Merge "update to v1.2.1" am: 84cdd3fa7d am: 0dd1133823 am: bc8db48636
Original change: https://android-review.googlesource.com/c/platform/external/webp/+/1780754
Change-Id: Id16c84da7bf7275b9529981958c9e1c214471502
diff --git a/Android.bp b/Android.bp
index 4889d18..e81468f 100644
--- a/Android.bp
+++ b/Android.bp
@@ -184,6 +184,7 @@
"src/dsp/lossless_msa.c",
"src/dsp/lossless_neon.c",
"src/dsp/lossless_sse2.c",
+ "src/dsp/lossless_sse41.c",
"src/dsp/rescaler.c",
"src/dsp/rescaler_mips32.c",
"src/dsp/rescaler_mips_dsp_r2.c",
diff --git a/README b/README
index bbe2c81..54c5175 100644
--- a/README
+++ b/README
@@ -4,7 +4,7 @@
\__\__/\____/\_____/__/ ____ ___
/ _/ / \ \ / _ \/ _/
/ \_/ / / \ \ __/ \__
- \____/____/\_____/_____/____/v1.2.0
+ \____/____/\_____/_____/____/v1.2.1
Description:
============
@@ -619,7 +619,7 @@
pic.width = width;
pic.height = height;
// allocated picture of dimension width x height
- if (!WebPPictureAllocate(&pic)) {
+ if (!WebPPictureAlloc(&pic)) {
return 0; // memory error
}
// at this point, 'pic' has been initialized as a container,
diff --git a/README.android b/README.android
index eb8a65e..3dd2d86 100644
--- a/README.android
+++ b/README.android
@@ -1,5 +1,5 @@
URL: https://chromium.googlesource.com/webm/libwebp
-Version: v1.2.0
+Version: v1.2.1
License: Google BSD like
Local modifications:
diff --git a/README.version b/README.version
index fa87b87..96c43aa 100644
--- a/README.version
+++ b/README.version
@@ -1,3 +1,3 @@
-URL: https://chromium.googlesource.com/webm/libwebp/+archive/v1.2.0.tar.gz
-Version: v1.2.0
+URL: https://chromium.googlesource.com/webm/libwebp/+archive/v1.2.1.tar.gz
+Version: v1.2.1
BugComponent: 20174
diff --git a/src/dec/alpha_dec.c b/src/dec/alpha_dec.c
index bce735b..0b93a30 100644
--- a/src/dec/alpha_dec.c
+++ b/src/dec/alpha_dec.c
@@ -183,7 +183,7 @@
assert(dec != NULL && io != NULL);
if (row < 0 || num_rows <= 0 || row + num_rows > height) {
- return NULL; // sanity check.
+ return NULL;
}
if (!dec->is_alpha_decoded_) {
diff --git a/src/dec/buffer_dec.c b/src/dec/buffer_dec.c
index 3cd94eb..4786cf0 100644
--- a/src/dec/buffer_dec.c
+++ b/src/dec/buffer_dec.c
@@ -102,7 +102,7 @@
int stride;
uint64_t size;
- if ((uint64_t)w * kModeBpp[mode] >= (1ull << 32)) {
+ if ((uint64_t)w * kModeBpp[mode] >= (1ull << 31)) {
return VP8_STATUS_INVALID_PARAM;
}
stride = w * kModeBpp[mode];
@@ -117,7 +117,6 @@
}
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;
@@ -156,11 +155,11 @@
}
if (WebPIsRGBMode(buffer->colorspace)) {
WebPRGBABuffer* const buf = &buffer->u.RGBA;
- buf->rgba += (buffer->height - 1) * buf->stride;
+ buf->rgba += (int64_t)(buffer->height - 1) * buf->stride;
buf->stride = -buf->stride;
} else {
WebPYUVABuffer* const buf = &buffer->u.YUVA;
- const int H = buffer->height;
+ const int64_t H = buffer->height;
buf->y += (H - 1) * buf->y_stride;
buf->y_stride = -buf->y_stride;
buf->u += ((H - 1) >> 1) * buf->u_stride;
@@ -188,8 +187,7 @@
const int ch = options->crop_height;
const int x = options->crop_left & ~1;
const int y = options->crop_top & ~1;
- if (x < 0 || y < 0 || cw <= 0 || ch <= 0 ||
- x + cw > width || y + ch > height) {
+ if (!WebPCheckCropDimensions(width, height, x, y, cw, ch)) {
return VP8_STATUS_INVALID_PARAM; // out of frame boundary.
}
width = cw;
diff --git a/src/dec/frame_dec.c b/src/dec/frame_dec.c
index 04609a8..91ca1f8 100644
--- a/src/dec/frame_dec.c
+++ b/src/dec/frame_dec.c
@@ -705,7 +705,7 @@
+ cache_size + alpha_size + WEBP_ALIGN_CST;
uint8_t* mem;
- if (needed != (size_t)needed) return 0; // check for overflow
+ if (!CheckSizeOverflow(needed)) return 0; // check for overflow
if (needed > dec->mem_size_) {
WebPSafeFree(dec->mem_);
dec->mem_size_ = 0;
diff --git a/src/dec/io_dec.c b/src/dec/io_dec.c
index 29dc634..5ef6298 100644
--- a/src/dec/io_dec.c
+++ b/src/dec/io_dec.c
@@ -298,46 +298,57 @@
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
+ // scratch memory for luma rescaler
+ const size_t work_size = 2 * (size_t)out_width;
const size_t uv_work_size = 2 * uv_out_width; // and for each u/v ones
- size_t tmp_size, rescaler_size;
+ uint64_t total_size;
+ size_t 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);
+ total_size = ((uint64_t)work_size + 2 * uv_work_size) * sizeof(*work);
if (has_alpha) {
- tmp_size += work_size * sizeof(*work);
+ total_size += (uint64_t)work_size * sizeof(*work);
}
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
+ total_size += rescaler_size;
+ if (!CheckSizeOverflow(total_size)) {
+ return 0;
+ }
- p->memory = WebPSafeMalloc(1ULL, tmp_size + rescaler_size);
+ p->memory = WebPSafeMalloc(1ULL, (size_t)total_size);
if (p->memory == NULL) {
return 0; // memory error
}
work = (rescaler_t*)p->memory;
- scalers = (WebPRescaler*)WEBP_ALIGN((const uint8_t*)work + tmp_size);
+ scalers = (WebPRescaler*)WEBP_ALIGN(
+ (const uint8_t*)work + total_size - rescaler_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);
+ if (!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)) {
+ return 0;
+ }
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);
+ if (!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)) {
+ return 0;
+ }
p->emit_alpha = EmitRescaledAlphaYUV;
WebPInitAlphaProcessing();
}
@@ -480,51 +491,58 @@
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
+ // scratch memory for one rescaler
+ const size_t work_size = 2 * (size_t)out_width;
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;
+ uint64_t tmp_size1, tmp_size2, total_size;
+ size_t 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;
- }
+ tmp_size1 = (uint64_t)num_rescalers * work_size;
+ tmp_size2 = (uint64_t)num_rescalers * out_width;
total_size = tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp);
rescaler_size = num_rescalers * sizeof(*p->scaler_y) + WEBP_ALIGN_CST;
+ total_size += rescaler_size;
+ if (!CheckSizeOverflow(total_size)) {
+ return 0;
+ }
- p->memory = WebPSafeMalloc(1ULL, total_size + rescaler_size);
+ p->memory = WebPSafeMalloc(1ULL, (size_t)total_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);
+ scalers = (WebPRescaler*)WEBP_ALIGN(
+ (const uint8_t*)work + total_size - rescaler_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);
+ if (!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)) {
+ return 0;
+ }
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);
+ if (!WebPRescalerInit(p->scaler_a, io->mb_w, io->mb_h,
+ tmp + 3 * out_width, out_width, out_height, 0, 1,
+ work + 3 * work_size)) {
+ return 0;
+ }
p->emit_alpha = EmitRescaledAlphaRGB;
if (p->output->colorspace == MODE_RGBA_4444 ||
p->output->colorspace == MODE_rgbA_4444) {
diff --git a/src/dec/vp8_dec.c b/src/dec/vp8_dec.c
index 8f73697..5f405e4 100644
--- a/src/dec/vp8_dec.c
+++ b/src/dec/vp8_dec.c
@@ -335,7 +335,7 @@
io->scaled_width = io->width;
io->scaled_height = io->height;
- io->mb_w = io->width; // sanity check
+ io->mb_w = io->width; // for soundness
io->mb_h = io->height; // ditto
VP8ResetProba(&dec->proba_);
diff --git a/src/dec/vp8i_dec.h b/src/dec/vp8i_dec.h
index a0c0af1..20526a8 100644
--- a/src/dec/vp8i_dec.h
+++ b/src/dec/vp8i_dec.h
@@ -32,7 +32,7 @@
// version numbers
#define DEC_MAJ_VERSION 1
#define DEC_MIN_VERSION 2
-#define DEC_REV_VERSION 0
+#define DEC_REV_VERSION 1
// YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
// Constraints are: We need to store one 16x16 block of luma samples (y),
diff --git a/src/dec/vp8l_dec.c b/src/dec/vp8l_dec.c
index 2d603b4..73c3b54 100644
--- a/src/dec/vp8l_dec.c
+++ b/src/dec/vp8l_dec.c
@@ -559,8 +559,11 @@
memory += work_size * sizeof(*work);
scaled_data = (uint32_t*)memory;
- WebPRescalerInit(dec->rescaler, in_width, in_height, (uint8_t*)scaled_data,
- out_width, out_height, 0, num_channels, work);
+ if (!WebPRescalerInit(dec->rescaler, in_width, in_height,
+ (uint8_t*)scaled_data, out_width, out_height,
+ 0, num_channels, work)) {
+ return 0;
+ }
return 1;
}
#endif // WEBP_REDUCE_SIZE
@@ -574,13 +577,14 @@
static int Export(WebPRescaler* const rescaler, WEBP_CSP_MODE colorspace,
int rgba_stride, uint8_t* const rgba) {
uint32_t* const src = (uint32_t*)rescaler->dst;
+ uint8_t* dst = rgba;
const int dst_width = rescaler->dst_width;
int num_lines_out = 0;
while (WebPRescalerHasPendingOutput(rescaler)) {
- uint8_t* const dst = rgba + num_lines_out * rgba_stride;
WebPRescalerExportRow(rescaler);
WebPMultARGBRow(src, dst_width, 1);
VP8LConvertFromBGRA(src, dst_width, colorspace, dst);
+ dst += rgba_stride;
++num_lines_out;
}
return num_lines_out;
@@ -594,8 +598,8 @@
int num_lines_in = 0;
int num_lines_out = 0;
while (num_lines_in < mb_h) {
- uint8_t* const row_in = in + num_lines_in * in_stride;
- uint8_t* const row_out = out + num_lines_out * out_stride;
+ uint8_t* const row_in = in + (uint64_t)num_lines_in * in_stride;
+ uint8_t* const row_out = out + (uint64_t)num_lines_out * out_stride;
const int lines_left = mb_h - num_lines_in;
const int needed_lines = WebPRescaleNeededLines(dec->rescaler, lines_left);
int lines_imported;
@@ -796,7 +800,8 @@
const WebPDecBuffer* const output = dec->output_;
if (WebPIsRGBMode(output->colorspace)) { // convert to RGBA
const WebPRGBABuffer* const buf = &output->u.RGBA;
- uint8_t* const rgba = buf->rgba + dec->last_out_row_ * buf->stride;
+ uint8_t* const rgba =
+ buf->rgba + (int64_t)dec->last_out_row_ * buf->stride;
const int num_rows_out =
#if !defined(WEBP_REDUCE_SIZE)
io->use_scaling ?
@@ -1514,7 +1519,7 @@
assert(dec->width_ <= final_width);
dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint32_t));
if (dec->pixels_ == NULL) {
- dec->argb_cache_ = NULL; // for sanity check
+ dec->argb_cache_ = NULL; // for soundness
dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
return 0;
}
@@ -1524,7 +1529,7 @@
static int AllocateInternalBuffers8b(VP8LDecoder* const dec) {
const uint64_t total_num_pixels = (uint64_t)dec->width_ * dec->height_;
- dec->argb_cache_ = NULL; // for sanity check
+ dec->argb_cache_ = NULL; // for soundness
dec->pixels_ = (uint32_t*)WebPSafeMalloc(total_num_pixels, sizeof(uint8_t));
if (dec->pixels_ == NULL) {
dec->status_ = VP8_STATUS_OUT_OF_MEMORY;
@@ -1666,7 +1671,6 @@
VP8Io* io = NULL;
WebPDecParams* params = NULL;
- // Sanity checks.
if (dec == NULL) return 0;
assert(dec->hdr_.huffman_tables_ != NULL);
diff --git a/src/dec/webp_dec.c b/src/dec/webp_dec.c
index 42d0988..77a54c5 100644
--- a/src/dec/webp_dec.c
+++ b/src/dec/webp_dec.c
@@ -785,6 +785,13 @@
//------------------------------------------------------------------------------
// Cropping and rescaling.
+int WebPCheckCropDimensions(int image_width, int image_height,
+ int x, int y, int w, int h) {
+ return !(x < 0 || y < 0 || w <= 0 || h <= 0 ||
+ x >= image_width || w > image_width || w > image_width - x ||
+ y >= image_height || h > image_height || h > image_height - y);
+}
+
int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
VP8Io* const io, WEBP_CSP_MODE src_colorspace) {
const int W = io->width;
@@ -792,7 +799,7 @@
int x = 0, y = 0, w = W, h = H;
// Cropping
- io->use_cropping = (options != NULL) && (options->use_cropping > 0);
+ io->use_cropping = (options != NULL) && options->use_cropping;
if (io->use_cropping) {
w = options->crop_width;
h = options->crop_height;
@@ -802,7 +809,7 @@
x &= ~1;
y &= ~1;
}
- if (x < 0 || y < 0 || w <= 0 || h <= 0 || x + w > W || y + h > H) {
+ if (!WebPCheckCropDimensions(W, H, x, y, w, h)) {
return 0; // out of frame boundary error
}
}
@@ -814,7 +821,7 @@
io->mb_h = h;
// Scaling
- io->use_scaling = (options != NULL) && (options->use_scaling > 0);
+ io->use_scaling = (options != NULL) && options->use_scaling;
if (io->use_scaling) {
int scaled_width = options->scaled_width;
int scaled_height = options->scaled_height;
@@ -835,8 +842,8 @@
if (io->use_scaling) {
// disable filter (only for large downscaling ratio).
- io->bypass_filtering = (io->scaled_width < W * 3 / 4) &&
- (io->scaled_height < H * 3 / 4);
+ io->bypass_filtering |= (io->scaled_width < W * 3 / 4) &&
+ (io->scaled_height < H * 3 / 4);
io->fancy_upsampling = 0;
}
return 1;
diff --git a/src/dec/webpi_dec.h b/src/dec/webpi_dec.h
index 24baff5..3b97388 100644
--- a/src/dec/webpi_dec.h
+++ b/src/dec/webpi_dec.h
@@ -77,6 +77,10 @@
//------------------------------------------------------------------------------
// Misc utils
+// Returns true if crop dimensions are within image bounds.
+int WebPCheckCropDimensions(int image_width, int image_height,
+ int x, int y, int w, int h);
+
// Initializes VP8Io with custom setup, io and teardown functions. The default
// hooks will use the supplied 'params' as io->opaque handle.
void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io);
diff --git a/src/demux/anim_decode.c b/src/demux/anim_decode.c
index 3dcacc3..2bf4dcf 100644
--- a/src/demux/anim_decode.c
+++ b/src/demux/anim_decode.c
@@ -87,11 +87,19 @@
int abi_version) {
WebPAnimDecoderOptions options;
WebPAnimDecoder* dec = NULL;
+ WebPBitstreamFeatures features;
if (webp_data == NULL ||
WEBP_ABI_IS_INCOMPATIBLE(abi_version, WEBP_DEMUX_ABI_VERSION)) {
return NULL;
}
+ // Validate the bitstream before doing expensive allocations. The demuxer may
+ // be more tolerant than the decoder.
+ if (WebPGetFeatures(webp_data->bytes, webp_data->size, &features) !=
+ VP8_STATUS_OK) {
+ return NULL;
+ }
+
// Note: calloc() so that the pointer members are initialized to NULL.
dec = (WebPAnimDecoder*)WebPSafeCalloc(1ULL, sizeof(*dec));
if (dec == NULL) goto Error;
@@ -145,7 +153,7 @@
uint32_t canvas_height) {
const uint64_t size =
(uint64_t)canvas_width * canvas_height * NUM_CHANNELS * sizeof(*buf);
- if (size != (size_t)size) return 0;
+ if (!CheckSizeOverflow(size)) return 0;
memset(buf, 0, (size_t)size);
return 1;
}
@@ -166,7 +174,7 @@
static int CopyCanvas(const uint8_t* src, uint8_t* dst,
uint32_t width, uint32_t height) {
const uint64_t size = (uint64_t)width * height * NUM_CHANNELS;
- if (size != (size_t)size) return 0;
+ if (!CheckSizeOverflow(size)) return 0;
assert(src != NULL && dst != NULL);
memcpy(dst, src, (size_t)size);
return 1;
diff --git a/src/demux/demux.c b/src/demux/demux.c
index 860e2ce..547a772 100644
--- a/src/demux/demux.c
+++ b/src/demux/demux.c
@@ -25,7 +25,7 @@
#define DMUX_MAJ_VERSION 1
#define DMUX_MIN_VERSION 2
-#define DMUX_REV_VERSION 0
+#define DMUX_REV_VERSION 1
typedef struct {
size_t start_; // start location of the data
@@ -221,12 +221,16 @@
const size_t chunk_start_offset = mem->start_;
const uint32_t fourcc = ReadLE32(mem);
const uint32_t payload_size = ReadLE32(mem);
- const uint32_t payload_size_padded = payload_size + (payload_size & 1);
- const size_t payload_available = (payload_size_padded > MemDataSize(mem))
- ? MemDataSize(mem) : payload_size_padded;
- const size_t chunk_size = CHUNK_HEADER_SIZE + payload_available;
+ uint32_t payload_size_padded;
+ size_t payload_available;
+ size_t chunk_size;
if (payload_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+
+ payload_size_padded = payload_size + (payload_size & 1);
+ payload_available = (payload_size_padded > MemDataSize(mem))
+ ? MemDataSize(mem) : payload_size_padded;
+ chunk_size = CHUNK_HEADER_SIZE + payload_available;
if (SizeIsInvalid(mem, payload_size_padded)) return PARSE_ERROR;
if (payload_size_padded > MemDataSize(mem)) status = PARSE_NEED_MORE_DATA;
@@ -451,9 +455,11 @@
const size_t chunk_start_offset = mem->start_;
const uint32_t fourcc = ReadLE32(mem);
const uint32_t chunk_size = ReadLE32(mem);
- const uint32_t chunk_size_padded = chunk_size + (chunk_size & 1);
+ uint32_t chunk_size_padded;
if (chunk_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+
+ chunk_size_padded = chunk_size + (chunk_size & 1);
if (SizeIsInvalid(mem, chunk_size_padded)) return PARSE_ERROR;
switch (fourcc) {
diff --git a/src/dsp/alpha_processing.c b/src/dsp/alpha_processing.c
index 3a27990..1892929 100644
--- a/src/dsp/alpha_processing.c
+++ b/src/dsp/alpha_processing.c
@@ -157,7 +157,8 @@
}
}
-void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
+void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr,
+ const uint8_t* WEBP_RESTRICT const alpha,
int width, int inverse) {
int x;
for (x = 0; x < width; ++x) {
@@ -178,7 +179,8 @@
#undef MFIX
void (*WebPMultARGBRow)(uint32_t* const ptr, int width, int inverse);
-void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr,
+ const uint8_t* WEBP_RESTRICT const alpha,
int width, int inverse);
//------------------------------------------------------------------------------
@@ -193,8 +195,8 @@
}
}
-void WebPMultRows(uint8_t* ptr, int stride,
- const uint8_t* alpha, int alpha_stride,
+void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride,
+ const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
int width, int num_rows, int inverse) {
int n;
for (n = 0; n < num_rows; ++n) {
@@ -290,9 +292,9 @@
}
#if !WEBP_NEON_OMIT_C_CODE
-static int DispatchAlpha_C(const uint8_t* alpha, int alpha_stride,
+static int DispatchAlpha_C(const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
int width, int height,
- uint8_t* dst, int dst_stride) {
+ uint8_t* WEBP_RESTRICT dst, int dst_stride) {
uint32_t alpha_mask = 0xff;
int i, j;
@@ -309,9 +311,10 @@
return (alpha_mask != 0xff);
}
-static void DispatchAlphaToGreen_C(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_C(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint32_t* WEBP_RESTRICT dst,
+ int dst_stride) {
int i, j;
for (j = 0; j < height; ++j) {
for (i = 0; i < width; ++i) {
@@ -322,9 +325,9 @@
}
}
-static int ExtractAlpha_C(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_C(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
int width, int height,
- uint8_t* alpha, int alpha_stride) {
+ uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
uint8_t alpha_mask = 0xff;
int i, j;
@@ -340,7 +343,8 @@
return (alpha_mask == 0xff);
}
-static void ExtractGreen_C(const uint32_t* argb, uint8_t* alpha, int size) {
+static void ExtractGreen_C(const uint32_t* WEBP_RESTRICT argb,
+ uint8_t* WEBP_RESTRICT alpha, int size) {
int i;
for (i = 0; i < size; ++i) alpha[i] = argb[i] >> 8;
}
@@ -372,8 +376,11 @@
}
#ifdef WORDS_BIGENDIAN
-static void PackARGB_C(const uint8_t* a, const uint8_t* r, const uint8_t* g,
- const uint8_t* b, int len, uint32_t* out) {
+static void PackARGB_C(const uint8_t* WEBP_RESTRICT a,
+ const uint8_t* WEBP_RESTRICT r,
+ const uint8_t* WEBP_RESTRICT g,
+ const uint8_t* WEBP_RESTRICT b,
+ int len, uint32_t* WEBP_RESTRICT out) {
int i;
for (i = 0; i < len; ++i) {
out[i] = MakeARGB32(a[4 * i], r[4 * i], g[4 * i], b[4 * i]);
@@ -381,8 +388,10 @@
}
#endif
-static void PackRGB_C(const uint8_t* r, const uint8_t* g, const uint8_t* b,
- int len, int step, uint32_t* out) {
+static void PackRGB_C(const uint8_t* WEBP_RESTRICT r,
+ const uint8_t* WEBP_RESTRICT g,
+ const uint8_t* WEBP_RESTRICT b,
+ int len, int step, uint32_t* WEBP_RESTRICT out) {
int i, offset = 0;
for (i = 0; i < len; ++i) {
out[i] = MakeARGB32(0xff, r[offset], g[offset], b[offset]);
@@ -392,16 +401,22 @@
void (*WebPApplyAlphaMultiply)(uint8_t*, int, int, int, int);
void (*WebPApplyAlphaMultiply4444)(uint8_t*, int, int, int);
-int (*WebPDispatchAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
-void (*WebPDispatchAlphaToGreen)(const uint8_t*, int, int, int, uint32_t*, int);
-int (*WebPExtractAlpha)(const uint8_t*, int, int, int, uint8_t*, int);
-void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int,
+ uint8_t* WEBP_RESTRICT, int);
+void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT, int, int, int,
+ uint32_t* WEBP_RESTRICT, int);
+int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT, int, int, int,
+ uint8_t* WEBP_RESTRICT, int);
+void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb,
+ uint8_t* WEBP_RESTRICT alpha, int size);
#ifdef WORDS_BIGENDIAN
void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r, const uint8_t* g,
const uint8_t* b, int, uint32_t*);
#endif
-void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
- int len, int step, uint32_t* out);
+void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r,
+ const uint8_t* WEBP_RESTRICT g,
+ const uint8_t* WEBP_RESTRICT b,
+ int len, int step, uint32_t* WEBP_RESTRICT out);
int (*WebPHasAlpha8b)(const uint8_t* src, int length);
int (*WebPHasAlpha32b)(const uint8_t* src, int length);
@@ -438,10 +453,10 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitAlphaProcessingSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
WebPInitAlphaProcessingSSE41();
}
@@ -455,7 +470,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
WebPInitAlphaProcessingNEON();
diff --git a/src/dsp/alpha_processing_neon.c b/src/dsp/alpha_processing_neon.c
index 9d55421..9e0ace9 100644
--- a/src/dsp/alpha_processing_neon.c
+++ b/src/dsp/alpha_processing_neon.c
@@ -80,9 +80,9 @@
//------------------------------------------------------------------------------
-static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint8_t* dst, int dst_stride) {
+static int DispatchAlpha_NEON(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint8_t* WEBP_RESTRICT dst, int dst_stride) {
uint32_t alpha_mask = 0xffffffffu;
uint8x8_t mask8 = vdup_n_u8(0xff);
uint32_t tmp[2];
@@ -112,9 +112,10 @@
return (alpha_mask != 0xffffffffu);
}
-static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_NEON(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint32_t* WEBP_RESTRICT dst,
+ int dst_stride) {
int i, j;
uint8x8x4_t greens; // leave A/R/B channels zero'd.
greens.val[0] = vdup_n_u8(0);
@@ -131,9 +132,9 @@
}
}
-static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_NEON(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
int width, int height,
- uint8_t* alpha, int alpha_stride) {
+ uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
uint32_t alpha_mask = 0xffffffffu;
uint8x8_t mask8 = vdup_n_u8(0xff);
uint32_t tmp[2];
@@ -161,8 +162,8 @@
return (alpha_mask == 0xffffffffu);
}
-static void ExtractGreen_NEON(const uint32_t* argb,
- uint8_t* alpha, int size) {
+static void ExtractGreen_NEON(const uint32_t* WEBP_RESTRICT argb,
+ uint8_t* WEBP_RESTRICT alpha, int size) {
int i;
for (i = 0; i + 16 <= size; i += 16) {
const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i));
diff --git a/src/dsp/alpha_processing_sse2.c b/src/dsp/alpha_processing_sse2.c
index f6c6e0f..a5f8c9f 100644
--- a/src/dsp/alpha_processing_sse2.c
+++ b/src/dsp/alpha_processing_sse2.c
@@ -18,9 +18,9 @@
//------------------------------------------------------------------------------
-static int DispatchAlpha_SSE2(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint8_t* dst, int dst_stride) {
+static int DispatchAlpha_SSE2(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint8_t* WEBP_RESTRICT dst, int dst_stride) {
// alpha_and stores an 'and' operation of all the alpha[] values. The final
// value is not 0xff if any of the alpha[] is not equal to 0xff.
uint32_t alpha_and = 0xff;
@@ -72,9 +72,10 @@
return (alpha_and != 0xff);
}
-static void DispatchAlphaToGreen_SSE2(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint32_t* dst, int dst_stride) {
+static void DispatchAlphaToGreen_SSE2(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint32_t* WEBP_RESTRICT dst,
+ int dst_stride) {
int i, j;
const __m128i zero = _mm_setzero_si128();
const int limit = width & ~15;
@@ -98,9 +99,9 @@
}
}
-static int ExtractAlpha_SSE2(const uint8_t* argb, int argb_stride,
+static int ExtractAlpha_SSE2(const uint8_t* WEBP_RESTRICT argb, int argb_stride,
int width, int height,
- uint8_t* alpha, int alpha_stride) {
+ uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
// alpha_and stores an 'and' operation of all the alpha[] values. The final
// value is not 0xff if any of the alpha[] is not equal to 0xff.
uint32_t alpha_and = 0xff;
@@ -317,7 +318,8 @@
if (width > 0) WebPMultARGBRow_C(ptr + x, width, inverse);
}
-static void MultRow_SSE2(uint8_t* const ptr, const uint8_t* const alpha,
+static void MultRow_SSE2(uint8_t* WEBP_RESTRICT const ptr,
+ const uint8_t* WEBP_RESTRICT const alpha,
int width, int inverse) {
int x = 0;
if (!inverse) {
diff --git a/src/dsp/alpha_processing_sse41.c b/src/dsp/alpha_processing_sse41.c
index 56040f9..cdf877c 100644
--- a/src/dsp/alpha_processing_sse41.c
+++ b/src/dsp/alpha_processing_sse41.c
@@ -19,9 +19,9 @@
//------------------------------------------------------------------------------
-static int ExtractAlpha_SSE41(const uint8_t* argb, int argb_stride,
- int width, int height,
- uint8_t* alpha, int alpha_stride) {
+static int ExtractAlpha_SSE41(const uint8_t* WEBP_RESTRICT argb,
+ int argb_stride, int width, int height,
+ uint8_t* WEBP_RESTRICT alpha, int alpha_stride) {
// alpha_and stores an 'and' operation of all the alpha[] values. The final
// value is not 0xff if any of the alpha[] is not equal to 0xff.
uint32_t alpha_and = 0xff;
diff --git a/src/dsp/cost.c b/src/dsp/cost.c
index cc681cd..460ec4f 100644
--- a/src/dsp/cost.c
+++ b/src/dsp/cost.c
@@ -395,12 +395,12 @@
VP8EncDspCostInitMIPSdspR2();
}
#endif
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8EncDspCostInitSSE2();
}
#endif
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (VP8GetCPUInfo(kNEON)) {
VP8EncDspCostInitNEON();
}
diff --git a/src/dsp/cpu.c b/src/dsp/cpu.c
index 4ca90d8..3145e19 100644
--- a/src/dsp/cpu.c
+++ b/src/dsp/cpu.c
@@ -189,17 +189,17 @@
// Use compile flags as an indicator of SIMD support instead of a runtime check.
static int wasmCPUInfo(CPUFeature feature) {
switch (feature) {
-#ifdef WEBP_USE_SSE2
+#ifdef WEBP_HAVE_SSE2
case kSSE2:
return 1;
#endif
-#ifdef WEBP_USE_SSE41
+#ifdef WEBP_HAVE_SSE41
case kSSE3:
case kSlowSSSE3:
case kSSE4_1:
return 1;
#endif
-#ifdef WEBP_USE_NEON
+#ifdef WEBP_HAVE_NEON
case kNEON:
return 1;
#endif
@@ -209,9 +209,10 @@
return 0;
}
VP8CPUInfo VP8GetCPUInfo = wasmCPUInfo;
-#elif defined(WEBP_USE_NEON)
-// define a dummy function to enable turning off NEON at runtime by setting
-// VP8DecGetCPUInfo = NULL
+#elif defined(WEBP_HAVE_NEON)
+// In most cases this function doesn't check for NEON support (it's assumed by
+// the configuration), but enables turning off NEON at runtime, for testing
+// purposes, by setting VP8DecGetCPUInfo = NULL.
static int armCPUInfo(CPUFeature feature) {
if (feature != kNEON) return 0;
#if defined(__linux__) && defined(WEBP_HAVE_NEON_RTCD)
diff --git a/src/dsp/dec.c b/src/dsp/dec.c
index 1119842..537c701 100644
--- a/src/dsp/dec.c
+++ b/src/dsp/dec.c
@@ -807,10 +807,10 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8DspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
VP8DspInitSSE41();
}
@@ -834,7 +834,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
VP8DspInitNEON();
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 298c721..513e159 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -27,6 +27,23 @@
#define BPS 32 // this is the common stride for enc/dec
//------------------------------------------------------------------------------
+// WEBP_RESTRICT
+
+// Declares a pointer with the restrict type qualifier if available.
+// This allows code to hint to the compiler that only this pointer references a
+// particular object or memory region within the scope of the block in which it
+// is declared. This may allow for improved optimizations due to the lack of
+// pointer aliasing. See also:
+// https://en.cppreference.com/w/c/language/restrict
+#if defined(__GNUC__)
+#define WEBP_RESTRICT __restrict__
+#elif defined(_MSC_VER)
+#define WEBP_RESTRICT __restrict
+#else
+#define WEBP_RESTRICT
+#endif
+
+//------------------------------------------------------------------------------
// CPU detection
#if defined(__GNUC__)
@@ -67,21 +84,31 @@
// files without intrinsics, allowing the corresponding Init() to be called.
// Files containing intrinsics will need to be built targeting the instruction
// set so should succeed on one of the earlier tests.
-#if defined(__SSE2__) || defined(WEBP_MSC_SSE2) || defined(WEBP_HAVE_SSE2)
+#if (defined(__SSE2__) || defined(WEBP_MSC_SSE2)) && \
+ (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE2))
#define WEBP_USE_SSE2
#endif
-#if defined(__SSE4_1__) || defined(WEBP_MSC_SSE41) || defined(WEBP_HAVE_SSE41)
+#if defined(WEBP_USE_SSE2) && !defined(WEBP_HAVE_SSE2)
+#define WEBP_HAVE_SSE2
+#endif
+
+#if (defined(__SSE4_1__) || defined(WEBP_MSC_SSE41)) && \
+ (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE41))
#define WEBP_USE_SSE41
#endif
+#if defined(WEBP_USE_SSE41) && !defined(WEBP_HAVE_SSE41)
+#define WEBP_HAVE_SSE41
+#endif
+
#undef WEBP_MSC_SSE41
#undef WEBP_MSC_SSE2
// The intrinsics currently cause compiler errors with arm-nacl-gcc and the
// inline assembly would need to be modified for use with Native Client.
-#if (defined(__ARM_NEON__) || \
- defined(__aarch64__) || defined(WEBP_HAVE_NEON)) && \
+#if ((defined(__ARM_NEON__) || defined(__aarch64__)) && \
+ (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_NEON))) && \
!defined(__native_client__)
#define WEBP_USE_NEON
#endif
@@ -97,6 +124,10 @@
#define WEBP_USE_INTRINSICS
#endif
+#if defined(WEBP_USE_NEON) && !defined(WEBP_HAVE_NEON)
+#define WEBP_HAVE_NEON
+#endif
+
#if defined(__mips__) && !defined(__mips64) && \
defined(__mips_isa_rev) && (__mips_isa_rev >= 1) && (__mips_isa_rev < 6)
#define WEBP_USE_MIPS32
@@ -116,7 +147,7 @@
#define WEBP_DSP_OMIT_C_CODE 1
#endif
-#if (defined(__aarch64__) || defined(__ARM_NEON__)) && WEBP_DSP_OMIT_C_CODE
+#if defined(WEBP_USE_NEON) && WEBP_DSP_OMIT_C_CODE
#define WEBP_NEON_OMIT_C_CODE 1
#else
#define WEBP_NEON_OMIT_C_CODE 0
@@ -578,26 +609,29 @@
// Dispatch the values from alpha[] plane to the ARGB destination 'dst'.
// Returns true if alpha[] plane has non-trivial values different from 0xff.
-extern int (*WebPDispatchAlpha)(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint8_t* dst, int dst_stride);
+extern int (*WebPDispatchAlpha)(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint8_t* WEBP_RESTRICT dst, int dst_stride);
// Transfer packed 8b alpha[] values to green channel in dst[], zero'ing the
// A/R/B values. 'dst_stride' is the stride for dst[] in uint32_t units.
-extern void (*WebPDispatchAlphaToGreen)(const uint8_t* alpha, int alpha_stride,
- int width, int height,
- uint32_t* dst, int dst_stride);
+extern void (*WebPDispatchAlphaToGreen)(const uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride, int width, int height,
+ uint32_t* WEBP_RESTRICT dst,
+ int dst_stride);
// Extract the alpha values from 32b values in argb[] and pack them into alpha[]
// (this is the opposite of WebPDispatchAlpha).
// Returns true if there's only trivial 0xff alpha values.
-extern int (*WebPExtractAlpha)(const uint8_t* argb, int argb_stride,
- int width, int height,
- uint8_t* alpha, int alpha_stride);
+extern int (*WebPExtractAlpha)(const uint8_t* WEBP_RESTRICT argb,
+ int argb_stride, int width, int height,
+ uint8_t* WEBP_RESTRICT alpha,
+ int alpha_stride);
// Extract the green values from 32b values in argb[] and pack them into alpha[]
// (this is the opposite of WebPDispatchAlphaToGreen).
-extern void (*WebPExtractGreen)(const uint32_t* argb, uint8_t* alpha, int size);
+extern void (*WebPExtractGreen)(const uint32_t* WEBP_RESTRICT argb,
+ uint8_t* WEBP_RESTRICT alpha, int size);
// Pre-Multiply operation transforms x into x * A / 255 (where x=Y,R,G or B).
// Un-Multiply operation transforms x into x * 255 / A.
@@ -610,29 +644,35 @@
int inverse);
// Same for a row of single values, with side alpha values.
-extern void (*WebPMultRow)(uint8_t* const ptr, const uint8_t* const alpha,
+extern void (*WebPMultRow)(uint8_t* WEBP_RESTRICT const ptr,
+ const uint8_t* WEBP_RESTRICT const alpha,
int width, int inverse);
// Same a WebPMultRow(), but for several 'num_rows' rows.
-void WebPMultRows(uint8_t* ptr, int stride,
- const uint8_t* alpha, int alpha_stride,
+void WebPMultRows(uint8_t* WEBP_RESTRICT ptr, int stride,
+ const uint8_t* WEBP_RESTRICT alpha, int alpha_stride,
int width, int num_rows, int inverse);
// Plain-C versions, used as fallback by some implementations.
-void WebPMultRow_C(uint8_t* const ptr, const uint8_t* const alpha,
+void WebPMultRow_C(uint8_t* WEBP_RESTRICT const ptr,
+ const uint8_t* WEBP_RESTRICT const alpha,
int width, int inverse);
void WebPMultARGBRow_C(uint32_t* const ptr, int width, int inverse);
#ifdef WORDS_BIGENDIAN
// ARGB packing function: a/r/g/b input is rgba or bgra order.
-extern void (*WebPPackARGB)(const uint8_t* a, const uint8_t* r,
- const uint8_t* g, const uint8_t* b, int len,
- uint32_t* out);
+extern void (*WebPPackARGB)(const uint8_t* WEBP_RESTRICT a,
+ const uint8_t* WEBP_RESTRICT r,
+ const uint8_t* WEBP_RESTRICT g,
+ const uint8_t* WEBP_RESTRICT b,
+ int len, uint32_t* WEBP_RESTRICT out);
#endif
// RGB packing function. 'step' can be 3 or 4. r/g/b input is rgb or bgr order.
-extern void (*WebPPackRGB)(const uint8_t* r, const uint8_t* g, const uint8_t* b,
- int len, int step, uint32_t* out);
+extern void (*WebPPackRGB)(const uint8_t* WEBP_RESTRICT r,
+ const uint8_t* WEBP_RESTRICT g,
+ const uint8_t* WEBP_RESTRICT b,
+ int len, int step, uint32_t* WEBP_RESTRICT out);
// This function returns true if src[i] contains a value different from 0xff.
extern int (*WebPHasAlpha8b)(const uint8_t* src, int length);
diff --git a/src/dsp/enc.c b/src/dsp/enc.c
index 2fddbc4..ea47a3f 100644
--- a/src/dsp/enc.c
+++ b/src/dsp/enc.c
@@ -773,10 +773,10 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8EncDspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
VP8EncDspInitSSE41();
}
@@ -800,7 +800,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
VP8EncDspInitNEON();
diff --git a/src/dsp/filters.c b/src/dsp/filters.c
index 9e910d9..4506567 100644
--- a/src/dsp/filters.c
+++ b/src/dsp/filters.c
@@ -254,7 +254,7 @@
#endif
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8FiltersInitSSE2();
}
@@ -271,7 +271,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
VP8FiltersInitNEON();
diff --git a/src/dsp/filters_sse2.c b/src/dsp/filters_sse2.c
index 4b3f2d0..5c33ec1 100644
--- a/src/dsp/filters_sse2.c
+++ b/src/dsp/filters_sse2.c
@@ -320,7 +320,12 @@
WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitSSE2(void) {
WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_SSE2;
+#if defined(CHROMIUM)
+ // TODO(crbug.com/654974)
+ (void)VerticalUnfilter_SSE2;
+#else
WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_SSE2;
+#endif
WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_SSE2;
WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_SSE2;
diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c
index 46b220e..d8bbb02 100644
--- a/src/dsp/lossless.c
+++ b/src/dsp/lossless.c
@@ -575,6 +575,7 @@
VP8LMapAlphaFunc VP8LMapColor8b;
extern void VP8LDspInitSSE2(void);
+extern void VP8LDspInitSSE41(void);
extern void VP8LDspInitNEON(void);
extern void VP8LDspInitMIPSdspR2(void);
extern void VP8LDspInitMSA(void);
@@ -621,9 +622,14 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8LDspInitSSE2();
+#if defined(WEBP_HAVE_SSE41)
+ if (VP8GetCPUInfo(kSSE4_1)) {
+ VP8LDspInitSSE41();
+ }
+#endif
}
#endif
#if defined(WEBP_USE_MIPS_DSP_R2)
@@ -638,7 +644,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
VP8LDspInitNEON();
diff --git a/src/dsp/lossless_enc.c b/src/dsp/lossless_enc.c
index a0c7ab9..c3e8537 100644
--- a/src/dsp/lossless_enc.c
+++ b/src/dsp/lossless_enc.c
@@ -329,6 +329,15 @@
static float FastSLog2Slow_C(uint32_t v) {
assert(v >= LOG_LOOKUP_IDX_MAX);
if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ)
+ // use clz if available
+ const int log_cnt = BitsLog2Floor(v) - 7;
+ const uint32_t y = 1 << log_cnt;
+ int correction = 0;
+ const float v_f = (float)v;
+ const uint32_t orig_v = v;
+ v >>= log_cnt;
+#else
int log_cnt = 0;
uint32_t y = 1;
int correction = 0;
@@ -339,6 +348,7 @@
v = v >> 1;
y = y << 1;
} while (v >= LOG_LOOKUP_IDX_MAX);
+#endif
// vf = (2^log_cnt) * Xf; where y = 2^log_cnt and Xf < 256
// Xf = floor(Xf) * (1 + (v % y) / v)
// log2(Xf) = log2(floor(Xf)) + log2(1 + (v % y) / v)
@@ -355,6 +365,14 @@
static float FastLog2Slow_C(uint32_t v) {
assert(v >= LOG_LOOKUP_IDX_MAX);
if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
+#if !defined(WEBP_HAVE_SLOW_CLZ_CTZ)
+ // use clz if available
+ const int log_cnt = BitsLog2Floor(v) - 7;
+ const uint32_t y = 1 << log_cnt;
+ const uint32_t orig_v = v;
+ double log_2;
+ v >>= log_cnt;
+#else
int log_cnt = 0;
uint32_t y = 1;
const uint32_t orig_v = v;
@@ -364,6 +382,7 @@
v = v >> 1;
y = y << 1;
} while (v >= LOG_LOOKUP_IDX_MAX);
+#endif
log_2 = kLog2Table[v] + log_cnt;
if (orig_v >= APPROX_LOG_MAX) {
// Since the division is still expensive, add this correction factor only
@@ -843,10 +862,10 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8LEncDspInitSSE2();
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
VP8LEncDspInitSSE41();
}
@@ -870,7 +889,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
VP8LEncDspInitNEON();
diff --git a/src/dsp/lossless_enc_sse2.c b/src/dsp/lossless_enc_sse2.c
index 90c2637..b2f83b8 100644
--- a/src/dsp/lossless_enc_sse2.c
+++ b/src/dsp/lossless_enc_sse2.c
@@ -232,79 +232,55 @@
//------------------------------------------------------------------------------
// Entropy
-// Checks whether the X or Y contribution is worth computing and adding.
-// Used in loop unrolling.
-#define ANALYZE_X_OR_Y(x_or_y, j) \
- do { \
- if ((x_or_y)[i + (j)] != 0) retval -= VP8LFastSLog2((x_or_y)[i + (j)]); \
- } while (0)
+// TODO(https://crbug.com/webp/499): this function produces different results
+// from the C code due to use of double/float resulting in output differences
+// when compared to -noasm.
+#if !(defined(WEBP_HAVE_SLOW_CLZ_CTZ) || defined(__i386__) || defined(_M_IX86))
-// Checks whether the X + Y contribution is worth computing and adding.
-// Used in loop unrolling.
-#define ANALYZE_XY(j) \
- do { \
- if (tmp[j] != 0) { \
- retval -= VP8LFastSLog2(tmp[j]); \
- ANALYZE_X_OR_Y(X, j); \
- } \
- } while (0)
-
-#if !(defined(__i386__) || defined(_M_IX86))
static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) {
int i;
double retval = 0.;
- int sumX, sumXY;
- int32_t tmp[4];
- __m128i zero = _mm_setzero_si128();
- // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
- __m128i sumXY_128 = zero;
- __m128i sumX_128 = zero;
+ int sumX = 0, sumXY = 0;
+ const __m128i zero = _mm_setzero_si128();
- for (i = 0; i < 256; i += 4) {
- const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
- const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
-
- // Check if any X is non-zero: this actually provides a speedup as X is
- // usually sparse.
- if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
- const __m128i xy_128 = _mm_add_epi32(x, y);
- sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
-
- sumX_128 = _mm_add_epi32(sumX_128, x);
-
- // Analyze the different X + Y.
- _mm_storeu_si128((__m128i*)tmp, xy_128);
-
- ANALYZE_XY(0);
- ANALYZE_XY(1);
- ANALYZE_XY(2);
- ANALYZE_XY(3);
- } else {
- // X is fully 0, so only deal with Y.
- sumXY_128 = _mm_add_epi32(sumXY_128, y);
-
- ANALYZE_X_OR_Y(Y, 0);
- ANALYZE_X_OR_Y(Y, 1);
- ANALYZE_X_OR_Y(Y, 2);
- ANALYZE_X_OR_Y(Y, 3);
+ for (i = 0; i < 256; i += 16) {
+ const __m128i x0 = _mm_loadu_si128((const __m128i*)(X + i + 0));
+ const __m128i y0 = _mm_loadu_si128((const __m128i*)(Y + i + 0));
+ const __m128i x1 = _mm_loadu_si128((const __m128i*)(X + i + 4));
+ const __m128i y1 = _mm_loadu_si128((const __m128i*)(Y + i + 4));
+ const __m128i x2 = _mm_loadu_si128((const __m128i*)(X + i + 8));
+ const __m128i y2 = _mm_loadu_si128((const __m128i*)(Y + i + 8));
+ const __m128i x3 = _mm_loadu_si128((const __m128i*)(X + i + 12));
+ const __m128i y3 = _mm_loadu_si128((const __m128i*)(Y + i + 12));
+ const __m128i x4 = _mm_packs_epi16(_mm_packs_epi32(x0, x1),
+ _mm_packs_epi32(x2, x3));
+ const __m128i y4 = _mm_packs_epi16(_mm_packs_epi32(y0, y1),
+ _mm_packs_epi32(y2, y3));
+ const int32_t mx = _mm_movemask_epi8(_mm_cmpgt_epi8(x4, zero));
+ int32_t my = _mm_movemask_epi8(_mm_cmpgt_epi8(y4, zero)) | mx;
+ while (my) {
+ const int32_t j = BitsCtz(my);
+ int xy;
+ if ((mx >> j) & 1) {
+ const int x = X[i + j];
+ sumXY += x;
+ retval -= VP8LFastSLog2(x);
+ }
+ xy = X[i + j] + Y[i + j];
+ sumX += xy;
+ retval -= VP8LFastSLog2(xy);
+ my &= my - 1;
}
}
-
- // Sum up sumX_128 to get sumX.
- _mm_storeu_si128((__m128i*)tmp, sumX_128);
- sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
-
- // Sum up sumXY_128 to get sumXY.
- _mm_storeu_si128((__m128i*)tmp, sumXY_128);
- sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
-
retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
return (float)retval;
}
-#endif // !(defined(__i386__) || defined(_M_IX86))
-#undef ANALYZE_X_OR_Y
-#undef ANALYZE_XY
+#else
+
+#define DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC // won't be faster
+
+#endif
//------------------------------------------------------------------------------
@@ -662,10 +638,7 @@
VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2;
VP8LAddVector = AddVector_SSE2;
VP8LAddVectorEq = AddVectorEq_SSE2;
- // TODO(https://crbug.com/webp/499): this function produces different results
- // from the C code due to use of double/float resulting in output differences
- // when compared to -noasm.
-#if !(defined(__i386__) || defined(_M_IX86))
+#if !defined(DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC)
VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2;
#endif
VP8LVectorMismatch = VectorMismatch_SSE2;
diff --git a/src/dsp/lossless_enc_sse41.c b/src/dsp/lossless_enc_sse41.c
index 719d8ed..ad358a6 100644
--- a/src/dsp/lossless_enc_sse41.c
+++ b/src/dsp/lossless_enc_sse41.c
@@ -44,46 +44,47 @@
//------------------------------------------------------------------------------
// Color Transform
-#define SPAN 8
+#define MK_CST_16(HI, LO) \
+ _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
+
static void CollectColorBlueTransforms_SSE41(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_blue, int red_to_blue,
int histo[]) {
- const __m128i mults_r = _mm_set1_epi16(CST_5b(red_to_blue));
- const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_blue));
- const __m128i mask_g = _mm_set1_epi16((short)0xff00); // green mask
- const __m128i mask_gb = _mm_set1_epi32(0xffff); // green/blue mask
- const __m128i mask_b = _mm_set1_epi16(0x00ff); // blue mask
- const __m128i shuffler_lo = _mm_setr_epi8(-1, 2, -1, 6, -1, 10, -1, 14, -1,
- -1, -1, -1, -1, -1, -1, -1);
- const __m128i shuffler_hi = _mm_setr_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1,
- 2, -1, 6, -1, 10, -1, 14);
- int y;
- for (y = 0; y < tile_height; ++y) {
- const uint32_t* const src = argb + y * stride;
- int i, x;
- for (x = 0; x + SPAN <= tile_width; x += SPAN) {
- uint16_t values[SPAN];
- const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
- const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
- const __m128i r0 = _mm_shuffle_epi8(in0, shuffler_lo);
- const __m128i r1 = _mm_shuffle_epi8(in1, shuffler_hi);
- const __m128i r = _mm_or_si128(r0, r1); // r 0
- const __m128i gb0 = _mm_and_si128(in0, mask_gb);
- const __m128i gb1 = _mm_and_si128(in1, mask_gb);
- const __m128i gb = _mm_packus_epi32(gb0, gb1); // g b
- const __m128i g = _mm_and_si128(gb, mask_g); // g 0
- const __m128i A = _mm_mulhi_epi16(r, mults_r); // x dbr
- const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dbg
- const __m128i C = _mm_sub_epi8(gb, B); // x b'
- const __m128i D = _mm_sub_epi8(C, A); // x b''
- const __m128i E = _mm_and_si128(D, mask_b); // 0 b''
- _mm_storeu_si128((__m128i*)values, E);
- for (i = 0; i < SPAN; ++i) ++histo[values[i]];
+ const __m128i mult =
+ MK_CST_16(CST_5b(red_to_blue) + 256,CST_5b(green_to_blue));
+ const __m128i perm =
+ _mm_setr_epi8(-1, 1, -1, 2, -1, 5, -1, 6, -1, 9, -1, 10, -1, 13, -1, 14);
+ if (tile_width >= 4) {
+ int y;
+ for (y = 0; y < tile_height; ++y) {
+ const uint32_t* const src = argb + y * stride;
+ const __m128i A1 = _mm_loadu_si128((const __m128i*)src);
+ const __m128i B1 = _mm_shuffle_epi8(A1, perm);
+ const __m128i C1 = _mm_mulhi_epi16(B1, mult);
+ const __m128i D1 = _mm_sub_epi16(A1, C1);
+ __m128i E = _mm_add_epi16(_mm_srli_epi32(D1, 16), D1);
+ int x;
+ for (x = 4; x + 4 <= tile_width; x += 4) {
+ const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x));
+ __m128i B2, C2, D2;
+ ++histo[_mm_extract_epi8(E, 0)];
+ B2 = _mm_shuffle_epi8(A2, perm);
+ ++histo[_mm_extract_epi8(E, 4)];
+ C2 = _mm_mulhi_epi16(B2, mult);
+ ++histo[_mm_extract_epi8(E, 8)];
+ D2 = _mm_sub_epi16(A2, C2);
+ ++histo[_mm_extract_epi8(E, 12)];
+ E = _mm_add_epi16(_mm_srli_epi32(D2, 16), D2);
+ }
+ ++histo[_mm_extract_epi8(E, 0)];
+ ++histo[_mm_extract_epi8(E, 4)];
+ ++histo[_mm_extract_epi8(E, 8)];
+ ++histo[_mm_extract_epi8(E, 12)];
}
}
{
- const int left_over = tile_width & (SPAN - 1);
+ const int left_over = tile_width & 3;
if (left_over > 0) {
VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride,
left_over, tile_height,
@@ -95,33 +96,37 @@
static void CollectColorRedTransforms_SSE41(const uint32_t* argb, int stride,
int tile_width, int tile_height,
int green_to_red, int histo[]) {
- const __m128i mults_g = _mm_set1_epi16(CST_5b(green_to_red));
- const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask
- const __m128i mask = _mm_set1_epi16(0xff);
- int y;
- for (y = 0; y < tile_height; ++y) {
- const uint32_t* const src = argb + y * stride;
- int i, x;
- for (x = 0; x + SPAN <= tile_width; x += SPAN) {
- uint16_t values[SPAN];
- const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
- const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
- const __m128i g0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0
- const __m128i g1 = _mm_and_si128(in1, mask_g);
- const __m128i g = _mm_packus_epi32(g0, g1); // g 0
- const __m128i A0 = _mm_srli_epi32(in0, 16); // 0 0 | x r
- const __m128i A1 = _mm_srli_epi32(in1, 16);
- const __m128i A = _mm_packus_epi32(A0, A1); // x r
- const __m128i B = _mm_mulhi_epi16(g, mults_g); // x dr
- const __m128i C = _mm_sub_epi8(A, B); // x r'
- const __m128i D = _mm_and_si128(C, mask); // 0 r'
- _mm_storeu_si128((__m128i*)values, D);
- for (i = 0; i < SPAN; ++i) ++histo[values[i]];
+ const __m128i mult = MK_CST_16(0, CST_5b(green_to_red));
+ const __m128i mask_g = _mm_set1_epi32(0x0000ff00);
+ if (tile_width >= 4) {
+ int y;
+ for (y = 0; y < tile_height; ++y) {
+ const uint32_t* const src = argb + y * stride;
+ const __m128i A1 = _mm_loadu_si128((const __m128i*)src);
+ const __m128i B1 = _mm_and_si128(A1, mask_g);
+ const __m128i C1 = _mm_madd_epi16(B1, mult);
+ __m128i D = _mm_sub_epi16(A1, C1);
+ int x;
+ for (x = 4; x + 4 <= tile_width; x += 4) {
+ const __m128i A2 = _mm_loadu_si128((const __m128i*)(src + x));
+ __m128i B2, C2;
+ ++histo[_mm_extract_epi8(D, 2)];
+ B2 = _mm_and_si128(A2, mask_g);
+ ++histo[_mm_extract_epi8(D, 6)];
+ C2 = _mm_madd_epi16(B2, mult);
+ ++histo[_mm_extract_epi8(D, 10)];
+ ++histo[_mm_extract_epi8(D, 14)];
+ D = _mm_sub_epi16(A2, C2);
+ }
+ ++histo[_mm_extract_epi8(D, 2)];
+ ++histo[_mm_extract_epi8(D, 6)];
+ ++histo[_mm_extract_epi8(D, 10)];
+ ++histo[_mm_extract_epi8(D, 14)];
}
}
{
- const int left_over = tile_width & (SPAN - 1);
+ const int left_over = tile_width & 3;
if (left_over > 0) {
VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride,
left_over, tile_height, green_to_red,
@@ -130,6 +135,8 @@
}
}
+#undef MK_CST_16
+
//------------------------------------------------------------------------------
// Entry point
diff --git a/src/dsp/lossless_sse2.c b/src/dsp/lossless_sse2.c
index aef0cee..3a0eb44 100644
--- a/src/dsp/lossless_sse2.c
+++ b/src/dsp/lossless_sse2.c
@@ -18,7 +18,6 @@
#include "src/dsp/common_sse2.h"
#include "src/dsp/lossless.h"
#include "src/dsp/lossless_common.h"
-#include <assert.h>
#include <emmintrin.h>
//------------------------------------------------------------------------------
diff --git a/src/dsp/lossless_sse41.c b/src/dsp/lossless_sse41.c
new file mode 100644
index 0000000..b0d6daa
--- /dev/null
+++ b/src/dsp/lossless_sse41.c
@@ -0,0 +1,132 @@
+// Copyright 2021 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// SSE41 variant of methods for lossless decoder
+
+#include "src/dsp/dsp.h"
+
+#if defined(WEBP_USE_SSE41)
+
+#include "src/dsp/common_sse41.h"
+#include "src/dsp/lossless.h"
+#include "src/dsp/lossless_common.h"
+
+//------------------------------------------------------------------------------
+// Color-space conversion functions
+
+static void TransformColorInverse_SSE41(const VP8LMultipliers* const m,
+ const uint32_t* const src,
+ int num_pixels, uint32_t* dst) {
+// sign-extended multiplying constants, pre-shifted by 5.
+#define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
+ const __m128i mults_rb = _mm_set1_epi32((uint32_t)CST(green_to_red_) << 16 |
+ (CST(green_to_blue_) & 0xffff));
+ const __m128i mults_b2 = _mm_set1_epi32(CST(red_to_blue_));
+#undef CST
+ const __m128i mask_ag = _mm_set1_epi32(0xff00ff00);
+ const __m128i perm1 = _mm_setr_epi8(-1, 1, -1, 1, -1, 5, -1, 5,
+ -1, 9, -1, 9, -1, 13, -1, 13);
+ const __m128i perm2 = _mm_setr_epi8(-1, 2, -1, -1, -1, 6, -1, -1,
+ -1, 10, -1, -1, -1, 14, -1, -1);
+ int i;
+ for (i = 0; i + 4 <= num_pixels; i += 4) {
+ const __m128i A = _mm_loadu_si128((const __m128i*)(src + i));
+ const __m128i B = _mm_shuffle_epi8(A, perm1); // argb -> g0g0
+ const __m128i C = _mm_mulhi_epi16(B, mults_rb);
+ const __m128i D = _mm_add_epi8(A, C);
+ const __m128i E = _mm_shuffle_epi8(D, perm2);
+ const __m128i F = _mm_mulhi_epi16(E, mults_b2);
+ const __m128i G = _mm_add_epi8(D, F);
+ const __m128i out = _mm_blendv_epi8(G, A, mask_ag);
+ _mm_storeu_si128((__m128i*)&dst[i], out);
+ }
+ // Fall-back to C-version for left-overs.
+ if (i != num_pixels) {
+ VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
+ }
+}
+
+//------------------------------------------------------------------------------
+
+#define ARGB_TO_RGB_SSE41 do { \
+ while (num_pixels >= 16) { \
+ const __m128i in0 = _mm_loadu_si128(in + 0); \
+ const __m128i in1 = _mm_loadu_si128(in + 1); \
+ const __m128i in2 = _mm_loadu_si128(in + 2); \
+ const __m128i in3 = _mm_loadu_si128(in + 3); \
+ const __m128i a0 = _mm_shuffle_epi8(in0, perm0); \
+ const __m128i a1 = _mm_shuffle_epi8(in1, perm1); \
+ const __m128i a2 = _mm_shuffle_epi8(in2, perm2); \
+ const __m128i a3 = _mm_shuffle_epi8(in3, perm3); \
+ const __m128i b0 = _mm_blend_epi16(a0, a1, 0xc0); \
+ const __m128i b1 = _mm_blend_epi16(a1, a2, 0xf0); \
+ const __m128i b2 = _mm_blend_epi16(a2, a3, 0xfc); \
+ _mm_storeu_si128(out + 0, b0); \
+ _mm_storeu_si128(out + 1, b1); \
+ _mm_storeu_si128(out + 2, b2); \
+ in += 4; \
+ out += 3; \
+ num_pixels -= 16; \
+ } \
+} while (0)
+
+static void ConvertBGRAToRGB_SSE41(const uint32_t* src, int num_pixels,
+ uint8_t* dst) {
+ const __m128i* in = (const __m128i*)src;
+ __m128i* out = (__m128i*)dst;
+ const __m128i perm0 = _mm_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9,
+ 8, 14, 13, 12, -1, -1, -1, -1);
+ const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39);
+ const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e);
+ const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93);
+
+ ARGB_TO_RGB_SSE41;
+
+ // left-overs
+ if (num_pixels > 0) {
+ VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
+ }
+}
+
+static void ConvertBGRAToBGR_SSE41(const uint32_t* src,
+ int num_pixels, uint8_t* dst) {
+ const __m128i* in = (const __m128i*)src;
+ __m128i* out = (__m128i*)dst;
+ const __m128i perm0 = _mm_setr_epi8(0, 1, 2, 4, 5, 6, 8, 9, 10,
+ 12, 13, 14, -1, -1, -1, -1);
+ const __m128i perm1 = _mm_shuffle_epi32(perm0, 0x39);
+ const __m128i perm2 = _mm_shuffle_epi32(perm0, 0x4e);
+ const __m128i perm3 = _mm_shuffle_epi32(perm0, 0x93);
+
+ ARGB_TO_RGB_SSE41;
+
+ // left-overs
+ if (num_pixels > 0) {
+ VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
+ }
+}
+
+#undef ARGB_TO_RGB_SSE41
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8LDspInitSSE41(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE41(void) {
+ VP8LTransformColorInverse = TransformColorInverse_SSE41;
+ VP8LConvertBGRAToRGB = ConvertBGRAToRGB_SSE41;
+ VP8LConvertBGRAToBGR = ConvertBGRAToBGR_SSE41;
+}
+
+#else // !WEBP_USE_SSE41
+
+WEBP_DSP_INIT_STUB(VP8LDspInitSSE41)
+
+#endif // WEBP_USE_SSE41
diff --git a/src/dsp/rescaler.c b/src/dsp/rescaler.c
index c5a01e8..14620ce 100644
--- a/src/dsp/rescaler.c
+++ b/src/dsp/rescaler.c
@@ -38,8 +38,9 @@
int x_out = channel;
// simple bilinear interpolation
int accum = wrk->x_add;
- int left = src[x_in];
- int right = (wrk->src_width > 1) ? src[x_in + x_stride] : left;
+ rescaler_t left = (rescaler_t)src[x_in];
+ rescaler_t right =
+ (wrk->src_width > 1) ? (rescaler_t)src[x_in + x_stride] : left;
x_in += x_stride;
while (1) {
wrk->frow[x_out] = right * wrk->x_add + (left - right) * accum;
@@ -50,7 +51,7 @@
left = right;
x_in += x_stride;
assert(x_in < wrk->src_width * x_stride);
- right = src[x_in];
+ right = (rescaler_t)src[x_in];
accum += wrk->x_add;
}
}
@@ -213,7 +214,7 @@
WebPRescalerImportRowShrink = WebPRescalerImportRowShrink_C;
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPRescalerDspInitSSE2();
}
@@ -235,7 +236,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
WebPRescalerDspInitNEON();
diff --git a/src/dsp/ssim.c b/src/dsp/ssim.c
index 989ce82..f85c2e6 100644
--- a/src/dsp/ssim.c
+++ b/src/dsp/ssim.c
@@ -150,7 +150,7 @@
#endif
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
VP8SSIMDspInitSSE2();
}
diff --git a/src/dsp/upsampling.c b/src/dsp/upsampling.c
index 9b60da5..87f771f 100644
--- a/src/dsp/upsampling.c
+++ b/src/dsp/upsampling.c
@@ -233,12 +233,12 @@
WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444_C;
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitYUV444ConvertersSSE2();
}
#endif
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
WebPInitYUV444ConvertersSSE41();
}
@@ -278,12 +278,12 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitUpsamplersSSE2();
}
#endif
-#if defined(WEBP_USE_SSE41)
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
WebPInitUpsamplersSSE41();
}
@@ -300,7 +300,7 @@
#endif
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
WebPInitUpsamplersNEON();
diff --git a/src/dsp/yuv.c b/src/dsp/yuv.c
index 14e67fc..48466f8 100644
--- a/src/dsp/yuv.c
+++ b/src/dsp/yuv.c
@@ -90,16 +90,16 @@
// If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitSamplersSSE2();
}
-#endif // WEBP_USE_SSE2
-#if defined(WEBP_USE_SSE41)
+#endif // WEBP_HAVE_SSE2
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
WebPInitSamplersSSE41();
}
-#endif // WEBP_USE_SSE41
+#endif // WEBP_HAVE_SSE41
#if defined(WEBP_USE_MIPS32)
if (VP8GetCPUInfo(kMIPS32)) {
WebPInitSamplersMIPS32();
@@ -276,26 +276,26 @@
#endif
if (VP8GetCPUInfo != NULL) {
-#if defined(WEBP_USE_SSE2)
+#if defined(WEBP_HAVE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
WebPInitConvertARGBToYUVSSE2();
WebPInitSharpYUVSSE2();
}
-#endif // WEBP_USE_SSE2
-#if defined(WEBP_USE_SSE41)
+#endif // WEBP_HAVE_SSE2
+#if defined(WEBP_HAVE_SSE41)
if (VP8GetCPUInfo(kSSE4_1)) {
WebPInitConvertARGBToYUVSSE41();
}
-#endif // WEBP_USE_SSE41
+#endif // WEBP_HAVE_SSE41
}
-#if defined(WEBP_USE_NEON)
+#if defined(WEBP_HAVE_NEON)
if (WEBP_NEON_OMIT_C_CODE ||
(VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) {
WebPInitConvertARGBToYUVNEON();
WebPInitSharpYUVNEON();
}
-#endif // WEBP_USE_NEON
+#endif // WEBP_HAVE_NEON
assert(WebPConvertARGBToY != NULL);
assert(WebPConvertARGBToUV != NULL);
diff --git a/src/enc/alpha_enc.c b/src/enc/alpha_enc.c
index dce9ca9..0b54f3e 100644
--- a/src/enc/alpha_enc.c
+++ b/src/enc/alpha_enc.c
@@ -303,7 +303,7 @@
int ok = 1;
const int reduce_levels = (quality < 100);
- // quick sanity checks
+ // quick correctness checks
assert((uint64_t)data_size == (uint64_t)width * height); // as per spec
assert(enc != NULL && pic != NULL && pic->a != NULL);
assert(output != NULL && output_size != NULL);
@@ -361,7 +361,7 @@
//------------------------------------------------------------------------------
// Main calls
-static int CompressAlphaJob(void* arg1, void* dummy) {
+static int CompressAlphaJob(void* arg1, void* unused) {
VP8Encoder* const enc = (VP8Encoder*)arg1;
const WebPConfig* config = enc->config_;
uint8_t* alpha_data = NULL;
@@ -375,13 +375,13 @@
filter, effort_level, &alpha_data, &alpha_size)) {
return 0;
}
- if (alpha_size != (uint32_t)alpha_size) { // Sanity check.
+ if (alpha_size != (uint32_t)alpha_size) { // Soundness check.
WebPSafeFree(alpha_data);
return 0;
}
enc->alpha_data_size_ = (uint32_t)alpha_size;
enc->alpha_data_ = alpha_data;
- (void)dummy;
+ (void)unused;
return 1;
}
diff --git a/src/enc/histogram_enc.c b/src/enc/histogram_enc.c
index edc6e4f..38a0ceb 100644
--- a/src/enc/histogram_enc.c
+++ b/src/enc/histogram_enc.c
@@ -1171,13 +1171,15 @@
int VP8LGetHistoImageSymbols(int xsize, int ysize,
const VP8LBackwardRefs* const refs,
int quality, int low_effort,
- int histo_bits, int cache_bits,
+ int histogram_bits, int cache_bits,
VP8LHistogramSet* const image_histo,
VP8LHistogram* const tmp_histo,
uint16_t* const histogram_symbols) {
int ok = 0;
- const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
- const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
+ const int histo_xsize =
+ histogram_bits ? VP8LSubSampleSize(xsize, histogram_bits) : 1;
+ const int histo_ysize =
+ histogram_bits ? VP8LSubSampleSize(ysize, histogram_bits) : 1;
const int image_histo_raw_size = histo_xsize * histo_ysize;
VP8LHistogramSet* const orig_histo =
VP8LAllocateHistogramSet(image_histo_raw_size, cache_bits);
@@ -1193,7 +1195,7 @@
if (orig_histo == NULL || map_tmp == NULL) goto Error;
// Construct the histograms from backward references.
- HistogramBuild(xsize, histo_bits, refs, orig_histo);
+ HistogramBuild(xsize, histogram_bits, refs, orig_histo);
// Copies the histograms and computes its bit_cost.
// histogram_symbols is optimized
HistogramCopyAndAnalyze(orig_histo, image_histo, &num_used,
diff --git a/src/enc/histogram_enc.h b/src/enc/histogram_enc.h
index 54c2d21..c3428b5 100644
--- a/src/enc/histogram_enc.h
+++ b/src/enc/histogram_enc.h
@@ -64,8 +64,8 @@
const VP8LBackwardRefs* const refs,
int palette_code_bits);
-// Return the size of the histogram for a given palette_code_bits.
-int VP8LGetHistogramSize(int palette_code_bits);
+// Return the size of the histogram for a given cache_bits.
+int VP8LGetHistogramSize(int cache_bits);
// Set the palette_code_bits and reset the stats.
// If init_arrays is true, the arrays are also filled with 0's.
@@ -110,7 +110,7 @@
const VP8LBackwardRefs* const refs,
int quality, int low_effort,
int histogram_bits, int cache_bits,
- VP8LHistogramSet* const image_in,
+ VP8LHistogramSet* const image_histo,
VP8LHistogram* const tmp_histo,
uint16_t* const histogram_symbols);
diff --git a/src/enc/picture_rescale_enc.c b/src/enc/picture_rescale_enc.c
index 58a6ae7..a75f5d9 100644
--- a/src/enc/picture_rescale_enc.c
+++ b/src/enc/picture_rescale_enc.c
@@ -164,22 +164,25 @@
//------------------------------------------------------------------------------
// Simple picture rescaler
-static void RescalePlane(const uint8_t* src,
- int src_width, int src_height, int src_stride,
- uint8_t* dst,
- int dst_width, int dst_height, int dst_stride,
- rescaler_t* const work,
- int num_channels) {
+static int RescalePlane(const uint8_t* src,
+ int src_width, int src_height, int src_stride,
+ uint8_t* dst,
+ int dst_width, int dst_height, int dst_stride,
+ rescaler_t* const work,
+ int num_channels) {
WebPRescaler rescaler;
int y = 0;
- WebPRescalerInit(&rescaler, src_width, src_height,
- dst, dst_width, dst_height, dst_stride,
- num_channels, work);
+ if (!WebPRescalerInit(&rescaler, src_width, src_height,
+ dst, dst_width, dst_height, dst_stride,
+ num_channels, work)) {
+ return 0;
+ }
while (y < src_height) {
y += WebPRescalerImport(&rescaler, src_height - y,
src + y * src_stride, src_stride);
WebPRescalerExport(&rescaler);
}
+ return 1;
}
static void AlphaMultiplyARGB(WebPPicture* const pic, int inverse) {
@@ -222,25 +225,28 @@
// If present, we need to rescale alpha first (for AlphaMultiplyY).
if (pic->a != NULL) {
WebPInitAlphaProcessing();
- RescalePlane(pic->a, prev_width, prev_height, pic->a_stride,
- tmp.a, width, height, tmp.a_stride, work, 1);
+ if (!RescalePlane(pic->a, prev_width, prev_height, pic->a_stride,
+ tmp.a, width, height, tmp.a_stride, work, 1)) {
+ return 0;
+ }
}
// We take transparency into account on the luma plane only. That's not
// totally exact blending, but still is a good approximation.
AlphaMultiplyY(pic, 0);
- RescalePlane(pic->y, prev_width, prev_height, pic->y_stride,
- tmp.y, width, height, tmp.y_stride, work, 1);
+ if (!RescalePlane(pic->y, prev_width, prev_height, pic->y_stride,
+ tmp.y, width, height, tmp.y_stride, work, 1) ||
+ !RescalePlane(pic->u,
+ HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+ tmp.u,
+ HALVE(width), HALVE(height), tmp.uv_stride, work, 1) ||
+ !RescalePlane(pic->v,
+ HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
+ tmp.v,
+ HALVE(width), HALVE(height), tmp.uv_stride, work, 1)) {
+ return 0;
+ }
AlphaMultiplyY(&tmp, 1);
-
- RescalePlane(pic->u,
- HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
- tmp.u,
- HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
- RescalePlane(pic->v,
- HALVE(prev_width), HALVE(prev_height), pic->uv_stride,
- tmp.v,
- HALVE(width), HALVE(height), tmp.uv_stride, work, 1);
} else {
work = (rescaler_t*)WebPSafeMalloc(2ULL * width * 4, sizeof(*work));
if (work == NULL) {
@@ -252,11 +258,12 @@
// the premultiplication afterward (while preserving the alpha channel).
WebPInitAlphaProcessing();
AlphaMultiplyARGB(pic, 0);
- RescalePlane((const uint8_t*)pic->argb, prev_width, prev_height,
- pic->argb_stride * 4,
- (uint8_t*)tmp.argb, width, height,
- tmp.argb_stride * 4,
- work, 4);
+ if (!RescalePlane((const uint8_t*)pic->argb, prev_width, prev_height,
+ pic->argb_stride * 4,
+ (uint8_t*)tmp.argb, width, height,
+ tmp.argb_stride * 4, work, 4)) {
+ return 0;
+ }
AlphaMultiplyARGB(&tmp, 1);
}
WebPPictureFree(pic);
diff --git a/src/enc/syntax_enc.c b/src/enc/syntax_enc.c
index a9e5a6c..e18cf65 100644
--- a/src/enc/syntax_enc.c
+++ b/src/enc/syntax_enc.c
@@ -349,7 +349,7 @@
(enc->alpha_data_size_ & 1);
riff_size += CHUNK_HEADER_SIZE + padded_alpha_size;
}
- // Sanity check.
+ // RIFF size should fit in 32-bits.
if (riff_size > 0xfffffffeU) {
return WebPEncodingSetError(pic, VP8_ENC_ERROR_FILE_TOO_BIG);
}
diff --git a/src/enc/vp8i_enc.h b/src/enc/vp8i_enc.h
index 0e35562..67e9509 100644
--- a/src/enc/vp8i_enc.h
+++ b/src/enc/vp8i_enc.h
@@ -32,7 +32,7 @@
// version numbers
#define ENC_MAJ_VERSION 1
#define ENC_MIN_VERSION 2
-#define ENC_REV_VERSION 0
+#define ENC_REV_VERSION 1
enum { MAX_LF_LEVELS = 64, // Maximum loop filter level
MAX_VARIABLE_LEVEL = 67, // last (inclusive) level with variable cost
@@ -286,8 +286,7 @@
// save the yuv_out_ boundary values to top_/left_ arrays for next iterations.
void VP8IteratorSaveBoundary(VP8EncIterator* const it);
// Report progression based on macroblock rows. Return 0 for user-abort request.
-int VP8IteratorProgress(const VP8EncIterator* const it,
- int final_delta_percent);
+int VP8IteratorProgress(const VP8EncIterator* const it, int delta);
// Intra4x4 iterations
void VP8IteratorStartI4(VP8EncIterator* const it);
// returns true if not done.
diff --git a/src/enc/vp8l_enc.c b/src/enc/vp8l_enc.c
index 0b44ebe..e330e71 100644
--- a/src/enc/vp8l_enc.c
+++ b/src/enc/vp8l_enc.c
@@ -65,25 +65,22 @@
*col2 = tmp;
}
-static void GreedyMinimizeDeltas(uint32_t palette[], int num_colors) {
- // Find greedily always the closest color of the predicted color to minimize
- // deltas in the palette. This reduces storage needs since the
- // palette is stored with delta encoding.
- uint32_t predict = 0x00000000;
- int i, k;
- for (i = 0; i < num_colors; ++i) {
- int best_ix = i;
- uint32_t best_score = ~0U;
- for (k = i; k < num_colors; ++k) {
- const uint32_t cur_score = PaletteColorDistance(palette[k], predict);
- if (best_score > cur_score) {
- best_score = cur_score;
- best_ix = k;
- }
+static WEBP_INLINE int SearchColorNoIdx(const uint32_t sorted[], uint32_t color,
+ int num_colors) {
+ int low = 0, hi = num_colors;
+ if (sorted[low] == color) return low; // loop invariant: sorted[low] != color
+ while (1) {
+ const int mid = (low + hi) >> 1;
+ if (sorted[mid] == color) {
+ return mid;
+ } else if (sorted[mid] < color) {
+ low = mid;
+ } else {
+ hi = mid;
}
- SwapColor(&palette[best_ix], &palette[i]);
- predict = palette[i];
}
+ assert(0);
+ return 0;
}
// The palette has been sorted by alpha. This function checks if the other
@@ -92,7 +89,8 @@
// no benefit to re-organize them greedily. A monotonic development
// would be spotted in green-only situations (like lossy alpha) or gray-scale
// images.
-static int PaletteHasNonMonotonousDeltas(uint32_t palette[], int num_colors) {
+static int PaletteHasNonMonotonousDeltas(const uint32_t* const palette,
+ int num_colors) {
uint32_t predict = 0x000000;
int i;
uint8_t sign_found = 0x00;
@@ -115,28 +113,215 @@
return (sign_found & (sign_found << 1)) != 0; // two consequent signs.
}
+static void PaletteSortMinimizeDeltas(const uint32_t* const palette_sorted,
+ int num_colors, uint32_t* const palette) {
+ uint32_t predict = 0x00000000;
+ int i, k;
+ memcpy(palette, palette_sorted, num_colors * sizeof(*palette));
+ if (!PaletteHasNonMonotonousDeltas(palette_sorted, num_colors)) return;
+ // Find greedily always the closest color of the predicted color to minimize
+ // deltas in the palette. This reduces storage needs since the
+ // palette is stored with delta encoding.
+ for (i = 0; i < num_colors; ++i) {
+ int best_ix = i;
+ uint32_t best_score = ~0U;
+ for (k = i; k < num_colors; ++k) {
+ const uint32_t cur_score = PaletteColorDistance(palette[k], predict);
+ if (best_score > cur_score) {
+ best_score = cur_score;
+ best_ix = k;
+ }
+ }
+ SwapColor(&palette[best_ix], &palette[i]);
+ predict = palette[i];
+ }
+}
+
+// Sort palette in increasing order and prepare an inverse mapping array.
+static void PrepareMapToPalette(const uint32_t palette[], uint32_t num_colors,
+ uint32_t sorted[], uint32_t idx_map[]) {
+ uint32_t i;
+ memcpy(sorted, palette, num_colors * sizeof(*sorted));
+ qsort(sorted, num_colors, sizeof(*sorted), PaletteCompareColorsForQsort);
+ for (i = 0; i < num_colors; ++i) {
+ idx_map[SearchColorNoIdx(sorted, palette[i], num_colors)] = i;
+ }
+}
+
+// -----------------------------------------------------------------------------
+// Modified Zeng method from "A Survey on Palette Reordering
+// Methods for Improving the Compression of Color-Indexed Images" by Armando J.
+// Pinho and Antonio J. R. Neves.
+
+// Finds the biggest cooccurrence in the matrix.
+static void CoOccurrenceFindMax(const uint32_t* const cooccurrence,
+ uint32_t num_colors, uint8_t* const c1,
+ uint8_t* const c2) {
+ // Find the index that is most frequently located adjacent to other
+ // (different) indexes.
+ uint32_t best_sum = 0u;
+ uint32_t i, j, best_cooccurrence;
+ *c1 = 0u;
+ for (i = 0; i < num_colors; ++i) {
+ uint32_t sum = 0;
+ for (j = 0; j < num_colors; ++j) sum += cooccurrence[i * num_colors + j];
+ if (sum > best_sum) {
+ best_sum = sum;
+ *c1 = i;
+ }
+ }
+ // Find the index that is most frequently found adjacent to *c1.
+ *c2 = 0u;
+ best_cooccurrence = 0u;
+ for (i = 0; i < num_colors; ++i) {
+ if (cooccurrence[*c1 * num_colors + i] > best_cooccurrence) {
+ best_cooccurrence = cooccurrence[*c1 * num_colors + i];
+ *c2 = i;
+ }
+ }
+ assert(*c1 != *c2);
+}
+
+// Builds the cooccurrence matrix
+static WebPEncodingError CoOccurrenceBuild(const WebPPicture* const pic,
+ const uint32_t* const palette,
+ uint32_t num_colors,
+ uint32_t* cooccurrence) {
+ uint32_t *lines, *line_top, *line_current, *line_tmp;
+ int x, y;
+ const uint32_t* src = pic->argb;
+ uint32_t prev_pix = ~src[0];
+ uint32_t prev_idx = 0u;
+ uint32_t idx_map[MAX_PALETTE_SIZE] = {0};
+ uint32_t palette_sorted[MAX_PALETTE_SIZE];
+ lines = (uint32_t*)WebPSafeMalloc(2 * pic->width, sizeof(*lines));
+ if (lines == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+ line_top = &lines[0];
+ line_current = &lines[pic->width];
+ PrepareMapToPalette(palette, num_colors, palette_sorted, idx_map);
+ for (y = 0; y < pic->height; ++y) {
+ for (x = 0; x < pic->width; ++x) {
+ const uint32_t pix = src[x];
+ if (pix != prev_pix) {
+ prev_idx = idx_map[SearchColorNoIdx(palette_sorted, pix, num_colors)];
+ prev_pix = pix;
+ }
+ line_current[x] = prev_idx;
+ // 4-connectivity is what works best as mentioned in "On the relation
+ // between Memon's and the modified Zeng's palette reordering methods".
+ if (x > 0 && prev_idx != line_current[x - 1]) {
+ const uint32_t left_idx = line_current[x - 1];
+ ++cooccurrence[prev_idx * num_colors + left_idx];
+ ++cooccurrence[left_idx * num_colors + prev_idx];
+ }
+ if (y > 0 && prev_idx != line_top[x]) {
+ const uint32_t top_idx = line_top[x];
+ ++cooccurrence[prev_idx * num_colors + top_idx];
+ ++cooccurrence[top_idx * num_colors + prev_idx];
+ }
+ }
+ line_tmp = line_top;
+ line_top = line_current;
+ line_current = line_tmp;
+ src += pic->argb_stride;
+ }
+ WebPSafeFree(lines);
+ return VP8_ENC_OK;
+}
+
+struct Sum {
+ uint8_t index;
+ uint32_t sum;
+};
+
+// Implements the modified Zeng method from "A Survey on Palette Reordering
+// Methods for Improving the Compression of Color-Indexed Images" by Armando J.
+// Pinho and Antonio J. R. Neves.
+static WebPEncodingError PaletteSortModifiedZeng(
+ const WebPPicture* const pic, const uint32_t* const palette_sorted,
+ uint32_t num_colors, uint32_t* const palette) {
+ uint32_t i, j, ind;
+ uint8_t remapping[MAX_PALETTE_SIZE];
+ uint32_t* cooccurrence;
+ struct Sum sums[MAX_PALETTE_SIZE];
+ uint32_t first, last;
+ uint32_t num_sums;
+ // TODO(vrabaud) check whether one color images should use palette or not.
+ if (num_colors <= 1) return VP8_ENC_OK;
+ // Build the co-occurrence matrix.
+ cooccurrence =
+ (uint32_t*)WebPSafeCalloc(num_colors * num_colors, sizeof(*cooccurrence));
+ if (cooccurrence == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+ if (CoOccurrenceBuild(pic, palette_sorted, num_colors, cooccurrence) !=
+ VP8_ENC_OK) {
+ WebPSafeFree(cooccurrence);
+ return VP8_ENC_ERROR_OUT_OF_MEMORY;
+ }
+
+ // Initialize the mapping list with the two best indices.
+ CoOccurrenceFindMax(cooccurrence, num_colors, &remapping[0], &remapping[1]);
+
+ // We need to append and prepend to the list of remapping. To this end, we
+ // actually define the next start/end of the list as indices in a vector (with
+ // a wrap around when the end is reached).
+ first = 0;
+ last = 1;
+ num_sums = num_colors - 2; // -2 because we know the first two values
+ if (num_sums > 0) {
+ // Initialize the sums with the first two remappings and find the best one
+ struct Sum* best_sum = &sums[0];
+ best_sum->index = 0u;
+ best_sum->sum = 0u;
+ for (i = 0, j = 0; i < num_colors; ++i) {
+ if (i == remapping[0] || i == remapping[1]) continue;
+ sums[j].index = i;
+ sums[j].sum = cooccurrence[i * num_colors + remapping[0]] +
+ cooccurrence[i * num_colors + remapping[1]];
+ if (sums[j].sum > best_sum->sum) best_sum = &sums[j];
+ ++j;
+ }
+
+ while (num_sums > 0) {
+ const uint8_t best_index = best_sum->index;
+ // Compute delta to know if we need to prepend or append the best index.
+ int32_t delta = 0;
+ const int32_t n = num_colors - num_sums;
+ for (ind = first, j = 0; (ind + j) % num_colors != last + 1; ++j) {
+ const uint16_t l_j = remapping[(ind + j) % num_colors];
+ delta += (n - 1 - 2 * (int32_t)j) *
+ (int32_t)cooccurrence[best_index * num_colors + l_j];
+ }
+ if (delta > 0) {
+ first = (first == 0) ? num_colors - 1 : first - 1;
+ remapping[first] = best_index;
+ } else {
+ ++last;
+ remapping[last] = best_index;
+ }
+ // Remove best_sum from sums.
+ *best_sum = sums[num_sums - 1];
+ --num_sums;
+ // Update all the sums and find the best one.
+ best_sum = &sums[0];
+ for (i = 0; i < num_sums; ++i) {
+ sums[i].sum += cooccurrence[best_index * num_colors + sums[i].index];
+ if (sums[i].sum > best_sum->sum) best_sum = &sums[i];
+ }
+ }
+ }
+ assert((last + 1) % num_colors == first);
+ WebPSafeFree(cooccurrence);
+
+ // Re-map the palette.
+ for (i = 0; i < num_colors; ++i) {
+ palette[i] = palette_sorted[remapping[(first + i) % num_colors]];
+ }
+ return VP8_ENC_OK;
+}
+
// -----------------------------------------------------------------------------
// Palette
-// If number of colors in the image is less than or equal to MAX_PALETTE_SIZE,
-// creates a palette and returns true, else returns false.
-static int AnalyzeAndCreatePalette(const WebPPicture* const pic,
- int low_effort,
- uint32_t palette[MAX_PALETTE_SIZE],
- int* const palette_size) {
- const int num_colors = WebPGetColorPalette(pic, palette);
- if (num_colors > MAX_PALETTE_SIZE) {
- *palette_size = 0;
- return 0;
- }
- *palette_size = num_colors;
- qsort(palette, num_colors, sizeof(*palette), PaletteCompareColorsForQsort);
- if (!low_effort && PaletteHasNonMonotonousDeltas(palette, num_colors)) {
- GreedyMinimizeDeltas(palette, num_colors);
- }
- return 1;
-}
-
// These five modes are evaluated and their respective entropy is computed.
typedef enum {
kDirect = 0,
@@ -149,6 +334,13 @@
} EntropyIx;
typedef enum {
+ kSortedDefault = 0,
+ kMinimizeDelta = 1,
+ kModifiedZeng = 2,
+ kUnusedPalette = 3,
+} PaletteSorting;
+
+typedef enum {
kHistoAlpha = 0,
kHistoAlphaPred,
kHistoGreen,
@@ -362,11 +554,14 @@
} CrunchSubConfig;
typedef struct {
int entropy_idx_;
+ PaletteSorting palette_sorting_type_;
CrunchSubConfig sub_configs_[CRUNCH_SUBCONFIGS_MAX];
int sub_configs_size_;
} CrunchConfig;
-#define CRUNCH_CONFIGS_MAX kNumEntropyIx
+// +2 because we add a palette sorting configuration for kPalette and
+// kPaletteAndSpatial.
+#define CRUNCH_CONFIGS_MAX (kNumEntropyIx + 2)
static int EncoderAnalyze(VP8LEncoder* const enc,
CrunchConfig crunch_configs[CRUNCH_CONFIGS_MAX],
@@ -386,9 +581,15 @@
int do_no_cache = 0;
assert(pic != NULL && pic->argb != NULL);
- use_palette =
- AnalyzeAndCreatePalette(pic, low_effort,
- enc->palette_, &enc->palette_size_);
+ // Check whether a palette is possible.
+ enc->palette_size_ = WebPGetColorPalette(pic, enc->palette_sorted_);
+ use_palette = (enc->palette_size_ <= MAX_PALETTE_SIZE);
+ if (!use_palette) {
+ enc->palette_size_ = 0;
+ } else {
+ qsort(enc->palette_sorted_, enc->palette_size_,
+ sizeof(*enc->palette_sorted_), PaletteCompareColorsForQsort);
+ }
// Empirical bit sizes.
enc->histo_bits_ = GetHistoBits(method, use_palette,
@@ -398,6 +599,8 @@
if (low_effort) {
// AnalyzeEntropy is somewhat slow.
crunch_configs[0].entropy_idx_ = use_palette ? kPalette : kSpatialSubGreen;
+ crunch_configs[0].palette_sorting_type_ =
+ use_palette ? kSortedDefault : kUnusedPalette;
n_lz77s = 1;
*crunch_configs_size = 1;
} else {
@@ -418,13 +621,28 @@
// a palette.
if ((i != kPalette && i != kPaletteAndSpatial) || use_palette) {
assert(*crunch_configs_size < CRUNCH_CONFIGS_MAX);
- crunch_configs[(*crunch_configs_size)++].entropy_idx_ = i;
+ crunch_configs[(*crunch_configs_size)].entropy_idx_ = i;
+ if (use_palette && (i == kPalette || i == kPaletteAndSpatial)) {
+ crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+ kMinimizeDelta;
+ ++*crunch_configs_size;
+ // Also add modified Zeng's method.
+ crunch_configs[(*crunch_configs_size)].entropy_idx_ = i;
+ crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+ kModifiedZeng;
+ } else {
+ crunch_configs[(*crunch_configs_size)].palette_sorting_type_ =
+ kUnusedPalette;
+ }
+ ++*crunch_configs_size;
}
}
} else {
// Only choose the guessed best transform.
*crunch_configs_size = 1;
crunch_configs[0].entropy_idx_ = min_entropy_ix;
+ crunch_configs[0].palette_sorting_type_ =
+ use_palette ? kMinimizeDelta : kUnusedPalette;
if (config->quality >= 75 && method == 5) {
// Test with and without color cache.
do_no_cache = 1;
@@ -432,6 +650,7 @@
if (min_entropy_ix == kPalette) {
*crunch_configs_size = 2;
crunch_configs[1].entropy_idx_ = kPaletteAndSpatial;
+ crunch_configs[1].palette_sorting_type_ = kMinimizeDelta;
}
}
}
@@ -1283,22 +1502,6 @@
// -----------------------------------------------------------------------------
-static WEBP_INLINE int SearchColorNoIdx(const uint32_t sorted[], uint32_t color,
- int hi) {
- int low = 0;
- if (sorted[low] == color) return low; // loop invariant: sorted[low] != color
- while (1) {
- const int mid = (low + hi) >> 1;
- if (sorted[mid] == color) {
- return mid;
- } else if (sorted[mid] < color) {
- low = mid;
- } else {
- hi = mid;
- }
- }
-}
-
#define APPLY_PALETTE_GREEDY_MAX 4
static WEBP_INLINE uint32_t SearchColorGreedy(const uint32_t palette[],
@@ -1333,17 +1536,6 @@
(32 - PALETTE_INV_SIZE_BITS);
}
-// Sort palette in increasing order and prepare an inverse mapping array.
-static void PrepareMapToPalette(const uint32_t palette[], int num_colors,
- uint32_t sorted[], uint32_t idx_map[]) {
- int i;
- memcpy(sorted, palette, num_colors * sizeof(*sorted));
- qsort(sorted, num_colors, sizeof(*sorted), PaletteCompareColorsForQsort);
- for (i = 0; i < num_colors; ++i) {
- idx_map[SearchColorNoIdx(sorted, palette[i], num_colors)] = i;
- }
-}
-
// Use 1 pixel cache for ARGB pixels.
#define APPLY_PALETTE_FOR(COLOR_INDEX) do { \
uint32_t prev_pix = palette[0]; \
@@ -1571,7 +1763,8 @@
enc->use_predict_ = (entropy_idx == kSpatial) ||
(entropy_idx == kSpatialSubGreen) ||
(entropy_idx == kPaletteAndSpatial);
- if (low_effort) {
+ // When using a palette, R/B==0, hence no need to test for cross-color.
+ if (low_effort || enc->use_palette_) {
enc->use_cross_color_ = 0;
} else {
enc->use_cross_color_ = red_and_blue_always_zero ? 0 : enc->use_predict_;
@@ -1603,6 +1796,19 @@
// Encode palette
if (enc->use_palette_) {
+ if (crunch_configs[idx].palette_sorting_type_ == kSortedDefault) {
+ // Nothing to do, we have already sorted the palette.
+ memcpy(enc->palette_, enc->palette_sorted_,
+ enc->palette_size_ * sizeof(*enc->palette_));
+ } else if (crunch_configs[idx].palette_sorting_type_ == kMinimizeDelta) {
+ PaletteSortMinimizeDeltas(enc->palette_sorted_, enc->palette_size_,
+ enc->palette_);
+ } else {
+ assert(crunch_configs[idx].palette_sorting_type_ == kModifiedZeng);
+ err = PaletteSortModifiedZeng(enc->pic_, enc->palette_sorted_,
+ enc->palette_size_, enc->palette_);
+ if (err != VP8_ENC_OK) goto Error;
+ }
err = EncodePalette(bw, low_effort, enc);
if (err != VP8_ENC_OK) goto Error;
err = MapImageFromPalette(enc, use_delta_palette);
@@ -1767,6 +1973,8 @@
enc_side->palette_size_ = enc_main->palette_size_;
memcpy(enc_side->palette_, enc_main->palette_,
sizeof(enc_main->palette_));
+ memcpy(enc_side->palette_sorted_, enc_main->palette_sorted_,
+ sizeof(enc_main->palette_sorted_));
param->enc_ = enc_side;
}
// Create the workers.
diff --git a/src/enc/vp8li_enc.h b/src/enc/vp8li_enc.h
index 94210ce..00de489 100644
--- a/src/enc/vp8li_enc.h
+++ b/src/enc/vp8li_enc.h
@@ -69,6 +69,8 @@
int use_palette_;
int palette_size_;
uint32_t palette_[MAX_PALETTE_SIZE];
+ // Sorted version of palette_ for cache purposes.
+ uint32_t palette_sorted_[MAX_PALETTE_SIZE];
// Some 'scratch' (potentially large) objects.
struct VP8LBackwardRefs refs_[4]; // Backward Refs array for temporaries.
diff --git a/src/mux/anim_encode.c b/src/mux/anim_encode.c
index 7be9906..7078d9a 100644
--- a/src/mux/anim_encode.c
+++ b/src/mux/anim_encode.c
@@ -248,9 +248,6 @@
enc = (WebPAnimEncoder*)WebPSafeCalloc(1, sizeof(*enc));
if (enc == NULL) return NULL;
- // sanity inits, so we can call WebPAnimEncoderDelete():
- enc->encoded_frames_ = NULL;
- enc->mux_ = NULL;
MarkNoError(enc);
// Dimensions and options.
@@ -421,7 +418,7 @@
const int max_allowed_diff_lossy = QualityToMaxDiff(quality);
const int max_allowed_diff = is_lossless ? 0 : max_allowed_diff_lossy;
- // Sanity checks.
+ // Assumption/correctness checks.
assert(src->width == dst->width && src->height == dst->height);
assert(rect->x_offset_ + rect->width_ <= dst->width);
assert(rect->y_offset_ + rect->height_ <= dst->height);
@@ -949,7 +946,8 @@
int new_duration;
assert(enc->count_ >= 1);
- assert(prev_enc_frame->sub_frame_.duration ==
+ assert(!prev_enc_frame->is_key_frame_ ||
+ prev_enc_frame->sub_frame_.duration ==
prev_enc_frame->key_frame_.duration);
assert(prev_enc_frame->sub_frame_.duration ==
(prev_enc_frame->sub_frame_.duration & (MAX_DURATION - 1)));
@@ -966,7 +964,7 @@
0x10, 0x88, 0x88, 0x08
};
const WebPData lossless_1x1 = {
- lossless_1x1_bytes, sizeof(lossless_1x1_bytes)
+ lossless_1x1_bytes, sizeof(lossless_1x1_bytes)
};
const uint8_t lossy_1x1_bytes[] = {
0x52, 0x49, 0x46, 0x46, 0x40, 0x00, 0x00, 0x00, 0x57, 0x45, 0x42, 0x50,
@@ -1358,6 +1356,12 @@
if (!IncreasePreviousDuration(enc, (int)prev_frame_duration)) {
return 0;
}
+ // IncreasePreviousDuration() may add a frame to avoid exceeding
+ // MAX_DURATION which could cause CacheFrame() to over read encoded_frames_
+ // before the next flush.
+ if (enc->count_ == enc->size_ && !FlushFrames(enc)) {
+ return 0;
+ }
} else {
enc->first_timestamp_ = timestamp;
}
diff --git a/src/mux/muxedit.c b/src/mux/muxedit.c
index ccf14b2..02c3ede 100644
--- a/src/mux/muxedit.c
+++ b/src/mux/muxedit.c
@@ -235,7 +235,6 @@
WebPMuxImage wpi;
WebPMuxError err;
- // Sanity checks.
if (mux == NULL || bitstream == NULL || bitstream->bytes == NULL ||
bitstream->size > MAX_CHUNK_PAYLOAD) {
return WEBP_MUX_INVALID_ARGUMENT;
@@ -267,7 +266,6 @@
WebPMuxImage wpi;
WebPMuxError err;
- // Sanity checks.
if (mux == NULL || info == NULL) return WEBP_MUX_INVALID_ARGUMENT;
if (info->id != WEBP_CHUNK_ANMF) return WEBP_MUX_INVALID_ARGUMENT;
diff --git a/src/mux/muxi.h b/src/mux/muxi.h
index 2289822..330da66 100644
--- a/src/mux/muxi.h
+++ b/src/mux/muxi.h
@@ -29,7 +29,7 @@
#define MUX_MAJ_VERSION 1
#define MUX_MIN_VERSION 2
-#define MUX_REV_VERSION 0
+#define MUX_REV_VERSION 1
// Chunk object.
typedef struct WebPChunk WebPChunk;
diff --git a/src/mux/muxread.c b/src/mux/muxread.c
index 0101fde..8005039 100644
--- a/src/mux/muxread.c
+++ b/src/mux/muxread.c
@@ -56,7 +56,7 @@
uint32_t chunk_size;
WebPData chunk_data;
- // Sanity checks.
+ // Correctness checks.
if (data_size < CHUNK_HEADER_SIZE) return WEBP_MUX_NOT_ENOUGH_DATA;
chunk_size = GetLE32(data + TAG_SIZE);
if (chunk_size > MAX_CHUNK_PAYLOAD) return WEBP_MUX_BAD_DATA;
@@ -186,7 +186,6 @@
WebPChunk** chunk_list_ends[WEBP_CHUNK_NIL + 1] = { NULL };
ChunkInit(&chunk);
- // Sanity checks.
if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_MUX_ABI_VERSION)) {
return NULL; // version mismatch
}
@@ -481,7 +480,6 @@
WebPMuxError err;
WebPMuxImage* wpi;
- // Sanity checks.
if (mux == NULL || frame == NULL) {
return WEBP_MUX_INVALID_ARGUMENT;
}
diff --git a/src/utils/bit_reader_inl_utils.h b/src/utils/bit_reader_inl_utils.h
index 46b3880..404b9a6 100644
--- a/src/utils/bit_reader_inl_utils.h
+++ b/src/utils/bit_reader_inl_utils.h
@@ -55,7 +55,7 @@
// makes sure br->value_ has at least BITS bits worth of data
static WEBP_UBSAN_IGNORE_UNDEF WEBP_INLINE
-void VP8LoadNewBytes(VP8BitReader* const br) {
+void VP8LoadNewBytes(VP8BitReader* WEBP_RESTRICT const br) {
assert(br != NULL && br->buf_ != NULL);
// Read 'BITS' bits at a time if possible.
if (br->buf_ < br->buf_max_) {
@@ -104,7 +104,7 @@
}
// Read a bit with proba 'prob'. Speed-critical function!
-static WEBP_INLINE int VP8GetBit(VP8BitReader* const br,
+static WEBP_INLINE int VP8GetBit(VP8BitReader* WEBP_RESTRICT const br,
int prob, const char label[]) {
// Don't move this declaration! It makes a big speed difference to store
// 'range' *before* calling VP8LoadNewBytes(), even if this function doesn't
@@ -137,7 +137,8 @@
// simplified version of VP8GetBit() for prob=0x80 (note shift is always 1 here)
static WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW WEBP_INLINE
-int VP8GetSigned(VP8BitReader* const br, int v, const char label[]) {
+int VP8GetSigned(VP8BitReader* WEBP_RESTRICT const br, int v,
+ const char label[]) {
if (br->bits_ < 0) {
VP8LoadNewBytes(br);
}
@@ -155,7 +156,7 @@
}
}
-static WEBP_INLINE int VP8GetBitAlt(VP8BitReader* const br,
+static WEBP_INLINE int VP8GetBitAlt(VP8BitReader* WEBP_RESTRICT const br,
int prob, const char label[]) {
// Don't move this declaration! It makes a big speed difference to store
// 'range' *before* calling VP8LoadNewBytes(), even if this function doesn't
diff --git a/src/utils/bit_writer_utils.c b/src/utils/bit_writer_utils.c
index bef0e31..2f40850 100644
--- a/src/utils/bit_writer_utils.c
+++ b/src/utils/bit_writer_utils.c
@@ -278,7 +278,7 @@
// If needed, make some room by flushing some bits out.
if (bw->cur_ + VP8L_WRITER_BYTES > bw->end_) {
const uint64_t extra_size = (bw->end_ - bw->buf_) + MIN_EXTRA_SIZE;
- if (extra_size != (size_t)extra_size ||
+ if (!CheckSizeOverflow(extra_size) ||
!VP8LBitWriterResize(bw, (size_t)extra_size)) {
bw->cur_ = bw->buf_;
bw->error_ = 1;
@@ -314,7 +314,7 @@
while (used >= VP8L_WRITER_BITS) {
if (bw->cur_ + VP8L_WRITER_BYTES > bw->end_) {
const uint64_t extra_size = (bw->end_ - bw->buf_) + MIN_EXTRA_SIZE;
- if (extra_size != (size_t)extra_size ||
+ if (!CheckSizeOverflow(extra_size) ||
!VP8LBitWriterResize(bw, (size_t)extra_size)) {
bw->cur_ = bw->buf_;
bw->error_ = 1;
diff --git a/src/utils/color_cache_utils.c b/src/utils/color_cache_utils.c
index b09f538..7b5222b 100644
--- a/src/utils/color_cache_utils.c
+++ b/src/utils/color_cache_utils.c
@@ -20,22 +20,22 @@
//------------------------------------------------------------------------------
// VP8LColorCache.
-int VP8LColorCacheInit(VP8LColorCache* const cc, int hash_bits) {
+int VP8LColorCacheInit(VP8LColorCache* const color_cache, int hash_bits) {
const int hash_size = 1 << hash_bits;
- assert(cc != NULL);
+ assert(color_cache != NULL);
assert(hash_bits > 0);
- cc->colors_ = (uint32_t*)WebPSafeCalloc((uint64_t)hash_size,
- sizeof(*cc->colors_));
- if (cc->colors_ == NULL) return 0;
- cc->hash_shift_ = 32 - hash_bits;
- cc->hash_bits_ = hash_bits;
+ color_cache->colors_ = (uint32_t*)WebPSafeCalloc(
+ (uint64_t)hash_size, sizeof(*color_cache->colors_));
+ if (color_cache->colors_ == NULL) return 0;
+ color_cache->hash_shift_ = 32 - hash_bits;
+ color_cache->hash_bits_ = hash_bits;
return 1;
}
-void VP8LColorCacheClear(VP8LColorCache* const cc) {
- if (cc != NULL) {
- WebPSafeFree(cc->colors_);
- cc->colors_ = NULL;
+void VP8LColorCacheClear(VP8LColorCache* const color_cache) {
+ if (color_cache != NULL) {
+ WebPSafeFree(color_cache->colors_);
+ color_cache->colors_ = NULL;
}
}
diff --git a/src/utils/huffman_encode_utils.c b/src/utils/huffman_encode_utils.c
index 6f3b1bb..fd7a47d 100644
--- a/src/utils/huffman_encode_utils.c
+++ b/src/utils/huffman_encode_utils.c
@@ -404,8 +404,7 @@
// Main entry point
void VP8LCreateHuffmanTree(uint32_t* const histogram, int tree_depth_limit,
- uint8_t* const buf_rle,
- HuffmanTree* const huff_tree,
+ uint8_t* const buf_rle, HuffmanTree* const huff_tree,
HuffmanTreeCode* const huff_code) {
const int num_symbols = huff_code->num_symbols;
memset(buf_rle, 0, num_symbols * sizeof(*buf_rle));
diff --git a/src/utils/huffman_encode_utils.h b/src/utils/huffman_encode_utils.h
index 3e6763c..3f7f1d8 100644
--- a/src/utils/huffman_encode_utils.h
+++ b/src/utils/huffman_encode_utils.h
@@ -51,7 +51,7 @@
// huffman code tree.
void VP8LCreateHuffmanTree(uint32_t* const histogram, int tree_depth_limit,
uint8_t* const buf_rle, HuffmanTree* const huff_tree,
- HuffmanTreeCode* const tree);
+ HuffmanTreeCode* const huff_code);
#ifdef __cplusplus
}
diff --git a/src/utils/rescaler_utils.c b/src/utils/rescaler_utils.c
index 4bcae24..a0581a1 100644
--- a/src/utils/rescaler_utils.c
+++ b/src/utils/rescaler_utils.c
@@ -12,66 +12,74 @@
// Author: Skal (pascal.massimino@gmail.com)
#include <assert.h>
+#include <limits.h>
#include <stdlib.h>
#include <string.h>
#include "src/dsp/dsp.h"
#include "src/utils/rescaler_utils.h"
+#include "src/utils/utils.h"
//------------------------------------------------------------------------------
-void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
- uint8_t* const dst,
- int dst_width, int dst_height, int dst_stride,
- int num_channels, rescaler_t* const work) {
+int WebPRescalerInit(WebPRescaler* const rescaler,
+ int src_width, int src_height,
+ uint8_t* const dst,
+ int dst_width, int dst_height, int dst_stride,
+ int num_channels, rescaler_t* const work) {
const int x_add = src_width, x_sub = dst_width;
const int y_add = src_height, y_sub = dst_height;
- wrk->x_expand = (src_width < dst_width);
- wrk->y_expand = (src_height < dst_height);
- wrk->src_width = src_width;
- wrk->src_height = src_height;
- wrk->dst_width = dst_width;
- wrk->dst_height = dst_height;
- wrk->src_y = 0;
- wrk->dst_y = 0;
- wrk->dst = dst;
- wrk->dst_stride = dst_stride;
- wrk->num_channels = num_channels;
+ const uint64_t total_size = 2ull * dst_width * num_channels * sizeof(*work);
+ if (!CheckSizeOverflow(total_size)) return 0;
+
+ rescaler->x_expand = (src_width < dst_width);
+ rescaler->y_expand = (src_height < dst_height);
+ rescaler->src_width = src_width;
+ rescaler->src_height = src_height;
+ rescaler->dst_width = dst_width;
+ rescaler->dst_height = dst_height;
+ rescaler->src_y = 0;
+ rescaler->dst_y = 0;
+ rescaler->dst = dst;
+ rescaler->dst_stride = dst_stride;
+ rescaler->num_channels = num_channels;
// for 'x_expand', we use bilinear interpolation
- wrk->x_add = wrk->x_expand ? (x_sub - 1) : x_add;
- wrk->x_sub = wrk->x_expand ? (x_add - 1) : x_sub;
- if (!wrk->x_expand) { // fx_scale is not used otherwise
- wrk->fx_scale = WEBP_RESCALER_FRAC(1, wrk->x_sub);
+ rescaler->x_add = rescaler->x_expand ? (x_sub - 1) : x_add;
+ rescaler->x_sub = rescaler->x_expand ? (x_add - 1) : x_sub;
+ if (!rescaler->x_expand) { // fx_scale is not used otherwise
+ rescaler->fx_scale = WEBP_RESCALER_FRAC(1, rescaler->x_sub);
}
// vertical scaling parameters
- wrk->y_add = wrk->y_expand ? y_add - 1 : y_add;
- wrk->y_sub = wrk->y_expand ? y_sub - 1 : y_sub;
- wrk->y_accum = wrk->y_expand ? wrk->y_sub : wrk->y_add;
- if (!wrk->y_expand) {
+ rescaler->y_add = rescaler->y_expand ? y_add - 1 : y_add;
+ rescaler->y_sub = rescaler->y_expand ? y_sub - 1 : y_sub;
+ rescaler->y_accum = rescaler->y_expand ? rescaler->y_sub : rescaler->y_add;
+ if (!rescaler->y_expand) {
// This is WEBP_RESCALER_FRAC(dst_height, x_add * y_add) without the cast.
- // Its value is <= WEBP_RESCALER_ONE, because dst_height <= wrk->y_add, and
- // wrk->x_add >= 1;
- const uint64_t ratio =
- (uint64_t)dst_height * WEBP_RESCALER_ONE / (wrk->x_add * wrk->y_add);
+ // Its value is <= WEBP_RESCALER_ONE, because dst_height <= rescaler->y_add
+ // and rescaler->x_add >= 1;
+ const uint64_t num = (uint64_t)dst_height * WEBP_RESCALER_ONE;
+ const uint64_t den = (uint64_t)rescaler->x_add * rescaler->y_add;
+ const uint64_t ratio = num / den;
if (ratio != (uint32_t)ratio) {
// When ratio == WEBP_RESCALER_ONE, we can't represent the ratio with the
// current fixed-point precision. This happens when src_height ==
- // wrk->y_add (which == src_height), and wrk->x_add == 1.
+ // rescaler->y_add (which == src_height), and rescaler->x_add == 1.
// => We special-case fxy_scale = 0, in WebPRescalerExportRow().
- wrk->fxy_scale = 0;
+ rescaler->fxy_scale = 0;
} else {
- wrk->fxy_scale = (uint32_t)ratio;
+ rescaler->fxy_scale = (uint32_t)ratio;
}
- wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->y_sub);
+ rescaler->fy_scale = WEBP_RESCALER_FRAC(1, rescaler->y_sub);
} else {
- wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->x_add);
- // wrk->fxy_scale is unused here.
+ rescaler->fy_scale = WEBP_RESCALER_FRAC(1, rescaler->x_add);
+ // rescaler->fxy_scale is unused here.
}
- wrk->irow = work;
- wrk->frow = work + num_channels * dst_width;
- memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
+ rescaler->irow = work;
+ rescaler->frow = work + num_channels * dst_width;
+ memset(work, 0, (size_t)total_size);
WebPRescalerDspInit();
+ return 1;
}
int WebPRescalerGetScaledDimensions(int src_width, int src_height,
@@ -82,6 +90,7 @@
{
int width = *scaled_width;
int height = *scaled_height;
+ const int max_size = INT_MAX / 2;
// if width is unspecified, scale original proportionally to height ratio.
if (width == 0 && src_height > 0) {
@@ -94,7 +103,7 @@
(int)(((uint64_t)src_height * width + src_width - 1) / src_width);
}
// Check if the overall dimensions still make sense.
- if (width <= 0 || height <= 0) {
+ if (width <= 0 || height <= 0 || width > max_size || height > max_size) {
return 0;
}
@@ -107,31 +116,34 @@
//------------------------------------------------------------------------------
// all-in-one calls
-int WebPRescaleNeededLines(const WebPRescaler* const wrk, int max_num_lines) {
- const int num_lines = (wrk->y_accum + wrk->y_sub - 1) / wrk->y_sub;
+int WebPRescaleNeededLines(const WebPRescaler* const rescaler,
+ int max_num_lines) {
+ const int num_lines =
+ (rescaler->y_accum + rescaler->y_sub - 1) / rescaler->y_sub;
return (num_lines > max_num_lines) ? max_num_lines : num_lines;
}
-int WebPRescalerImport(WebPRescaler* const wrk, int num_lines,
+int WebPRescalerImport(WebPRescaler* const rescaler, int num_lines,
const uint8_t* src, int src_stride) {
int total_imported = 0;
- while (total_imported < num_lines && !WebPRescalerHasPendingOutput(wrk)) {
- if (wrk->y_expand) {
- rescaler_t* const tmp = wrk->irow;
- wrk->irow = wrk->frow;
- wrk->frow = tmp;
+ while (total_imported < num_lines &&
+ !WebPRescalerHasPendingOutput(rescaler)) {
+ if (rescaler->y_expand) {
+ rescaler_t* const tmp = rescaler->irow;
+ rescaler->irow = rescaler->frow;
+ rescaler->frow = tmp;
}
- WebPRescalerImportRow(wrk, src);
- if (!wrk->y_expand) { // Accumulate the contribution of the new row.
+ WebPRescalerImportRow(rescaler, src);
+ if (!rescaler->y_expand) { // Accumulate the contribution of the new row.
int x;
- for (x = 0; x < wrk->num_channels * wrk->dst_width; ++x) {
- wrk->irow[x] += wrk->frow[x];
+ for (x = 0; x < rescaler->num_channels * rescaler->dst_width; ++x) {
+ rescaler->irow[x] += rescaler->frow[x];
}
}
- ++wrk->src_y;
+ ++rescaler->src_y;
src += src_stride;
++total_imported;
- wrk->y_accum -= wrk->y_sub;
+ rescaler->y_accum -= rescaler->y_sub;
}
return total_imported;
}
diff --git a/src/utils/rescaler_utils.h b/src/utils/rescaler_utils.h
index ca41e42..ef201ef 100644
--- a/src/utils/rescaler_utils.h
+++ b/src/utils/rescaler_utils.h
@@ -47,12 +47,13 @@
};
// Initialize a rescaler given scratch area 'work' and dimensions of src & dst.
-void WebPRescalerInit(WebPRescaler* const rescaler,
- int src_width, int src_height,
- uint8_t* const dst,
- int dst_width, int dst_height, int dst_stride,
- int num_channels,
- rescaler_t* const work);
+// Returns false in case of error.
+int WebPRescalerInit(WebPRescaler* const rescaler,
+ int src_width, int src_height,
+ uint8_t* const dst,
+ int dst_width, int dst_height, int dst_stride,
+ int num_channels,
+ rescaler_t* const work);
// If either 'scaled_width' or 'scaled_height' (but not both) is 0 the value
// will be calculated preserving the aspect ratio, otherwise the values are
diff --git a/src/utils/utils.c b/src/utils/utils.c
index 6080e19..9e464c1 100644
--- a/src/utils/utils.c
+++ b/src/utils/utils.c
@@ -101,6 +101,9 @@
#if defined(MALLOC_LIMIT)
{
const char* const malloc_limit_str = getenv("MALLOC_LIMIT");
+#if MALLOC_LIMIT > 1
+ mem_limit = (size_t)MALLOC_LIMIT;
+#endif
if (malloc_limit_str != NULL) {
mem_limit = atoi(malloc_limit_str);
}
@@ -169,16 +172,16 @@
const uint64_t total_size = nmemb * size;
if (nmemb == 0) return 1;
if ((uint64_t)size > WEBP_MAX_ALLOCABLE_MEMORY / nmemb) return 0;
- if (total_size != (size_t)total_size) return 0;
+ if (!CheckSizeOverflow(total_size)) return 0;
#if defined(PRINT_MEM_INFO) && defined(MALLOC_FAIL_AT)
if (countdown_to_fail > 0 && --countdown_to_fail == 0) {
return 0; // fake fail!
}
#endif
-#if defined(MALLOC_LIMIT)
+#if defined(PRINT_MEM_INFO) && defined(MALLOC_LIMIT)
if (mem_limit > 0) {
const uint64_t new_total_mem = (uint64_t)total_mem + total_size;
- if (new_total_mem != (size_t)new_total_mem ||
+ if (!CheckSizeOverflow(new_total_mem) ||
new_total_mem > mem_limit) {
return 0; // fake fail!
}
diff --git a/src/utils/utils.h b/src/utils/utils.h
index 2a3ec92..ef04f10 100644
--- a/src/utils/utils.h
+++ b/src/utils/utils.h
@@ -42,6 +42,10 @@
#endif
#endif // WEBP_MAX_ALLOCABLE_MEMORY
+static WEBP_INLINE int CheckSizeOverflow(uint64_t size) {
+ return size == (size_t)size;
+}
+
// size-checking safe malloc/calloc: verify that the requested size is not too
// large, or return NULL. You don't need to call these for constructs like
// malloc(sizeof(foo)), but only if there's picture-dependent size involved
@@ -107,24 +111,33 @@
PutLE16(data + 2, (int)(val >> 16));
}
-// Returns (int)floor(log2(n)). n must be > 0.
// use GNU builtins where available.
#if defined(__GNUC__) && \
((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
+// Returns (int)floor(log2(n)). n must be > 0.
static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
return 31 ^ __builtin_clz(n);
}
+// counts the number of trailing zero
+static WEBP_INLINE int BitsCtz(uint32_t n) { return __builtin_ctz(n); }
#elif defined(_MSC_VER) && _MSC_VER > 1310 && \
(defined(_M_X64) || defined(_M_IX86))
#include <intrin.h>
#pragma intrinsic(_BitScanReverse)
+#pragma intrinsic(_BitScanForward)
static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
- unsigned long first_set_bit;
+ unsigned long first_set_bit; // NOLINT (runtime/int)
_BitScanReverse(&first_set_bit, n);
return first_set_bit;
}
-#else // default: use the C-version.
+static WEBP_INLINE int BitsCtz(uint32_t n) {
+ unsigned long first_set_bit; // NOLINT (runtime/int)
+ _BitScanForward(&first_set_bit, n);
+ return first_set_bit;
+}
+#else // default: use the (slow) C-version.
+#define WEBP_HAVE_SLOW_CLZ_CTZ // signal that the Clz/Ctz function are slow
// Returns 31 ^ clz(n) = log2(n). This is the default C-implementation, either
// based on table or not. Can be used as fallback if clz() is not available.
#define WEBP_NEED_LOG_TABLE_8BIT
@@ -139,6 +152,15 @@
}
static WEBP_INLINE int BitsLog2Floor(uint32_t n) { return WebPLog2FloorC(n); }
+
+static WEBP_INLINE int BitsCtz(uint32_t n) {
+ int i;
+ for (i = 0; i < 32; ++i, n >>= 1) {
+ if (n & 1) return i;
+ }
+ return 32;
+}
+
#endif
//------------------------------------------------------------------------------
