dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-11-19 20:08:28 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

multi-thread decoding: ~25-30% faster

Browse Source 
To be enabled with the flag WEBP_USE_THREAD.
For now it's only available on unix (pthread), when using Makefile.unix
Will be switched on more generally later.

In-loop filtering and output (=rescaling/yuv->rgb conversion)
is done in parallel to bitstream decoding, lagging 1 row behind.

Example:
examples/dwebp bryce.webp -v
Time to decode picture: 0.680s

examples/dwebp bryce.webp -v -mt
Time to decode picture: 0.515s

Change-Id: Ic30a897423137a3bdace9c4e30465ef758fe53f2
This commit is contained in:
Pascal Massimino 2011-07-22 13:09:10 -07:00
parent acd8ba4229
commit fc7815d692
16 changed files with 519 additions and 120 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
Android.mk
View File

@ -15,6 +15,7 @@ LOCAL_SRC_FILES := \
src/dec/io.c \
src/dec/buffer.c \
src/dec/yuv.c \
src/dec/thread.c \
src/enc/alpha.c \
src/enc/analysis.c \
src/enc/bit_writer.c \
@ -30,7 +31,7 @@ LOCAL_SRC_FILES := \
src/enc/tree.c \
src/enc/webpenc.c
LOCAL_CFLAGS := -Wall -DANDROID -DHAVE_MALLOC_H -DHAVE_PTHREAD \
LOCAL_CFLAGS := -Wall -DANDROID -DHAVE_MALLOC_H -DHAVE_PTHREAD -DWEBP_USE_THREAD \
-finline-functions -frename-registers -ffast-math \
-s -fomit-frame-pointer -Isrc/webp

1
Makefile.vc
View File

@ -146,6 +146,7 @@ X_OBJS= \
$(DIROBJ)\dec\idec.obj \
$(DIROBJ)\dec\alpha.obj \
$(DIROBJ)\dec\layer.obj \
$(DIROBJ)\dec\thread.obj \
$(DIROBJ)\enc\analysis.obj \
$(DIROBJ)\enc\bit_writer.obj \
$(DIROBJ)\enc\config.obj \

1
README
View File

@ -217,6 +217,7 @@ Use following options to convert into alternate image formats:
-version .... print version number and exit.
-nofancy ..... don't use the fancy YUV420 upscaler.
-nofilter .... disable in-loop filtering.
-mt .......... use multi-threading
-crop <x> <y> <w> <h> ... crop output with the given rectangle
-scale <w> <h> .......... scale the output (*after* any cropping)
-h ....... this help message.

3
examples/dwebp.c
View File

@ -318,6 +318,7 @@ static void Help(void) {
" -version .... print version number and exit.\n"
" -nofancy ..... don't use the fancy YUV420 upscaler.\n"
" -nofilter .... disable in-loop filtering.\n"
" -mt .......... use multi-threading\n"
" -crop <x> <y> <w> <h> ... crop output with the given rectangle\n"
" -scale <w> <h> .......... scale the output (*after* any cropping)\n"
#ifdef WEBP_EXPERIMENTAL_FEATURES
@ -372,6 +373,8 @@ int main(int argc, const char *argv[]) {
return 0;
} else if (!strcmp(argv[c], "-pgm")) {
format = PGM;
} else if (!strcmp(argv[c], "-mt")) {
config.options.use_threads = 1;
} else if (!strcmp(argv[c], "-crop") && c < argc - 4) {
config.options.use_cropping = 1;
config.options.crop_left = strtol(argv[++c], NULL, 0);

8
makefile.unix
View File

@ -36,12 +36,16 @@ endif
# Extra flags to enable experimental features and code
# EXTRA_FLAGS += -DWEBP_EXPERIMENTAL_FEATURES
# Extra flags to enable multi-threading
EXTRA_FLAGS += -DWEBP_USE_THREAD
EXTRA_LIBS += -lpthread
# Extra flags to emulate C89 strictness with the full ANSI
EXTRA_FLAGS += -Wextra -Wold-style-definition
EXTRA_FLAGS += -Wmissing-prototypes
EXTRA_FLAGS += -Wmissing-declarations
EXTRA_FLAGS += -Wdeclaration-after-statement
# EXTRA_FLAGS += -Wvla
EXTRA_FLAGS += -Wvla
#### Nothing should normally be changed below this line ####
@ -58,7 +62,7 @@ OBJS = src/enc/webpenc.o src/enc/bit_writer.o src/enc/syntax.o \
src/dec/bits.o src/dec/dsp.o src/dec/dsp_sse2.o src/dec/frame.o \
src/dec/webp.o src/dec/quant.o src/dec/tree.o src/dec/vp8.o \
src/dec/yuv.o src/dec/idec.o src/dec/alpha.o src/dec/layer.o \
src/dec/io.o src/dec/io_sse2.o src/dec/buffer.o
src/dec/io.o src/dec/io_sse2.o src/dec/buffer.o src/dec/thread.o
HDRS = src/webp/encode.h src/enc/vp8enci.h src/enc/bit_writer.h \
src/enc/cost.h src/dec/bits.h src/dec/vp8i.h src/dec/yuv.h
OUTPUT = examples/cwebp examples/dwebp src/libwebp.a

3
man/dwebp.1
View File

@ -41,6 +41,9 @@ Don't use the in-loop filtering process even if it is required by
the bitstream. This may produce visible blocks on the non-compliant output,
but will make the decoding faster.
.TP
.B \-mt
Use multi-threading for decoding, if possible.
.TP
.B \-crop x_position y_position width height
Crop the decoded picture to a rectangle with top-left corner at coordinates
(\fBx_position\fP, \fBy_position\fP) and size \fBwidth\fP x \fBheight\fP.

4
src/dec/Makefile.am
View File

@ -2,13 +2,13 @@ AM_CPPFLAGS = -I$(top_srcdir)/src
libwebpdecode_la_SOURCES = bits.h vp8i.h yuv.h bits.c dsp.c dsp_sse2.c frame.c \
quant.c tree.c vp8.c webp.c yuv.c idec.c alpha.c \
layer.c io.c io_sse2.c buffer.c
layer.c io.c io_sse2.c buffer.c thread.c
libwebpdecode_la_LDFLAGS = -version-info 0:0:0
libwebpdecode_la_CPPFLAGS = $(USE_EXPERIMENTAL_CODE)
libwebpdecodeinclude_HEADERS = ../webp/decode.h ../webp/decode_vp8.h ../webp/types.h
libwebpdecodeincludedir = $(includedir)/webp
noinst_HEADERS = bits.h vp8i.h webpi.h yuv.h
noinst_HEADERS = bits.h vp8i.h webpi.h yuv.h thread.h
noinst_LTLIBRARIES = libwebpdecode.la
# uncomment the following line (and comment the above) if you want

239
src/dec/frame.c
View File

@ -18,6 +18,53 @@ extern "C" {
#define ALIGN_MASK (32 - 1)
//------------------------------------------------------------------------------
// For multi-threaded decoding we need to use 3 rows of 16 pixels as delay line.
//
// Reason is: the deblocking filter cannot deblock the bottom horizontal edges
// immediately, and needs to wait for first few rows of the next macroblock to
// be decoded. Hence, deblocking is lagging behind by 4 or 8 pixels (depending
// on strength).
// With two threads, the vertical positions of the rows being decoded are:
// Decode: [ 0..15][16..31][32..47][48..63][64..79][...
// Deblock: [ 0..11][12..27][28..43][44..59][...
// If we use two threads and two caches of 16 pixels, the sequence would be:
// Decode: [ 0..15][16..31][ 0..15!!][16..31][ 0..15][...
// Deblock: [ 0..11][12..27!!][-4..11][12..27][...
// The problem occurs during row [12..15!!] that both the decoding and
// deblocking threads are writing simultaneously.
// With 3 cache lines, one get a safe write pattern:
// Decode: [ 0..15][16..31][32..47][ 0..15][16..31][32..47][0..
// Deblock: [ 0..11][12..27][28..43][-4..11][12..27][28...
// Note that multi-threaded output _without_ deblocking can make use of two
// cache lines of 16 pixels only, since there's no lagging behind. The decoding
// and output process have non-concurrent writing:
// Decode: [ 0..15][16..31][ 0..15][16..31][...
// io->put: [ 0..15][16..31][ 0..15][...
#define MT_CACHE_LINES 3
#define ST_CACHE_LINES 1 // 1 cache row only for single-threaded case
// Initialize multi/single-thread worker
static int InitThreadContext(VP8Decoder* const dec) {
dec->cache_id_ = 0;
if (dec->use_threads_) {
WebPWorker* const worker = &dec->worker_;
if (!WebPWorkerReset(worker)) {
return VP8SetError(dec, VP8_STATUS_OUT_OF_MEMORY,
"thread initialization failed.");
}
worker->data1 = dec;
worker->data2 = (void*)&dec->thread_ctx_.io_;
worker->hook = (WebPWorkerHook)VP8FinishRow;
dec->num_caches_ =
(dec->filter_type_ > 0) ? MT_CACHE_LINES : MT_CACHE_LINES - 1;
} else {
dec->num_caches_ = ST_CACHE_LINES;
}
return 1;
}
//------------------------------------------------------------------------------
// Memory setup
@ -28,19 +75,25 @@ extern "C" {
// U/V, so it's 8 samples total (because of the 2x upsampling).
static const uint8_t kFilterExtraRows[3] = { 0, 2, 8 };
int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
static int AllocateMemory(VP8Decoder* const dec) {
const int num_caches = dec->num_caches_;
const int mb_w = dec->mb_w_;
const int intra_pred_mode_size = 4 * mb_w * sizeof(uint8_t);
const int top_size = (16 + 8 + 8) * mb_w;
const int info_size = (mb_w + 1) * sizeof(VP8MB);
const int mb_info_size = (mb_w + 1) * sizeof(VP8MB);
const int f_info_size =
(dec->filter_type_ > 0) ?
mb_w * (dec->use_threads_ ? 2 : 1) * sizeof(VP8FInfo)
: 0;
const int yuv_size = YUV_SIZE * sizeof(*dec->yuv_b_);
const int coeffs_size = 384 * sizeof(*dec->coeffs_);
const int cache_height = (16 + kFilterExtraRows[dec->filter_type_]) * 3 / 2;
const int cache_height = (16 * num_caches
+ kFilterExtraRows[dec->filter_type_]) * 3 / 2;
const int cache_size = top_size * cache_height;
const int alpha_size =
dec->alpha_data_ ? (dec->pic_hdr_.width_ * dec->pic_hdr_.height_) : 0;
const int needed = intra_pred_mode_size
+ top_size + info_size
+ top_size + mb_info_size + f_info_size
+ yuv_size + coeffs_size
+ cache_size + alpha_size + ALIGN_MASK;
uint8_t* mem;
@ -68,7 +121,18 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
mem += 8 * mb_w;
dec->mb_info_ = ((VP8MB*)mem) + 1;
mem += info_size;
mem += mb_info_size;
dec->f_info_ = f_info_size ? (VP8FInfo*)mem : NULL;
mem += f_info_size;
dec->thread_ctx_.id_ = 0;
dec->thread_ctx_.f_info_ = dec->f_info_;
if (dec->use_threads_) {
// secondary cache line. The deblocking process need to make use of the
// filtering strength from previous macroblock row, while the new ones
// are being decoded in parallel. We'll just swap the pointers.
dec->thread_ctx_.f_info_ += mb_w;
}
mem = (uint8_t*)((uintptr_t)(mem + ALIGN_MASK) & ~ALIGN_MASK);
assert((yuv_size & ALIGN_MASK) == 0);
@ -85,8 +149,11 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
const int extra_y = extra_rows * dec->cache_y_stride_;
const int extra_uv = (extra_rows / 2) * dec->cache_uv_stride_;
dec->cache_y_ = ((uint8_t*)mem) + extra_y;
dec->cache_u_ = dec->cache_y_ + 16 * dec->cache_y_stride_ + extra_uv;
dec->cache_v_ = dec->cache_u_ + 8 * dec->cache_uv_stride_ + extra_uv;
dec->cache_u_ = dec->cache_y_
+ 16 * num_caches * dec->cache_y_stride_ + extra_uv;
dec->cache_v_ = dec->cache_u_
+ 8 * num_caches * dec->cache_uv_stride_ + extra_uv;
dec->cache_id_ = 0;
}
mem += cache_size;
@ -95,11 +162,15 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
mem += alpha_size;
// note: left-info is initialized once for all.
memset(dec->mb_info_ - 1, 0, (mb_w + 1) * sizeof(*dec->mb_info_));
memset(dec->mb_info_ - 1, 0, mb_info_size);
// initialize top
memset(dec->intra_t_, B_DC_PRED, intra_pred_mode_size);
return 1;
}
static void InitIo(VP8Decoder* const dec, VP8Io* io) {
// prepare 'io'
io->mb_y = 0;
io->y = dec->cache_y_;
@ -109,11 +180,14 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
io->uv_stride = dec->cache_uv_stride_;
io->fancy_upsampling = 0; // default
io->a = NULL;
}
// Init critical function pointers and look-up tables.
VP8DspInitTables();
int VP8InitFrame(VP8Decoder* const dec, VP8Io* io) {
if (!InitThreadContext(dec)) return 0; // call first. Sets dec->num_caches_.
if (!AllocateMemory(dec)) return 0;
InitIo(dec, io);
VP8DspInitTables(); // Init critical function pointers and look-up tables.
VP8DspInit();
return 1;
}
@ -129,11 +203,12 @@ static inline int hev_thresh_from_level(int level, int keyframe) {
}
static void DoFilter(const VP8Decoder* const dec, int mb_x, int mb_y) {
VP8MB* const mb = dec->mb_info_ + mb_x;
uint8_t* const y_dst = dec->cache_y_ + mb_x * 16;
const VP8ThreadContext* const ctx = &dec->thread_ctx_;
const int y_bps = dec->cache_y_stride_;
const int level = mb->f_level_;
const int ilevel = mb->f_ilevel_;
VP8FInfo* const f_info = ctx->f_info_ + mb_x;
uint8_t* const y_dst = dec->cache_y_ + ctx->id_ * 16 * y_bps + mb_x * 16;
const int level = f_info->f_level_;
const int ilevel = f_info->f_ilevel_;
const int limit = 2 * level + ilevel;
if (level == 0) {
return;
@ -142,26 +217,26 @@ static void DoFilter(const VP8Decoder* const dec, int mb_x, int mb_y) {
if (mb_x > 0) {
VP8SimpleHFilter16(y_dst, y_bps, limit + 4);
}
if (mb->f_inner_) {
if (f_info->f_inner_) {
VP8SimpleHFilter16i(y_dst, y_bps, limit);
}
if (mb_y > 0) {
VP8SimpleVFilter16(y_dst, y_bps, limit + 4);
}
if (mb->f_inner_) {
if (f_info->f_inner_) {
VP8SimpleVFilter16i(y_dst, y_bps, limit);
}
} else { // complex
uint8_t* const u_dst = dec->cache_u_ + mb_x * 8;
uint8_t* const v_dst = dec->cache_v_ + mb_x * 8;
const int uv_bps = dec->cache_uv_stride_;
uint8_t* const u_dst = dec->cache_u_ + ctx->id_ * 8 * uv_bps + mb_x * 8;
uint8_t* const v_dst = dec->cache_v_ + ctx->id_ * 8 * uv_bps + mb_x * 8;
const int hev_thresh =
hev_thresh_from_level(level, dec->frm_hdr_.key_frame_);
if (mb_x > 0) {
VP8HFilter16(y_dst, y_bps, limit + 4, ilevel, hev_thresh);
VP8HFilter8(u_dst, v_dst, uv_bps, limit + 4, ilevel, hev_thresh);
}
if (mb->f_inner_) {
if (f_info->f_inner_) {
VP8HFilter16i(y_dst, y_bps, limit, ilevel, hev_thresh);
VP8HFilter8i(u_dst, v_dst, uv_bps, limit, ilevel, hev_thresh);
}
@ -169,21 +244,20 @@ static void DoFilter(const VP8Decoder* const dec, int mb_x, int mb_y) {
VP8VFilter16(y_dst, y_bps, limit + 4, ilevel, hev_thresh);
VP8VFilter8(u_dst, v_dst, uv_bps, limit + 4, ilevel, hev_thresh);
}
if (mb->f_inner_) {
if (f_info->f_inner_) {
VP8VFilter16i(y_dst, y_bps, limit, ilevel, hev_thresh);
VP8VFilter8i(u_dst, v_dst, uv_bps, limit, ilevel, hev_thresh);
}
}
}
void VP8FilterRow(const VP8Decoder* const dec) {
// Filter the decoded macroblock row (if needed)
static void FilterRow(const VP8Decoder* const dec) {
int mb_x;
assert(dec->filter_type_ > 0);
if (dec->mb_y_ < dec->tl_mb_y_ || dec->mb_y_ > dec->br_mb_y_) {
return;
}
const int mb_y = dec->thread_ctx_.mb_y_;
assert(dec->thread_ctx_.filter_row_);
for (mb_x = dec->tl_mb_x_; mb_x < dec->br_mb_x_; ++mb_x) {
DoFilter(dec, mb_x, dec->mb_y_);
DoFilter(dec, mb_x, mb_y);
}
}
@ -191,7 +265,8 @@ void VP8FilterRow(const VP8Decoder* const dec) {
void VP8StoreBlock(VP8Decoder* const dec) {
if (dec->filter_type_ > 0) {
VP8MB* const info = dec->mb_info_ + dec->mb_x_;
VP8FInfo* const info = dec->f_info_ + dec->mb_x_;
const int skip = dec->mb_info_[dec->mb_x_].skip_;
int level = dec->filter_levels_[dec->segment_];
if (dec->filter_hdr_.use_lf_delta_) {
// TODO(skal): only CURRENT is handled for now.
@ -215,14 +290,16 @@ void VP8StoreBlock(VP8Decoder* const dec) {
}
info->f_ilevel_ = (level < 1) ? 1 : level;
info->f_inner_ = (!info->skip_ || dec->is_i4x4_);
info->f_inner_ = (!skip || dec->is_i4x4_);
}
{
// Transfer samples to row cache
int y;
uint8_t* const ydst = dec->cache_y_ + dec->mb_x_ * 16;
uint8_t* const udst = dec->cache_u_ + dec->mb_x_ * 8;
uint8_t* const vdst = dec->cache_v_ + dec->mb_x_ * 8;
const int y_offset = dec->cache_id_ * 16 * dec->cache_y_stride_;
const int uv_offset = dec->cache_id_ * 8 * dec->cache_uv_stride_;
uint8_t* const ydst = dec->cache_y_ + dec->mb_x_ * 16 + y_offset;
uint8_t* const udst = dec->cache_u_ + dec->mb_x_ * 8 + uv_offset;
uint8_t* const vdst = dec->cache_v_ + dec->mb_x_ * 8 + uv_offset;
for (y = 0; y < 16; ++y) {
memcpy(ydst + y * dec->cache_y_stride_,
dec->yuv_b_ + Y_OFF + y * BPS, 16);
@ -249,17 +326,27 @@ void VP8StoreBlock(VP8Decoder* const dec) {
#define MACROBLOCK_VPOS(mb_y) ((mb_y) * 16) // vertical position of a MB
// Finalize and transmit a complete row. Return false in case of user-abort.
int VP8FinishRow(VP8Decoder* const dec, VP8Io* io) {
int ok = 1;
const VP8ThreadContext* const ctx = &dec->thread_ctx_;
const int extra_y_rows = kFilterExtraRows[dec->filter_type_];
const int ysize = extra_y_rows * dec->cache_y_stride_;
const int uvsize = (extra_y_rows / 2) * dec->cache_uv_stride_;
uint8_t* const ydst = dec->cache_y_ - ysize;
uint8_t* const udst = dec->cache_u_ - uvsize;
uint8_t* const vdst = dec->cache_v_ - uvsize;
const int first_row = (dec->mb_y_ == 0);
const int last_row = (dec->mb_y_ >= dec->br_mb_y_ - 1);
int y_start = MACROBLOCK_VPOS(dec->mb_y_);
int y_end = MACROBLOCK_VPOS(dec->mb_y_ + 1);
const int y_offset = ctx->id_ * 16 * dec->cache_y_stride_;
const int uv_offset = ctx->id_ * 8 * dec->cache_uv_stride_;
uint8_t* const ydst = dec->cache_y_ - ysize + y_offset;
uint8_t* const udst = dec->cache_u_ - uvsize + uv_offset;
uint8_t* const vdst = dec->cache_v_ - uvsize + uv_offset;
const int first_row = (ctx->mb_y_ == 0);
const int last_row = (ctx->mb_y_ >= dec->br_mb_y_ - 1);
int y_start = MACROBLOCK_VPOS(ctx->mb_y_);
int y_end = MACROBLOCK_VPOS(ctx->mb_y_ + 1);
if (ctx->filter_row_) {
FilterRow(dec);
}
if (io->put) {
if (!first_row) {
y_start -= extra_y_rows;
@ -267,9 +354,9 @@ int VP8FinishRow(VP8Decoder* const dec, VP8Io* io) {
io->u = udst;
io->v = vdst;
} else {
io->y = dec->cache_y_;
io->u = dec->cache_u_;
io->v = dec->cache_v_;
io->y = dec->cache_y_ + y_offset;
io->u = dec->cache_u_ + uv_offset;
io->v = dec->cache_v_ + uv_offset;
}
if (!last_row) {
@ -309,27 +396,63 @@ int VP8FinishRow(VP8Decoder* const dec, VP8Io* io) {
io->mb_y = y_start - io->crop_top;
io->mb_w = io->crop_right - io->crop_left;
io->mb_h = y_end - y_start;
if (!io->put(io)) {
return 0;
}
ok = io->put(io);
}
}
// rotate top samples
if (!last_row) {
memcpy(ydst, ydst + 16 * dec->cache_y_stride_, ysize);
memcpy(udst, udst + 8 * dec->cache_uv_stride_, uvsize);
memcpy(vdst, vdst + 8 * dec->cache_uv_stride_, uvsize);
// rotate top samples if needed
if (ctx->id_ + 1 == dec->num_caches_) {
if (!last_row) {
memcpy(dec->cache_y_ - ysize, ydst + 16 * dec->cache_y_stride_, ysize);
memcpy(dec->cache_u_ - uvsize, udst + 8 * dec->cache_uv_stride_, uvsize);
memcpy(dec->cache_v_ - uvsize, vdst + 8 * dec->cache_uv_stride_, uvsize);
}
}
return 1;
return ok;
}
#undef MACROBLOCK_VPOS
//------------------------------------------------------------------------------
int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io) {
int ok = 1;
VP8ThreadContext* const ctx = &dec->thread_ctx_;
if (!dec->use_threads_) {
// ctx->id_ and ctx->f_info_ are already set
ctx->mb_y_ = dec->mb_y_;
ctx->filter_row_ = dec->filter_row_;
ok = VP8FinishRow(dec, io);
} else {
WebPWorker* const worker = &dec->worker_;
// Finish previous job *before* updating context
ok &= WebPWorkerSync(worker);
assert(worker->status_ == OK);
if (ok) { // spawn a new deblocking/output job
ctx->io_ = *io;
ctx->id_ = dec->cache_id_;
ctx->mb_y_ = dec->mb_y_;
ctx->filter_row_ = dec->filter_row_;
if (ctx->filter_row_) { // just swap filter info
VP8FInfo* const tmp = ctx->f_info_;
ctx->f_info_ = dec->f_info_;
dec->f_info_ = tmp;
}
WebPWorkerLaunch(worker);
if (++dec->cache_id_ == dec->num_caches_) {
dec->cache_id_ = 0;
}
}
}
return ok;
}
//------------------------------------------------------------------------------
// Finish setting up the decoding parameter once user's setup() is called.
VP8StatusCode VP8FinishFrameSetup(VP8Decoder* const dec, VP8Io* const io) {
VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
// Call setup() first. This may trigger additional decoding features on 'io'.
// Note: Afterward, we must call teardown() not matter what.
if (io->setup && !io->setup(io)) {
VP8SetError(dec, VP8_STATUS_USER_ABORT, "Frame setup failed");
return dec->status_;
@ -376,6 +499,18 @@ VP8StatusCode VP8FinishFrameSetup(VP8Decoder* const dec, VP8Io* const io) {
return VP8_STATUS_OK;
}
int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io) {
int ok = 1;
if (dec->use_threads_) {
ok = WebPWorkerSync(&dec->worker_);
}
if (io->teardown) {
io->teardown(io);
}
return ok;
}
//------------------------------------------------------------------------------
// Main reconstruction function.

40
src/dec/idec.c
View File

@ -324,28 +324,28 @@ static VP8StatusCode DecodePartition0(WebPIDecoder* const idec) {
return IDecError(idec, dec->status_);
}
// Allocate memory and prepare everything.
if (!VP8InitFrame(dec, io)) {
return IDecError(idec, dec->status_);
}
if (!CopyParts0Data(idec)) {
return IDecError(idec, VP8_STATUS_OUT_OF_MEMORY);
}
// Finish setting up the decoding parameters.
if (VP8FinishFrameSetup(dec, io) != VP8_STATUS_OK) {
// Finish setting up the decoding parameters. Will call io->setup().
if (VP8EnterCritical(dec, io) != VP8_STATUS_OK) {
return IDecError(idec, dec->status_);
}
// Note: past this point, teardown() must always be called
// in case of error.
idec->state_ = STATE_DATA;
// Allocate memory and prepare everything.
if (!VP8InitFrame(dec, io)) {
return IDecError(idec, dec->status_);
}
return VP8_STATUS_OK;
}
// Remaining partitions
static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
VP8BitReader* br;
VP8BitReader* br;
VP8Decoder* const dec = idec->dec_;
VP8Io* const io = &idec->io_;
@ -355,12 +355,8 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
for (; dec->mb_y_ < dec->mb_h_; ++dec->mb_y_) {
VP8BitReader* token_br = &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
if (dec->mb_x_ == 0) {
VP8MB* const left = dec->mb_info_ - 1;
left->nz_ = 0;
left->dc_nz_ = 0;
memset(dec->intra_l_, B_DC_PRED, sizeof(dec->intra_l_));
VP8InitScanline(dec);
}
for (; dec->mb_x_ < dec->mb_w_; dec->mb_x_++) {
MBContext context;
SaveContext(dec, token_br, &context);
@ -383,17 +379,14 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
assert(idec->mem_.start_ <= idec->mem_.end_);
}
}
if (dec->filter_type_ > 0) {
VP8FilterRow(dec);
}
if (!VP8FinishRow(dec, io)) {
if (!VP8ProcessRow(dec, io)) {
return IDecError(idec, VP8_STATUS_USER_ABORT);
}
dec->mb_x_ = 0;
}
if (io->teardown) {
io->teardown(io);
// Synchronize the thread and check for errors.
if (!VP8ExitCritical(dec, io)) {
return IDecError(idec, VP8_STATUS_USER_ABORT);
}
dec->ready_ = 0;
idec->state_ = STATE_DONE;
@ -443,6 +436,13 @@ WebPIDecoder* WebPINewDecoder(WebPDecBuffer* const output_buffer) {
idec->params_.output = output_buffer ? output_buffer : &idec->output_;
WebPInitCustomIo(&idec->params_, &idec->io_); // Plug the I/O functions.
#ifdef WEBP_USE_THREAD
idec->dec_->use_threads_ = idec->params_.options &&
(idec->params_.options->use_threads > 0);
#else
idec->dec_->use_threads_ = 0;
#endif
return idec;
}

128
src/dec/thread.c Normal file
View File

@ -0,0 +1,128 @@
// Copyright 2011 Google Inc.
//
// This code is licensed under the same terms as WebM:
// Software License Agreement: http://www.webmproject.org/license/software/
// Additional IP Rights Grant: http://www.webmproject.org/license/additional/
// -----------------------------------------------------------------------------
//
// Multi-threaded worker
//
// Author: skal@google.com (Pascal Massimino)
#include "./vp8i.h"
#include "./thread.h"
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif
#ifdef WEBP_USE_THREAD
static void *WebPWorkerThreadLoop(void *ptr) { // thread loop
WebPWorker* const worker = (WebPWorker*)ptr;
int done = 0;
while (!done) {
pthread_mutex_lock(&worker->mutex_);
while (worker->status_ == OK) { // wait in idling mode
pthread_cond_wait(&worker->condition_, &worker->mutex_);
}
if (worker->status_ == WORK) {
if (worker->hook) {
worker->had_error |= !worker->hook(worker->data1, worker->data2);
}
worker->status_ = OK;
} else if (worker->status_ == NOT_OK) { // finish the worker
done = 1;
}
// signal to the main thread that we're done (for Sync())
pthread_cond_signal(&worker->condition_);
pthread_mutex_unlock(&worker->mutex_);
}
return NULL; // Thread is finished
}
// main thread state control
static void WebPWorkerChangeState(WebPWorker* const worker,
WebPWorkerStatus new_status) {
pthread_mutex_lock(&worker->mutex_);
assert(worker->status_ >= OK);
// wait for the worker to finish
while (worker->status_ != OK) {
pthread_cond_wait(&worker->condition_, &worker->mutex_);
}
// assign new status and release the working thread if needed
if (new_status != OK) {
worker->status_ = new_status;
pthread_cond_signal(&worker->condition_);
}
pthread_mutex_unlock(&worker->mutex_);
}
#endif
//-----------------------------------------------------------------------------
void WebPWorkerInit(WebPWorker* const worker) {
memset(worker, 0, sizeof(*worker));
worker->status_ = NOT_OK;
}
int WebPWorkerSync(WebPWorker* const worker) {
#ifdef WEBP_USE_THREAD
WebPWorkerChangeState(worker, OK);
#endif
assert(worker->status_ == OK);
return !worker->had_error;
}
int WebPWorkerReset(WebPWorker* const worker) {
int ok = 1;
worker->had_error = 0;
if (worker->status_ < OK) {
#ifdef WEBP_USE_THREAD
if (pthread_mutex_init(&worker->mutex_, 0) ||
pthread_cond_init(&worker->condition_, 0)) {
return 0;
}
pthread_mutex_lock(&worker->mutex_);
ok = !pthread_create(&worker->thread_, 0, WebPWorkerThreadLoop, worker);
if (ok) worker->status_ = OK;
pthread_mutex_unlock(&worker->mutex_);
#else
worker->status_ = OK;
#endif
} else if (worker->status_ > OK) {
ok = WebPWorkerSync(worker);
}
assert(!ok || (worker->status_ == OK));
return ok;
}
void WebPWorkerLaunch(WebPWorker* const worker) {
#ifdef WEBP_USE_THREAD
WebPWorkerChangeState(worker, WORK);
#else
if (worker->hook)
worker->had_error |= !worker->hook(worker->data1, worker->data2);
#endif
}
void WebPWorkerEnd(WebPWorker* const worker) {
if (worker->status_ >= OK) {
#ifdef WEBP_USE_THREAD
WebPWorkerChangeState(worker, NOT_OK);
pthread_join(worker->thread_, NULL);
pthread_mutex_destroy(&worker->mutex_);
pthread_cond_destroy(&worker->condition_);
#else
worker->status_ = NOT_OK;
#endif
}
assert(worker->status_ == NOT_OK);
}
//------------------------------------------------------------------------------
#if defined(__cplusplus) || defined(c_plusplus)
} // extern "C"
#endif

79
src/dec/thread.h Normal file
View File

@ -0,0 +1,79 @@
// Copyright 2011 Google Inc.
//
// This code is licensed under the same terms as WebM:
// Software License Agreement: http://www.webmproject.org/license/software/
// Additional IP Rights Grant: http://www.webmproject.org/license/additional/
// -----------------------------------------------------------------------------
//
// Multi-threaded worker
//
// Author: skal@google.com (Pascal Massimino)
#ifndef WEBP_DEC_THREAD_H
#define WEBP_DEC_THREAD_H
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif
#if WEBP_USE_THREAD
#if defined(_WIN32)
#undef WEBP_USE_THREAD
#else
#include <pthread.h>
#endif // _WIN32
#endif // WEBP_USE_THREAD
// State of the worker thread object
typedef enum {
NOT_OK = 0, // object is unusable
OK, // ready to work
WORK // busy finishing the current task
} WebPWorkerStatus;
// Function to be called by the worker thread. Takes two opaque pointers as
// arguments (data1 and data2), and should return false in case of error.
typedef int (*WebPWorkerHook)(void*, void*);
// Synchronize object used to launch job in the worker thread
typedef struct {
#if WEBP_USE_THREAD
pthread_mutex_t mutex_;
pthread_cond_t condition_;
pthread_t thread_;
#endif
WebPWorkerStatus status_;
WebPWorkerHook hook; // hook to call
void* data1; // first argument passed to 'hook'
void* data2; // second argument passed to 'hook'
int had_error; // return value of the last call to 'hook'
} WebPWorker;
// Must be called first, before any other method.
void WebPWorkerInit(WebPWorker* const worker);
// Must be called initialize the object and spawn the thread. Re-entrant.
// Will potentially launch the thread. Returns false in case of error.
int WebPWorkerReset(WebPWorker* const worker);
// Make sure the previous work is finished. Returns true if worker->had_error
// was not set and not error condition was triggered by the working thread.
int WebPWorkerSync(WebPWorker* const worker);
// Trigger the thread to call hook() with data1 and data2 argument. These
// hook/data1/data2 can be changed at any time before calling this function,
// but not be changed afterward until the next call to WebPWorkerSync().
void WebPWorkerLaunch(WebPWorker* const worker);
// Kill the thread and terminate the object. To use the object again, one
// must call WebPWorkerReset() again.
void WebPWorkerEnd(WebPWorker* const worker);
//------------------------------------------------------------------------------
#if defined(__cplusplus) || defined(c_plusplus)
} // extern "C"
#endif
#endif // WEBP_DEC_THREAD_H

59
src/dec/vp8.c
View File

@ -43,6 +43,7 @@ VP8Decoder* VP8New(void) {
VP8Decoder* dec = (VP8Decoder*)calloc(1, sizeof(VP8Decoder));
if (dec) {
SetOk(dec);
WebPWorkerInit(&dec->worker_);
dec->ready_ = 0;
}
return dec;
@ -682,15 +683,21 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
return (!token_br->eof_);
}
void VP8InitScanline(VP8Decoder* const dec) {
VP8MB* const left = dec->mb_info_ - 1;
left->nz_ = 0;
left->dc_nz_ = 0;
memset(dec->intra_l_, B_DC_PRED, sizeof(dec->intra_l_));
dec->filter_row_ =
(dec->filter_type_ > 0) &&
(dec->mb_y_ >= dec->tl_mb_y_) && (dec->mb_y_ <= dec->br_mb_y_);
}
static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
for (dec->mb_y_ = 0; dec->mb_y_ < dec->br_mb_y_; ++dec->mb_y_) {
VP8MB* const left = dec->mb_info_ - 1;
VP8BitReader* const token_br =
&dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
left->nz_ = 0;
left->dc_nz_ = 0;
memset(dec->intra_l_, B_DC_PRED, sizeof(dec->intra_l_));
VP8InitScanline(dec);
for (dec->mb_x_ = 0; dec->mb_x_ < dec->mb_w_; dec->mb_x_++) {
if (!VP8DecodeMB(dec, token_br)) {
return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
@ -701,13 +708,13 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
// Store data and save block's filtering params
VP8StoreBlock(dec);
}
if (dec->filter_type_ > 0) {
VP8FilterRow(dec);
}
if (!VP8FinishRow(dec, io)) {
if (!VP8ProcessRow(dec, io)) {
return VP8SetError(dec, VP8_STATUS_USER_ABORT, "Output aborted.");
}
}
if (dec->use_threads_ && !WebPWorkerSync(&dec->worker_)) {
return 0;
}
// Finish
#ifndef ONLY_KEYFRAME_CODE
@ -729,6 +736,7 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
// Main entry point
int VP8Decode(VP8Decoder* const dec, VP8Io* const io) {
int ok = 0;
if (dec == NULL) {
return 0;
}
@ -744,30 +752,24 @@ int VP8Decode(VP8Decoder* const dec, VP8Io* const io) {
}
assert(dec->ready_);
// Will allocate memory and prepare everything.
if (!VP8InitFrame(dec, io)) {
VP8Clear(dec);
return 0;
// Finish setting up the decoding parameter. Will call io->setup().
ok = (VP8EnterCritical(dec, io) == VP8_STATUS_OK);
if (ok) { // good to go.
// Will allocate memory and prepare everything.
if (ok) ok = VP8InitFrame(dec, io);
// Main decoding loop
if (ok) ok = ParseFrame(dec, io);
// Exit.
ok &= VP8ExitCritical(dec, io);
}
// Finish setting up the decoding parameter
if (VP8FinishFrameSetup(dec, io) != VP8_STATUS_OK) {
if (!ok) {
VP8Clear(dec);
return 0;
}
// Main decoding loop
{
const int ret = ParseFrame(dec, io);
if (io->teardown) {
io->teardown(io);
}
if (!ret) {
VP8Clear(dec);
return 0;
}
}
dec->ready_ = 0;
return 1;
}
@ -776,6 +778,9 @@ void VP8Clear(VP8Decoder* const dec) {
if (dec == NULL) {
return;
}
if (dec->use_threads_) {
WebPWorkerEnd(&dec->worker_);
}
if (dec->mem_) {
free(dec->mem_);
}

57
src/dec/vp8i.h
View File

@ -13,7 +13,8 @@
#define WEBP_DEC_VP8I_H_
#include <string.h> // for memcpy()
#include "bits.h"
#include "./bits.h"
#include "./thread.h"
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
@ -147,16 +148,16 @@ typedef struct {
//-----------------------------------------------------------------------------
// Informations about the macroblocks.
typedef struct {
// block type
uint8_t skip_:1;
// filter specs
uint8_t f_level_:6; // filter strength: 0..63
uint8_t f_ilevel_:6; // inner limit: 1..63
uint8_t f_inner_:1; // do inner filtering?
// cbp
uint8_t nz_; // non-zero AC/DC coeffs
uint8_t dc_nz_; // non-zero DC coeffs
typedef struct { // filter specs
unsigned int f_level_:6; // filter strength: 0..63
unsigned int f_ilevel_:6; // inner limit: 1..63
unsigned int f_inner_:1; // do inner filtering?
} VP8FInfo;
typedef struct { // used for syntax-parsing
unsigned int nz_; // non-zero AC/DC coeffs
unsigned int dc_nz_:1; // non-zero DC coeffs
unsigned int skip_:1; // block type
} VP8MB;
// Dequantization matrices
@ -164,6 +165,15 @@ typedef struct {
uint16_t y1_mat_[2], y2_mat_[2], uv_mat_[2]; // [DC / AC]
} VP8QuantMatrix;
// Persistent information needed by the parallel processing
typedef struct {
int id_; // cache row to process (in [0..2])
int mb_y_; // macroblock position of the row
int filter_row_; // true if row-filtering is needed
VP8FInfo* f_info_; // filter strengths
VP8Io io_; // copy of the VP8Io to pass to put()
} VP8ThreadContext;
//-----------------------------------------------------------------------------
// VP8Decoder: the main opaque structure handed over to user
@ -181,6 +191,13 @@ struct VP8Decoder {
VP8FilterHeader filter_hdr_;
VP8SegmentHeader segment_hdr_;
// Worker
WebPWorker worker_;
int use_threads_; // use multi-thread
int cache_id_; // current cache row
int num_caches_; // number of cached rows of 16 pixels (1, 2 or 3)
VP8ThreadContext thread_ctx_; // Thread context
// dimension, in macroblock units.
int mb_w_, mb_h_;
@ -219,7 +236,8 @@ struct VP8Decoder {
uint8_t* y_t_; // top luma samples: 16 * mb_w_
uint8_t* u_t_, *v_t_; // top u/v samples: 8 * mb_w_ each
VP8MB* mb_info_; // contextual macroblock infos (mb_w_ + 1)
VP8MB* mb_info_; // contextual macroblock info (mb_w_ + 1)
VP8FInfo* f_info_; // filter strength info
uint8_t* yuv_b_; // main block for Y/U/V (size = YUV_SIZE)
int16_t* coeffs_; // 384 coeffs = (16+8+8) * 4*4
@ -249,6 +267,7 @@ struct VP8Decoder {
// Filtering side-info
int filter_type_; // 0=off, 1=simple, 2=complex
int filter_row_; // per-row flag
uint8_t filter_levels_[NUM_MB_SEGMENTS]; // precalculated per-segment
// extensions
@ -288,13 +307,23 @@ int VP8InitFrame(VP8Decoder* const dec, VP8Io* io);
// Predict a block and add residual
void VP8ReconstructBlock(VP8Decoder* const dec);
// Call io->setup() and finish setting up scan parameters.
VP8StatusCode VP8FinishFrameSetup(VP8Decoder* const dec, VP8Io* const io);
// After this call returns, one must always call VP8ExitCritical() with the
// same parameters. Both functions should be used in pair. Returns VP8_STATUS_OK
// if ok, otherwise sets and returns the error status on *dec.
VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io);
// Must always be called in pair with VP8EnterCritical().
// Returns false in case of error.
int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io);
// Filter the decoded macroblock row (if needed)
void VP8FilterRow(const VP8Decoder* const dec);
int VP8FinishRow(VP8Decoder* const dec, VP8Io* io); // multi threaded call
// Process the last decoded row (filtering + output)
int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io);
// Store a block, along with filtering params
void VP8StoreBlock(VP8Decoder* const dec);
// Finalize and transmit a complete row. Return false in case of user-abort.
int VP8FinishRow(VP8Decoder* const dec, VP8Io* const io);
// To be called at the start of a new scanline, to initialize predictors.
void VP8InitScanline(VP8Decoder* const dec);
// Decode one macroblock. Returns false if there is not enough data.
int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br);

6
src/dec/webp.c
View File

@ -92,6 +92,12 @@ static VP8StatusCode DecodeInto(const uint8_t* data, uint32_t data_size,
io.data_size = data_size;
WebPInitCustomIo(params, &io); // Plug the I/O functions.
#ifdef WEBP_USE_THREAD
dec->use_threads_ = params->options && (params->options->use_threads > 0);
#else
dec->use_threads_ = 0;
#endif
// Decode bitstream header, update io->width/io->height.
if (!VP8GetHeaders(dec, &io)) {
status = VP8_STATUS_BITSTREAM_ERROR;

1
src/webp/decode.h
View File

@ -351,6 +351,7 @@ typedef struct {
int scaled_width, scaled_height; // final resolution
int force_rotation; // forced rotation (to be applied _last_)
int no_enhancement; // if true, discard enhancement layer
int use_threads; // if true, use multi-threaded decoding
} WebPDecoderOptions;
// Main object storing the configuration for advanced decoding.

7
src/webp/decode_vp8.h
View File

@ -67,10 +67,13 @@ struct VP8Io {
VP8IoPutHook put;
// called just before starting to decode the blocks.
// Should returns 0 in case of error.
// Must return false in case of setup error, true otherwise. If false is
// returned, teardown() will NOT be called. But if the setup succeeded
// and true is returned, then teardown() will always be called afterward.
VP8IoSetupHook setup;
// called just after block decoding is finished (or when an error occurred).
// Called just after block decoding is finished (or when an error occurred
// during put()). Is NOT called if setup() failed.
VP8IoTeardownHook teardown;
// this is a recommendation for the user-side yuv->rgb converter. This flag