update the Unfilter API in dsp to process one row independently

This will allow to work in-place on cropped area later.

Also sped up the inverse gradient filtering in SSE2 (~4%)

Change-Id: I463149eee95d36984328f163a1e17f8cabd87441
This commit is contained in:
Pascal Massimino
2016-04-14 19:12:09 +02:00
parent 602f344a36
commit 2102ccd091
8 changed files with 297 additions and 308 deletions

View File

@ -111,17 +111,29 @@ static int ALPHDecode(VP8Decoder* const dec, int row, int num_rows) {
ALPHDecoder* const alph_dec = dec->alph_dec_;
const int width = alph_dec->width_;
const int height = alph_dec->io_.crop_bottom;
uint8_t* const output = dec->alpha_plane_;
if (alph_dec->method_ == ALPHA_NO_COMPRESSION) {
const size_t offset = row * width;
const size_t num_pixels = num_rows * width;
assert(dec->alpha_data_size_ >= ALPHA_HEADER_LEN + offset + num_pixels);
memcpy(dec->alpha_plane_ + offset,
dec->alpha_data_ + ALPHA_HEADER_LEN + offset, num_pixels);
if (WebPUnfilters[alph_dec->filter_] != NULL) {
WebPUnfilters[alph_dec->filter_](width, height, width,
row, num_rows, output);
int y;
const uint8_t* prev_line = dec->alpha_prev_line_;
const uint8_t* deltas = dec->alpha_data_ + ALPHA_HEADER_LEN + row * width;
uint8_t* dst = dec->alpha_plane_ + row * width;
assert(deltas <= &dec->alpha_data_[dec->alpha_data_size_]);
if (alph_dec->filter_ != WEBP_FILTER_NONE) {
assert(WebPUnfilters[alph_dec->filter_] != NULL);
for (y = 0; y < num_rows; ++y) {
WebPUnfilters[alph_dec->filter_](prev_line, deltas, dst, width);
prev_line = dst;
dst += width;
deltas += width;
}
} else {
for (y = 0; y < num_rows; ++y) {
memcpy(dst, deltas, width * sizeof(*dst));
prev_line = dst;
dst += width;
deltas += width;
}
}
dec->alpha_prev_line_ = prev_line;
} else { // alph_dec->method_ == ALPHA_LOSSLESS_COMPRESSION
assert(alph_dec->vp8l_dec_ != NULL);
if (!VP8LDecodeAlphaImageStream(alph_dec, row + num_rows)) {
@ -146,6 +158,7 @@ static int AllocateAlphaPlane(VP8Decoder* const dec, const VP8Io* const io) {
return 0;
}
dec->alpha_plane_ = dec->alpha_plane_mem_;
dec->alpha_prev_line_ = NULL;
return 1;
}

View File

@ -36,6 +36,7 @@ struct ALPHDecoder {
// pixel, sometimes VP8LDecoder may need to allocate
// 4 bytes per pixel internally during decode.
uint8_t* output_;
const uint8_t* prev_line_; // last output row (or NULL)
};
//------------------------------------------------------------------------------

View File

@ -261,6 +261,7 @@ struct VP8Decoder {
int is_alpha_decoded_; // true if alpha_data_ is decoded in alpha_plane_
uint8_t* alpha_plane_mem_; // memory allocated for alpha_plane_
uint8_t* alpha_plane_; // output. Persistent, contains the whole data.
const uint8_t* alpha_prev_line_; // last decoded alpha row (or NULL)
int alpha_dithering_; // derived from decoding options (0=off, 100=full)
};

View File

@ -771,35 +771,46 @@ static int Is8bOptimizable(const VP8LMetadata* const hdr) {
return 1;
}
static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int row) {
static void AlphaApplyFilter(ALPHDecoder* const alph_dec,
int first_row, int last_row,
uint8_t* out, int stride) {
if (alph_dec->filter_ != WEBP_FILTER_NONE) {
int y;
const uint8_t* prev_line = alph_dec->prev_line_;
assert(WebPUnfilters[alph_dec->filter_] != NULL);
for (y = first_row; y < last_row; ++y) {
WebPUnfilters[alph_dec->filter_](prev_line, out, out, stride);
prev_line = out;
out += stride;
}
alph_dec->prev_line_ = prev_line;
}
}
static void ExtractPalettedAlphaRows(VP8LDecoder* const dec, int last_row) {
// For vertical and gradient filtering, we need to decode the part above the
// crop_top row, in order to have the correct spatial predictors.
const ALPHDecoder* const alph_dec = (const ALPHDecoder*)dec->io_->opaque;
ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io_->opaque;
const int top_row =
(alph_dec->filter_ == WEBP_FILTER_NONE ||
alph_dec->filter_ == WEBP_FILTER_HORIZONTAL) ? dec->io_->crop_top
: dec->last_row_;
const int first_row = (dec->last_row_ < top_row) ? top_row : dec->last_row_;
assert(row <= dec->io_->crop_bottom);
if (row > first_row) {
assert(last_row <= dec->io_->crop_bottom);
if (last_row > first_row) {
// Special method for paletted alpha data. We only process the cropped area.
const int width = dec->io_->width;
uint8_t* const out = alph_dec->output_ + width * first_row;
uint8_t* out = alph_dec->output_ + width * first_row;
const uint8_t* const in =
(uint8_t*)dec->pixels_ + dec->width_ * first_row;
VP8LTransform* const transform = &dec->transforms_[0];
assert(dec->next_transform_ == 1);
assert(transform->type_ == COLOR_INDEXING_TRANSFORM);
VP8LColorIndexInverseTransformAlpha(transform, first_row, row,
VP8LColorIndexInverseTransformAlpha(transform, first_row, last_row,
in, out);
if (alph_dec->filter_ != WEBP_FILTER_NONE) {
assert(WebPUnfilters[alph_dec->filter_] != NULL);
WebPUnfilters[alph_dec->filter_](width, dec->io_->height, width,
first_row, row - first_row,
alph_dec->output_);
}
AlphaApplyFilter(alph_dec, first_row, last_row, out, width);
}
dec->last_row_ = dec->last_out_row_ = row;
dec->last_row_ = dec->last_out_row_ = last_row;
}
//------------------------------------------------------------------------------
@ -1452,29 +1463,25 @@ static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
//------------------------------------------------------------------------------
// 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_;
static void ExtractAlphaRows(VP8LDecoder* const dec, int last_row) {
const int num_rows = last_row - dec->last_row_;
const uint32_t* const in = dec->pixels_ + dec->width_ * dec->last_row_;
assert(row <= dec->io_->crop_bottom);
assert(last_row <= dec->io_->crop_bottom);
if (num_rows > 0) {
// Extract alpha (which is stored in the green plane).
const ALPHDecoder* const alph_dec = (const ALPHDecoder*)dec->io_->opaque;
ALPHDecoder* const alph_dec = (ALPHDecoder*)dec->io_->opaque;
uint8_t* const output = alph_dec->output_;
const int width = dec->io_->width; // the final width (!= dec->width_)
const int cache_pixs = width * num_rows;
uint8_t* const dst = output + width * dec->last_row_;
uint8_t* dst = output + width * dec->last_row_;
const uint32_t* const src = dec->argb_cache_;
int i;
ApplyInverseTransforms(dec, num_rows, in);
for (i = 0; i < cache_pixs; ++i) dst[i] = (src[i] >> 8) & 0xff;
if (alph_dec->filter_ != WEBP_FILTER_NONE) {
assert(WebPUnfilters[alph_dec->filter_] != NULL);
WebPUnfilters[alph_dec->filter_](width, dec->io_->height, width,
dec->last_row_, num_rows, output);
}
AlphaApplyFilter(alph_dec, dec->last_row_, last_row, dst, width);
}
dec->last_row_ = dec->last_out_row_ = row;
dec->last_row_ = dec->last_out_row_ = last_row;
}
int VP8LDecodeAlphaHeader(ALPHDecoder* const alph_dec,