generic-library/webp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Further reduce memory to decode lossy+alpha images

Browse Source 
Earlier such images were using roughly 9 * width * height bytes for
decoding. Now, they take 6 * width * height memory.

Change-Id: Ie4a681ca5074d96d64f30b2597fafdca648dd8f7
(cherry picked from commit 64c844863a)
This commit is contained in:
Urvang Joshi 2013-05-13 16:24:49 -07:00 committed by James Zern
parent 07db70d20f
commit cc128e0bfc
4 changed files with 253 additions and 166 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

322
src/dec/vp8l.c
View File

@ -625,10 +625,24 @@ static void ApplyInverseTransforms(VP8LDecoder* const dec, int num_rows,
}
}
// Special method for paletted alpha data.
static void ApplyInverseTransformsAlpha(VP8LDecoder* const dec, int num_rows,
const uint8_t* const rows) {
const int start_row = dec->last_row_;
const int end_row = start_row + num_rows;
const uint8_t* rows_in = rows;
uint8_t* rows_out = (uint8_t*)dec->io_->opaque + dec->io_->width * start_row;
VP8LTransform* const transform = &dec->transforms_[0];
assert(dec->next_transform_ == 1);
assert(transform->type_ == COLOR_INDEXING_TRANSFORM);
VP8LColorIndexInverseTransformAlpha(transform, start_row, end_row, rows_in,
rows_out);
}
// Processes (transforms, scales & color-converts) the rows decoded after the
// last call.
static void ProcessRows(VP8LDecoder* const dec, int row) {
const uint32_t* const rows = dec->argb_ + dec->width_ * dec->last_row_;
const uint32_t* const rows = dec->pixels_ + dec->width_ * dec->last_row_;
const int num_rows = row - dec->last_row_;
if (num_rows <= 0) return; // Nothing to be done.
@ -667,121 +681,135 @@ static void ProcessRows(VP8LDecoder* const dec, int row) {
assert(dec->last_row_ <= dec->height_);
}
static int DecodeImageData(VP8LDecoder* const dec,
uint32_t* const data, int width, int height,
ProcessRowsFunc process_func) {
int ok = 1;
int col = 0, row = 0;
VP8LBitReader* const br = &dec->br_;
VP8LMetadata* const hdr = &dec->hdr_;
HTreeGroup* htree_group = hdr->htree_groups_;
uint32_t* src = data;
uint32_t* last_cached = data;
uint32_t* const src_end = data + width * height;
const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES;
const int color_cache_limit = len_code_limit + hdr->color_cache_size_;
VP8LColorCache* const color_cache =
(hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL;
const int mask = hdr->huffman_mask_;
assert(htree_group != NULL);
while (!br->eos_ && src < src_end) {
int code;
// Only update when changing tile. Note we could use the following test:
// if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed
// but that's actually slower and requires storing the previous col/row
if ((col & mask) == 0) {
htree_group = GetHtreeGroupForPos(hdr, col, row);
}
VP8LFillBitWindow(br);
code = ReadSymbol(&htree_group->htrees_[GREEN], br);
if (code < NUM_LITERAL_CODES) { // Literal.
int red, green, blue, alpha;
red = ReadSymbol(&htree_group->htrees_[RED], br);
green = code;
VP8LFillBitWindow(br);
blue = ReadSymbol(&htree_group->htrees_[BLUE], br);
alpha = ReadSymbol(&htree_group->htrees_[ALPHA], br);
*src = (alpha << 24) + (red << 16) + (green << 8) + blue;
AdvanceByOne:
++src;
++col;
if (col >= width) {
col = 0;
++row;
if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) {
process_func(dec, row);
}
if (color_cache != NULL) {
while (last_cached < src) {
VP8LColorCacheInsert(color_cache, *last_cached++);
}
}
}
} else if (code < len_code_limit) { // Backward reference
int dist_code, dist;
const int length_sym = code - NUM_LITERAL_CODES;
const int length = GetCopyLength(length_sym, br);
const int dist_symbol = ReadSymbol(&htree_group->htrees_[DIST], br);
VP8LFillBitWindow(br);
dist_code = GetCopyDistance(dist_symbol, br);
dist = PlaneCodeToDistance(width, dist_code);
if (src - data < dist || src_end - src < length) {
ok = 0;
goto End;
}
{
int i;
for (i = 0; i < length; ++i) src[i] = src[i - dist];
src += length;
}
col += length;
while (col >= width) {
col -= width;
++row;
if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) {
process_func(dec, row);
}
}
if (src < src_end) {
htree_group = GetHtreeGroupForPos(hdr, col, row);
if (color_cache != NULL) {
while (last_cached < src) {
VP8LColorCacheInsert(color_cache, *last_cached++);
}
}
}
} else if (code < color_cache_limit) { // Color cache.
const int key = code - len_code_limit;
assert(color_cache != NULL);
while (last_cached < src) {
VP8LColorCacheInsert(color_cache, *last_cached++);
}
*src = VP8LColorCacheLookup(color_cache, key);
goto AdvanceByOne;
} else { // Not reached.
ok = 0;
goto End;
}
ok = !br->error_;
if (!ok) goto End;
}
// Process the remaining rows corresponding to last row-block.
if (process_func != NULL) process_func(dec, row);
End:
if (br->error_ || !ok || (br->eos_ && src < src_end)) {
ok = 0;
dec->status_ = (!br->eos_) ?
VP8_STATUS_BITSTREAM_ERROR : VP8_STATUS_SUSPENDED;
} else if (src == src_end) {
dec->state_ = READ_DATA;
}
return ok;
#define DECODE_DATA_FUNC(FUNC_NAME, TYPE, STORE_PIXEL) \
static int FUNC_NAME(VP8LDecoder* const dec, TYPE* const data, int width, \
int height, ProcessRowsFunc process_func) { \
int ok = 1; \
int col = 0, row = 0; \
VP8LBitReader* const br = &dec->br_; \
VP8LMetadata* const hdr = &dec->hdr_; \
HTreeGroup* htree_group = hdr->htree_groups_; \
TYPE* src = data; \
TYPE* last_cached = data; \
TYPE* const src_end = data + width * height; \
const int len_code_limit = NUM_LITERAL_CODES + NUM_LENGTH_CODES; \
const int color_cache_limit = len_code_limit + hdr->color_cache_size_; \
VP8LColorCache* const color_cache = \
(hdr->color_cache_size_ > 0) ? &hdr->color_cache_ : NULL; \
const int mask = hdr->huffman_mask_; \
assert(htree_group != NULL); \
while (!br->eos_ && src < src_end) { \
int code; \
/* Only update when changing tile. Note we could use this test: */ \
/* if "((((prev_col ^ col) | prev_row ^ row)) > mask)" -> tile changed */ \
/* but that's actually slower and needs storing the previous col/row. */ \
if ((col & mask) == 0) { \
htree_group = GetHtreeGroupForPos(hdr, col, row); \
} \
VP8LFillBitWindow(br); \
code = ReadSymbol(&htree_group->htrees_[GREEN], br); \
if (code < NUM_LITERAL_CODES) { /* Literal*/ \
int red, green, blue, alpha; \
red = ReadSymbol(&htree_group->htrees_[RED], br); \
green = code; \
VP8LFillBitWindow(br); \
blue = ReadSymbol(&htree_group->htrees_[BLUE], br); \
alpha = ReadSymbol(&htree_group->htrees_[ALPHA], br); \
*src = STORE_PIXEL(alpha, red, green, blue); \
AdvanceByOne: \
++src; \
++col; \
if (col >= width) { \
col = 0; \
++row; \
if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) { \
process_func(dec, row); \
} \
if (color_cache != NULL) { \
while (last_cached < src) { \
VP8LColorCacheInsert(color_cache, *last_cached++); \
} \
} \
} \
} else if (code < len_code_limit) { /* Backward reference */ \
int dist_code, dist; \
const int length_sym = code - NUM_LITERAL_CODES; \
const int length = GetCopyLength(length_sym, br); \
const int dist_symbol = ReadSymbol(&htree_group->htrees_[DIST], br); \
VP8LFillBitWindow(br); \
dist_code = GetCopyDistance(dist_symbol, br); \
dist = PlaneCodeToDistance(width, dist_code); \
if (src - data < dist || src_end - src < length) { \
ok = 0; \
goto End; \
} \
{ \
int i; \
for (i = 0; i < length; ++i) src[i] = src[i - dist]; \
src += length; \
} \
col += length; \
while (col >= width) { \
col -= width; \
++row; \
if ((process_func != NULL) && (row % NUM_ARGB_CACHE_ROWS == 0)) { \
process_func(dec, row); \
} \
} \
if (src < src_end) { \
htree_group = GetHtreeGroupForPos(hdr, col, row); \
if (color_cache != NULL) { \
while (last_cached < src) { \
VP8LColorCacheInsert(color_cache, *last_cached++); \
} \
} \
} \
} else if (code < color_cache_limit) { /* Color cache */ \
const int key = code - len_code_limit; \
assert(color_cache != NULL); \
while (last_cached < src) { \
VP8LColorCacheInsert(color_cache, *last_cached++); \
} \
*src = VP8LColorCacheLookup(color_cache, key); \
goto AdvanceByOne; \
} else { /* Not reached */ \
ok = 0; \
goto End; \
} \
ok = !br->error_; \
if (!ok) goto End; \
} \
/* Process the remaining rows corresponding to last row-block. */ \
if (process_func != NULL) process_func(dec, row); \
End: \
if (br->error_ || !ok || (br->eos_ && src < src_end)) { \
ok = 0; \
dec->status_ = \
(!br->eos_) ? VP8_STATUS_BITSTREAM_ERROR : VP8_STATUS_SUSPENDED; \
} else if (src == src_end) { \
dec->state_ = READ_DATA; \
} \
return ok; \
}
static WEBP_INLINE uint32_t GetARGBPixel(int alpha, int red, int green,
int blue) {
return (alpha << 24) | (red << 16) | (green << 8) | blue;
}
static WEBP_INLINE uint8_t GetAlphaPixel(int alpha, int red, int green,
int blue) {
(void)alpha;
(void)red;
(void)blue;
return green; // Alpha value is stored in green channel.
}
DECODE_DATA_FUNC(DecodeImageData, uint32_t, GetARGBPixel)
DECODE_DATA_FUNC(DecodeAlphaData, uint8_t, GetAlphaPixel)
#undef DECODE_DATA_FUNC
// -----------------------------------------------------------------------------
// VP8LTransform
@ -903,8 +931,8 @@ void VP8LClear(VP8LDecoder* const dec) {
if (dec == NULL) return;
ClearMetadata(&dec->hdr_);
free(dec->argb_);
dec->argb_ = NULL;
free(dec->pixels_);
dec->pixels_ = NULL;
for (i = 0; i < dec->next_transform_; ++i) {
ClearTransform(&dec->transforms_[i]);
}
@ -1028,35 +1056,38 @@ static int DecodeImageStream(int xsize, int ysize,
}
//------------------------------------------------------------------------------
// Allocate dec->argb_ and dec->argb_cache_ using dec->width_ and dec->height_
static int AllocateARGBBuffers(VP8LDecoder* const dec, int final_width) {
// Allocate internal buffers dec->pixels_ and dec->argb_cache_.
static int AllocateInternalBuffers(VP8LDecoder* const dec, int final_width,
size_t bytes_per_pixel) {
const int argb_cache_needed = (bytes_per_pixel == sizeof(uint32_t));
const uint64_t num_pixels = (uint64_t)dec->width_ * dec->height_;
// Scratch buffer corresponding to top-prediction row for transforming the
// first row in the row-blocks.
const uint64_t cache_top_pixels = final_width;
// Scratch buffer for temporary BGRA storage.
const uint64_t cache_pixels = (uint64_t)final_width * NUM_ARGB_CACHE_ROWS;
// first row in the row-blocks. Not needed for paletted alpha.
const uint64_t cache_top_pixels = argb_cache_needed ? final_width : 0ULL;
// Scratch buffer for temporary BGRA storage. Not needed for paletted alpha.
const uint64_t cache_pixels =
argb_cache_needed ? (uint64_t)final_width * NUM_ARGB_CACHE_ROWS : 0ULL;
const uint64_t total_num_pixels =
num_pixels + cache_top_pixels + cache_pixels;
assert(dec->width_ <= final_width);
dec->argb_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(*dec->argb_));
if (dec->argb_ == NULL) {
dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, bytes_per_pixel);
if (dec->pixels_ == NULL) {
dec->argb_cache_ = NULL; // for sanity check
dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
return 0;
}
dec->argb_cache_ = dec->argb_ + num_pixels + cache_top_pixels;
dec->argb_cache_ =
argb_cache_needed ? dec->pixels_ + num_pixels + cache_top_pixels : NULL;
return 1;
}
//------------------------------------------------------------------------------
// Special row-processing that only stores the alpha data.
// Special row-processing that only stores the alpha data.
static void ExtractAlphaRows(VP8LDecoder* const dec, int row) {
const int num_rows = row - dec->last_row_;
const uint32_t* const in = dec->argb_ + dec->width_ * dec->last_row_;
const uint32_t* const in = dec->pixels_ + dec->width_ * dec->last_row_;
if (num_rows <= 0) return; // Nothing to be done.
ApplyInverseTransforms(dec, num_rows, in);
@ -1070,7 +1101,17 @@ static void ExtractAlphaRows(VP8LDecoder* const dec, int row) {
int i;
for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
}
dec->last_row_ = dec->last_out_row_ = row;
}
// Row-processing for the special case when alpha data contains only one
// transform: color indexing.
static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int row) {
const int num_rows = row - dec->last_row_;
const uint8_t* const in =
(uint8_t*)dec->pixels_ + dec->width_ * dec->last_row_;
if (num_rows <= 0) return; // Nothing to be done.
ApplyInverseTransformsAlpha(dec, num_rows, in);
dec->last_row_ = dec->last_out_row_ = row;
}
@ -1079,6 +1120,7 @@ int VP8LDecodeAlphaImageStream(int width, int height, const uint8_t* const data,
VP8Io io;
int ok = 0;
VP8LDecoder* const dec = VP8LNew();
size_t bytes_per_pixel = sizeof(uint32_t); // Default: BGRA mode.
if (dec == NULL) return 0;
dec->width_ = width;
@ -1097,13 +1139,24 @@ int VP8LDecodeAlphaImageStream(int width, int height, const uint8_t* const data,
dec->action_ = READ_HDR;
if (!DecodeImageStream(width, height, 1, dec, NULL)) goto Err;
// Allocate output (note that dec->width_ may have changed here).
if (!AllocateARGBBuffers(dec, width)) goto Err;
// Special case: if alpha data contains only the color indexing transform
// (a frequent case), we will use DecodeAlphaData() method that only needs
// allocation of 1 byte per pixel (alpha channel).
if (dec->next_transform_ == 1 &&
dec->transforms_[0].type_ == COLOR_INDEXING_TRANSFORM) {
bytes_per_pixel = sizeof(uint8_t);
}
// Allocate internal buffers (note that dec->width_ may have changed here).
if (!AllocateInternalBuffers(dec, width, bytes_per_pixel)) goto Err;
// Decode (with special row processing).
dec->action_ = READ_DATA;
ok = DecodeImageData(dec, dec->argb_, dec->width_, dec->height_,
ExtractAlphaRows);
ok = (bytes_per_pixel == sizeof(uint8_t)) ?
DecodeAlphaData(dec, (uint8_t*)dec->pixels_, dec->width_, dec->height_,
ExtractPalettedAlphaRows) :
DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
ExtractAlphaRows);
Err:
VP8LDelete(dec);
@ -1143,6 +1196,7 @@ int VP8LDecodeHeader(VP8LDecoder* const dec, VP8Io* const io) {
}
int VP8LDecodeImage(VP8LDecoder* const dec) {
const size_t bytes_per_pixel = sizeof(uint32_t);
VP8Io* io = NULL;
WebPDecParams* params = NULL;
@ -1162,13 +1216,13 @@ int VP8LDecodeImage(VP8LDecoder* const dec) {
goto Err;
}
if (!AllocateARGBBuffers(dec, io->width)) goto Err;
if (!AllocateInternalBuffers(dec, io->width, bytes_per_pixel)) goto Err;
if (io->use_scaling && !AllocateAndInitRescaler(dec, io)) goto Err;
// Decode.
dec->action_ = READ_DATA;
if (!DecodeImageData(dec, dec->argb_, dec->width_, dec->height_,
if (!DecodeImageData(dec, dec->pixels_, dec->width_, dec->height_,
ProcessRows)) {
goto Err;
}

3
src/dec/vp8li.h
View File

@ -63,7 +63,8 @@ typedef struct {
const WebPDecBuffer *output_; // shortcut to io->opaque->output
uint32_t *argb_; // Internal data: always in BGRA color mode.
uint32_t *pixels_; // Internal data: either uint8_t* for alpha
// or uint32_t* for BGRA.
uint32_t *argb_cache_; // Scratch buffer for temporary BGRA storage.
VP8LBitReader br_;

87
src/dsp/lossless.c
View File

@ -1093,39 +1093,64 @@ static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
}
// Separate out pixels packed together using pixel-bundling.
static void ColorIndexInverseTransform(
const VP8LTransform* const transform,
int y_start, int y_end, const uint32_t* src, uint32_t* dst) {
int y;
const int bits_per_pixel = 8 >> transform->bits_;
const int width = transform->xsize_;
const uint32_t* const color_map = transform->data_;
if (bits_per_pixel < 8) {
const int pixels_per_byte = 1 << transform->bits_;
const int count_mask = pixels_per_byte - 1;
const uint32_t bit_mask = (1 << bits_per_pixel) - 1;
for (y = y_start; y < y_end; ++y) {
uint32_t packed_pixels = 0;
int x;
for (x = 0; x < width; ++x) {
// We need to load fresh 'packed_pixels' once every 'pixels_per_byte'
// increments of x. Fortunately, pixels_per_byte is a power of 2, so
// can just use a mask for that, instead of decrementing a counter.
if ((x & count_mask) == 0) packed_pixels = ((*src++) >> 8) & 0xff;
*dst++ = color_map[packed_pixels & bit_mask];
packed_pixels >>= bits_per_pixel;
}
}
} else {
for (y = y_start; y < y_end; ++y) {
int x;
for (x = 0; x < width; ++x) {
*dst++ = color_map[((*src++) >> 8) & 0xff];
}
}
}
// We define two methods for ARGB data (uint32_t) and alpha-only data (uint8_t).
#define COLOR_INDEX_INVERSE(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE) \
void FUNC_NAME(const VP8LTransform* const transform, \
int y_start, int y_end, const TYPE* src, TYPE* dst) { \
int y; \
const int bits_per_pixel = 8 >> transform->bits_; \
const int width = transform->xsize_; \
const uint32_t* const color_map = transform->data_; \
if (bits_per_pixel < 8) { \
const int pixels_per_byte = 1 << transform->bits_; \
const int count_mask = pixels_per_byte - 1; \
const uint32_t bit_mask = (1 << bits_per_pixel) - 1; \
for (y = y_start; y < y_end; ++y) { \
uint32_t packed_pixels = 0; \
int x; \
for (x = 0; x < width; ++x) { \
/* We need to load fresh 'packed_pixels' once every */ \
/* 'pixels_per_byte' increments of x. Fortunately, pixels_per_byte */ \
/* is a power of 2, so can just use a mask for that, instead of */ \
/* decrementing a counter. */ \
if ((x & count_mask) == 0) packed_pixels = GET_INDEX(*src++); \
*dst++ = GET_VALUE(color_map[packed_pixels & bit_mask]); \
packed_pixels >>= bits_per_pixel; \
} \
} \
} else { \
for (y = y_start; y < y_end; ++y) { \
int x; \
for (x = 0; x < width; ++x) { \
*dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]); \
} \
} \
} \
}
static WEBP_INLINE uint32_t GetARGBIndex(uint32_t index) {
return (index >> 8) & 0xff;
}
static WEBP_INLINE uint8_t GetAlphaIndex(uint8_t index) {
return index;
}
static WEBP_INLINE uint32_t GetARGBValue(uint32_t val) {
return val;
}
static WEBP_INLINE uint8_t GetAlphaValue(uint32_t val) {
return (val >> 8) & 0xff;
}
static COLOR_INDEX_INVERSE(ColorIndexInverseTransform, uint32_t, GetARGBIndex,
GetARGBValue)
COLOR_INDEX_INVERSE(VP8LColorIndexInverseTransformAlpha, uint8_t, GetAlphaIndex,
GetAlphaValue)
#undef COLOR_INDEX_INVERSE
void VP8LInverseTransform(const VP8LTransform* const transform,
int row_start, int row_end,
const uint32_t* const in, uint32_t* const out) {

7
src/dsp/lossless.h
View File

@ -33,6 +33,13 @@ void VP8LInverseTransform(const struct VP8LTransform* const transform,
int row_start, int row_end,
const uint32_t* const in, uint32_t* const out);
// Similar to the static method ColorIndexInverseTransform() that is part of
// lossless.c, but used only for alpha decoding. It takes uint8_t (rather than
// uint32_t) arguments for 'src' and 'dst'.
void VP8LColorIndexInverseTransformAlpha(
const struct VP8LTransform* const transform, int y_start, int y_end,
const uint8_t* src, uint8_t* dst);
// Subtracts green from blue and red channels.
void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs);