dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-18 23:09:52 +02:00
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
Download Patch File Download Diff File Expand all files Collapse all files

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.