should have received a copy of the GNU Lesser General Public
|
|
|
|
|
* License along with FFmpeg; if not, write to the Free Software
|
|
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* @file
|
|
|
|
|
* VC-1 and WMV3 parser
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include "libavutil/attributes.h"
|
|
|
|
|
#include "parser.h"
|
|
|
|
|
#include "vc1.h"
|
|
|
|
|
#include "get_bits.h"
|
|
|
|
|
#include "internal.h"
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
/** The maximum number of bytes of a sequence, entry point or
|
|
|
|
|
* frame header whose values we pay any attention to */
|
|
|
|
|
#define UNESCAPED_THRESHOLD 37
|
|
|
|
|
|
|
|
|
|
/** The maximum number of bytes of a sequence, entry point or
|
|
|
|
|
* frame header which must be valid memory (because they are
|
|
|
|
|
* used to update the bitstream cache in skip_bits() calls)
|
|
|
|
|
*/
|
|
|
|
|
#define UNESCAPED_LIMIT 144
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
NO_MATCH,
|
|
|
|
|
ONE_ZERO,
|
|
|
|
|
TWO_ZEROS,
|
|
|
|
|
ONE
|
|
|
|
|
} VC1ParseSearchState;
|
|
|
|
|
|
|
|
|
|
typedef struct VC1ParseContext {
|
|
|
|
|
ParseContext pc;
|
|
|
|
|
VC1Context v;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
uint8_t prev_start_code;
|
|
|
|
|
size_t bytes_to_skip;
|
|
|
|
|
uint8_t unesc_buffer[UNESCAPED_LIMIT];
|
|
|
|
|
size_t unesc_index;
|
|
|
|
|
VC1ParseSearchState search_state;
|
|
|
|
|
} VC1ParseContext;
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
static void vc1_extract_header(AVCodecParserContext *s, AVCodecContext *avctx,
|
|
|
|
|
const uint8_t *buf, int buf_size)
|
|
|
|
|
{
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
/* Parse the header we just finished unescaping */
|
|
|
|
|
VC1ParseContext *vpc = s->priv_data;
|
|
|
|
|
GetBitContext gb;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
int ret;
|
|
|
|
|
vpc->v.s.avctx = avctx;
|
|
|
|
|
vpc->v.parse_only = 1;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
init_get_bits(&gb, buf, buf_size * 8);
|
|
|
|
|
switch (vpc->prev_start_code) {
|
|
|
|
|
case VC1_CODE_SEQHDR & 0xFF:
|
|
|
|
|
ff_vc1_decode_sequence_header(avctx, &vpc->v, &gb);
|
|
|
|
|
break;
|
|
|
|
|
case VC1_CODE_ENTRYPOINT & 0xFF:
|
|
|
|
|
ff_vc1_decode_entry_point(avctx, &vpc->v, &gb);
|
|
|
|
|
break;
|
|
|
|
|
case VC1_CODE_FRAME & 0xFF:
|
|
|
|
|
if(vpc->v.profile < PROFILE_ADVANCED)
|
|
|
|
|
ret = ff_vc1_parse_frame_header (&vpc->v, &gb);
|
|
|
|
|
else
|
|
|
|
|
ret = ff_vc1_parse_frame_header_adv(&vpc->v, &gb);
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
if (ret < 0)
|
|
|
|
|
break;
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
/* keep AV_PICTURE_TYPE_BI internal to VC1 */
|
|
|
|
|
if (vpc->v.s.pict_type == AV_PICTURE_TYPE_BI)
|
|
|
|
|
s->pict_type = AV_PICTURE_TYPE_B;
|
|
|
|
|
else
|
|
|
|
|
s->pict_type = vpc->v.s.pict_type;
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
if (avctx->ticks_per_frame > 1){
|
|
|
|
|
// process pulldown flags
|
|
|
|
|
s->repeat_pict = 1;
|
|
|
|
|
// Pulldown flags are only valid when 'broadcast' has been set.
|
|
|
|
|
// So ticks_per_frame will be 2
|
|
|
|
|
if (vpc->v.rff){
|
|
|
|
|
// repeat field
|
|
|
|
|
s->repeat_pict = 2;
|
|
|
|
|
}else if (vpc->v.rptfrm){
|
|
|
|
|
// repeat frames
|
|
|
|
|
s->repeat_pict = vpc->v.rptfrm * 2 + 1;
|
|
|
|
|
}
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
}else{
|
|
|
|
|
s->repeat_pict = 0;
|
|
|
|
|
}
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
if (vpc->v.broadcast && vpc->v.interlace && !vpc->v.psf)
|
|
|
|
|
s->field_order = vpc->v.tff ? AV_FIELD_TT : AV_FIELD_BB;
|
|
|
|
|
else
|
|
|
|
|
s->field_order = AV_FIELD_PROGRESSIVE;
|
|
|
|
|
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (avctx->framerate.num)
|
|
|
|
|
avctx->time_base = av_inv_q(av_mul_q(avctx->framerate, (AVRational){avctx->ticks_per_frame, 1}));
|
|
|
|
|
s->format = vpc->v.chromaformat == 1 ? AV_PIX_FMT_YUV420P
|
|
|
|
|
: AV_PIX_FMT_NONE;
|
|
|
|
|
if (avctx->width && avctx->height) {
|
|
|
|
|
s->width = avctx->width;
|
|
|
|
|
s->height = avctx->height;
|
|
|
|
|
s->coded_width = FFALIGN(avctx->coded_width, 16);
|
|
|
|
|
s->coded_height = FFALIGN(avctx->coded_height, 16);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int vc1_parse(AVCodecParserContext *s,
|
|
|
|
|
AVCodecContext *avctx,
|
|
|
|
|
const uint8_t **poutbuf, int *poutbuf_size,
|
|
|
|
|
const uint8_t *buf, int buf_size)
|
|
|
|
|
{
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
/* Here we do the searching for frame boundaries and headers at
|
|
|
|
|
* the same time. Only a minimal amount at the start of each
|
|
|
|
|
* header is unescaped. */
|
|
|
|
|
VC1ParseContext *vpc = s->priv_data;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
int pic_found = vpc->pc.frame_start_found;
|
|
|
|
|
uint8_t *unesc_buffer = vpc->unesc_buffer;
|
|
|
|
|
size_t unesc_index = vpc->unesc_index;
|
|
|
|
|
VC1ParseSearchState search_state = vpc->search_state;
|
|
|
|
|
int start_code_found = 0;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
int next = END_NOT_FOUND;
|
|
|
|
|
int i = vpc->bytes_to_skip;
|
|
|
|
|
|
|
|
|
|
if (pic_found && buf_size == 0) {
|
|
|
|
|
/* EOF considered as end of frame */
|
|
|
|
|
memset(unesc_buffer + unesc_index, 0, UNESCAPED_THRESHOLD - unesc_index);
|
|
|
|
|
vc1_extract_header(s, avctx, unesc_buffer, unesc_index);
|
|
|
|
|
next = 0;
|
|
|
|
|
}
|
|
|
|
|
while (i < buf_size) {
|
|
|
|
|
uint8_t b;
|
|
|
|
|
start_code_found = 0;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
while (i < buf_size && unesc_index < UNESCAPED_THRESHOLD) {
|
|
|
|
|
b = buf[i++];
|
|
|
|
|
unesc_buffer[unesc_index++] = b;
|
|
|
|
|
if (search_state <= ONE_ZERO)
|
|
|
|
|
search_state = b ? NO_MATCH : search_state + 1;
|
|
|
|
|
else if (search_state == TWO_ZEROS) {
|
|
|
|
|
if (b == 1)
|
|
|
|
|
search_state = ONE;
|
|
|
|
|
else if (b > 1) {
|
|
|
|
|
if (b == 3)
|
|
|
|
|
unesc_index--; // swallow emulation prevention byte
|
|
|
|
|
search_state = NO_MATCH;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
else { // search_state == ONE
|
|
|
|
|
// Header unescaping terminates early due to detection of next start code
|
|
|
|
|
search_state = NO_MATCH;
|
|
|
|
|
start_code_found = 1;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if ((s->flags & PARSER_FLAG_COMPLETE_FRAMES) &&
|
|
|
|
|
unesc_index >= UNESCAPED_THRESHOLD &&
|
|
|
|
|
vpc->prev_start_code == (VC1_CODE_FRAME & 0xFF))
|
|
|
|
|
{
|
|
|
|
|
// No need to keep scanning the rest of the buffer for
|
|
|
|
|
// start codes if we know it contains a complete frame and
|
|
|
|
|
// we've already unescaped all we need of the frame header
|
|
|
|
|
vc1_extract_header(s, avctx, unesc_buffer, unesc_index);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
if (unesc_index >= UNESCAPED_THRESHOLD && !start_code_found) {
|
|
|
|
|
while (i < buf_size) {
|
|
|
|
|
if (search_state == NO_MATCH) {
|
|
|
|
|
i += vpc->v.vc1dsp.startcode_find_candidate(buf + i, buf_size - i);
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
if (i < buf_size) {
|
|
|
|
|
search_state = ONE_ZERO;
|
|
|
|
|
}
|
|
|
|
|
i++;
|
|
|
|
|
} else {
|
|
|
|
|
b = buf[i++];
|
|
|
|
|
if (search_state == ONE_ZERO)
|
|
|
|
|
search_state = b ? NO_MATCH : TWO_ZEROS;
|
|
|
|
|
else if (search_state == TWO_ZEROS) {
|
|
|
|
|
if (b >= 1)
|
|
|
|
|
search_state = b == 1 ? ONE : NO_MATCH;
|
|
|
|
|
}
|
|
|
|
|
else { // search_state == ONE
|
|
|
|
|
search_state = NO_MATCH;
|
|
|
|
|
start_code_found = 1;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
if (start_code_found) {
|
|
|
|
|
vc1_extract_header(s, avctx, unesc_buffer, unesc_index);
|
|
|
|
|
|
|
|
|
|
vpc->prev_start_code = b;
|
|
|
|
|
unesc_index = 0;
|
|
|
|
|
|
|
|
|
|
if (!(s->flags & PARSER_FLAG_COMPLETE_FRAMES)) {
|
|
|
|
|
if (!pic_found && (b == (VC1_CODE_FRAME & 0xFF) || b == (VC1_CODE_FIELD & 0xFF))) {
|
|
|
|
|
pic_found = 1;
|
|
|
|
|
}
|
|
|
|
|
else if (pic_found && b != (VC1_CODE_FIELD & 0xFF) && b != (VC1_CODE_SLICE & 0xFF)) {
|
|
|
|
|
next = i - 4;
|
|
|
|
|
pic_found = b == (VC1_CODE_FRAME & 0xFF);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
vpc->pc.frame_start_found = pic_found;
|
|
|
|
|
vpc->unesc_index = unesc_index;
|
|
|
|
|
vpc->search_state = search_state;
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
if (s->flags & PARSER_FLAG_COMPLETE_FRAMES) {
|
|
|
|
|
next = buf_size;
|
|
|
|
|
} else {
|
|
|
|
|
if (ff_combine_frame(&vpc->pc, next, &buf, &buf_size) < 0) {
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
vpc->bytes_to_skip = 0;
|
|
|
|
|
*poutbuf = NULL;
|
|
|
|
|
*poutbuf_size = 0;
|
|
|
|
|
return buf_size;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
/* If we return with a valid pointer to a combined frame buffer
|
|
|
|
|
* then on the next call then we'll have been unhelpfully rewound
|
|
|
|
|
* by up to 4 bytes (depending upon whether the start code
|
|
|
|
|
* overlapped the input buffer, and if so by how much). We don't
|
|
|
|
|
* want this: it will either cause spurious second detections of
|
|
|
|
|
* the start code we've already seen, or cause extra bytes to be
|
|
|
|
|
* inserted at the start of the unescaped buffer. */
|
|
|
|
|
vpc->bytes_to_skip = 4;
|
|
|
|
|
if (next < 0 && next != END_NOT_FOUND)
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
vpc->bytes_to_skip += next;
|
|
|
|
|
|
|
|
|
|
*poutbuf = buf;
|
|
|
|
|
*poutbuf_size = buf_size;
|
|
|
|
|
return next;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int vc1_split(AVCodecContext *avctx,
|
|
|
|
|
const uint8_t *buf, int buf_size)
|
|
|
|
|
{
|
|
|
|
|
uint32_t state = -1;
|
|
|
|
|
int charged = 0;
|
|
|
|
|
const uint8_t *ptr = buf, *end = buf + buf_size;
|
|
|
|
|
|
|
|
|
|
while (ptr < end) {
|
|
|
|
|
ptr = avpriv_find_start_code(ptr, end, &state);
|
|
|
|
|
if (state == VC1_CODE_SEQHDR || state == VC1_CODE_ENTRYPOINT) {
|
|
|
|
|
charged = 1;
|
|
|
|
|
} else if (charged && IS_MARKER(state))
|
|
|
|
|
return ptr - 4 - buf;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static av_cold int vc1_parse_init(AVCodecParserContext *s)
|
|
|
|
|
{
|
|
|
|
|
VC1ParseContext *vpc = s->priv_data;
|
|
|
|
|
vpc->v.s.slice_context_count = 1;
|
vc-1: Optimise parser (with special attention to ARM)
The previous implementation of the parser made four passes over each input
buffer (reduced to two if the container format already guaranteed the input
buffer corresponded to frames, such as with MKV). But these buffers are
often 200K in size, certainly enough to flush the data out of L1 cache, and
for many CPUs, all the way out to main memory. The passes were:
1) locate frame boundaries (not needed for MKV etc)
2) copy the data into a contiguous block (not needed for MKV etc)
3) locate the start codes within each frame
4) unescape the data between start codes
After this, the unescaped data was parsed to extract certain header fields,
but because the unescape operation was so large, this was usually also
effectively operating on uncached memory. Most of the unescaped data was
simply thrown away and never processed further. Only step 2 - because it
used memcpy - was using prefetch, making things even worse.
This patch reorganises these steps so that, aside from the copying, the
operations are performed in parallel, maximising cache utilisation. No more
than the worst-case number of bytes needed for header parsing is unescaped.
Most of the data is, in practice, only read in order to search for a start
code, for which optimised implementations already existed in the H264 codec
(notably the ARM version uses prefetch, so we end up doing both remaining
passes at maximum speed). For MKV files, we know when we've found the last
start code of interest in a given frame, so we are able to avoid doing even
that one remaining pass for most of the buffer.
In some use-cases (such as the Raspberry Pi) video decode is handled by the
GPU, but the entire elementary stream is still fed through the parser to
pick out certain elements of the header which are necessary to manage the
decode process. As you might expect, in these cases, the performance of the
parser is significant.
To measure parser performance, I used the same VC-1 elementary stream in
either an MPEG-2 transport stream or a MKV file, and fed it through ffmpeg
with -c:v copy -c:a copy -f null. These are the gperftools counts for
those streams, both filtered to only include vc1_parse() and its callees,
and unfiltered (to include the whole binary). Lower numbers are better:
Before After
File Filtered Mean StdDev Mean StdDev Confidence Change
M2TS No 861.7 8.2 650.5 8.1 100.0% +32.5%
MKV No 868.9 7.4 731.7 9.0 100.0% +18.8%
M2TS Yes 250.0 11.2 27.2 3.4 100.0% +817.9%
MKV Yes 149.0 12.8 1.7 0.8 100.0% +8526.3%
Yes, that last case shows vc1_parse() running 86 times faster! The M2TS
case does show a larger absolute improvement though, since it was worse
to begin with.
This patch has been tested with the FATE suite (albeit on x86 for speed).
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
11 years ago
|
|
|
vpc->v.first_pic_header_flag = 1;
|
|
|
|
|
vpc->prev_start_code = 0;
|
|
|
|
|
vpc->bytes_to_skip = 0;
|
|
|
|
|
vpc->unesc_index = 0;
|
|
|
|
|
vpc->search_state = NO_MATCH;
|
|
|
|
|
return ff_vc1_init_common(&vpc->v);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
AVCodecParser ff_vc1_parser = {
|
|
|
|
|
.codec_ids = { AV_CODEC_ID_VC1 },
|
|
|
|
|
.priv_data_size = sizeof(VC1ParseContext),
|
|
|
|
|
.parser_init = vc1_parse_init,
|
|
|
|
|
.parser_parse = vc1_parse,
|
|
|
|
|
.parser_close = ff_parse_close,
|
|
|
|
|
.split = vc1_split,
|
|
|
|
|
};
|
