You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

711 lines
23 KiB

/*
* MagicYUV decoder
* Copyright (c) 2016 Paul B Mahol
*
* 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
*/
#include <stdlib.h>
#include <string.h>
#define CACHED_BITSTREAM_READER !ARCH_X86_32
#include "libavutil/pixdesc.h"
#include "avcodec.h"
#include "bytestream.h"
#include "codec_internal.h"
#include "decode.h"
#include "get_bits.h"
#include "lossless_videodsp.h"
#include "thread.h"
#define VLC_BITS 12
typedef struct Slice {
uint32_t start;
uint32_t size;
} Slice;
typedef enum Prediction {
LEFT = 1,
GRADIENT,
MEDIAN,
} Prediction;
typedef struct HuffEntry {
uint8_t len;
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
uint16_t sym;
} HuffEntry;
typedef struct MagicYUVContext {
AVFrame *p;
int max;
int bps;
int slice_height;
int nb_slices;
int planes; // number of encoded planes in bitstream
int decorrelate; // postprocessing work
int color_matrix; // video color matrix
int flags;
int interlaced; // video is interlaced
const uint8_t *buf; // pointer to AVPacket->data
int hshift[4];
int vshift[4];
Slice *slices[4]; // slice bitstream positions for each plane
unsigned int slices_size[4]; // slice sizes for each plane
VLC vlc[4]; // VLC for each plane
VLC_MULTI multi[4]; // Buffer for joint VLC data
int (*magy_decode_slice)(AVCodecContext *avctx, void *tdata,
int j, int threadnr);
LLVidDSPContext llviddsp;
} MagicYUVContext;
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
static int huff_build(const uint8_t len[], uint16_t codes_pos[33],
VLC *vlc, VLC_MULTI *multi, int nb_elems, void *logctx)
{
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
HuffEntry he[4096];
for (int i = 31; i > 0; i--)
codes_pos[i] += codes_pos[i + 1];
for (unsigned i = nb_elems; i-- > 0;)
he[--codes_pos[len[i]]] = (HuffEntry){ len[i], i };
ff_free_vlc(vlc);
ff_free_vlc_multi(multi);
return ff_init_vlc_multi_from_lengths(vlc, multi, FFMIN(he[0].len, VLC_BITS), nb_elems, nb_elems,
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
&he[0].len, sizeof(he[0]),
&he[0].sym, sizeof(he[0]), sizeof(he[0].sym),
0, 0, logctx);
}
static void magicyuv_median_pred16(uint16_t *dst, const uint16_t *src1,
const uint16_t *diff, intptr_t w,
int *left, int *left_top, int max)
{
int i;
uint16_t l, lt;
l = *left;
lt = *left_top;
for (i = 0; i < w; i++) {
l = mid_pred(l, src1[i], (l + src1[i] - lt)) + diff[i];
l &= max;
lt = src1[i];
dst[i] = l;
}
*left = l;
*left_top = lt;
}
#define READ_PLANE(dst, plane, b, c) \
{ \
x = 0; \
for (; CACHED_BITSTREAM_READER && x < width-c && get_bits_left(&gb) > 0;) {\
ret = get_vlc_multi(&gb, (uint8_t *)dst + x * b, multi, \
vlc, vlc_bits, 3); \
if (ret > 0) \
x += ret; \
if (ret <= 0) \
return AVERROR_INVALIDDATA; \
} \
for (; x < width && get_bits_left(&gb) > 0; x++) \
dst[x] = get_vlc2(&gb, vlc, vlc_bits, 3); \
dst += stride; \
}
static int magy_decode_slice10(AVCodecContext *avctx, void *tdata,
int j, int threadnr)
{
const MagicYUVContext *s = avctx->priv_data;
int interlaced = s->interlaced;
const int bps = s->bps;
const int max = s->max - 1;
AVFrame *p = s->p;
int i, k, x;
GetBitContext gb;
uint16_t *dst;
for (i = 0; i < s->planes; i++) {
int left, lefttop, top;
int height = AV_CEIL_RSHIFT(FFMIN(s->slice_height, avctx->coded_height - j * s->slice_height), s->vshift[i]);
int width = AV_CEIL_RSHIFT(avctx->coded_width, s->hshift[i]);
int sheight = AV_CEIL_RSHIFT(s->slice_height, s->vshift[i]);
ptrdiff_t fake_stride = (p->linesize[i] / 2) * (1 + interlaced);
ptrdiff_t stride = p->linesize[i] / 2;
const VLC_MULTI_ELEM *const multi = s->multi[i].table;
const VLCElem *const vlc = s->vlc[i].table;
const int vlc_bits = s->vlc[i].bits;
int flags, pred;
int ret = init_get_bits8(&gb, s->buf + s->slices[i][j].start,
s->slices[i][j].size);
if (ret < 0)
return ret;
flags = get_bits(&gb, 8);
pred = get_bits(&gb, 8);
dst = (uint16_t *)p->data[i] + j * sheight * stride;
if (flags & 1) {
if (get_bits_left(&gb) < bps * width * height)
return AVERROR_INVALIDDATA;
for (k = 0; k < height; k++) {
for (x = 0; x < width; x++)
dst[x] = get_bits(&gb, bps);
dst += stride;
}
} else {
for (k = 0; k < height; k++)
READ_PLANE(dst, i, 2, 3)
}
switch (pred) {
case LEFT:
dst = (uint16_t *)p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
}
for (k = 1 + interlaced; k < height; k++) {
s->llviddsp.add_left_pred_int16(dst, dst, max, width, dst[-fake_stride]);
dst += stride;
}
break;
case GRADIENT:
dst = (uint16_t *)p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
}
for (k = 1 + interlaced; k < height; k++) {
top = dst[-fake_stride];
left = top + dst[0];
dst[0] = left & max;
for (x = 1; x < width; x++) {
top = dst[x - fake_stride];
lefttop = dst[x - (fake_stride + 1)];
left += top - lefttop + dst[x];
dst[x] = left & max;
}
dst += stride;
}
break;
case MEDIAN:
dst = (uint16_t *)p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred_int16(dst, dst, max, width, 0);
dst += stride;
}
lefttop = left = dst[0];
for (k = 1 + interlaced; k < height; k++) {
magicyuv_median_pred16(dst, dst - fake_stride, dst, width, &left, &lefttop, max);
lefttop = left = dst[0];
dst += stride;
}
break;
default:
avpriv_request_sample(avctx, "Unknown prediction: %d", pred);
}
}
if (s->decorrelate) {
int height = FFMIN(s->slice_height, avctx->coded_height - j * s->slice_height);
int width = avctx->coded_width;
uint16_t *r = (uint16_t *)p->data[0] + j * s->slice_height * p->linesize[0] / 2;
uint16_t *g = (uint16_t *)p->data[1] + j * s->slice_height * p->linesize[1] / 2;
uint16_t *b = (uint16_t *)p->data[2] + j * s->slice_height * p->linesize[2] / 2;
for (i = 0; i < height; i++) {
for (k = 0; k < width; k++) {
b[k] = (b[k] + g[k]) & max;
r[k] = (r[k] + g[k]) & max;
}
b += p->linesize[0] / 2;
g += p->linesize[1] / 2;
r += p->linesize[2] / 2;
}
}
return 0;
}
static int magy_decode_slice(AVCodecContext *avctx, void *tdata,
int j, int threadnr)
{
const MagicYUVContext *s = avctx->priv_data;
int interlaced = s->interlaced;
AVFrame *p = s->p;
int i, k, x, min_width;
GetBitContext gb;
uint8_t *dst;
for (i = 0; i < s->planes; i++) {
int left, lefttop, top;
int height = AV_CEIL_RSHIFT(FFMIN(s->slice_height, avctx->coded_height - j * s->slice_height), s->vshift[i]);
int width = AV_CEIL_RSHIFT(avctx->coded_width, s->hshift[i]);
int sheight = AV_CEIL_RSHIFT(s->slice_height, s->vshift[i]);
ptrdiff_t fake_stride = p->linesize[i] * (1 + interlaced);
ptrdiff_t stride = p->linesize[i];
const uint8_t *slice = s->buf + s->slices[i][j].start;
const VLC_MULTI_ELEM *const multi = s->multi[i].table;
const VLCElem *const vlc = s->vlc[i].table;
const int vlc_bits = s->vlc[i].bits;
int flags, pred;
flags = bytestream_get_byte(&slice);
pred = bytestream_get_byte(&slice);
dst = p->data[i] + j * sheight * stride;
if (flags & 1) {
if (s->slices[i][j].size - 2 < width * height)
return AVERROR_INVALIDDATA;
for (k = 0; k < height; k++) {
bytestream_get_buffer(&slice, dst, width);
dst += stride;
}
} else {
int ret = init_get_bits8(&gb, slice, s->slices[i][j].size - 2);
if (ret < 0)
return ret;
for (k = 0; k < height; k++)
READ_PLANE(dst, i, 1, 5)
}
switch (pred) {
case LEFT:
dst = p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
}
for (k = 1 + interlaced; k < height; k++) {
s->llviddsp.add_left_pred(dst, dst, width, dst[-fake_stride]);
dst += stride;
}
break;
case GRADIENT:
dst = p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
}
min_width = FFMIN(width, 32);
for (k = 1 + interlaced; k < height; k++) {
top = dst[-fake_stride];
left = top + dst[0];
dst[0] = left;
for (x = 1; x < min_width; x++) { /* dsp need aligned 32 */
top = dst[x - fake_stride];
lefttop = dst[x - (fake_stride + 1)];
left += top - lefttop + dst[x];
dst[x] = left;
}
if (width > 32)
s->llviddsp.add_gradient_pred(dst + 32, fake_stride, width - 32);
dst += stride;
}
break;
case MEDIAN:
dst = p->data[i] + j * sheight * stride;
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
if (interlaced) {
s->llviddsp.add_left_pred(dst, dst, width, 0);
dst += stride;
}
lefttop = left = dst[0];
for (k = 1 + interlaced; k < height; k++) {
s->llviddsp.add_median_pred(dst, dst - fake_stride,
dst, width, &left, &lefttop);
lefttop = left = dst[0];
dst += stride;
}
break;
default:
avpriv_request_sample(avctx, "Unknown prediction: %d", pred);
}
}
if (s->decorrelate) {
int height = FFMIN(s->slice_height, avctx->coded_height - j * s->slice_height);
int width = avctx->coded_width;
uint8_t *b = p->data[0] + j * s->slice_height * p->linesize[0];
uint8_t *g = p->data[1] + j * s->slice_height * p->linesize[1];
uint8_t *r = p->data[2] + j * s->slice_height * p->linesize[2];
for (i = 0; i < height; i++) {
s->llviddsp.add_bytes(b, g, width);
s->llviddsp.add_bytes(r, g, width);
b += p->linesize[0];
g += p->linesize[1];
r += p->linesize[2];
}
}
return 0;
}
static int build_huffman(AVCodecContext *avctx, const uint8_t *table,
int table_size, int max)
{
MagicYUVContext *s = avctx->priv_data;
GetByteContext gb;
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
uint8_t len[4096];
avcodec/magicyuv: Avoid AV_QSORT when creating Huffman table The MagicYUV format stores Huffman tables in its bitstream by coding the length of a given symbol; it does not code the actual code directly, instead this is to be inferred by the rule that a symbol is to the left of every shorter symbol in the Huffman tree and that for symbols of the same length the symbol is ascending from left to right. Our decoder implemented this by first sorting the array containing length and symbol of each element according to descending length and for equal length, according to ascending symbol. Afterwards, the current state in the tree got encoded in a variable code; if the next array entry had length len, then the len most significant bits of code contained the code of this entry. Whenever an entry of the array of length len was processed, code was incremented by 1U << (32 - len). So two entries of length len have the same effect as incrementing code by 1U << (32 - (len - 1)), which corresponds to the parent node of length len - 1 of the two nodes of length len etc. This commit modifies this to avoid sorting the entries before calculating the codes. This is done by calculating how many non-leaf nodes there are on each level of the tree before calculating the codes. Afterwards every leaf node on this level gets assigned the number of nodes already on this level as code. This of course works only because the entries are already sorted by their symbol initially, so that this algorithm indeed gives ascending symbols from left to right on every level. This offers both speed- as well as (obvious) codesize advantages. With Clang 10 the number of decicycles for build_huffman decreased from 1561987 to 1228405; for GCC 9 it went from 1825096 decicyles to 1429921. These tests were carried out with a sample with 150 frames that was looped 13 times; and this was iterated 10 times. The earlier reference point here is from the point when the loop generating the codes was traversed in reverse order (as the patch reversing the order led to performance penalties). Reviewed-by: Paul B Mahol <onemda@gmail.com> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
uint16_t length_count[33] = { 0 };
int i = 0, j = 0, k;
bytestream2_init(&gb, table, table_size);
while (bytestream2_get_bytes_left(&gb) > 0) {
int b = bytestream2_peek_byteu(&gb) & 0x80;
int x = bytestream2_get_byteu(&gb) & ~0x80;
int l = 1;
if (b) {
if (bytestream2_get_bytes_left(&gb) <= 0)
break;
l += bytestream2_get_byteu(&gb);
}
k = j + l;
if (k > max || x == 0 || x > 32) {
av_log(avctx, AV_LOG_ERROR, "Invalid Huffman codes\n");
return AVERROR_INVALIDDATA;
}
avcodec/magicyuv: Avoid AV_QSORT when creating Huffman table The MagicYUV format stores Huffman tables in its bitstream by coding the length of a given symbol; it does not code the actual code directly, instead this is to be inferred by the rule that a symbol is to the left of every shorter symbol in the Huffman tree and that for symbols of the same length the symbol is ascending from left to right. Our decoder implemented this by first sorting the array containing length and symbol of each element according to descending length and for equal length, according to ascending symbol. Afterwards, the current state in the tree got encoded in a variable code; if the next array entry had length len, then the len most significant bits of code contained the code of this entry. Whenever an entry of the array of length len was processed, code was incremented by 1U << (32 - len). So two entries of length len have the same effect as incrementing code by 1U << (32 - (len - 1)), which corresponds to the parent node of length len - 1 of the two nodes of length len etc. This commit modifies this to avoid sorting the entries before calculating the codes. This is done by calculating how many non-leaf nodes there are on each level of the tree before calculating the codes. Afterwards every leaf node on this level gets assigned the number of nodes already on this level as code. This of course works only because the entries are already sorted by their symbol initially, so that this algorithm indeed gives ascending symbols from left to right on every level. This offers both speed- as well as (obvious) codesize advantages. With Clang 10 the number of decicycles for build_huffman decreased from 1561987 to 1228405; for GCC 9 it went from 1825096 decicyles to 1429921. These tests were carried out with a sample with 150 frames that was looped 13 times; and this was iterated 10 times. The earlier reference point here is from the point when the loop generating the codes was traversed in reverse order (as the patch reversing the order led to performance penalties). Reviewed-by: Paul B Mahol <onemda@gmail.com> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
length_count[x] += l;
for (; j < k; j++)
avcodec/magicyuv: Optimize creating Huffman tables MagicYUV transmits its Huffman trees by providing the length of the code corresponding to each symbol; then the decoder has to assemble the table in such a way that (i) longer codes are to the left of the tree and (ii) for codes of the same length the symbols are ascending from left to right. Up until now the decoder did this as follows: It counted the number of codes of each length and derived the first code of a given length via (ii). Then the array of lengths is traversed a second time to create the codes; there is one running counter for each length to do so. This process creates a default symbol table (that is omitted). This commit changes this as follows: Everything is indexed by the position in the tree (with codes to the left first); given (i), we can calculate the ranges occupied by the codes of each length; and with (ii) we can derive the actual symbols of each code; the running counters for each length are now used for the symbols and not for the codes. Doing so allows us to switch to ff_init_vlc_from_lengths(); this has the advantage that the codes table needs only be traversed once and that the codes need not be sorted any more (right now, the codes that are so long that they will be put into subtables need to be sorted so that codes that end up in the same subtable are contiguous). For a sample produced by our encoder (natural content, 4000 frames, YUV420p, ten iterations, GCC 9.3) this decreased the amount of decicycles for each call to build_huffman() from 1336049 to 1309401. Notice that our encoder restricts the code lengths to 12 and our decoder only uses subtables when the code is longer than 12 bits, so the sorting that can be avoided does not happen at the moment. If one reduces the decoder's tables to nine bits, the performance improvement becomes more apparent: The amount of decicycles for build_huffman() decreased from 1165210 to 654055. Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
len[j] = x;
if (j == max) {
j = 0;
if (huff_build(len, length_count, &s->vlc[i], &s->multi[i], max, avctx)) {
av_log(avctx, AV_LOG_ERROR, "Cannot build Huffman codes\n");
return AVERROR_INVALIDDATA;
}
i++;
if (i == s->planes) {
break;
}
avcodec/magicyuv: Avoid AV_QSORT when creating Huffman table The MagicYUV format stores Huffman tables in its bitstream by coding the length of a given symbol; it does not code the actual code directly, instead this is to be inferred by the rule that a symbol is to the left of every shorter symbol in the Huffman tree and that for symbols of the same length the symbol is ascending from left to right. Our decoder implemented this by first sorting the array containing length and symbol of each element according to descending length and for equal length, according to ascending symbol. Afterwards, the current state in the tree got encoded in a variable code; if the next array entry had length len, then the len most significant bits of code contained the code of this entry. Whenever an entry of the array of length len was processed, code was incremented by 1U << (32 - len). So two entries of length len have the same effect as incrementing code by 1U << (32 - (len - 1)), which corresponds to the parent node of length len - 1 of the two nodes of length len etc. This commit modifies this to avoid sorting the entries before calculating the codes. This is done by calculating how many non-leaf nodes there are on each level of the tree before calculating the codes. Afterwards every leaf node on this level gets assigned the number of nodes already on this level as code. This of course works only because the entries are already sorted by their symbol initially, so that this algorithm indeed gives ascending symbols from left to right on every level. This offers both speed- as well as (obvious) codesize advantages. With Clang 10 the number of decicycles for build_huffman decreased from 1561987 to 1228405; for GCC 9 it went from 1825096 decicyles to 1429921. These tests were carried out with a sample with 150 frames that was looped 13 times; and this was iterated 10 times. The earlier reference point here is from the point when the loop generating the codes was traversed in reverse order (as the patch reversing the order led to performance penalties). Reviewed-by: Paul B Mahol <onemda@gmail.com> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
4 years ago
memset(length_count, 0, sizeof(length_count));
}
}
if (i != s->planes) {
av_log(avctx, AV_LOG_ERROR, "Huffman tables too short\n");
return AVERROR_INVALIDDATA;
}
return 0;
}
static int magy_decode_frame(AVCodecContext *avctx, AVFrame *p,
int *got_frame, AVPacket *avpkt)
{
MagicYUVContext *s = avctx->priv_data;
GetByteContext gb;
uint32_t first_offset, offset, next_offset, header_size, slice_width;
int width, height, format, version, table_size;
int ret, i, j;
if (avpkt->size < 36)
return AVERROR_INVALIDDATA;
bytestream2_init(&gb, avpkt->data, avpkt->size);
if (bytestream2_get_le32u(&gb) != MKTAG('M', 'A', 'G', 'Y'))
return AVERROR_INVALIDDATA;
header_size = bytestream2_get_le32u(&gb);
if (header_size < 32 || header_size >= avpkt->size) {
av_log(avctx, AV_LOG_ERROR,
"header or packet too small %"PRIu32"\n", header_size);
return AVERROR_INVALIDDATA;
}
version = bytestream2_get_byteu(&gb);
if (version != 7) {
avpriv_request_sample(avctx, "Version %d", version);
return AVERROR_PATCHWELCOME;
}
s->hshift[1] =
s->vshift[1] =
s->hshift[2] =
s->vshift[2] = 0;
s->decorrelate = 0;
s->bps = 8;
format = bytestream2_get_byteu(&gb);
switch (format) {
case 0x65:
avctx->pix_fmt = AV_PIX_FMT_GBRP;
s->decorrelate = 1;
break;
case 0x66:
avctx->pix_fmt = AV_PIX_FMT_GBRAP;
s->decorrelate = 1;
break;
case 0x67:
avctx->pix_fmt = AV_PIX_FMT_YUV444P;
break;
case 0x68:
avctx->pix_fmt = AV_PIX_FMT_YUV422P;
s->hshift[1] =
s->hshift[2] = 1;
break;
case 0x69:
avctx->pix_fmt = AV_PIX_FMT_YUV420P;
s->hshift[1] =
s->vshift[1] =
s->hshift[2] =
s->vshift[2] = 1;
break;
case 0x6a:
avctx->pix_fmt = AV_PIX_FMT_YUVA444P;
break;
case 0x6b:
avctx->pix_fmt = AV_PIX_FMT_GRAY8;
break;
case 0x6c:
avctx->pix_fmt = AV_PIX_FMT_YUV422P10;
s->hshift[1] =
s->hshift[2] = 1;
s->bps = 10;
break;
case 0x76:
avctx->pix_fmt = AV_PIX_FMT_YUV444P10;
s->bps = 10;
break;
case 0x6d:
avctx->pix_fmt = AV_PIX_FMT_GBRP10;
s->decorrelate = 1;
s->bps = 10;
break;
case 0x6e:
avctx->pix_fmt = AV_PIX_FMT_GBRAP10;
s->decorrelate = 1;
s->bps = 10;
break;
case 0x6f:
avctx->pix_fmt = AV_PIX_FMT_GBRP12;
s->decorrelate = 1;
s->bps = 12;
break;
case 0x70:
avctx->pix_fmt = AV_PIX_FMT_GBRAP12;
s->decorrelate = 1;
s->bps = 12;
break;
case 0x73:
avctx->pix_fmt = AV_PIX_FMT_GRAY10;
s->bps = 10;
break;
case 0x7b:
avctx->pix_fmt = AV_PIX_FMT_YUV420P10;
s->hshift[1] =
s->vshift[1] =
s->hshift[2] =
s->vshift[2] = 1;
s->bps = 10;
break;
default:
avpriv_request_sample(avctx, "Format 0x%X", format);
return AVERROR_PATCHWELCOME;
}
s->max = 1 << s->bps;
s->magy_decode_slice = s->bps == 8 ? magy_decode_slice : magy_decode_slice10;
s->planes = av_pix_fmt_count_planes(avctx->pix_fmt);
bytestream2_skipu(&gb, 1);
s->color_matrix = bytestream2_get_byteu(&gb);
s->flags = bytestream2_get_byteu(&gb);
s->interlaced = !!(s->flags & 2);
bytestream2_skipu(&gb, 3);
width = bytestream2_get_le32u(&gb);
height = bytestream2_get_le32u(&gb);
ret = ff_set_dimensions(avctx, width, height);
if (ret < 0)
return ret;
slice_width = bytestream2_get_le32u(&gb);
if (slice_width != avctx->coded_width) {
avpriv_request_sample(avctx, "Slice width %"PRIu32, slice_width);
return AVERROR_PATCHWELCOME;
}
s->slice_height = bytestream2_get_le32u(&gb);
if (s->slice_height <= 0 || s->slice_height > INT_MAX - avctx->coded_height) {
av_log(avctx, AV_LOG_ERROR,
"invalid slice height: %d\n", s->slice_height);
return AVERROR_INVALIDDATA;
}
bytestream2_skipu(&gb, 4);
s->nb_slices = (avctx->coded_height + s->slice_height - 1) / s->slice_height;
if (s->nb_slices > INT_MAX / FFMAX(sizeof(Slice), 4 * 5)) {
av_log(avctx, AV_LOG_ERROR,
"invalid number of slices: %d\n", s->nb_slices);
return AVERROR_INVALIDDATA;
}
if (s->interlaced) {
if ((s->slice_height >> s->vshift[1]) < 2) {
av_log(avctx, AV_LOG_ERROR, "impossible slice height\n");
return AVERROR_INVALIDDATA;
}
if ((avctx->coded_height % s->slice_height) && ((avctx->coded_height % s->slice_height) >> s->vshift[1]) < 2) {
av_log(avctx, AV_LOG_ERROR, "impossible height\n");
return AVERROR_INVALIDDATA;
}
}
if (bytestream2_get_bytes_left(&gb) <= s->nb_slices * s->planes * 5)
return AVERROR_INVALIDDATA;
for (i = 0; i < s->planes; i++) {
av_fast_malloc(&s->slices[i], &s->slices_size[i], s->nb_slices * sizeof(Slice));
if (!s->slices[i])
return AVERROR(ENOMEM);
offset = bytestream2_get_le32u(&gb);
if (offset >= avpkt->size - header_size)
return AVERROR_INVALIDDATA;
if (i == 0)
first_offset = offset;
for (j = 0; j < s->nb_slices - 1; j++) {
s->slices[i][j].start = offset + header_size;
next_offset = bytestream2_get_le32u(&gb);
if (next_offset <= offset || next_offset >= avpkt->size - header_size)
return AVERROR_INVALIDDATA;
s->slices[i][j].size = next_offset - offset;
if (s->slices[i][j].size < 2)
return AVERROR_INVALIDDATA;
offset = next_offset;
}
s->slices[i][j].start = offset + header_size;
s->slices[i][j].size = avpkt->size - s->slices[i][j].start;
if (s->slices[i][j].size < 2)
return AVERROR_INVALIDDATA;
}
if (bytestream2_get_byteu(&gb) != s->planes)
return AVERROR_INVALIDDATA;
bytestream2_skipu(&gb, s->nb_slices * s->planes);
table_size = header_size + first_offset - bytestream2_tell(&gb);
if (table_size < 2)
return AVERROR_INVALIDDATA;
ret = build_huffman(avctx, avpkt->data + bytestream2_tell(&gb),
table_size, s->max);
if (ret < 0)
return ret;
p->pict_type = AV_PICTURE_TYPE_I;
p->flags |= AV_FRAME_FLAG_KEY;
if ((ret = ff_thread_get_buffer(avctx, p, 0)) < 0)
return ret;
s->buf = avpkt->data;
s->p = p;
avctx->execute2(avctx, s->magy_decode_slice, NULL, NULL, s->nb_slices);
if (avctx->pix_fmt == AV_PIX_FMT_GBRP ||
avctx->pix_fmt == AV_PIX_FMT_GBRAP ||
avctx->pix_fmt == AV_PIX_FMT_GBRP10 ||
avctx->pix_fmt == AV_PIX_FMT_GBRAP10||
avctx->pix_fmt == AV_PIX_FMT_GBRAP12||
avctx->pix_fmt == AV_PIX_FMT_GBRP12) {
FFSWAP(uint8_t*, p->data[0], p->data[1]);
FFSWAP(int, p->linesize[0], p->linesize[1]);
} else {
switch (s->color_matrix) {
case 1:
p->colorspace = AVCOL_SPC_BT470BG;
break;
case 2:
p->colorspace = AVCOL_SPC_BT709;
break;
}
p->color_range = (s->flags & 4) ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
}
*got_frame = 1;
return avpkt->size;
}
static av_cold int magy_decode_init(AVCodecContext *avctx)
{
MagicYUVContext *s = avctx->priv_data;
ff_llviddsp_init(&s->llviddsp);
return 0;
}
static av_cold int magy_decode_end(AVCodecContext *avctx)
{
MagicYUVContext * const s = avctx->priv_data;
int i;
for (i = 0; i < FF_ARRAY_ELEMS(s->slices); i++) {
av_freep(&s->slices[i]);
s->slices_size[i] = 0;
ff_free_vlc(&s->vlc[i]);
ff_free_vlc_multi(&s->multi[i]);
}
return 0;
}
const FFCodec ff_magicyuv_decoder = {
.p.name = "magicyuv",
CODEC_LONG_NAME("MagicYUV video"),
.p.type = AVMEDIA_TYPE_VIDEO,
.p.id = AV_CODEC_ID_MAGICYUV,
.priv_data_size = sizeof(MagicYUVContext),
.init = magy_decode_init,
.close = magy_decode_end,
FF_CODEC_DECODE_CB(magy_decode_frame),
.p.capabilities = AV_CODEC_CAP_DR1 |
AV_CODEC_CAP_FRAME_THREADS |
AV_CODEC_CAP_SLICE_THREADS,
};