Introduce a 'fast' alpha mode

.. where only 2 filtering modes are potentially
tried, instead of all of them. This is fast than the exhaustive 'best'
mode, and not much worse.

Options for cwebp are:
 -alpha_filter none
 -alpha_filter fast      (<- default)
 -alpha_filter best      (<- slow)

Change-Id: I8cb90ee11b8f981811e013ea4ad5bf72ba3ea7d4

README

@@ -149,6 +149,8 @@ options:
   -map <int> ............. print map of extra info.
   -d <file.pgm> .......... dump the compressed output (PGM file).
   -alpha_method <int> .... Transparency-compression method (0..1)
+  -alpha_filter <string> . predictive filtering for alpha plane.
+                           One of: none, fast (default) or best.
   -noalpha ............... discard any transparency information.
 
   -short ................. condense printed message

examples/cwebp.c

@@ -700,7 +700,8 @@ static void HelpLong(void) {
   printf("  -map <int> ............. print map of extra info.\n");
   printf("  -d <file.pgm> .......... dump the compressed output (PGM file).\n");
   printf("  -alpha_method <int> .... Transparency-compression method (0..1)\n");
-  printf("  -alpha_filter <int> .... predictive filtering for Alpha (0..5)\n");
+  printf("  -alpha_filter <string> . predictive filtering for alpha plane.\n");
+  printf("                           One of: none, fast (default) or best.\n");
   printf("  -noalpha ............... discard any transparency information.\n");
 
   printf("\n");
@@ -795,7 +796,17 @@ int main(int argc, const char *argv[]) {
     } else if (!strcmp(argv[c], "-alpha_method") && c < argc - 1) {
       config.alpha_compression = strtol(argv[++c], NULL, 0);
     } else if (!strcmp(argv[c], "-alpha_filter") && c < argc - 1) {
-      config.alpha_filtering = strtol(argv[++c], NULL, 0);
+      ++c;
+      if (!strcmp(argv[c], "none")) {
+        config.alpha_filtering = 0;
+      } else if (!strcmp(argv[c], "fast")) {
+        config.alpha_filtering = 1;
+      } else if (!strcmp(argv[c], "best")) {
+        config.alpha_filtering = 2;
+      } else {
+        fprintf(stderr, "Error! Unrecognized alpha filter: %s\n", argv[c]);
+        goto Error;
+      }
     } else if (!strcmp(argv[c], "-noalpha")) {
       keep_alpha = 0;
     } else if (!strcmp(argv[c], "-size") && c < argc - 1) {

man/cwebp.1

@@ -41,11 +41,13 @@ Specify the compression factor for alpha compression between 0 and 100.
 Lossless compression of alpha is achieved using a value of 100, while the lower
 values result in a lossy compression. The default is 100.
 .TP
-.B \-alpha_filter int
+.B \-alpha_filter string
-Specify the predictive filtering method (between 0 and 5) for alpha plane.
+Specify the predictive filtering method for alpha plane. One of 'none', 'fast'
-These correspond to prediction modes none, horizontal, vertical, gradient and
+or 'best', in increasing complexity and slowness order. Default is 'fast'.
-paeth filters. The prediction mode 5 will try all the prediction modes (0 to 4)
+Internally, alpha filtering is performed using four possible predictions (none,
-and pick the best prediction mode. The default value is 0 (no prediction).
+horizontal, vertical, gradient). The 'best' mode will try each modes in turn and
+pick the one which gives the smaller size. The 'fast' mode will just try to
+form an a-priori guess without testing all modes.
 .TP
 .B \-f int
 Specify the strength of the deblocking filter, between 0 (no filtering)

src/enc/alpha.c

@@ -14,6 +14,7 @@
 
 #include "./vp8enci.h"
 #include "../utils/alpha.h"
+#include "../utils/filters.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
@@ -33,10 +34,15 @@ int VP8EncFinishAlpha(VP8Encoder* enc) {
     const WebPPicture* pic = enc->pic_;
     uint8_t* tmp_data = NULL;
     size_t tmp_size = 0;
+    const WEBP_FILTER_TYPE filter =
+        (config->alpha_filtering == 0) ? WEBP_FILTER_NONE :
+        (config->alpha_filtering == 1) ? WEBP_FILTER_FAST :
+                                         WEBP_FILTER_BEST;
+
     assert(pic->a);
     if (!EncodeAlpha(pic->a, pic->width, pic->height, pic->a_stride,
                      config->alpha_quality, config->alpha_compression,
-                     config->alpha_filtering, &tmp_data, &tmp_size)) {
+                     filter, &tmp_data, &tmp_size)) {
       return 0;
     }
     if (tmp_size != (uint32_t)tmp_size) {  // Sanity check.

src/enc/config.c

@@ -42,7 +42,7 @@ int WebPConfigInitInternal(WebPConfig* const config,
   config->autofilter = 0;
   config->partition_limit = 0;
   config->alpha_compression = 1;
-  config->alpha_filtering = 0;
+  config->alpha_filtering = 1;
   config->alpha_quality = 100;
 
   // TODO(skal): tune.

src/utils/alpha.c

@@ -203,14 +203,35 @@ static int EncodeZlibTCoder(const uint8_t* data, int width, int height,
   return ok && !bw->error_;
 }
 
+// -----------------------------------------------------------------------------
+
 static int EncodeAlphaInternal(const uint8_t* data, int width, int height,
-                               int method, VP8BitWriter* const bw) {
+                               int method, int filter, size_t data_size,
+                               uint8_t* tmp_alpha, VP8BitWriter* const bw) {
   int ok = 0;
+  const uint8_t* alpha_src;
+  WebPFilterFunc filter_func;
+  uint8_t header[ALPHA_HEADER_LEN];
+  const size_t expected_size = (method == 0) ?
+          (ALPHA_HEADER_LEN + data_size) : (data_size >> 5);
+  header[0] = (filter << 4) | method;
+  header[1] = 0;                // reserved byte for later use
+  VP8BitWriterInit(bw, expected_size);
+  VP8BitWriterAppend(bw, header, sizeof(header));
+
+  filter_func = WebPFilters[filter];
+  if (filter_func) {
+    filter_func(data, width, height, 1, width, tmp_alpha);
+    alpha_src = tmp_alpha;
+  }  else {
+    alpha_src = data;
+  }
+
   if (method == 0) {
-    ok = VP8BitWriterAppend(bw, data, width * height);
+    ok = VP8BitWriterAppend(bw, alpha_src, width * height);
     ok = ok && !bw->error_;
-  } else if (method == 1) {
+  } else {
-    ok = EncodeZlibTCoder(data, width, height, bw);
+    ok = EncodeZlibTCoder(alpha_src, width, height, bw);
     VP8BitWriterFinish(bw);
   }
   return ok;
@@ -239,7 +260,7 @@ int EncodeAlpha(const uint8_t* data, int width, int height, int stride,
   assert(data != NULL && output != NULL && output_size != NULL);
   assert(width > 0 && height > 0);
   assert(stride >= width);
-  assert(filter < WEBP_FILTER_LAST);
+  assert(filter >= WEBP_FILTER_NONE && filter <= WEBP_FILTER_FAST);
 
   if (quality < 0 || quality > 100) {
     return 0;
@@ -267,62 +288,67 @@ int EncodeAlpha(const uint8_t* data, int width, int height, int stride,
   }
 
   if (ok) {
-    WEBP_FILTER_TYPE this_filter;
-    size_t best_size = 1 << 30;
-    uint8_t* tmp_out = NULL;
-    uint8_t* const filtered_alpha = (uint8_t*)malloc(data_size);
-    if (filtered_alpha == NULL) {
-      free(quant_alpha);
-      return 0;
-    }
-    // Filtering.
-    for (this_filter = WEBP_FILTER_NONE; this_filter < WEBP_FILTER_LAST;
-         ++this_filter) {
-      uint8_t header[ALPHA_HEADER_LEN];
     VP8BitWriter bw;
-      WebPFilterFunc filter_func = NULL;
+    size_t best_score;
-      const size_t expected_size = (method == 0) ?
+    int test_filter;
-          (ALPHA_HEADER_LEN + data_size) : (data_size >> 5);
+    uint8_t* filtered_alpha = NULL;
-      if (this_filter == WEBP_FILTER_BEST) {
+
-        continue;
+    // We always test WEBP_FILTER_NONE first.
-      } else if (this_filter != filter && filter != WEBP_FILTER_BEST) {
+    ok = EncodeAlphaInternal(quant_alpha, width, height, method,
+                             WEBP_FILTER_NONE, data_size, NULL, &bw);
+    if (!ok) {
+      VP8BitWriterWipeOut(&bw);
+      goto End;
+    }
+    best_score = VP8BitWriterSize(&bw);
+
+    if (filter == WEBP_FILTER_FAST) {  // Quick estimate of a second candidate?
+      filter = EstimateBestFilter(quant_alpha, width, height, width);
+    }
+    // Stop?
+    if (filter == WEBP_FILTER_NONE) {
+      goto Ok;
+    }
+
+    filtered_alpha = (uint8_t*)malloc(data_size);
+    ok = (filtered_alpha != NULL);
+    if (!ok) {
+      goto End;
+    }
+
+    // Try the other mode(s).
+    for (test_filter = WEBP_FILTER_HORIZONTAL;
+         ok && (test_filter <= WEBP_FILTER_GRADIENT);
+         ++test_filter) {
+      VP8BitWriter tmp_bw;
+      if (filter != WEBP_FILTER_BEST && test_filter != filter) {
         continue;
       }
 
-      header[0] = ((this_filter & 0x0f) << 4) | (method & 0x0f);
+      ok = EncodeAlphaInternal(quant_alpha, width, height, method, test_filter,
-      header[1] = 0;                // reserved byte for later use
+                               data_size, filtered_alpha, &tmp_bw);
-      VP8BitWriterInit(&bw, expected_size);
-      VP8BitWriterAppend(&bw, header, sizeof(header));
-
-      filter_func = WebPFilters[this_filter];
-      if (filter_func) {
-        filter_func(quant_alpha, width, height, 1, width, filtered_alpha);
-        ok = EncodeAlphaInternal(filtered_alpha, width, height, method, &bw);
-      } else {
-        ok = EncodeAlphaInternal(quant_alpha, width, height, method, &bw);
-      }
       if (ok) {
-        const size_t this_size = VP8BitWriterSize(&bw);
+        const size_t score = VP8BitWriterSize(&tmp_bw);
-        if (this_size < best_size) {
+        if (score < best_score) {
-          free(tmp_out);
+          // swap bitwriter objects.
-          tmp_out = VP8BitWriterBuf(&bw);
+          VP8BitWriter tmp = tmp_bw;
-          best_size = this_size;
+          tmp_bw = bw;
+          bw = tmp;
+          best_score = score;
+        }
       } else {
         VP8BitWriterWipeOut(&bw);
       }
-      } else {
+      VP8BitWriterWipeOut(&tmp_bw);
-        free(tmp_out);
-        VP8BitWriterWipeOut(&bw);
-        break;
-      }
     }
+ Ok:
     if (ok) {
-      *output_size = best_size;
+      *output_size = VP8BitWriterSize(&bw);
-      *output = tmp_out;
+      *output = VP8BitWriterBuf(&bw);
     }
     free(filtered_alpha);
   }
-
+ End:
   free(quant_alpha);
   return ok;
 }
@@ -379,7 +405,7 @@ int DecodeAlpha(const uint8_t* data, size_t data_size,
   uint8_t* decoded_data = NULL;
   const size_t decoded_size = height * width;
   uint8_t* unfiltered_data = NULL;
-  int filter;
+  WEBP_FILTER_TYPE filter;
   int ok = 0;
   int method;
 
@@ -394,7 +420,7 @@ int DecodeAlpha(const uint8_t* data, size_t data_size,
   filter = data[0] >> 4;
   ok = (data[1] == 0);
   if (method < 0 || method > 1 ||
-      filter < WEBP_FILTER_NONE || filter > WEBP_FILTER_PAETH || !ok) {
+      filter > WEBP_FILTER_GRADIENT || !ok) {
     return 0;
   }
 

src/utils/filters.c

@@ -115,6 +115,11 @@ static void VerticalUnfilter(const uint8_t* data, int width, int height,
 //------------------------------------------------------------------------------
 // Gradient filter.
 
+static WEBP_INLINE uint8_t GradientPredictor(uint8_t a, uint8_t b, uint8_t c) {
+  const int g = a + b - c;
+  return (g < 0) ? 0 : (g > 255) ? 255 : (uint8_t)g;
+}
+
 static void GradientFilter(const uint8_t* data, int width, int height,
                            int bpp, int stride, uint8_t* filtered_data) {
   int h;
@@ -131,9 +136,9 @@ static void GradientFilter(const uint8_t* data, int width, int height,
     const uint8_t* const prev_line = scan_line - stride;
     memcpy((void*)out, (const void*)scan_line, bpp);
     for (w = bpp; w < width * bpp; ++w) {
-      const uint8_t predictor = scan_line[w - bpp] + prev_line[w] -
+      out[w] = scan_line[w] - GradientPredictor(scan_line[w - bpp],
-                                prev_line[w - bpp];
+                                                prev_line[w],
-      out[w] = scan_line[w] - predictor;
+                                                prev_line[w - bpp]);
     }
   }
 }
@@ -154,110 +159,81 @@ static void GradientUnfilter(const uint8_t* data, int width, int height,
     const uint8_t* const out_prev_line = out - stride;
     memcpy((void*)out, (const void*)scan_line, bpp);
     for (w = bpp; w < width * bpp; ++w) {
-      const uint8_t predictor = out[w - bpp] + out_prev_line[w] -
+      out[w] = scan_line[w] + GradientPredictor(out[w - bpp],
-                                out_prev_line[w - bpp];
+                                                out_prev_line[w],
-      out[w] = scan_line[w] + predictor;
+                                                out_prev_line[w - bpp]);
-    }
-  }
-}
-
-//------------------------------------------------------------------------------
-// Paeth filter.
-
-static WEBP_INLINE int AbsDiff(int a, int b) {
-  return (a > b) ? a - b : b - a;
-}
-
-static WEBP_INLINE uint8_t PaethPredictor(uint8_t a, uint8_t b, uint8_t c) {
-  const int p = a + b - c;  // Base.
-  const int pa = AbsDiff(p, a);
-  const int pb = AbsDiff(p, b);
-  const int pc = AbsDiff(p, c);
-
-  // Return nearest to base of a, b, c.
-  return (pa <= pb && pa <= pc) ? a : (pb <= pc) ? b : c;
-}
-
-static void PaethFilter(const uint8_t* data, int width, int height,
-                        int bpp, int stride, uint8_t* filtered_data) {
-  int w;
-  int h;
-  SANITY_CHECK(data, filtered_data);
-
-  // Top scan line (special case).
-  memcpy((void*)filtered_data, (const void*)data, bpp);
-  for (w = bpp; w < width * bpp; ++w) {
-    // Note: PaethPredictor(scan_line[w - bpp], 0, 0) == scan_line[w - bpp].
-    filtered_data[w] = data[w] - data[w - bpp];
-  }
-
-  // Filter line-by-line.
-  for (h = 1; h < height; ++h) {
-    int w;
-    const uint8_t* const scan_line = data + h * stride;
-    uint8_t* const out = filtered_data + h * stride;
-    const uint8_t* const prev_line = scan_line - stride;
-    for (w = 0; w < bpp; ++w) {
-      // Note: PaethPredictor(0, prev_line[w], 0) == prev_line[w].
-      out[w] = scan_line[w] - prev_line[w];
-    }
-    for (w = bpp; w < width * bpp; ++w) {
-      out[w] = scan_line[w] - PaethPredictor(scan_line[w - bpp], prev_line[w],
-                                             prev_line[w - bpp]);
-    }
-  }
-}
-
-static void PaethUnfilter(const uint8_t* data, int width, int height,
-                          int bpp, int stride, uint8_t* recon_data) {
-  int w;
-  int h;
-  SANITY_CHECK(data, recon_data);
-
-  // Top scan line (special case).
-  memcpy((void*)recon_data, (const void*)data, bpp);
-  for (w = bpp; w < width * bpp; ++w) {
-    // Note: PaethPredictor(out[w - bpp], 0, 0) == out[w - bpp].
-    recon_data[w] = data[w] + recon_data[w - bpp];
-  }
-
-  // Unfilter line-by-line.
-  for (h = 1; h < height; ++h) {
-    int w;
-    const uint8_t* const scan_line = data + h * stride;
-    uint8_t* const out = recon_data + h * stride;
-    const uint8_t* const out_prev = out - stride;
-    for (w = 0; w < bpp; ++w) {
-      // Note: PaethPredictor(0, out_prev[w], 0) == out_prev[w].
-      out[w] = scan_line[w] + out_prev[w];
-    }
-    for (w = bpp; w < width * bpp; ++w) {
-      out[w] = scan_line[w] + PaethPredictor(out[w - bpp], out_prev[w],
-                                             out_prev[w - bpp]);
     }
   }
 }
 
 #undef SANITY_CHECK
 
+// -----------------------------------------------------------------------------
+// Quick estimate of a potentially interesting filter mode to try, in addition
+// to the default NONE.
+
+#define SMAX 16
+#define SDIFF(a, b) (abs((a) - (b)) >> 4)   // Scoring diff, in [0..SMAX)
+
+WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
+                                    int width, int height, int stride) {
+  int i, j;
+  int bins[WEBP_FILTER_LAST][SMAX];
+  memset(bins, 0, sizeof(bins));
+  // We only sample every other pixels. That's enough.
+  for (j = 2; j < height - 1; j += 2) {
+    const uint8_t* const p = data + j * stride;
+    int mean = p[0];
+    for (i = 2; i < width - 1; i += 2) {
+      const int diff0 = SDIFF(p[i], mean);
+      const int diff1 = SDIFF(p[i], p[i - 1]);
+      const int diff2 = SDIFF(p[i], p[i - width]);
+      const int grad_pred =
+          GradientPredictor(p[i - 1], p[i - width], p[i - width - 1]);
+      const int diff3 = SDIFF(p[i], grad_pred);
+      bins[WEBP_FILTER_NONE][diff0] = 1;
+      bins[WEBP_FILTER_HORIZONTAL][diff1] = 1;
+      bins[WEBP_FILTER_VERTICAL][diff2] = 1;
+      bins[WEBP_FILTER_GRADIENT][diff3] = 1;
+      mean = (3 * mean + p[i] + 2) >> 2;
+    }
+  }
+  {
+    WEBP_FILTER_TYPE filter, best_filter = WEBP_FILTER_NONE;
+    int best_score = 0x7fffffff;
+    for (filter = WEBP_FILTER_NONE; filter < WEBP_FILTER_LAST; ++filter) {
+      int score = 0;
+      for (i = 0; i < SMAX; ++i) {
+        if (bins[filter][i] > 0) {
+          score += i;
+        }
+      }
+      if (score < best_score) {
+        best_score = score;
+        best_filter = filter;
+      }
+    }
+    return best_filter;
+  }
+}
+
+#undef SMAX
+#undef SDIFF
+
 //------------------------------------------------------------------------------
 
 const WebPFilterFunc WebPFilters[WEBP_FILTER_LAST] = {
     NULL,              // WEBP_FILTER_NONE
     HorizontalFilter,  // WEBP_FILTER_HORIZONTAL
     VerticalFilter,    // WEBP_FILTER_VERTICAL
-    GradientFilter,    // WEBP_FILTER_GRADIENT
+    GradientFilter     // WEBP_FILTER_GRADIENT
-    PaethFilter,       // WEBP_FILTER_PAETH
-    NULL               // WEBP_FILTER_BEST
 };
 
 const WebPFilterFunc WebPUnfilters[WEBP_FILTER_LAST] = {
     NULL,                // WEBP_FILTER_NONE
     HorizontalUnfilter,  // WEBP_FILTER_HORIZONTAL
     VerticalUnfilter,    // WEBP_FILTER_VERTICAL
-    GradientUnfilter,    // WEBP_FILTER_GRADIENT
+    GradientUnfilter     // WEBP_FILTER_GRADIENT
-    PaethUnfilter,       // WEBP_FILTER_PAETH
-    NULL                 // WEBP_FILTER_BEST
 };
 
 //------------------------------------------------------------------------------

src/utils/filters.h

@@ -24,9 +24,9 @@ typedef enum {
   WEBP_FILTER_HORIZONTAL,
   WEBP_FILTER_VERTICAL,
   WEBP_FILTER_GRADIENT,
-  WEBP_FILTER_PAETH,
+  WEBP_FILTER_LAST = WEBP_FILTER_GRADIENT + 1,  // end marker
   WEBP_FILTER_BEST,
-  WEBP_FILTER_LAST,
+  WEBP_FILTER_FAST
 } WEBP_FILTER_TYPE;
 
 typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height,
@@ -38,10 +38,14 @@ typedef void (*WebPFilterFunc)(const uint8_t* in, int width, int height,
 // 'bpp' is number of bytes per pixel, and
 // 'stride' is number of bytes per scan line (with possible padding).
 // 'out' should be pre-allocated.
-extern const WebPFilterFunc WebPFilters[/*WEBP_FILTER_LAST*/];
+extern const WebPFilterFunc WebPFilters[WEBP_FILTER_LAST];
 
 // Reconstruct the original data from the given filtered data.
-extern const WebPFilterFunc WebPUnfilters[/*WEBP_FILTER_LAST*/];
+extern const WebPFilterFunc WebPUnfilters[WEBP_FILTER_LAST];
+
+// Fast estimate of a potentially good filter.
+extern WEBP_FILTER_TYPE EstimateBestFilter(const uint8_t* data,
+                                           int width, int height, int stride);
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"

src/webp/encode.h

@@ -75,8 +75,7 @@ typedef struct {
                           // 1 = backward reference counts encoded with
                           // arithmetic encoder). Default is 1.
   int alpha_filtering;    // Predictive filtering method for alpha plane.
-                          // (0 = none, 1 = horizontal, 2 = vertical, 3 = grad,
+                          //  0: none, 1: fast, 2: best. Default if 1.
-                          // 4 = Paeth and 5 = Best of (0 .. 4).
   int alpha_quality;      // Between 0 (smallest size) and 100 (lossless).
                           // Default is 100.
 } WebPConfig;