// Assemble a single image WebP bitstream from 'wpi'.
static WebPMuxError SynthesizeBitstream(const WebPMuxImage* const wpi,
WebPData* const bitstream) {
uint8_t* dst;
// Allocate data.
const int need_vp8x = (wpi->alpha_ != NULL);
const size_t vp8x_size = need_vp8x ? CHUNK_HEADER_SIZE + VP8X_CHUNK_SIZE : 0;
const size_t alpha_size = need_vp8x ? ChunkDiskSize(wpi->alpha_) : 0;
// Note: No need to output ANMF/FRGM chunk for a single image.
const size_t size = RIFF_HEADER_SIZE + vp8x_size + alpha_size +
ChunkDiskSize(wpi->img_);
uint8_t* const data = (uint8_t*)WebPSafeMalloc(1ULL, size);
if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
// Main RIFF header.
dst = MuxEmitRiffHeader(data, size);
if (need_vp8x) {
dst = EmitVP8XChunk(dst, wpi->width_, wpi->height_, ALPHA_FLAG); // VP8X.
dst = ChunkListEmit(wpi->alpha_, dst); // ALPH.
}
// Bitstream.
dst = ChunkListEmit(wpi->img_, dst);
assert(dst == data + size);
// Output.
bitstream->bytes = data;
bitstream->size = size;
return WEBP_MUX_OK;
}
// Create data for frame/fragment given image data, offsets and duration.
static WebPMuxError CreateFrameFragmentData(
int width, int height, const WebPMuxFrameInfo* const info, int is_frame,
WebPData* const frame_frgm) {
uint8_t* frame_frgm_bytes;
const size_t frame_frgm_size = kChunks[is_frame ? IDX_ANMF : IDX_FRGM].size;
assert(width > 0 && height > 0 && info->duration >= 0);
assert(info->dispose_method == (info->dispose_method & 1));
// Note: assertion on upper bounds is done in PutLE24().
frame_frgm_bytes = (uint8_t*)WebPSafeMalloc(1ULL, frame_frgm_size);
if (frame_frgm_bytes == NULL) return WEBP_MUX_MEMORY_ERROR;
PutLE24(frame_frgm_bytes + 0, info->x_offset / 2);
PutLE24(frame_frgm_bytes + 3, info->y_offset / 2);
if (is_frame) {
PutLE24(frame_frgm_bytes + 6, width - 1);
PutLE24(frame_frgm_bytes + 9, height - 1);
PutLE24(frame_frgm_bytes + 12, info->duration);
frame_frgm_bytes[15] =
(info->blend_method == WEBP_MUX_NO_BLEND ? 2 : 0) |
(info->dispose_method == WEBP_MUX_DISPOSE_BACKGROUND ? 1 : 0);
}
frame_frgm->bytes = frame_frgm_bytes;
frame_frgm->size = frame_frgm_size;
return WEBP_MUX_OK;
}
static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
const int w = buffer->width;
const int h = buffer->height;
const WEBP_CSP_MODE mode = buffer->colorspace;
if (w <= 0 || h <= 0 || !IsValidColorspace(mode)) {
return VP8_STATUS_INVALID_PARAM;
}
if (!buffer->is_external_memory && buffer->private_memory == NULL) {
uint8_t* output;
int uv_stride = 0, a_stride = 0;
uint64_t uv_size = 0, a_size = 0, total_size;
// We need memory and it hasn't been allocated yet.
// => initialize output buffer, now that dimensions are known.
const int stride = w * kModeBpp[mode];
const uint64_t size = (uint64_t)stride * h;
if (!WebPIsRGBMode(mode)) {
uv_stride = (w + 1) / 2;
uv_size = (uint64_t)uv_stride * ((h + 1) / 2);
if (mode == MODE_YUVA) {
a_stride = w;
a_size = (uint64_t)a_stride * h;
}
}
total_size = size + 2 * uv_size + a_size;
// Security/sanity checks
output = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*output));
if (output == NULL) {
return VP8_STATUS_OUT_OF_MEMORY;
}
buffer->private_memory = output;
if (!WebPIsRGBMode(mode)) { // YUVA initialization
WebPYUVABuffer* const buf = &buffer->u.YUVA;
buf->y = output;
buf->y_stride = stride;
buf->y_size = (size_t)size;
buf->u = output + size;
buf->u_stride = uv_stride;
buf->u_size = (size_t)uv_size;
buf->v = output + size + uv_size;
buf->v_stride = uv_stride;
buf->v_size = (size_t)uv_size;
if (mode == MODE_YUVA) {
buf->a = output + size + 2 * uv_size;
}
buf->a_size = (size_t)a_size;
buf->a_stride = a_stride;
} else { // RGBA initialization
WebPRGBABuffer* const buf = &buffer->u.RGBA;
buf->rgba = output;
buf->stride = stride;
buf->size = (size_t)size;
}
}
return CheckDecBuffer(buffer);
}
int HuffmanTreeBuildImplicit(HuffmanTree* const tree,
const int* const code_lengths,
int code_lengths_size) {
int symbol;
int num_symbols = 0;
int root_symbol = 0;
assert(tree != NULL);
assert(code_lengths != NULL);
// Find out number of symbols and the root symbol.
for (symbol = 0; symbol < code_lengths_size; ++symbol) {
if (code_lengths[symbol] > 0) {
// Note: code length = 0 indicates non-existent symbol.
++num_symbols;
root_symbol = symbol;
}
}
// Initialize the tree. Will fail for num_symbols = 0
if (!TreeInit(tree, num_symbols)) return 0;
// Build tree.
if (num_symbols == 1) { // Trivial case.
const int max_symbol = code_lengths_size;
if (root_symbol < 0 || root_symbol >= max_symbol) {
HuffmanTreeRelease(tree);
return 0;
}
return TreeAddSymbol(tree, root_symbol, 0, 0);
} else { // Normal case.
int ok = 0;
// Get Huffman codes from the code lengths.
int* const codes =
(int*)WebPSafeMalloc((uint64_t)code_lengths_size, sizeof(*codes));
if (codes == NULL) goto End;
if (!HuffmanCodeLengthsToCodes(code_lengths, code_lengths_size, codes)) {
goto End;
}
// Add symbols one-by-one.
for (symbol = 0; symbol < code_lengths_size; ++symbol) {
if (code_lengths[symbol] > 0) {
if (!TreeAddSymbol(tree, symbol, codes[symbol], code_lengths[symbol])) {
goto End;
}
}
}
ok = 1;
End:
free(codes);
ok = ok && IsFull(tree);
if (!ok) HuffmanTreeRelease(tree);
return ok;
}
}
static int CustomSetup(VP8Io* io) {
WebPDecParams* const p = (WebPDecParams*)io->opaque;
const WEBP_CSP_MODE colorspace = p->output->colorspace;
const int is_rgb = WebPIsRGBMode(colorspace);
const int is_alpha = WebPIsAlphaMode(colorspace);
p->memory = NULL;
p->emit = NULL;
p->emit_alpha = NULL;
p->emit_alpha_row = NULL;
if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
return 0;
}
if (is_alpha && WebPIsPremultipliedMode(colorspace)) {
WebPInitUpsamplers();
}
if (io->use_scaling) {
#if !defined(WEBP_REDUCE_SIZE)
const int ok = is_rgb ? InitRGBRescaler(io, p) : InitYUVRescaler(io, p);
if (!ok) {
return 0; // memory error
}
#else
return 0; // rescaling support not compiled
#endif
} else {
if (is_rgb) {
WebPInitSamplers();
p->emit = EmitSampledRGB; // default
if (io->fancy_upsampling) {
#ifdef FANCY_UPSAMPLING
const int uv_width = (io->mb_w + 1) >> 1;
p->memory = WebPSafeMalloc(1ULL, (size_t)(io->mb_w + 2 * uv_width));
if (p->memory == NULL) {
return 0; // memory error.
}
p->tmp_y = (uint8_t*)p->memory;
p->tmp_u = p->tmp_y + io->mb_w;
p->tmp_v = p->tmp_u + uv_width;
p->emit = EmitFancyRGB;
WebPInitUpsamplers();
#endif
}
} else {
p->emit = EmitYUV;
}
if (is_alpha) { // need transparency output
p->emit_alpha =
(colorspace == MODE_RGBA_4444 || colorspace == MODE_rgbA_4444) ?
EmitAlphaRGBA4444
: is_rgb ? EmitAlphaRGB
: EmitAlphaYUV;
if (is_rgb) {
WebPInitAlphaProcessing();
}
}
}
HTreeGroup* VP8LHtreeGroupsNew(int num_htree_groups) {
HTreeGroup* const htree_groups =
(HTreeGroup*)WebPSafeMalloc(num_htree_groups, sizeof(*htree_groups));
if (htree_groups == NULL) {
return NULL;
}
assert(num_htree_groups <= MAX_HTREE_GROUPS);
return htree_groups;
}
WebPMux* WebPNewInternal(int version) {
if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_MUX_ABI_VERSION)) {
return NULL;
} else {
WebPMux* const mux = (WebPMux*)WebPSafeMalloc(1ULL, sizeof(WebPMux));
if (mux != NULL) MuxInit(mux);
return mux;
}
}
WebPMuxError WebPMuxAssemble(WebPMux* mux, WebPData* assembled_data) {
size_t size = 0;
uint8_t* data = NULL;
uint8_t* dst = NULL;
WebPMuxError err;
if (assembled_data == NULL) {
return WEBP_MUX_INVALID_ARGUMENT;
}
// Clean up returned data, in case something goes wrong.
memset(assembled_data, 0, sizeof(*assembled_data));
if (mux == NULL) {
return WEBP_MUX_INVALID_ARGUMENT;
}
// Finalize mux.
err = MuxCleanup(mux);
if (err != WEBP_MUX_OK) return err;
err = CreateVP8XChunk(mux);
if (err != WEBP_MUX_OK) return err;
// Allocate data.
size = ChunkListDiskSize(mux->vp8x_) + ChunkListDiskSize(mux->iccp_)
+ ChunkListDiskSize(mux->anim_) + ImageListDiskSize(mux->images_)
+ ChunkListDiskSize(mux->exif_) + ChunkListDiskSize(mux->xmp_)
+ ChunkListDiskSize(mux->unknown_) + RIFF_HEADER_SIZE;
data = (uint8_t*)WebPSafeMalloc(1ULL, size);
if (data == NULL) return WEBP_MUX_MEMORY_ERROR;
// Emit header & chunks.
dst = MuxEmitRiffHeader(data, size);
dst = ChunkListEmit(mux->vp8x_, dst);
dst = ChunkListEmit(mux->iccp_, dst);
dst = ChunkListEmit(mux->anim_, dst);
dst = ImageListEmit(mux->images_, dst);
dst = ChunkListEmit(mux->exif_, dst);
dst = ChunkListEmit(mux->xmp_, dst);
dst = ChunkListEmit(mux->unknown_, dst);
assert(dst == data + size);
// Validate mux.
err = MuxValidate(mux);
if (err != WEBP_MUX_OK) {
WebPSafeFree(data);
data = NULL;
size = 0;
}
// Finalize data.
assembled_data->bytes = data;
assembled_data->size = size;
return err;
}
static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
const WebPYUVABuffer* const buf = &p->output->u.YUVA;
const int out_width = io->scaled_width;
const int out_height = io->scaled_height;
const int uv_out_width = (out_width + 1) >> 1;
const int uv_out_height = (out_height + 1) >> 1;
const int uv_in_width = (io->mb_w + 1) >> 1;
const int uv_in_height = (io->mb_h + 1) >> 1;
const size_t work_size = 2 * out_width; // scratch memory for luma rescaler
const size_t uv_work_size = 2 * uv_out_width; // and for each u/v ones
size_t tmp_size, rescaler_size;
rescaler_t* work;
WebPRescaler* scalers;
const int num_rescalers = has_alpha ? 4 : 3;
tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
if (has_alpha) {
tmp_size += work_size * sizeof(*work);
}
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
p->memory = WebPSafeMalloc(1ULL, tmp_size + rescaler_size);
if (p->memory == NULL) {
return 0; // memory error
}
work = (rescaler_t*)p->memory;
scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + tmp_size);
p->scaler_y = &scalers[0];
p->scaler_u = &scalers[1];
p->scaler_v = &scalers[2];
p->scaler_a = has_alpha ? &scalers[3] : NULL;
WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
buf->y, out_width, out_height, buf->y_stride, 1,
work);
WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
buf->u, uv_out_width, uv_out_height, buf->u_stride, 1,
work + work_size);
WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
buf->v, uv_out_width, uv_out_height, buf->v_stride, 1,
work + work_size + uv_work_size);
p->emit = EmitRescaledYUV;
if (has_alpha) {
WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
buf->a, out_width, out_height, buf->a_stride, 1,
work + work_size + 2 * uv_work_size);
p->emit_alpha = EmitRescaledAlphaYUV;
WebPInitAlphaProcessing();
}
return 1;
}
int VP8LHashChainInit(VP8LHashChain* const p, int size) {
assert(p->size_ == 0);
assert(p->offset_length_ == NULL);
assert(size > 0);
p->offset_length_ =
(uint32_t*)WebPSafeMalloc(size, sizeof(*p->offset_length_));
if (p->offset_length_ == NULL) return 0;
p->size_ = size;
return 1;
}
static int TreeInit(HuffmanTree* const tree, int num_leaves) {
assert(tree != NULL);
if (num_leaves == 0) return 0;
// We allocate maximum possible nodes in the tree at once.
// Note that a Huffman tree is a full binary tree; and in a full binary tree
// with L leaves, the total number of nodes N = 2 * L - 1.
tree->max_nodes_ = 2 * num_leaves - 1;
tree->root_ = (HuffmanTreeNode*)WebPSafeMalloc((uint64_t)tree->max_nodes_,
sizeof(*tree->root_));
if (tree->root_ == NULL) return 0;
TreeNodeInit(tree->root_); // Initialize root.
tree->num_nodes_ = 1;
return 1;
}
static int TBufferNewPage(VP8TBuffer* const b) {
VP8Tokens* page = NULL;
if (!b->error_) {
const size_t size = sizeof(*page) + b->page_size_ * sizeof(token_t);
page = (VP8Tokens*)WebPSafeMalloc(1ULL, size);
}
if (page == NULL) {
b->error_ = 1;
return 0;
}
page->next_ = NULL;
*b->last_page_ = page;
b->last_page_ = &page->next_;
b->left_ = b->page_size_;
b->tokens_ = (token_t*)TOKEN_DATA(page);
return 1;
}
int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
int type, float result[5]) {
VP8DistoStats stats[5];
int w, h;
memset(stats, 0, sizeof(stats));
VP8SSIMDspInit();
if (src == NULL || ref == NULL ||
src->width != ref->width || src->height != ref->height ||
src->use_argb != ref->use_argb || result == NULL) {
return 0;
}
w = src->width;
h = src->height;
if (src->use_argb == 1) {
if (src->argb == NULL || ref->argb == NULL) {
return 0;
} else {
int i, j, c;
uint8_t* tmp1, *tmp2;
uint8_t* const tmp_plane =
(uint8_t*)WebPSafeMalloc(2ULL * w * h, sizeof(*tmp_plane));
if (tmp_plane == NULL) return 0;
tmp1 = tmp_plane;
tmp2 = tmp_plane + w * h;
for (c = 0; c < 4; ++c) {
for (j = 0; j < h; ++j) {
for (i = 0; i < w; ++i) {
tmp1[j * w + i] = src->argb[i + j * src->argb_stride] >> (c * 8);
tmp2[j * w + i] = ref->argb[i + j * ref->argb_stride] >> (c * 8);
}
}
if (type >= 2) {
AccumulateLSIM(tmp1, w, tmp2, w, w, h, &stats[c]);
} else {
VP8SSIMAccumulatePlane(tmp1, w, tmp2, w, w, h, &stats[c]);
}
}
free(tmp_plane);
}
} else {
static int CustomSetup(VP8Io* io) {
WebPDecParams* const p = (WebPDecParams*)io->opaque;
const WEBP_CSP_MODE colorspace = p->output->colorspace;
const int is_rgb = WebPIsRGBMode(colorspace);
const int is_alpha = WebPIsAlphaMode(colorspace);
p->memory = NULL;
p->emit = NULL;
p->emit_alpha = NULL;
p->emit_alpha_row = NULL;
if (!WebPIoInitFromOptions(p->options, io, is_alpha ? MODE_YUV : MODE_YUVA)) {
return 0;
}
if (is_alpha && WebPIsPremultipliedMode(colorspace)) {
WebPInitUpsamplers();
}
if (io->use_scaling) {
const int ok = is_rgb ? InitRGBRescaler(io, p) : InitYUVRescaler(io, p);
if (!ok) {
return 0; // memory error
}
} else {
if (is_rgb) {
p->emit = EmitSampledRGB; // default
if (io->fancy_upsampling) {
#ifdef FANCY_UPSAMPLING
const int uv_width = (io->mb_w + 1) >> 1;
p->memory = WebPSafeMalloc(1ULL, (size_t)(io->mb_w + 2 * uv_width));
if (p->memory == NULL) {
return 0; // memory error.
}
p->tmp_y = (uint8_t*)p->memory;
p->tmp_u = p->tmp_y + io->mb_w;
p->tmp_v = p->tmp_u + uv_width;
p->emit = EmitFancyRGB;
WebPInitUpsamplers();
#endif
} else {
WebPInitSamplers();
}
} else {
// Initialize all params.
static int InitParams(uint8_t* const data, int width, int height, int stride,
int radius, SmoothParams* const p) {
const int R = 2 * radius + 1; // total size of the kernel
const size_t size_scratch_m = (R + 1) * width * sizeof(*p->start_);
const size_t size_m = width * sizeof(*p->average_);
const size_t size_lut = (1 + 2 * LUT_SIZE) * sizeof(*p->correction_);
const size_t total_size = size_scratch_m + size_m + size_lut;
uint8_t* mem = (uint8_t*)WebPSafeMalloc(1U, total_size);
if (mem == NULL) return 0;
p->mem_ = (void*)mem;
p->start_ = (uint16_t*)mem;
p->cur_ = p->start_;
p->end_ = p->start_ + R * width;
p->top_ = p->end_ - width;
memset(p->top_, 0, width * sizeof(*p->top_));
mem += size_scratch_m;
p->average_ = (uint16_t*)mem;
mem += size_m;
p->width_ = width;
p->height_ = height;
p->stride_ = stride;
p->src_ = data;
p->dst_ = data;
p->radius_ = radius;
p->scale_ = (1 << (FIX + LFIX)) / (R * R); // normalization constant
p->row_ = -radius;
// analyze the input distribution so we can best-fit the threshold
CountLevels(p);
// correction table
p->correction_ = ((int16_t*)mem) + LUT_SIZE;
InitCorrectionLUT(p->correction_, p->min_level_dist_);
return 1;
}
// Create a new block, either from the free list or allocated
static PixOrCopyBlock* BackwardRefsNewBlock(VP8LBackwardRefs* const refs) {
PixOrCopyBlock* b = refs->free_blocks_;
if (b == NULL) { // allocate new memory chunk
const size_t total_size =
sizeof(*b) + refs->block_size_ * sizeof(*b->start_);
b = (PixOrCopyBlock*)WebPSafeMalloc(1ULL, total_size);
if (b == NULL) {
refs->error_ |= 1;
return NULL;
}
b->start_ = (PixOrCopy*)((uint8_t*)b + sizeof(*b)); // not always aligned
} else { // recycle from free-list
refs->free_blocks_ = b->next_;
}
*refs->tail_ = b;
refs->tail_ = &b->next_;
refs->last_block_ = b;
b->next_ = NULL;
b->size_ = 0;
return b;
}
int WebPPlaneDistortion(const uint8_t* src, size_t src_stride,
const uint8_t* ref, size_t ref_stride,
int width, int height, size_t x_step,
int type, float* distortion, float* result) {
uint8_t* allocated = NULL;
const AccumulateFunc metric = (type == 0) ? AccumulateSSE :
(type == 1) ? AccumulateSSIM :
AccumulateLSIM;
if (src == NULL || ref == NULL ||
src_stride < x_step * width || ref_stride < x_step * width ||
result == NULL || distortion == NULL) {
return 0;
}
VP8SSIMDspInit();
if (x_step != 1) { // extract a packed plane if needed
int x, y;
uint8_t* tmp1;
uint8_t* tmp2;
allocated =
(uint8_t*)WebPSafeMalloc(2ULL * width * height, sizeof(*allocated));
if (allocated == NULL) return 0;
tmp1 = allocated;
tmp2 = tmp1 + (size_t)width * height;
for (y = 0; y < height; ++y) {
for (x = 0; x < width; ++x) {
tmp1[x + y * width] = src[x * x_step + y * src_stride];
tmp2[x + y * width] = ref[x * x_step + y * ref_stride];
}
}
src = tmp1;
ref = tmp2;
}
*distortion = (float)metric(src, width, ref, width, width, height);
WebPSafeFree(allocated);
*result = (type == 1) ? (float)GetLogSSIM(*distortion, (double)width * height)
: (float)GetPSNR(*distortion, (double)width * height);
return 1;
}
static void SmoothSegmentMap(VP8Encoder* const enc) {
int n, x, y;
const int w = enc->mb_w_;
const int h = enc->mb_h_;
const int majority_cnt_3_x_3_grid = 5;
uint8_t* const tmp = (uint8_t*)WebPSafeMalloc(w * h, sizeof(*tmp));
assert((uint64_t)(w * h) == (uint64_t)w * h); // no overflow, as per spec
if (tmp == NULL) return;
for (y = 1; y < h - 1; ++y) {
for (x = 1; x < w - 1; ++x) {
int cnt[NUM_MB_SEGMENTS] = { 0 };
const VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
int majority_seg = mb->segment_;
// Check the 8 neighbouring segment values.
cnt[mb[-w - 1].segment_]++; // top-left
cnt[mb[-w + 0].segment_]++; // top
cnt[mb[-w + 1].segment_]++; // top-right
cnt[mb[ - 1].segment_]++; // left
cnt[mb[ + 1].segment_]++; // right
cnt[mb[ w - 1].segment_]++; // bottom-left
cnt[mb[ w + 0].segment_]++; // bottom
cnt[mb[ w + 1].segment_]++; // bottom-right
for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
if (cnt[n] >= majority_cnt_3_x_3_grid) {
majority_seg = n;
break;
}
}
tmp[x + y * w] = majority_seg;
}
}
for (y = 1; y < h - 1; ++y) {
for (x = 1; x < w - 1; ++x) {
VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
mb->segment_ = tmp[x + y * w];
}
}
WebPSafeFree(tmp);
}
static void SmoothSegmentMap(VP8Encoder* const enc) {
int n, x, y;
const int w = enc->mb_w_;
const int h = enc->mb_h_;
const int majority_cnt_3_x_3_grid = 5;
uint8_t* const tmp = (uint8_t*)WebPSafeMalloc((uint64_t)w * h, sizeof(*tmp));
assert((uint64_t)(w * h) == (uint64_t)w * h);
if (tmp == NULL) return;
for (y = 1; y < h - 1; ++y) {
for (x = 1; x < w - 1; ++x) {
int cnt[NUM_MB_SEGMENTS] = { 0 };
const VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
int majority_seg = mb->segment_;
cnt[mb[-w - 1].segment_]++;
cnt[mb[-w + 0].segment_]++;
cnt[mb[-w + 1].segment_]++;
cnt[mb[ - 1].segment_]++;
cnt[mb[ + 1].segment_]++;
cnt[mb[ w - 1].segment_]++;
cnt[mb[ w + 0].segment_]++;
cnt[mb[ w + 1].segment_]++;
for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
if (cnt[n] >= majority_cnt_3_x_3_grid) {
majority_seg = n;
}
}
tmp[x + y * w] = majority_seg;
}
}
for (y = 1; y < h - 1; ++y) {
for (x = 1; x < w - 1; ++x) {
VP8MBInfo* const mb = &enc->mb_info_[x + w * y];
mb->segment_ = tmp[x + y * w];
}
}
free(tmp);
}
static int BackwardReferencesLz77Box(int xsize, int ysize,
const uint32_t* const argb, int cache_bits,
const VP8LHashChain* const hash_chain_best,
VP8LHashChain* hash_chain,
VP8LBackwardRefs* const refs) {
int i;
const int pix_count = xsize * ysize;
uint16_t* counts;
int window_offsets[WINDOW_OFFSETS_SIZE_MAX] = {0};
int window_offsets_new[WINDOW_OFFSETS_SIZE_MAX] = {0};
int window_offsets_size = 0;
int window_offsets_new_size = 0;
uint16_t* const counts_ini =
(uint16_t*)WebPSafeMalloc(xsize * ysize, sizeof(*counts_ini));
int best_offset_prev = -1, best_length_prev = -1;
if (counts_ini == NULL) return 0;
// counts[i] counts how many times a pixel is repeated starting at position i.
i = pix_count - 2;
counts = counts_ini + i;
counts[1] = 1;
for (; i >= 0; --i, --counts) {
if (argb[i] == argb[i + 1]) {
// Max out the counts to MAX_LENGTH.
counts[0] = counts[1] + (counts[1] != MAX_LENGTH);
} else {
counts[0] = 1;
}
}
// Figure out the window offsets around a pixel. They are stored in a
// spiraling order around the pixel as defined by VP8LDistanceToPlaneCode.
{
int x, y;
for (y = 0; y <= 6; ++y) {
for (x = -6; x <= 6; ++x) {
const int offset = y * xsize + x;
int plane_code;
// Ignore offsets that bring us after the pixel.
if (offset <= 0) continue;
plane_code = VP8LDistanceToPlaneCode(xsize, offset) - 1;
if (plane_code >= WINDOW_OFFSETS_SIZE_MAX) continue;
window_offsets[plane_code] = offset;
}
}
// For narrow images, not all plane codes are reached, so remove those.
for (i = 0; i < WINDOW_OFFSETS_SIZE_MAX; ++i) {
if (window_offsets[i] == 0) continue;
window_offsets[window_offsets_size++] = window_offsets[i];
}
// Given a pixel P, find the offsets that reach pixels unreachable from P-1
// with any of the offsets in window_offsets[].
for (i = 0; i < window_offsets_size; ++i) {
int j;
int is_reachable = 0;
for (j = 0; j < window_offsets_size && !is_reachable; ++j) {
is_reachable |= (window_offsets[i] == window_offsets[j] + 1);
}
if (!is_reachable) {
window_offsets_new[window_offsets_new_size] = window_offsets[i];
++window_offsets_new_size;
}
}
}
hash_chain->offset_length_[0] = 0;
for (i = 1; i < pix_count; ++i) {
int ind;
int best_length = VP8LHashChainFindLength(hash_chain_best, i);
int best_offset;
int do_compute = 1;
if (best_length >= MAX_LENGTH) {
// Do not recompute the best match if we already have a maximal one in the
// window.
best_offset = VP8LHashChainFindOffset(hash_chain_best, i);
for (ind = 0; ind < window_offsets_size; ++ind) {
if (best_offset == window_offsets[ind]) {
do_compute = 0;
break;
}
}
}
if (do_compute) {
// Figure out if we should use the offset/length from the previous pixel
// as an initial guess and therefore only inspect the offsets in
// window_offsets_new[].
const int use_prev =
(best_length_prev > 1) && (best_length_prev < MAX_LENGTH);
const int num_ind =
use_prev ? window_offsets_new_size : window_offsets_size;
best_length = use_prev ? best_length_prev - 1 : 0;
best_offset = use_prev ? best_offset_prev : 0;
// Find the longest match in a window around the pixel.
for (ind = 0; ind < num_ind; ++ind) {
int curr_length = 0;
int j = i;
int j_offset =
use_prev ? i - window_offsets_new[ind] : i - window_offsets[ind];
if (j_offset < 0 || argb[j_offset] != argb[i]) continue;
// The longest match is the sum of how many times each pixel is
// repeated.
do {
const int counts_j_offset = counts_ini[j_offset];
const int counts_j = counts_ini[j];
if (counts_j_offset != counts_j) {
curr_length +=
(counts_j_offset < counts_j) ? counts_j_offset : counts_j;
break;
}
// The same color is repeated counts_pos times at j_offset and j.
curr_length += counts_j_offset;
j_offset += counts_j_offset;
j += counts_j_offset;
} while (curr_length <= MAX_LENGTH && j < pix_count &&
argb[j_offset] == argb[j]);
if (best_length < curr_length) {
best_offset =
use_prev ? window_offsets_new[ind] : window_offsets[ind];
if (curr_length >= MAX_LENGTH) {
best_length = MAX_LENGTH;
break;
} else {
best_length = curr_length;
}
}
}
}
assert(i + best_length <= pix_count);
assert(best_length <= MAX_LENGTH);
if (best_length <= MIN_LENGTH) {
hash_chain->offset_length_[i] = 0;
best_offset_prev = 0;
best_length_prev = 0;
} else {
hash_chain->offset_length_[i] =
(best_offset << MAX_LENGTH_BITS) | (uint32_t)best_length;
best_offset_prev = best_offset;
best_length_prev = best_length;
}
}
hash_chain->offset_length_[0] = 0;
WebPSafeFree(counts_ini);
return BackwardReferencesLz77(xsize, ysize, argb, cache_bits, hash_chain,
refs);
}
int VP8LHashChainFill(VP8LHashChain* const p, int quality,
const uint32_t* const argb, int xsize, int ysize,
int low_effort) {
const int size = xsize * ysize;
const int iter_max = GetMaxItersForQuality(quality);
const uint32_t window_size = GetWindowSizeForHashChain(quality, xsize);
int pos;
int argb_comp;
uint32_t base_position;
int32_t* hash_to_first_index;
// Temporarily use the p->offset_length_ as a hash chain.
int32_t* chain = (int32_t*)p->offset_length_;
assert(size > 0);
assert(p->size_ != 0);
assert(p->offset_length_ != NULL);
if (size <= 2) {
p->offset_length_[0] = p->offset_length_[size - 1] = 0;
return 1;
}
hash_to_first_index =
(int32_t*)WebPSafeMalloc(HASH_SIZE, sizeof(*hash_to_first_index));
if (hash_to_first_index == NULL) return 0;
// Set the int32_t array to -1.
memset(hash_to_first_index, 0xff, HASH_SIZE * sizeof(*hash_to_first_index));
// Fill the chain linking pixels with the same hash.
argb_comp = (argb[0] == argb[1]);
for (pos = 0; pos < size - 2;) {
uint32_t hash_code;
const int argb_comp_next = (argb[pos + 1] == argb[pos + 2]);
if (argb_comp && argb_comp_next) {
// Consecutive pixels with the same color will share the same hash.
// We therefore use a different hash: the color and its repetition
// length.
uint32_t tmp[2];
uint32_t len = 1;
tmp[0] = argb[pos];
// Figure out how far the pixels are the same.
// The last pixel has a different 64 bit hash, as its next pixel does
// not have the same color, so we just need to get to the last pixel equal
// to its follower.
while (pos + (int)len + 2 < size && argb[pos + len + 2] == argb[pos]) {
++len;
}
if (len > MAX_LENGTH) {
// Skip the pixels that match for distance=1 and length>MAX_LENGTH
// because they are linked to their predecessor and we automatically
// check that in the main for loop below. Skipping means setting no
// predecessor in the chain, hence -1.
memset(chain + pos, 0xff, (len - MAX_LENGTH) * sizeof(*chain));
pos += len - MAX_LENGTH;
len = MAX_LENGTH;
}
// Process the rest of the hash chain.
while (len) {
tmp[1] = len--;
hash_code = GetPixPairHash64(tmp);
chain[pos] = hash_to_first_index[hash_code];
hash_to_first_index[hash_code] = pos++;
}
argb_comp = 0;
} else {
// Just move one pixel forward.
hash_code = GetPixPairHash64(argb + pos);
chain[pos] = hash_to_first_index[hash_code];
hash_to_first_index[hash_code] = pos++;
argb_comp = argb_comp_next;
}
}
// Process the penultimate pixel.
chain[pos] = hash_to_first_index[GetPixPairHash64(argb + pos)];
WebPSafeFree(hash_to_first_index);
// Find the best match interval at each pixel, defined by an offset to the
// pixel and a length. The right-most pixel cannot match anything to the right
// (hence a best length of 0) and the left-most pixel nothing to the left
// (hence an offset of 0).
assert(size > 2);
p->offset_length_[0] = p->offset_length_[size - 1] = 0;
for (base_position = size - 2; base_position > 0;) {
const int max_len = MaxFindCopyLength(size - 1 - base_position);
const uint32_t* const argb_start = argb + base_position;
int iter = iter_max;
int best_length = 0;
uint32_t best_distance = 0;
uint32_t best_argb;
const int min_pos =
(base_position > window_size) ? base_position - window_size : 0;
const int length_max = (max_len < 256) ? max_len : 256;
uint32_t max_base_position;
pos = chain[base_position];
if (!low_effort) {
int curr_length;
// Heuristic: use the comparison with the above line as an initialization.
if (base_position >= (uint32_t)xsize) {
curr_length = FindMatchLength(argb_start - xsize, argb_start,
best_length, max_len);
if (curr_length > best_length) {
best_length = curr_length;
best_distance = xsize;
}
--iter;
}
// Heuristic: compare to the previous pixel.
curr_length =
FindMatchLength(argb_start - 1, argb_start, best_length, max_len);
if (curr_length > best_length) {
best_length = curr_length;
best_distance = 1;
}
--iter;
// Skip the for loop if we already have the maximum.
if (best_length == MAX_LENGTH) pos = min_pos - 1;
}
best_argb = argb_start[best_length];
for (; pos >= min_pos && --iter; pos = chain[pos]) {
int curr_length;
assert(base_position > (uint32_t)pos);
if (argb[pos + best_length] != best_argb) continue;
curr_length = VP8LVectorMismatch(argb + pos, argb_start, max_len);
if (best_length < curr_length) {
best_length = curr_length;
best_distance = base_position - pos;
best_argb = argb_start[best_length];
// Stop if we have reached a good enough length.
if (best_length >= length_max) break;
}
}
// We have the best match but in case the two intervals continue matching
// to the left, we have the best matches for the left-extended pixels.
max_base_position = base_position;
while (1) {
assert(best_length <= MAX_LENGTH);
assert(best_distance <= WINDOW_SIZE);
p->offset_length_[base_position] =
(best_distance << MAX_LENGTH_BITS) | (uint32_t)best_length;
--base_position;
// Stop if we don't have a match or if we are out of bounds.
if (best_distance == 0 || base_position == 0) break;
// Stop if we cannot extend the matching intervals to the left.
if (base_position < best_distance ||
argb[base_position - best_distance] != argb[base_position]) {
break;
}
// Stop if we are matching at its limit because there could be a closer
// matching interval with the same maximum length. Then again, if the
// matching interval is as close as possible (best_distance == 1), we will
// never find anything better so let's continue.
if (best_length == MAX_LENGTH && best_distance != 1 &&
base_position + MAX_LENGTH < max_base_position) {
break;
}
if (best_length < MAX_LENGTH) {
++best_length;
max_base_position = base_position;
}
}
}
return 1;
}
int VP8LBuildHuffmanTable(HuffmanCode* const root_table, int root_bits,
const int code_lengths[], int code_lengths_size) {
HuffmanCode* table = root_table; // next available space in table
int total_size = 1 << root_bits; // total size root table + 2nd level table
int* sorted = NULL; // symbols sorted by code length
int len; // current code length
int symbol; // symbol index in original or sorted table
// number of codes of each length:
int count[MAX_ALLOWED_CODE_LENGTH + 1] = { 0 };
// offsets in sorted table for each length:
int offset[MAX_ALLOWED_CODE_LENGTH + 1];
assert(code_lengths_size != 0);
assert(code_lengths != NULL);
assert(root_table != NULL);
assert(root_bits > 0);
// Build histogram of code lengths.
for (symbol = 0; symbol < code_lengths_size; ++symbol) {
if (code_lengths[symbol] > MAX_ALLOWED_CODE_LENGTH) {
return 0;
}
++count[code_lengths[symbol]];
}
// Error, all code lengths are zeros.
if (count[0] == code_lengths_size) {
return 0;
}
// Generate offsets into sorted symbol table by code length.
offset[1] = 0;
for (len = 1; len < MAX_ALLOWED_CODE_LENGTH; ++len) {
if (count[len] > (1 << len)) {
return 0;
}
offset[len + 1] = offset[len] + count[len];
}
sorted = (int*)WebPSafeMalloc(code_lengths_size, sizeof(*sorted));
if (sorted == NULL) {
return 0;
}
// Sort symbols by length, by symbol order within each length.
for (symbol = 0; symbol < code_lengths_size; ++symbol) {
const int symbol_code_length = code_lengths[symbol];
if (code_lengths[symbol] > 0) {
sorted[offset[symbol_code_length]++] = symbol;
}
}
// Special case code with only one value.
if (offset[MAX_ALLOWED_CODE_LENGTH] == 1) {
HuffmanCode code;
code.bits = 0;
code.value = (uint16_t)sorted[0];
ReplicateValue(table, 1, total_size, code);
WebPSafeFree(sorted);
return total_size;
}
{
int step; // step size to replicate values in current table
uint32_t low = -1; // low bits for current root entry
uint32_t mask = total_size - 1; // mask for low bits
uint32_t key = 0; // reversed prefix code
int num_nodes = 1; // number of Huffman tree nodes
int num_open = 1; // number of open branches in current tree level
int table_bits = root_bits; // key length of current table
int table_size = 1 << table_bits; // size of current table
symbol = 0;
// Fill in root table.
for (len = 1, step = 2; len <= root_bits; ++len, step <<= 1) {
num_open <<= 1;
num_nodes += num_open;
num_open -= count[len];
if (num_open < 0) {
WebPSafeFree(sorted);
return 0;
}
for (; count[len] > 0; --count[len]) {
HuffmanCode code;
code.bits = (uint8_t)len;
code.value = (uint16_t)sorted[symbol++];
ReplicateValue(&table[key], step, table_size, code);
key = GetNextKey(key, len);
}
}
// Fill in 2nd level tables and add pointers to root table.
for (len = root_bits + 1, step = 2; len <= MAX_ALLOWED_CODE_LENGTH;
++len, step <<= 1) {
num_open <<= 1;
num_nodes += num_open;
num_open -= count[len];
if (num_open < 0) {
WebPSafeFree(sorted);
return 0;
}
for (; count[len] > 0; --count[len]) {
HuffmanCode code;
if ((key & mask) != low) {
table += table_size;
table_bits = NextTableBitSize(count, len, root_bits);
table_size = 1 << table_bits;
total_size += table_size;
low = key & mask;
root_table[low].bits = (uint8_t)(table_bits + root_bits);
root_table[low].value = (uint16_t)((table - root_table) - low);
}
code.bits = (uint8_t)(len - root_bits);
code.value = (uint16_t)sorted[symbol++];
ReplicateValue(&table[key >> root_bits], step, table_size, code);
key = GetNextKey(key, len);
}
}
// Check if tree is full.
if (num_nodes != 2 * offset[MAX_ALLOWED_CODE_LENGTH] - 1) {
WebPSafeFree(sorted);
return 0;
}
}
WebPSafeFree(sorted);
return total_size;
}
static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
WebPPicture* const picture) {
const int use_filter =
(config->filter_strength > 0) || (config->autofilter > 0);
const int mb_w = (picture->width + 15) >> 4;
const int mb_h = (picture->height + 15) >> 4;
const int preds_w = 4 * mb_w + 1;
const int preds_h = 4 * mb_h + 1;
const size_t preds_size = preds_w * preds_h * sizeof(uint8_t);
const int top_stride = mb_w * 16;
const size_t nz_size = (mb_w + 1) * sizeof(uint32_t);
const size_t cache_size = (3 * YUV_SIZE + PRED_SIZE) * sizeof(uint8_t);
const size_t info_size = mb_w * mb_h * sizeof(VP8MBInfo);
const size_t samples_size = (2 * top_stride + // top-luma/u/v
16 + 16 + 16 + 8 + 1 + // left y/u/v
2 * ALIGN_CST) // align all
* sizeof(uint8_t);
const size_t lf_stats_size =
config->autofilter ? sizeof(LFStats) + ALIGN_CST : 0;
VP8Encoder* enc;
uint8_t* mem;
const uint64_t size = (uint64_t)sizeof(VP8Encoder) // main struct
+ ALIGN_CST // cache alignment
+ cache_size // working caches
+ info_size // modes info
+ preds_size // prediction modes
+ samples_size // top/left samples
+ nz_size // coeff context bits
+ lf_stats_size; // autofilter stats
#ifdef PRINT_MEMORY_INFO
printf("===================================\n");
printf("Memory used:\n"
" encoder: %ld\n"
" block cache: %ld\n"
" info: %ld\n"
" preds: %ld\n"
" top samples: %ld\n"
" non-zero: %ld\n"
" lf-stats: %ld\n"
" total: %ld\n",
sizeof(VP8Encoder) + ALIGN_CST, cache_size, info_size,
preds_size, samples_size, nz_size, lf_stats_size, size);
printf("Transcient object sizes:\n"
" VP8EncIterator: %ld\n"
" VP8ModeScore: %ld\n"
" VP8SegmentInfo: %ld\n"
" VP8Proba: %ld\n"
" LFStats: %ld\n",
sizeof(VP8EncIterator), sizeof(VP8ModeScore),
sizeof(VP8SegmentInfo), sizeof(VP8Proba),
sizeof(LFStats));
printf("Picture size (yuv): %ld\n",
mb_w * mb_h * 384 * sizeof(uint8_t));
printf("===================================\n");
#endif
mem = (uint8_t*)WebPSafeMalloc(size, sizeof(*mem));
if (mem == NULL) {
WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
return NULL;
}
enc = (VP8Encoder*)mem;
mem = (uint8_t*)DO_ALIGN(mem + sizeof(*enc));
memset(enc, 0, sizeof(*enc));
enc->num_parts_ = 1 << config->partitions;
enc->mb_w_ = mb_w;
enc->mb_h_ = mb_h;
enc->preds_w_ = preds_w;
enc->yuv_in_ = (uint8_t*)mem;
mem += YUV_SIZE;
enc->yuv_out_ = (uint8_t*)mem;
mem += YUV_SIZE;
enc->yuv_out2_ = (uint8_t*)mem;
mem += YUV_SIZE;
enc->yuv_p_ = (uint8_t*)mem;
mem += PRED_SIZE;
enc->mb_info_ = (VP8MBInfo*)mem;
mem += info_size;
enc->preds_ = ((uint8_t*)mem) + 1 + enc->preds_w_;
mem += preds_w * preds_h * sizeof(uint8_t);
enc->nz_ = 1 + (uint32_t*)mem;
mem += nz_size;
enc->lf_stats_ = lf_stats_size ? (LFStats*)DO_ALIGN(mem) : NULL;
mem += lf_stats_size;
// top samples (all 16-aligned)
mem = (uint8_t*)DO_ALIGN(mem);
enc->y_top_ = (uint8_t*)mem;
enc->uv_top_ = enc->y_top_ + top_stride;
mem += 2 * top_stride;
mem = (uint8_t*)DO_ALIGN(mem + 1);
enc->y_left_ = (uint8_t*)mem;
mem += 16 + 16;
enc->u_left_ = (uint8_t*)mem;
mem += 16;
enc->v_left_ = (uint8_t*)mem;
mem += 8;
enc->config_ = config;
enc->profile_ = use_filter ? ((config->filter_type == 1) ? 0 : 1) : 2;
enc->pic_ = picture;
enc->percent_ = 0;
MapConfigToTools(enc);
VP8EncDspInit();
VP8DefaultProbas(enc);
ResetSegmentHeader(enc);
ResetFilterHeader(enc);
ResetBoundaryPredictions(enc);
VP8EncInitAlpha(enc);
#ifdef WEBP_EXPERIMENTAL_FEATURES
VP8EncInitLayer(enc);
#endif
VP8TBufferInit(&enc->tokens_);
return enc;
}
// Create an optimal Huffman tree.
//
// (data,length): population counts.
// tree_limit: maximum bit depth (inclusive) of the codes.
// bit_depths[]: how many bits are used for the symbol.
//
// Returns 0 when an error has occurred.
//
// The catch here is that the tree cannot be arbitrarily deep
//
// count_limit is the value that is to be faked as the minimum value
// and this minimum value is raised until the tree matches the
// maximum length requirement.
//
// This algorithm is not of excellent performance for very long data blocks,
// especially when population counts are longer than 2**tree_limit, but
// we are not planning to use this with extremely long blocks.
//
// See http://en.wikipedia.org/wiki/Huffman_coding
static int GenerateOptimalTree(const int* const histogram, int histogram_size,
int tree_depth_limit,
uint8_t* const bit_depths) {
int count_min;
HuffmanTree* tree_pool;
HuffmanTree* tree;
int tree_size_orig = 0;
int i;
for (i = 0; i < histogram_size; ++i) {
if (histogram[i] != 0) {
++tree_size_orig;
}
}
if (tree_size_orig == 0) { // pretty optimal already!
return 1;
}
// 3 * tree_size is enough to cover all the nodes representing a
// population and all the inserted nodes combining two existing nodes.
// The tree pool needs 2 * (tree_size_orig - 1) entities, and the
// tree needs exactly tree_size_orig entities.
tree = (HuffmanTree*)WebPSafeMalloc(3ULL * tree_size_orig, sizeof(*tree));
if (tree == NULL) return 0;
tree_pool = tree + tree_size_orig;
// For block sizes with less than 64k symbols we never need to do a
// second iteration of this loop.
// If we actually start running inside this loop a lot, we would perhaps
// be better off with the Katajainen algorithm.
assert(tree_size_orig <= (1 << (tree_depth_limit - 1)));
for (count_min = 1; ; count_min *= 2) {
int tree_size = tree_size_orig;
// We need to pack the Huffman tree in tree_depth_limit bits.
// So, we try by faking histogram entries to be at least 'count_min'.
int idx = 0;
int j;
for (j = 0; j < histogram_size; ++j) {
if (histogram[j] != 0) {
const int count =
(histogram[j] < count_min) ? count_min : histogram[j];
tree[idx].total_count_ = count;
tree[idx].value_ = j;
tree[idx].pool_index_left_ = -1;
tree[idx].pool_index_right_ = -1;
++idx;
}
}
// Build the Huffman tree.
qsort(tree, tree_size, sizeof(*tree), CompareHuffmanTrees);
if (tree_size > 1) { // Normal case.
int tree_pool_size = 0;
while (tree_size > 1) { // Finish when we have only one root.
int count;
tree_pool[tree_pool_size++] = tree[tree_size - 1];
tree_pool[tree_pool_size++] = tree[tree_size - 2];
count = tree_pool[tree_pool_size - 1].total_count_ +
tree_pool[tree_pool_size - 2].total_count_;
tree_size -= 2;
{
// Search for the insertion point.
int k;
for (k = 0; k < tree_size; ++k) {
if (tree[k].total_count_ <= count) {
break;
}
}
memmove(tree + (k + 1), tree + k, (tree_size - k) * sizeof(*tree));
tree[k].total_count_ = count;
tree[k].value_ = -1;
tree[k].pool_index_left_ = tree_pool_size - 1;
tree[k].pool_index_right_ = tree_pool_size - 2;
tree_size = tree_size + 1;
}
}
SetBitDepths(&tree[0], tree_pool, bit_depths, 0);
} else if (tree_size == 1) { // Trivial case: only one element.
bit_depths[tree[0].value_] = 1;
}
{
// Test if this Huffman tree satisfies our 'tree_depth_limit' criteria.
int max_depth = bit_depths[0];
for (j = 1; j < histogram_size; ++j) {
if (max_depth < bit_depths[j]) {
max_depth = bit_depths[j];
}
}
if (max_depth <= tree_depth_limit) {
break;
}
}
}
free(tree);
return 1;
}
WebPMux* WebPMuxCreateInternal(const WebPData* bitstream, int copy_data,
int version) {
size_t riff_size;
uint32_t tag;
const uint8_t* end;
WebPMux* mux = NULL;
WebPMuxImage* wpi = NULL;
const uint8_t* data;
size_t size;
WebPChunk chunk;
ChunkInit(&chunk);
// Sanity checks.
if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_MUX_ABI_VERSION)) {
return NULL; // version mismatch
}
if (bitstream == NULL) return NULL;
data = bitstream->bytes;
size = bitstream->size;
if (data == NULL) return NULL;
if (size < RIFF_HEADER_SIZE) return NULL;
if (GetLE32(data + 0) != MKFOURCC('R', 'I', 'F', 'F') ||
GetLE32(data + CHUNK_HEADER_SIZE) != MKFOURCC('W', 'E', 'B', 'P')) {
return NULL;
}
mux = WebPMuxNew();
if (mux == NULL) return NULL;
if (size < RIFF_HEADER_SIZE + TAG_SIZE) goto Err;
tag = GetLE32(data + RIFF_HEADER_SIZE);
if (tag != kChunks[IDX_VP8].tag &&
tag != kChunks[IDX_VP8L].tag &&
tag != kChunks[IDX_VP8X].tag) {
goto Err; // First chunk should be VP8, VP8L or VP8X.
}
riff_size = SizeWithPadding(GetLE32(data + TAG_SIZE));
if (riff_size > MAX_CHUNK_PAYLOAD || riff_size > size) {
goto Err;
} else {
if (riff_size < size) { // Redundant data after last chunk.
size = riff_size; // To make sure we don't read any data beyond mux_size.
}
}
end = data + size;
data += RIFF_HEADER_SIZE;
size -= RIFF_HEADER_SIZE;
wpi = (WebPMuxImage*)WebPSafeMalloc(1ULL, sizeof(*wpi));
if (wpi == NULL) goto Err;
MuxImageInit(wpi);
// Loop over chunks.
while (data != end) {
size_t data_size;
WebPChunkId id;
WebPChunk** chunk_list;
if (ChunkVerifyAndAssign(&chunk, data, size, riff_size,
copy_data) != WEBP_MUX_OK) {
goto Err;
}
data_size = ChunkDiskSize(&chunk);
id = ChunkGetIdFromTag(chunk.tag_);
switch (id) {
case WEBP_CHUNK_ALPHA:
if (wpi->alpha_ != NULL) goto Err; // Consecutive ALPH chunks.
if (ChunkSetNth(&chunk, &wpi->alpha_, 1) != WEBP_MUX_OK) goto Err;
wpi->is_partial_ = 1; // Waiting for a VP8 chunk.
break;
case WEBP_CHUNK_IMAGE:
if (ChunkSetNth(&chunk, &wpi->img_, 1) != WEBP_MUX_OK) goto Err;
if (!MuxImageFinalize(wpi)) goto Err;
wpi->is_partial_ = 0; // wpi is completely filled.
PushImage:
// Add this to mux->images_ list.
if (MuxImagePush(wpi, &mux->images_) != WEBP_MUX_OK) goto Err;
MuxImageInit(wpi); // Reset for reading next image.
break;
case WEBP_CHUNK_ANMF:
if (wpi->is_partial_) goto Err; // Previous wpi is still incomplete.
if (!MuxImageParse(&chunk, copy_data, wpi)) goto Err;
ChunkRelease(&chunk);
goto PushImage;
break;
default: // A non-image chunk.
if (wpi->is_partial_) goto Err; // Encountered a non-image chunk before
// getting all chunks of an image.
chunk_list = MuxGetChunkListFromId(mux, id); // List to add this chunk.
if (ChunkSetNth(&chunk, chunk_list, 0) != WEBP_MUX_OK) goto Err;
if (id == WEBP_CHUNK_VP8X) { // grab global specs
mux->canvas_width_ = GetLE24(data + 12) + 1;
mux->canvas_height_ = GetLE24(data + 15) + 1;
}
break;
}
data += data_size;
size -= data_size;
ChunkInit(&chunk);
}
// Validate mux if complete.
if (MuxValidate(mux) != WEBP_MUX_OK) goto Err;
MuxImageDelete(wpi);
return mux; // All OK;
Err: // Something bad happened.
ChunkRelease(&chunk);
MuxImageDelete(wpi);
WebPMuxDelete(mux);
return NULL;
}
static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
WebPPicture* const picture) {
const int use_filter =
(config->filter_strength > 0) || (config->autofilter > 0);
const int mb_w = (picture->width + 15) >> 4;
const int mb_h = (picture->height + 15) >> 4;
const int preds_w = 4 * mb_w + 1;
const int preds_h = 4 * mb_h + 1;
const size_t preds_size = preds_w * preds_h * sizeof(uint8_t);
const int top_stride = mb_w * 16;
const size_t nz_size = (mb_w + 1) * sizeof(uint32_t) + ALIGN_CST;
const size_t info_size = mb_w * mb_h * sizeof(VP8MBInfo);
const size_t samples_size = 2 * top_stride * sizeof(uint8_t) // top-luma/u/v
+ ALIGN_CST; // align all
const size_t lf_stats_size =
config->autofilter ? sizeof(LFStats) + ALIGN_CST : 0;
VP8Encoder* enc;
uint8_t* mem;
const uint64_t size = (uint64_t)sizeof(VP8Encoder) // main struct
+ ALIGN_CST // cache alignment
+ info_size // modes info
+ preds_size // prediction modes
+ samples_size // top/left samples
+ nz_size // coeff context bits
+ lf_stats_size; // autofilter stats
#ifdef PRINT_MEMORY_INFO
printf("===================================\n");
printf("Memory used:\n"
" encoder: %ld\n"
" info: %ld\n"
" preds: %ld\n"
" top samples: %ld\n"
" non-zero: %ld\n"
" lf-stats: %ld\n"
" total: %ld\n",
sizeof(VP8Encoder) + ALIGN_CST, info_size,
preds_size, samples_size, nz_size, lf_stats_size, size);
printf("Transient object sizes:\n"
" VP8EncIterator: %ld\n"
" VP8ModeScore: %ld\n"
" VP8SegmentInfo: %ld\n"
" VP8Proba: %ld\n"
" LFStats: %ld\n",
sizeof(VP8EncIterator), sizeof(VP8ModeScore),
sizeof(VP8SegmentInfo), sizeof(VP8Proba),
sizeof(LFStats));
printf("Picture size (yuv): %ld\n",
mb_w * mb_h * 384 * sizeof(uint8_t));
printf("===================================\n");
#endif
mem = (uint8_t*)WebPSafeMalloc(size, sizeof(*mem));
if (mem == NULL) {
WebPEncodingSetError(picture, VP8_ENC_ERROR_OUT_OF_MEMORY);
return NULL;
}
enc = (VP8Encoder*)mem;
mem = (uint8_t*)DO_ALIGN(mem + sizeof(*enc));
memset(enc, 0, sizeof(*enc));
enc->num_parts_ = 1 << config->partitions;
enc->mb_w_ = mb_w;
enc->mb_h_ = mb_h;
enc->preds_w_ = preds_w;
enc->mb_info_ = (VP8MBInfo*)mem;
mem += info_size;
enc->preds_ = ((uint8_t*)mem) + 1 + enc->preds_w_;
mem += preds_w * preds_h * sizeof(uint8_t);
enc->nz_ = 1 + (uint32_t*)DO_ALIGN(mem);
mem += nz_size;
enc->lf_stats_ = lf_stats_size ? (LFStats*)DO_ALIGN(mem) : NULL;
mem += lf_stats_size;
// top samples (all 16-aligned)
mem = (uint8_t*)DO_ALIGN(mem);
enc->y_top_ = (uint8_t*)mem;
enc->uv_top_ = enc->y_top_ + top_stride;
mem += 2 * top_stride;
assert(mem <= (uint8_t*)enc + size);
enc->config_ = config;
enc->profile_ = use_filter ? ((config->filter_type == 1) ? 0 : 1) : 2;
enc->pic_ = picture;
enc->percent_ = 0;
MapConfigToTools(enc);
VP8EncDspInit();
VP8DefaultProbas(enc);
ResetSegmentHeader(enc);
ResetFilterHeader(enc);
ResetBoundaryPredictions(enc);
VP8GetResidualCostInit();
VP8SetResidualCoeffsInit();
VP8EncInitAlpha(enc);
// lower quality means smaller output -> we modulate a little the page
// size based on quality. This is just a crude 1rst-order prediction.
{
const float scale = 1.f + config->quality * 5.f / 100.f; // in [1,6]
VP8TBufferInit(&enc->tokens_, (int)(mb_w * mb_h * 4 * scale));
}
return enc;
}
static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
const int has_alpha = WebPIsAlphaMode(p->output->colorspace);
const int out_width = io->scaled_width;
const int out_height = io->scaled_height;
const int uv_in_width = (io->mb_w + 1) >> 1;
const int uv_in_height = (io->mb_h + 1) >> 1;
const size_t work_size = 2 * out_width; // scratch memory for one rescaler
rescaler_t* work; // rescalers work area
uint8_t* tmp; // tmp storage for scaled YUV444 samples before RGB conversion
size_t tmp_size1, tmp_size2, total_size, rescaler_size;
WebPRescaler* scalers;
const int num_rescalers = has_alpha ? 4 : 3;
tmp_size1 = 3 * work_size;
tmp_size2 = 3 * out_width;
if (has_alpha) {
tmp_size1 += work_size;
tmp_size2 += out_width;
}
total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
p->memory = WebPSafeMalloc(1ULL, total_size + rescaler_size);
if (p->memory == NULL) {
return 0; // memory error
}
work = (rescaler_t*)p->memory;
tmp = (uint8_t*)(work + tmp_size1);
scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + total_size);
p->scaler_y = &scalers[0];
p->scaler_u = &scalers[1];
p->scaler_v = &scalers[2];
p->scaler_a = has_alpha ? &scalers[3] : NULL;
WebPRescalerInit(p->scaler_y, io->mb_w, io->mb_h,
tmp + 0 * out_width, out_width, out_height, 0, 1,
work + 0 * work_size);
WebPRescalerInit(p->scaler_u, uv_in_width, uv_in_height,
tmp + 1 * out_width, out_width, out_height, 0, 1,
work + 1 * work_size);
WebPRescalerInit(p->scaler_v, uv_in_width, uv_in_height,
tmp + 2 * out_width, out_width, out_height, 0, 1,
work + 2 * work_size);
p->emit = EmitRescaledRGB;
WebPInitYUV444Converters();
if (has_alpha) {
WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
tmp + 3 * out_width, out_width, out_height, 0, 1,
work + 3 * work_size);
p->emit_alpha = EmitRescaledAlphaRGB;
if (p->output->colorspace == MODE_RGBA_4444 ||
p->output->colorspace == MODE_rgbA_4444) {
p->emit_alpha_row = ExportAlphaRGBA4444;
} else {
p->emit_alpha_row = ExportAlpha;
}
WebPInitAlphaProcessing();
}
return 1;
}
