Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(386)

Issues Search
    My Issues | Starred     Open | Closed | All

Unified Diff: third_party/libwebp/dsp/lossless_enc.c

Issue 1546003002: libwebp: update to 0.5.0 (Closed) Base URL: https://chromium.googlesource.com/chromium/src.git@master
Patch Set: rebase Created 5 years ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
« no previous file with comments | « third_party/libwebp/dsp/lossless.c ('k') | third_party/libwebp/dsp/lossless_enc_mips32.c » ('j') | no next file with comments »
Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
Index: third_party/libwebp/dsp/lossless_enc.c
diff --git a/third_party/libwebp/dsp/lossless.c b/third_party/libwebp/dsp/lossless_enc.c
similarity index 53%
copy from third_party/libwebp/dsp/lossless.c
copy to third_party/libwebp/dsp/lossless_enc.c
index ee334bceb0b9669bc7791c21ef2da255eb9445c7..2eafa3da7d23d69bc7e084e0e0d519b9bbd0a9e1 100644
--- a/third_party/libwebp/dsp/lossless.c
+++ b/third_party/libwebp/dsp/lossless_enc.c
@@ -1,4 +1,4 @@
-// Copyright 2012 Google Inc. All Rights Reserved.
+// Copyright 2015 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
@@ -7,7 +7,7 @@
// be found in the AUTHORS file in the root of the source tree.
// -----------------------------------------------------------------------------
//
-// Image transforms and color space conversion methods for lossless decoder.
+// Image transform methods for lossless encoder.
//
// Authors: Vikas Arora (vikaas.arora@gmail.com)
// Jyrki Alakuijala (jyrki@google.com)
@@ -24,6 +24,9 @@
#define MAX_DIFF_COST (1e30f)
+static const int kPredLowEffort = 11;
+static const uint32_t kMaskAlpha = 0xff000000;
+
// lookup table for small values of log2(int)
const float kLog2Table[LOG_LOOKUP_IDX_MAX] = {
0.0000000000000000f, 0.0000000000000000f,
@@ -326,13 +329,6 @@ const uint8_t kPrefixEncodeExtraBitsValue[PREFIX_LOOKUP_IDX_MAX] = {
112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126
};
-// The threshold till approximate version of log_2 can be used.
-// Practically, we can get rid of the call to log() as the two values match to
-// very high degree (the ratio of these two is 0.99999x).
-// Keeping a high threshold for now.
-#define APPROX_LOG_WITH_CORRECTION_MAX 65536
-#define APPROX_LOG_MAX 4096
-#define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
static float FastSLog2Slow(uint32_t v) {
assert(v >= LOG_LOOKUP_IDX_MAX);
if (v < APPROX_LOG_WITH_CORRECTION_MAX) {
@@ -384,166 +380,11 @@ static float FastLog2Slow(uint32_t v) {
}
}
-//------------------------------------------------------------------------------
-// Image transforms.
-
// Mostly used to reduce code size + readability
static WEBP_INLINE int GetMin(int a, int b) { return (a > b) ? b : a; }
-// In-place sum of each component with mod 256.
-static WEBP_INLINE void AddPixelsEq(uint32_t* a, uint32_t b) {
- const uint32_t alpha_and_green = (*a & 0xff00ff00u) + (b & 0xff00ff00u);
- const uint32_t red_and_blue = (*a & 0x00ff00ffu) + (b & 0x00ff00ffu);
- *a = (alpha_and_green & 0xff00ff00u) | (red_and_blue & 0x00ff00ffu);
-}
-
-static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) {
- return (((a0 ^ a1) & 0xfefefefeL) >> 1) + (a0 & a1);
-}
-
-static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) {
- return Average2(Average2(a0, a2), a1);
-}
-
-static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
- uint32_t a2, uint32_t a3) {
- return Average2(Average2(a0, a1), Average2(a2, a3));
-}
-
-static WEBP_INLINE uint32_t Clip255(uint32_t a) {
- if (a < 256) {
- return a;
- }
- // return 0, when a is a negative integer.
- // return 255, when a is positive.
- return ~a >> 24;
-}
-
-static WEBP_INLINE int AddSubtractComponentFull(int a, int b, int c) {
- return Clip255(a + b - c);
-}
-
-static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
- uint32_t c2) {
- const int a = AddSubtractComponentFull(c0 >> 24, c1 >> 24, c2 >> 24);
- const int r = AddSubtractComponentFull((c0 >> 16) & 0xff,
- (c1 >> 16) & 0xff,
- (c2 >> 16) & 0xff);
- const int g = AddSubtractComponentFull((c0 >> 8) & 0xff,
- (c1 >> 8) & 0xff,
- (c2 >> 8) & 0xff);
- const int b = AddSubtractComponentFull(c0 & 0xff, c1 & 0xff, c2 & 0xff);
- return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b;
-}
-
-static WEBP_INLINE int AddSubtractComponentHalf(int a, int b) {
- return Clip255(a + (a - b) / 2);
-}
-
-static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
- uint32_t c2) {
- const uint32_t ave = Average2(c0, c1);
- const int a = AddSubtractComponentHalf(ave >> 24, c2 >> 24);
- const int r = AddSubtractComponentHalf((ave >> 16) & 0xff, (c2 >> 16) & 0xff);
- const int g = AddSubtractComponentHalf((ave >> 8) & 0xff, (c2 >> 8) & 0xff);
- const int b = AddSubtractComponentHalf((ave >> 0) & 0xff, (c2 >> 0) & 0xff);
- return ((uint32_t)a << 24) | (r << 16) | (g << 8) | b;
-}
-
-// gcc-4.9 on ARM generates incorrect code in Select() when Sub3() is inlined.
-#if defined(__arm__) && LOCAL_GCC_VERSION == 0x409
-# define LOCAL_INLINE __attribute__ ((noinline))
-#else
-# define LOCAL_INLINE WEBP_INLINE
-#endif
-
-static LOCAL_INLINE int Sub3(int a, int b, int c) {
- const int pb = b - c;
- const int pa = a - c;
- return abs(pb) - abs(pa);
-}
-
-#undef LOCAL_INLINE
-
-static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
- const int pa_minus_pb =
- Sub3((a >> 24) , (b >> 24) , (c >> 24) ) +
- Sub3((a >> 16) & 0xff, (b >> 16) & 0xff, (c >> 16) & 0xff) +
- Sub3((a >> 8) & 0xff, (b >> 8) & 0xff, (c >> 8) & 0xff) +
- Sub3((a ) & 0xff, (b ) & 0xff, (c ) & 0xff);
- return (pa_minus_pb <= 0) ? a : b;
-}
-
//------------------------------------------------------------------------------
-// Predictors
-
-static uint32_t Predictor0(uint32_t left, const uint32_t* const top) {
- (void)top;
- (void)left;
- return ARGB_BLACK;
-}
-static uint32_t Predictor1(uint32_t left, const uint32_t* const top) {
- (void)top;
- return left;
-}
-static uint32_t Predictor2(uint32_t left, const uint32_t* const top) {
- (void)left;
- return top[0];
-}
-static uint32_t Predictor3(uint32_t left, const uint32_t* const top) {
- (void)left;
- return top[1];
-}
-static uint32_t Predictor4(uint32_t left, const uint32_t* const top) {
- (void)left;
- return top[-1];
-}
-static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average3(left, top[0], top[1]);
- return pred;
-}
-static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average2(left, top[-1]);
- return pred;
-}
-static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average2(left, top[0]);
- return pred;
-}
-static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average2(top[-1], top[0]);
- (void)left;
- return pred;
-}
-static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average2(top[0], top[1]);
- (void)left;
- return pred;
-}
-static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Average4(left, top[-1], top[0], top[1]);
- return pred;
-}
-static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = Select(top[0], left, top[-1]);
- return pred;
-}
-static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]);
- return pred;
-}
-static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
- const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]);
- return pred;
-}
-
-static const VP8LPredictorFunc kPredictorsC[16] = {
- Predictor0, Predictor1, Predictor2, Predictor3,
- Predictor4, Predictor5, Predictor6, Predictor7,
- Predictor8, Predictor9, Predictor10, Predictor11,
- Predictor12, Predictor13,
- Predictor0, Predictor0 // <- padding security sentinels
-};
+// Methods to calculate Entropy (Shannon).
static float PredictionCostSpatial(const int counts[256], int weight_0,
double exp_val) {
@@ -565,15 +406,15 @@ static float CombinedShannonEntropy(const int X[256], const int Y[256]) {
int sumX = 0, sumXY = 0;
for (i = 0; i < 256; ++i) {
const int x = X[i];
- const int xy = x + Y[i];
if (x != 0) {
+ const int xy = x + Y[i];
sumX += x;
retval -= VP8LFastSLog2(x);
sumXY += xy;
retval -= VP8LFastSLog2(xy);
- } else if (xy != 0) {
- sumXY += xy;
- retval -= VP8LFastSLog2(xy);
+ } else if (Y[i] != 0) {
+ sumXY += Y[i];
+ retval -= VP8LFastSLog2(Y[i]);
}
}
retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
@@ -587,11 +428,107 @@ static float PredictionCostSpatialHistogram(const int accumulated[4][256],
for (i = 0; i < 4; ++i) {
const double kExpValue = 0.94;
retval += PredictionCostSpatial(tile[i], 1, kExpValue);
- retval += CombinedShannonEntropy(tile[i], accumulated[i]);
+ retval += VP8LCombinedShannonEntropy(tile[i], accumulated[i]);
}
return (float)retval;
}
+void VP8LBitEntropyInit(VP8LBitEntropy* const entropy) {
+ entropy->entropy = 0.;
+ entropy->sum = 0;
+ entropy->nonzeros = 0;
+ entropy->max_val = 0;
+ entropy->nonzero_code = VP8L_NON_TRIVIAL_SYM;
+}
+
+void VP8LBitsEntropyUnrefined(const uint32_t* const array, int n,
+ VP8LBitEntropy* const entropy) {
+ int i;
+
+ VP8LBitEntropyInit(entropy);
+
+ for (i = 0; i < n; ++i) {
+ if (array[i] != 0) {
+ entropy->sum += array[i];
+ entropy->nonzero_code = i;
+ ++entropy->nonzeros;
+ entropy->entropy -= VP8LFastSLog2(array[i]);
+ if (entropy->max_val < array[i]) {
+ entropy->max_val = array[i];
+ }
+ }
+ }
+ entropy->entropy += VP8LFastSLog2(entropy->sum);
+}
+
+static WEBP_INLINE void GetEntropyUnrefinedHelper(
+ uint32_t val, int i, uint32_t* const val_prev, int* const i_prev,
+ VP8LBitEntropy* const bit_entropy, VP8LStreaks* const stats) {
+ const int streak = i - *i_prev;
+
+ // Gather info for the bit entropy.
+ if (*val_prev != 0) {
+ bit_entropy->sum += (*val_prev) * streak;
+ bit_entropy->nonzeros += streak;
+ bit_entropy->nonzero_code = *i_prev;
+ bit_entropy->entropy -= VP8LFastSLog2(*val_prev) * streak;
+ if (bit_entropy->max_val < *val_prev) {
+ bit_entropy->max_val = *val_prev;
+ }
+ }
+
+ // Gather info for the Huffman cost.
+ stats->counts[*val_prev != 0] += (streak > 3);
+ stats->streaks[*val_prev != 0][(streak > 3)] += streak;
+
+ *val_prev = val;
+ *i_prev = i;
+}
+
+void VP8LGetEntropyUnrefined(const uint32_t* const X, int length,
+ VP8LBitEntropy* const bit_entropy,
+ VP8LStreaks* const stats) {
+ int i;
+ int i_prev = 0;
+ uint32_t x_prev = X[0];
+
+ memset(stats, 0, sizeof(*stats));
+ VP8LBitEntropyInit(bit_entropy);
+
+ for (i = 1; i < length; ++i) {
+ const uint32_t x = X[i];
+ if (x != x_prev) {
+ VP8LGetEntropyUnrefinedHelper(x, i, &x_prev, &i_prev, bit_entropy, stats);
+ }
+ }
+ VP8LGetEntropyUnrefinedHelper(0, i, &x_prev, &i_prev, bit_entropy, stats);
+
+ bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum);
+}
+
+void VP8LGetCombinedEntropyUnrefined(const uint32_t* const X,
+ const uint32_t* const Y, int length,
+ VP8LBitEntropy* const bit_entropy,
+ VP8LStreaks* const stats) {
+ int i = 1;
+ int i_prev = 0;
+ uint32_t xy_prev = X[0] + Y[0];
+
+ memset(stats, 0, sizeof(*stats));
+ VP8LBitEntropyInit(bit_entropy);
+
+ for (i = 1; i < length; ++i) {
+ const uint32_t xy = X[i] + Y[i];
+ if (xy != xy_prev) {
+ VP8LGetEntropyUnrefinedHelper(xy, i, &xy_prev, &i_prev, bit_entropy,
+ stats);
+ }
+ }
+ VP8LGetEntropyUnrefinedHelper(0, i, &xy_prev, &i_prev, bit_entropy, stats);
+
+ bit_entropy->entropy += VP8LFastSLog2(bit_entropy->sum);
+}
+
static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) {
++histo_argb[0][argb >> 24];
++histo_argb[1][(argb >> 16) & 0xff];
@@ -599,10 +536,27 @@ static WEBP_INLINE void UpdateHisto(int histo_argb[4][256], uint32_t argb) {
++histo_argb[3][argb & 0xff];
}
+//------------------------------------------------------------------------------
+
+static WEBP_INLINE uint32_t Predict(VP8LPredictorFunc pred_func,
+ int x, int y,
+ const uint32_t* current_row,
+ const uint32_t* upper_row) {
+ if (y == 0) {
+ return (x == 0) ? ARGB_BLACK : current_row[x - 1]; // Left.
+ } else if (x == 0) {
+ return upper_row[x]; // Top.
+ } else {
+ return pred_func(current_row[x - 1], upper_row + x);
+ }
+}
+
+// Returns best predictor and updates the accumulated histogram.
static int GetBestPredictorForTile(int width, int height,
int tile_x, int tile_y, int bits,
- const int accumulated[4][256],
- const uint32_t* const argb_scratch) {
+ int accumulated[4][256],
+ const uint32_t* const argb_scratch,
+ int exact) {
const int kNumPredModes = 14;
const int col_start = tile_x << bits;
const int row_start = tile_y << bits;
@@ -612,13 +566,19 @@ static int GetBestPredictorForTile(int width, int height,
float best_diff = MAX_DIFF_COST;
int best_mode = 0;
int mode;
+ int histo_stack_1[4][256];
+ int histo_stack_2[4][256];
+ // Need pointers to be able to swap arrays.
+ int (*histo_argb)[256] = histo_stack_1;
+ int (*best_histo)[256] = histo_stack_2;
+
+ int i, j;
for (mode = 0; mode < kNumPredModes; ++mode) {
const uint32_t* current_row = argb_scratch;
const VP8LPredictorFunc pred_func = VP8LPredictors[mode];
float cur_diff;
int y;
- int histo_argb[4][256];
- memset(histo_argb, 0, sizeof(histo_argb));
+ memset(histo_argb, 0, sizeof(histo_stack_1));
for (y = 0; y < max_y; ++y) {
int x;
const int row = row_start + y;
@@ -626,65 +586,93 @@ static int GetBestPredictorForTile(int width, int height,
current_row = upper_row + width;
for (x = 0; x < max_x; ++x) {
const int col = col_start + x;
- uint32_t predict;
- if (row == 0) {
- predict = (col == 0) ? ARGB_BLACK : current_row[col - 1]; // Left.
- } else if (col == 0) {
- predict = upper_row[col]; // Top.
- } else {
- predict = pred_func(current_row[col - 1], upper_row + col);
+ const uint32_t predict =
+ Predict(pred_func, col, row, current_row, upper_row);
+ uint32_t residual = VP8LSubPixels(current_row[col], predict);
+ if (!exact && (current_row[col] & kMaskAlpha) == 0) {
+ residual &= kMaskAlpha; // See CopyTileWithPrediction.
}
- UpdateHisto(histo_argb, VP8LSubPixels(current_row[col], predict));
+ UpdateHisto(histo_argb, residual);
}
}
cur_diff = PredictionCostSpatialHistogram(
- accumulated, (const int (*)[256])histo_argb);
+ (const int (*)[256])accumulated, (const int (*)[256])histo_argb);
if (cur_diff < best_diff) {
+ int (*tmp)[256] = histo_argb;
+ histo_argb = best_histo;
+ best_histo = tmp;
best_diff = cur_diff;
best_mode = mode;
}
}
+ for (i = 0; i < 4; i++) {
+ for (j = 0; j < 256; j++) {
+ accumulated[i][j] += best_histo[i][j];
+ }
+ }
+
return best_mode;
}
-static void CopyTileWithPrediction(int width, int height,
- int tile_x, int tile_y, int bits, int mode,
- const uint32_t* const argb_scratch,
- uint32_t* const argb) {
- const int col_start = tile_x << bits;
- const int row_start = tile_y << bits;
- const int tile_size = 1 << bits;
- const int max_y = GetMin(tile_size, height - row_start);
- const int max_x = GetMin(tile_size, width - col_start);
- const VP8LPredictorFunc pred_func = VP8LPredictors[mode];
- const uint32_t* current_row = argb_scratch;
-
+static void CopyImageWithPrediction(int width, int height,
+ int bits, uint32_t* const modes,
+ uint32_t* const argb_scratch,
+ uint32_t* const argb,
+ int low_effort, int exact) {
+ const int tiles_per_row = VP8LSubSampleSize(width, bits);
+ const int mask = (1 << bits) - 1;
+ // The row size is one pixel longer to allow the top right pixel to point to
+ // the leftmost pixel of the next row when at the right edge.
+ uint32_t* current_row = argb_scratch;
+ uint32_t* upper_row = argb_scratch + width + 1;
int y;
- for (y = 0; y < max_y; ++y) {
+ VP8LPredictorFunc pred_func =
+ low_effort ? VP8LPredictors[kPredLowEffort] : NULL;
+
+ for (y = 0; y < height; ++y) {
int x;
- const int row = row_start + y;
- const uint32_t* const upper_row = current_row;
- current_row = upper_row + width;
- for (x = 0; x < max_x; ++x) {
- const int col = col_start + x;
- const int pix = row * width + col;
- uint32_t predict;
- if (row == 0) {
- predict = (col == 0) ? ARGB_BLACK : current_row[col - 1]; // Left.
- } else if (col == 0) {
- predict = upper_row[col]; // Top.
- } else {
- predict = pred_func(current_row[col - 1], upper_row + col);
+ uint32_t* tmp = upper_row;
+ upper_row = current_row;
+ current_row = tmp;
+ memcpy(current_row, argb + y * width, sizeof(*current_row) * width);
+ current_row[width] = (y + 1 < height) ? argb[(y + 1) * width] : ARGB_BLACK;
+
+ if (low_effort) {
+ for (x = 0; x < width; ++x) {
+ const uint32_t predict =
+ Predict(pred_func, x, y, current_row, upper_row);
+ argb[y * width + x] = VP8LSubPixels(current_row[x], predict);
+ }
+ } else {
+ for (x = 0; x < width; ++x) {
+ uint32_t predict, residual;
+ if ((x & mask) == 0) {
+ const int mode =
+ (modes[(y >> bits) * tiles_per_row + (x >> bits)] >> 8) & 0xff;
+ pred_func = VP8LPredictors[mode];
+ }
+ predict = Predict(pred_func, x, y, current_row, upper_row);
+ residual = VP8LSubPixels(current_row[x], predict);
+ if (!exact && (current_row[x] & kMaskAlpha) == 0) {
+ // If alpha is 0, cleanup RGB. We can choose the RGB values of the
+ // residual for best compression. The prediction of alpha itself can
+ // be non-zero and must be kept though. We choose RGB of the residual
+ // to be 0.
+ residual &= kMaskAlpha;
+ // Update input image so that next predictions use correct RGB value.
+ current_row[x] = predict & ~kMaskAlpha;
+ if (x == 0 && y != 0) upper_row[width] = current_row[x];
+ }
+ argb[y * width + x] = residual;
}
- argb[pix] = VP8LSubPixels(current_row[col], predict);
}
}
}
-void VP8LResidualImage(int width, int height, int bits,
+void VP8LResidualImage(int width, int height, int bits, int low_effort,
uint32_t* const argb, uint32_t* const argb_scratch,
- uint32_t* const image) {
+ uint32_t* const image, int exact) {
const int max_tile_size = 1 << bits;
const int tiles_per_row = VP8LSubSampleSize(width, bits);
const int tiles_per_col = VP8LSubSampleSize(height, bits);
@@ -692,104 +680,34 @@ void VP8LResidualImage(int width, int height, int bits,
uint32_t* const current_tile_rows = argb_scratch + width;
int tile_y;
int histo[4][256];
- memset(histo, 0, sizeof(histo));
- for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
- const int tile_y_offset = tile_y * max_tile_size;
- const int this_tile_height =
- (tile_y < tiles_per_col - 1) ? max_tile_size : height - tile_y_offset;
- int tile_x;
- if (tile_y > 0) {
- memcpy(upper_row, current_tile_rows + (max_tile_size - 1) * width,
- width * sizeof(*upper_row));
+ if (low_effort) {
+ int i;
+ for (i = 0; i < tiles_per_row * tiles_per_col; ++i) {
+ image[i] = ARGB_BLACK | (kPredLowEffort << 8);
}
- memcpy(current_tile_rows, &argb[tile_y_offset * width],
- this_tile_height * width * sizeof(*current_tile_rows));
- for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
- int pred;
- int y;
- const int tile_x_offset = tile_x * max_tile_size;
- int all_x_max = tile_x_offset + max_tile_size;
- if (all_x_max > width) {
- all_x_max = width;
+ } else {
+ memset(histo, 0, sizeof(histo));
+ for (tile_y = 0; tile_y < tiles_per_col; ++tile_y) {
+ const int tile_y_offset = tile_y * max_tile_size;
+ const int this_tile_height =
+ (tile_y < tiles_per_col - 1) ? max_tile_size : height - tile_y_offset;
+ int tile_x;
+ if (tile_y > 0) {
+ memcpy(upper_row, current_tile_rows + (max_tile_size - 1) * width,
+ width * sizeof(*upper_row));
}
- pred = GetBestPredictorForTile(width, height, tile_x, tile_y, bits,
- (const int (*)[256])histo,
- argb_scratch);
- image[tile_y * tiles_per_row + tile_x] = 0xff000000u | (pred << 8);
- CopyTileWithPrediction(width, height, tile_x, tile_y, bits, pred,
- argb_scratch, argb);
- for (y = 0; y < max_tile_size; ++y) {
- int ix;
- int all_x;
- int all_y = tile_y_offset + y;
- if (all_y >= height) {
- break;
- }
- ix = all_y * width + tile_x_offset;
- for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
- UpdateHisto(histo, argb[ix]);
- }
+ memcpy(current_tile_rows, &argb[tile_y_offset * width],
+ this_tile_height * width * sizeof(*current_tile_rows));
+ for (tile_x = 0; tile_x < tiles_per_row; ++tile_x) {
+ const int pred = GetBestPredictorForTile(width, height, tile_x, tile_y,
+ bits, (int (*)[256])histo, argb_scratch, exact);
+ image[tile_y * tiles_per_row + tile_x] = ARGB_BLACK | (pred << 8);
}
}
}
-}
-
-// Inverse prediction.
-static void PredictorInverseTransform(const VP8LTransform* const transform,
- int y_start, int y_end, uint32_t* data) {
- const int width = transform->xsize_;
- if (y_start == 0) { // First Row follows the L (mode=1) mode.
- int x;
- const uint32_t pred0 = Predictor0(data[-1], NULL);
- AddPixelsEq(data, pred0);
- for (x = 1; x < width; ++x) {
- const uint32_t pred1 = Predictor1(data[x - 1], NULL);
- AddPixelsEq(data + x, pred1);
- }
- data += width;
- ++y_start;
- }
- {
- int y = y_start;
- const int tile_width = 1 << transform->bits_;
- const int mask = tile_width - 1;
- const int safe_width = width & ~mask;
- const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
- const uint32_t* pred_mode_base =
- transform->data_ + (y >> transform->bits_) * tiles_per_row;
-
- while (y < y_end) {
- const uint32_t pred2 = Predictor2(data[-1], data - width);
- const uint32_t* pred_mode_src = pred_mode_base;
- VP8LPredictorFunc pred_func;
- int x = 1;
- int t = 1;
- // First pixel follows the T (mode=2) mode.
- AddPixelsEq(data, pred2);
- // .. the rest:
- while (x < safe_width) {
- pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
- for (; t < tile_width; ++t, ++x) {
- const uint32_t pred = pred_func(data[x - 1], data + x - width);
- AddPixelsEq(data + x, pred);
- }
- t = 0;
- }
- if (x < width) {
- pred_func = VP8LPredictors[((*pred_mode_src++) >> 8) & 0xf];
- for (; x < width; ++x) {
- const uint32_t pred = pred_func(data[x - 1], data + x - width);
- AddPixelsEq(data + x, pred);
- }
- }
- data += width;
- ++y;
- if ((y & mask) == 0) { // Use the same mask, since tiles are squares.
- pred_mode_base += tiles_per_row;
- }
- }
- }
+ CopyImageWithPrediction(width, height, bits,
+ image, argb_scratch, argb, low_effort, exact);
}
void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) {
@@ -803,20 +721,6 @@ void VP8LSubtractGreenFromBlueAndRed_C(uint32_t* argb_data, int num_pixels) {
}
}
-// Add green to blue and red channels (i.e. perform the inverse transform of
-// 'subtract green').
-void VP8LAddGreenToBlueAndRed_C(uint32_t* data, int num_pixels) {
- int i;
- for (i = 0; i < num_pixels; ++i) {
- const uint32_t argb = data[i];
- const uint32_t green = ((argb >> 8) & 0xff);
- uint32_t red_blue = (argb & 0x00ff00ffu);
- red_blue += (green << 16) | green;
- red_blue &= 0x00ff00ffu;
- data[i] = (argb & 0xff00ff00u) | red_blue;
- }
-}
-
static WEBP_INLINE void MultipliersClear(VP8LMultipliers* const m) {
m->green_to_red_ = 0;
m->green_to_blue_ = 0;
@@ -861,24 +765,6 @@ void VP8LTransformColor_C(const VP8LMultipliers* const m, uint32_t* data,
}
}
-void VP8LTransformColorInverse_C(const VP8LMultipliers* const m, uint32_t* data,
- int num_pixels) {
- int i;
- for (i = 0; i < num_pixels; ++i) {
- const uint32_t argb = data[i];
- const uint32_t green = argb >> 8;
- const uint32_t red = argb >> 16;
- uint32_t new_red = red;
- uint32_t new_blue = argb;
- new_red += ColorTransformDelta(m->green_to_red_, green);
- new_red &= 0xff;
- new_blue += ColorTransformDelta(m->green_to_blue_, green);
- new_blue += ColorTransformDelta(m->red_to_blue_, new_red);
- new_blue &= 0xff;
- data[i] = (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
- }
-}
-
static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
uint32_t argb) {
const uint32_t green = argb >> 8;
@@ -903,24 +789,32 @@ static float PredictionCostCrossColor(const int accumulated[256],
// Favor low entropy, locally and globally.
// Favor small absolute values for PredictionCostSpatial
static const double kExpValue = 2.4;
- return CombinedShannonEntropy(counts, accumulated) +
+ return VP8LCombinedShannonEntropy(counts, accumulated) +
PredictionCostSpatial(counts, 3, kExpValue);
}
+void VP8LCollectColorRedTransforms_C(const uint32_t* argb, int stride,
+ int tile_width, int tile_height,
+ int green_to_red, int histo[]) {
+ while (tile_height-- > 0) {
+ int x;
+ for (x = 0; x < tile_width; ++x) {
+ ++histo[TransformColorRed(green_to_red, argb[x])];
+ }
+ argb += stride;
+ }
+}
+
static float GetPredictionCostCrossColorRed(
- int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
- int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red,
- const int accumulated_red_histo[256], const uint32_t* const argb) {
- int all_y;
+ const uint32_t* argb, int stride, int tile_width, int tile_height,
+ VP8LMultipliers prev_x, VP8LMultipliers prev_y, int green_to_red,
+ const int accumulated_red_histo[256]) {
int histo[256] = { 0 };
float cur_diff;
- for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
- int ix = all_y * xsize + tile_x_offset;
- int all_x;
- for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
- ++histo[TransformColorRed(green_to_red, argb[ix])]; // red.
- }
- }
+
+ VP8LCollectColorRedTransforms(argb, stride, tile_width, tile_height,
+ green_to_red, histo);
+
cur_diff = PredictionCostCrossColor(accumulated_red_histo, histo);
if ((uint8_t)green_to_red == prev_x.green_to_red_) {
cur_diff -= 3; // favor keeping the areas locally similar
@@ -935,59 +829,58 @@ static float GetPredictionCostCrossColorRed(
}
static void GetBestGreenToRed(
- int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
- int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y,
- const int accumulated_red_histo[256], const uint32_t* const argb,
- VP8LMultipliers* const best_tx) {
- int min_green_to_red = -64;
- int max_green_to_red = 64;
- int green_to_red = 0;
- int eval_min = 1;
- int eval_max = 1;
- float cur_diff_min = MAX_DIFF_COST;
- float cur_diff_max = MAX_DIFF_COST;
- // Do a binary search to find the optimal green_to_red color transform.
- while (max_green_to_red - min_green_to_red > 2) {
- if (eval_min) {
- cur_diff_min = GetPredictionCostCrossColorRed(
- tile_x_offset, tile_y_offset, all_x_max, all_y_max, xsize,
- prev_x, prev_y, min_green_to_red, accumulated_red_histo, argb);
- eval_min = 0;
- }
- if (eval_max) {
- cur_diff_max = GetPredictionCostCrossColorRed(
- tile_x_offset, tile_y_offset, all_x_max, all_y_max, xsize,
- prev_x, prev_y, max_green_to_red, accumulated_red_histo, argb);
- eval_max = 0;
+ const uint32_t* argb, int stride, int tile_width, int tile_height,
+ VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
+ const int accumulated_red_histo[256], VP8LMultipliers* const best_tx) {
+ const int kMaxIters = 4 + ((7 * quality) >> 8); // in range [4..6]
+ int green_to_red_best = 0;
+ int iter, offset;
+ float best_diff = GetPredictionCostCrossColorRed(
+ argb, stride, tile_width, tile_height, prev_x, prev_y,
+ green_to_red_best, accumulated_red_histo);
+ for (iter = 0; iter < kMaxIters; ++iter) {
+ // ColorTransformDelta is a 3.5 bit fixed point, so 32 is equal to
+ // one in color computation. Having initial delta here as 1 is sufficient
+ // to explore the range of (-2, 2).
+ const int delta = 32 >> iter;
+ // Try a negative and a positive delta from the best known value.
+ for (offset = -delta; offset <= delta; offset += 2 * delta) {
+ const int green_to_red_cur = offset + green_to_red_best;
+ const float cur_diff = GetPredictionCostCrossColorRed(
+ argb, stride, tile_width, tile_height, prev_x, prev_y,
+ green_to_red_cur, accumulated_red_histo);
+ if (cur_diff < best_diff) {
+ best_diff = cur_diff;
+ green_to_red_best = green_to_red_cur;
+ }
}
- if (cur_diff_min < cur_diff_max) {
- green_to_red = min_green_to_red;
- max_green_to_red = (max_green_to_red + min_green_to_red) / 2;
- eval_max = 1;
- } else {
- green_to_red = max_green_to_red;
- min_green_to_red = (max_green_to_red + min_green_to_red) / 2;
- eval_min = 1;
+ }
+ best_tx->green_to_red_ = green_to_red_best;
+}
+
+void VP8LCollectColorBlueTransforms_C(const uint32_t* argb, int stride,
+ int tile_width, int tile_height,
+ int green_to_blue, int red_to_blue,
+ int histo[]) {
+ while (tile_height-- > 0) {
+ int x;
+ for (x = 0; x < tile_width; ++x) {
+ ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[x])];
}
+ argb += stride;
}
- best_tx->green_to_red_ = green_to_red;
}
static float GetPredictionCostCrossColorBlue(
- int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
- int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y,
- int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256],
- const uint32_t* const argb) {
- int all_y;
+ const uint32_t* argb, int stride, int tile_width, int tile_height,
+ VP8LMultipliers prev_x, VP8LMultipliers prev_y,
+ int green_to_blue, int red_to_blue, const int accumulated_blue_histo[256]) {
int histo[256] = { 0 };
float cur_diff;
- for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
- int all_x;
- int ix = all_y * xsize + tile_x_offset;
- for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
- ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[ix])];
- }
- }
+
+ VP8LCollectColorBlueTransforms(argb, stride, tile_width, tile_height,
+ green_to_blue, red_to_blue, histo);
+
cur_diff = PredictionCostCrossColor(accumulated_blue_histo, histo);
if ((uint8_t)green_to_blue == prev_x.green_to_blue_) {
cur_diff -= 3; // favor keeping the areas locally similar
@@ -1010,49 +903,55 @@ static float GetPredictionCostCrossColorBlue(
return cur_diff;
}
+#define kGreenRedToBlueNumAxis 8
+#define kGreenRedToBlueMaxIters 7
static void GetBestGreenRedToBlue(
- int tile_x_offset, int tile_y_offset, int all_x_max, int all_y_max,
- int xsize, VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
- const int accumulated_blue_histo[256], const uint32_t* const argb,
+ const uint32_t* argb, int stride, int tile_width, int tile_height,
+ VP8LMultipliers prev_x, VP8LMultipliers prev_y, int quality,
+ const int accumulated_blue_histo[256],
VP8LMultipliers* const best_tx) {
- float best_diff = MAX_DIFF_COST;
- float cur_diff;
- const int step = (quality < 25) ? 32 : (quality > 50) ? 8 : 16;
- const int min_green_to_blue = -32;
- const int max_green_to_blue = 32;
- const int min_red_to_blue = -32;
- const int max_red_to_blue = 32;
- const int num_iters =
- (1 + (max_green_to_blue - min_green_to_blue) / step) *
- (1 + (max_red_to_blue - min_red_to_blue) / step);
- // Number of tries to get optimal green_to_blue & red_to_blue color transforms
- // after finding a local minima.
- const int max_tries_after_min = 4 + (num_iters >> 2);
- int num_tries_after_min = 0;
- int green_to_blue;
- for (green_to_blue = min_green_to_blue;
- green_to_blue <= max_green_to_blue &&
- num_tries_after_min < max_tries_after_min;
- green_to_blue += step) {
- int red_to_blue;
- for (red_to_blue = min_red_to_blue;
- red_to_blue <= max_red_to_blue &&
- num_tries_after_min < max_tries_after_min;
- red_to_blue += step) {
- cur_diff = GetPredictionCostCrossColorBlue(
- tile_x_offset, tile_y_offset, all_x_max, all_y_max, xsize, prev_x,
- prev_y, green_to_blue, red_to_blue, accumulated_blue_histo, argb);
+ const int8_t offset[kGreenRedToBlueNumAxis][2] =
+ {{0, -1}, {0, 1}, {-1, 0}, {1, 0}, {-1, -1}, {-1, 1}, {1, -1}, {1, 1}};
+ const int8_t delta_lut[kGreenRedToBlueMaxIters] = { 16, 16, 8, 4, 2, 2, 2 };
+ const int iters =
+ (quality < 25) ? 1 : (quality > 50) ? kGreenRedToBlueMaxIters : 4;
+ int green_to_blue_best = 0;
+ int red_to_blue_best = 0;
+ int iter;
+ // Initial value at origin:
+ float best_diff = GetPredictionCostCrossColorBlue(
+ argb, stride, tile_width, tile_height, prev_x, prev_y,
+ green_to_blue_best, red_to_blue_best, accumulated_blue_histo);
+ for (iter = 0; iter < iters; ++iter) {
+ const int delta = delta_lut[iter];
+ int axis;
+ for (axis = 0; axis < kGreenRedToBlueNumAxis; ++axis) {
+ const int green_to_blue_cur =
+ offset[axis][0] * delta + green_to_blue_best;
+ const int red_to_blue_cur = offset[axis][1] * delta + red_to_blue_best;
+ const float cur_diff = GetPredictionCostCrossColorBlue(
+ argb, stride, tile_width, tile_height, prev_x, prev_y,
+ green_to_blue_cur, red_to_blue_cur, accumulated_blue_histo);
if (cur_diff < best_diff) {
best_diff = cur_diff;
- best_tx->green_to_blue_ = green_to_blue;
- best_tx->red_to_blue_ = red_to_blue;
- num_tries_after_min = 0;
- } else {
- ++num_tries_after_min;
+ green_to_blue_best = green_to_blue_cur;
+ red_to_blue_best = red_to_blue_cur;
+ }
+ if (quality < 25 && iter == 4) {
+ // Only axis aligned diffs for lower quality.
+ break; // next iter.
}
}
+ if (delta == 2 && green_to_blue_best == 0 && red_to_blue_best == 0) {
+ // Further iterations would not help.
+ break; // out of iter-loop.
+ }
}
+ best_tx->green_to_blue_ = green_to_blue_best;
+ best_tx->red_to_blue_ = red_to_blue_best;
}
+#undef kGreenRedToBlueMaxIters
+#undef kGreenRedToBlueNumAxis
static VP8LMultipliers GetBestColorTransformForTile(
int tile_x, int tile_y, int bits,
@@ -1067,14 +966,18 @@ static VP8LMultipliers GetBestColorTransformForTile(
const int tile_x_offset = tile_x * max_tile_size;
const int all_x_max = GetMin(tile_x_offset + max_tile_size, xsize);
const int all_y_max = GetMin(tile_y_offset + max_tile_size, ysize);
+ const int tile_width = all_x_max - tile_x_offset;
+ const int tile_height = all_y_max - tile_y_offset;
+ const uint32_t* const tile_argb = argb + tile_y_offset * xsize
+ + tile_x_offset;
VP8LMultipliers best_tx;
MultipliersClear(&best_tx);
- GetBestGreenToRed(tile_x_offset, tile_y_offset, all_x_max, all_y_max, xsize,
- prev_x, prev_y, accumulated_red_histo, argb, &best_tx);
- GetBestGreenRedToBlue(tile_x_offset, tile_y_offset, all_x_max, all_y_max,
- xsize, prev_x, prev_y, quality, accumulated_blue_histo,
- argb, &best_tx);
+ GetBestGreenToRed(tile_argb, xsize, tile_width, tile_height,
+ prev_x, prev_y, quality, accumulated_red_histo, &best_tx);
+ GetBestGreenRedToBlue(tile_argb, xsize, tile_width, tile_height,
+ prev_x, prev_y, quality, accumulated_blue_histo,
+ &best_tx);
return best_tx;
}
@@ -1149,293 +1052,6 @@ void VP8LColorSpaceTransform(int width, int height, int bits, int quality,
}
}
-// Color space inverse transform.
-static void ColorSpaceInverseTransform(const VP8LTransform* const transform,
- int y_start, int y_end, uint32_t* data) {
- const int width = transform->xsize_;
- const int tile_width = 1 << transform->bits_;
- const int mask = tile_width - 1;
- const int safe_width = width & ~mask;
- const int remaining_width = width - safe_width;
- const int tiles_per_row = VP8LSubSampleSize(width, transform->bits_);
- int y = y_start;
- const uint32_t* pred_row =
- transform->data_ + (y >> transform->bits_) * tiles_per_row;
-
- while (y < y_end) {
- const uint32_t* pred = pred_row;
- VP8LMultipliers m = { 0, 0, 0 };
- const uint32_t* const data_safe_end = data + safe_width;
- const uint32_t* const data_end = data + width;
- while (data < data_safe_end) {
- ColorCodeToMultipliers(*pred++, &m);
- VP8LTransformColorInverse(&m, data, tile_width);
- data += tile_width;
- }
- if (data < data_end) { // Left-overs using C-version.
- ColorCodeToMultipliers(*pred++, &m);
- VP8LTransformColorInverse(&m, data, remaining_width);
- data += remaining_width;
- }
- ++y;
- if ((y & mask) == 0) pred_row += tiles_per_row;
- }
-}
-
-// Separate out pixels packed together using pixel-bundling.
-// We define two methods for ARGB data (uint32_t) and alpha-only data (uint8_t).
-#define COLOR_INDEX_INVERSE(FUNC_NAME, TYPE, GET_INDEX, GET_VALUE) \
-void FUNC_NAME(const VP8LTransform* const transform, \
- int y_start, int y_end, const TYPE* src, TYPE* dst) { \
- int y; \
- const int bits_per_pixel = 8 >> transform->bits_; \
- const int width = transform->xsize_; \
- const uint32_t* const color_map = transform->data_; \
- if (bits_per_pixel < 8) { \
- const int pixels_per_byte = 1 << transform->bits_; \
- const int count_mask = pixels_per_byte - 1; \
- const uint32_t bit_mask = (1 << bits_per_pixel) - 1; \
- for (y = y_start; y < y_end; ++y) { \
- uint32_t packed_pixels = 0; \
- int x; \
- for (x = 0; x < width; ++x) { \
- /* We need to load fresh 'packed_pixels' once every */ \
- /* 'pixels_per_byte' increments of x. Fortunately, pixels_per_byte */ \
- /* is a power of 2, so can just use a mask for that, instead of */ \
- /* decrementing a counter. */ \
- if ((x & count_mask) == 0) packed_pixels = GET_INDEX(*src++); \
- *dst++ = GET_VALUE(color_map[packed_pixels & bit_mask]); \
- packed_pixels >>= bits_per_pixel; \
- } \
- } \
- } else { \
- for (y = y_start; y < y_end; ++y) { \
- int x; \
- for (x = 0; x < width; ++x) { \
- *dst++ = GET_VALUE(color_map[GET_INDEX(*src++)]); \
- } \
- } \
- } \
-}
-
-static WEBP_INLINE uint32_t GetARGBIndex(uint32_t idx) {
- return (idx >> 8) & 0xff;
-}
-
-static WEBP_INLINE uint8_t GetAlphaIndex(uint8_t idx) {
- return idx;
-}
-
-static WEBP_INLINE uint32_t GetARGBValue(uint32_t val) {
- return val;
-}
-
-static WEBP_INLINE uint8_t GetAlphaValue(uint32_t val) {
- return (val >> 8) & 0xff;
-}
-
-static COLOR_INDEX_INVERSE(ColorIndexInverseTransform, uint32_t, GetARGBIndex,
- GetARGBValue)
-COLOR_INDEX_INVERSE(VP8LColorIndexInverseTransformAlpha, uint8_t, GetAlphaIndex,
- GetAlphaValue)
-
-#undef COLOR_INDEX_INVERSE
-
-void VP8LInverseTransform(const VP8LTransform* const transform,
- int row_start, int row_end,
- const uint32_t* const in, uint32_t* const out) {
- const int width = transform->xsize_;
- assert(row_start < row_end);
- assert(row_end <= transform->ysize_);
- switch (transform->type_) {
- case SUBTRACT_GREEN:
- VP8LAddGreenToBlueAndRed(out, (row_end - row_start) * width);
- break;
- case PREDICTOR_TRANSFORM:
- PredictorInverseTransform(transform, row_start, row_end, out);
- if (row_end != transform->ysize_) {
- // The last predicted row in this iteration will be the top-pred row
- // for the first row in next iteration.
- memcpy(out - width, out + (row_end - row_start - 1) * width,
- width * sizeof(*out));
- }
- break;
- case CROSS_COLOR_TRANSFORM:
- ColorSpaceInverseTransform(transform, row_start, row_end, out);
- break;
- case COLOR_INDEXING_TRANSFORM:
- if (in == out && transform->bits_ > 0) {
- // Move packed pixels to the end of unpacked region, so that unpacking
- // can occur seamlessly.
- // Also, note that this is the only transform that applies on
- // the effective width of VP8LSubSampleSize(xsize_, bits_). All other
- // transforms work on effective width of xsize_.
- const int out_stride = (row_end - row_start) * width;
- const int in_stride = (row_end - row_start) *
- VP8LSubSampleSize(transform->xsize_, transform->bits_);
- uint32_t* const src = out + out_stride - in_stride;
- memmove(src, out, in_stride * sizeof(*src));
- ColorIndexInverseTransform(transform, row_start, row_end, src, out);
- } else {
- ColorIndexInverseTransform(transform, row_start, row_end, in, out);
- }
- break;
- }
-}
-
-//------------------------------------------------------------------------------
-// Color space conversion.
-
-static int is_big_endian(void) {
- static const union {
- uint16_t w;
- uint8_t b[2];
- } tmp = { 1 };
- return (tmp.b[0] != 1);
-}
-
-void VP8LConvertBGRAToRGB_C(const uint32_t* src,
- int num_pixels, uint8_t* dst) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
- *dst++ = (argb >> 16) & 0xff;
- *dst++ = (argb >> 8) & 0xff;
- *dst++ = (argb >> 0) & 0xff;
- }
-}
-
-void VP8LConvertBGRAToRGBA_C(const uint32_t* src,
- int num_pixels, uint8_t* dst) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
- *dst++ = (argb >> 16) & 0xff;
- *dst++ = (argb >> 8) & 0xff;
- *dst++ = (argb >> 0) & 0xff;
- *dst++ = (argb >> 24) & 0xff;
- }
-}
-
-void VP8LConvertBGRAToRGBA4444_C(const uint32_t* src,
- int num_pixels, uint8_t* dst) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
- const uint8_t rg = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
- const uint8_t ba = ((argb >> 0) & 0xf0) | ((argb >> 28) & 0xf);
-#ifdef WEBP_SWAP_16BIT_CSP
- *dst++ = ba;
- *dst++ = rg;
-#else
- *dst++ = rg;
- *dst++ = ba;
-#endif
- }
-}
-
-void VP8LConvertBGRAToRGB565_C(const uint32_t* src,
- int num_pixels, uint8_t* dst) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
- const uint8_t rg = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
- const uint8_t gb = ((argb >> 5) & 0xe0) | ((argb >> 3) & 0x1f);
-#ifdef WEBP_SWAP_16BIT_CSP
- *dst++ = gb;
- *dst++ = rg;
-#else
- *dst++ = rg;
- *dst++ = gb;
-#endif
- }
-}
-
-void VP8LConvertBGRAToBGR_C(const uint32_t* src,
- int num_pixels, uint8_t* dst) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
- *dst++ = (argb >> 0) & 0xff;
- *dst++ = (argb >> 8) & 0xff;
- *dst++ = (argb >> 16) & 0xff;
- }
-}
-
-static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
- int swap_on_big_endian) {
- if (is_big_endian() == swap_on_big_endian) {
- const uint32_t* const src_end = src + num_pixels;
- while (src < src_end) {
- const uint32_t argb = *src++;
-
-#if !defined(WORDS_BIGENDIAN)
-#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
- *(uint32_t*)dst = BSwap32(argb);
-#else // WEBP_REFERENCE_IMPLEMENTATION
- dst[0] = (argb >> 24) & 0xff;
- dst[1] = (argb >> 16) & 0xff;
- dst[2] = (argb >> 8) & 0xff;
- dst[3] = (argb >> 0) & 0xff;
-#endif
-#else // WORDS_BIGENDIAN
- dst[0] = (argb >> 0) & 0xff;
- dst[1] = (argb >> 8) & 0xff;
- dst[2] = (argb >> 16) & 0xff;
- dst[3] = (argb >> 24) & 0xff;
-#endif
- dst += sizeof(argb);
- }
- } else {
- memcpy(dst, src, num_pixels * sizeof(*src));
- }
-}
-
-void VP8LConvertFromBGRA(const uint32_t* const in_data, int num_pixels,
- WEBP_CSP_MODE out_colorspace, uint8_t* const rgba) {
- switch (out_colorspace) {
- case MODE_RGB:
- VP8LConvertBGRAToRGB(in_data, num_pixels, rgba);
- break;
- case MODE_RGBA:
- VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba);
- break;
- case MODE_rgbA:
- VP8LConvertBGRAToRGBA(in_data, num_pixels, rgba);
- WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
- break;
- case MODE_BGR:
- VP8LConvertBGRAToBGR(in_data, num_pixels, rgba);
- break;
- case MODE_BGRA:
- CopyOrSwap(in_data, num_pixels, rgba, 1);
- break;
- case MODE_bgrA:
- CopyOrSwap(in_data, num_pixels, rgba, 1);
- WebPApplyAlphaMultiply(rgba, 0, num_pixels, 1, 0);
- break;
- case MODE_ARGB:
- CopyOrSwap(in_data, num_pixels, rgba, 0);
- break;
- case MODE_Argb:
- CopyOrSwap(in_data, num_pixels, rgba, 0);
- WebPApplyAlphaMultiply(rgba, 1, num_pixels, 1, 0);
- break;
- case MODE_RGBA_4444:
- VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
- break;
- case MODE_rgbA_4444:
- VP8LConvertBGRAToRGBA4444(in_data, num_pixels, rgba);
- WebPApplyAlphaMultiply4444(rgba, num_pixels, 1, 0);
- break;
- case MODE_RGB_565:
- VP8LConvertBGRAToRGB565(in_data, num_pixels, rgba);
- break;
- default:
- assert(0); // Code flow should not reach here.
- }
-}
-
//------------------------------------------------------------------------------
// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
void VP8LBundleColorMap(const uint8_t* const row, int width,
@@ -1478,53 +1094,6 @@ static double ExtraCostCombined(const uint32_t* X, const uint32_t* Y,
return cost;
}
-// Returns the various RLE counts
-static VP8LStreaks HuffmanCostCount(const uint32_t* population, int length) {
- int i;
- int streak = 0;
- VP8LStreaks stats;
- memset(&stats, 0, sizeof(stats));
- for (i = 0; i < length - 1; ++i) {
- ++streak;
- if (population[i] == population[i + 1]) {
- continue;
- }
- stats.counts[population[i] != 0] += (streak > 3);
- stats.streaks[population[i] != 0][(streak > 3)] += streak;
- streak = 0;
- }
- ++streak;
- stats.counts[population[i] != 0] += (streak > 3);
- stats.streaks[population[i] != 0][(streak > 3)] += streak;
- return stats;
-}
-
-static VP8LStreaks HuffmanCostCombinedCount(const uint32_t* X,
- const uint32_t* Y, int length) {
- int i;
- int streak = 0;
- VP8LStreaks stats;
- memset(&stats, 0, sizeof(stats));
- for (i = 0; i < length - 1; ++i) {
- const int xy = X[i] + Y[i];
- const int xy_next = X[i + 1] + Y[i + 1];
- ++streak;
- if (xy == xy_next) {
- continue;
- }
- stats.counts[xy != 0] += (streak > 3);
- stats.streaks[xy != 0][(streak > 3)] += streak;
- streak = 0;
- }
- {
- const int xy = X[i] + Y[i];
- ++streak;
- stats.counts[xy != 0] += (streak > 3);
- stats.streaks[xy != 0][(streak > 3)] += streak;
- }
- return stats;
-}
-
//------------------------------------------------------------------------------
static void HistogramAdd(const VP8LHistogram* const a,
@@ -1563,61 +1132,52 @@ static void HistogramAdd(const VP8LHistogram* const a,
//------------------------------------------------------------------------------
VP8LProcessBlueAndRedFunc VP8LSubtractGreenFromBlueAndRed;
-VP8LProcessBlueAndRedFunc VP8LAddGreenToBlueAndRed;
-VP8LPredictorFunc VP8LPredictors[16];
VP8LTransformColorFunc VP8LTransformColor;
-VP8LTransformColorFunc VP8LTransformColorInverse;
-VP8LConvertFunc VP8LConvertBGRAToRGB;
-VP8LConvertFunc VP8LConvertBGRAToRGBA;
-VP8LConvertFunc VP8LConvertBGRAToRGBA4444;
-VP8LConvertFunc VP8LConvertBGRAToRGB565;
-VP8LConvertFunc VP8LConvertBGRAToBGR;
+VP8LCollectColorBlueTransformsFunc VP8LCollectColorBlueTransforms;
+VP8LCollectColorRedTransformsFunc VP8LCollectColorRedTransforms;
VP8LFastLog2SlowFunc VP8LFastLog2Slow;
VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
VP8LCostFunc VP8LExtraCost;
VP8LCostCombinedFunc VP8LExtraCostCombined;
+VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy;
-VP8LCostCountFunc VP8LHuffmanCostCount;
-VP8LCostCombinedCountFunc VP8LHuffmanCostCombinedCount;
+GetEntropyUnrefinedHelperFunc VP8LGetEntropyUnrefinedHelper;
VP8LHistogramAddFunc VP8LHistogramAdd;
-extern void VP8LDspInitSSE2(void);
-extern void VP8LDspInitNEON(void);
-extern void VP8LDspInitMIPS32(void);
+extern void VP8LEncDspInitSSE2(void);
+extern void VP8LEncDspInitSSE41(void);
+extern void VP8LEncDspInitNEON(void);
+extern void VP8LEncDspInitMIPS32(void);
+extern void VP8LEncDspInitMIPSdspR2(void);
-static volatile VP8CPUInfo lossless_last_cpuinfo_used =
- (VP8CPUInfo)&lossless_last_cpuinfo_used;
+static volatile VP8CPUInfo lossless_enc_last_cpuinfo_used =
+ (VP8CPUInfo)&lossless_enc_last_cpuinfo_used;
-void VP8LDspInit(void) {
- if (lossless_last_cpuinfo_used == VP8GetCPUInfo) return;
+WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInit(void) {
+ if (lossless_enc_last_cpuinfo_used == VP8GetCPUInfo) return;
- memcpy(VP8LPredictors, kPredictorsC, sizeof(VP8LPredictors));
+ VP8LDspInit();
VP8LSubtractGreenFromBlueAndRed = VP8LSubtractGreenFromBlueAndRed_C;
- VP8LAddGreenToBlueAndRed = VP8LAddGreenToBlueAndRed_C;
VP8LTransformColor = VP8LTransformColor_C;
- VP8LTransformColorInverse = VP8LTransformColorInverse_C;
- VP8LConvertBGRAToRGB = VP8LConvertBGRAToRGB_C;
- VP8LConvertBGRAToRGBA = VP8LConvertBGRAToRGBA_C;
- VP8LConvertBGRAToRGBA4444 = VP8LConvertBGRAToRGBA4444_C;
- VP8LConvertBGRAToRGB565 = VP8LConvertBGRAToRGB565_C;
- VP8LConvertBGRAToBGR = VP8LConvertBGRAToBGR_C;
+ VP8LCollectColorBlueTransforms = VP8LCollectColorBlueTransforms_C;
+ VP8LCollectColorRedTransforms = VP8LCollectColorRedTransforms_C;
VP8LFastLog2Slow = FastLog2Slow;
VP8LFastSLog2Slow = FastSLog2Slow;
VP8LExtraCost = ExtraCost;
VP8LExtraCostCombined = ExtraCostCombined;
+ VP8LCombinedShannonEntropy = CombinedShannonEntropy;
- VP8LHuffmanCostCount = HuffmanCostCount;
- VP8LHuffmanCostCombinedCount = HuffmanCostCombinedCount;
+ VP8LGetEntropyUnrefinedHelper = GetEntropyUnrefinedHelper;
VP8LHistogramAdd = HistogramAdd;
@@ -1625,21 +1185,31 @@ void VP8LDspInit(void) {
if (VP8GetCPUInfo != NULL) {
#if defined(WEBP_USE_SSE2)
if (VP8GetCPUInfo(kSSE2)) {
- VP8LDspInitSSE2();
+ VP8LEncDspInitSSE2();
+#if defined(WEBP_USE_SSE41)
+ if (VP8GetCPUInfo(kSSE4_1)) {
+ VP8LEncDspInitSSE41();
+ }
+#endif
}
#endif
#if defined(WEBP_USE_NEON)
if (VP8GetCPUInfo(kNEON)) {
- VP8LDspInitNEON();
+ VP8LEncDspInitNEON();
}
#endif
#if defined(WEBP_USE_MIPS32)
if (VP8GetCPUInfo(kMIPS32)) {
- VP8LDspInitMIPS32();
+ VP8LEncDspInitMIPS32();
+ }
+#endif
+#if defined(WEBP_USE_MIPS_DSP_R2)
+ if (VP8GetCPUInfo(kMIPSdspR2)) {
+ VP8LEncDspInitMIPSdspR2();
}
#endif
}
- lossless_last_cpuinfo_used = VP8GetCPUInfo;
+ lossless_enc_last_cpuinfo_used = VP8GetCPUInfo;
}
//------------------------------------------------------------------------------
« no previous file with comments | « third_party/libwebp/dsp/lossless.c ('k') | third_party/libwebp/dsp/lossless_enc_mips32.c » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698