add a -partition_limit option to limit the number of bits used by intra4x4

Although it degrades quality, this option is useful to avoid the 512k
limit for partition #0.
If not enough to reach the lower bound of 4bits per macroblock header,
one should also limit the number of segments used (down to -segments 1)

See the man file for extra details.

Change-Id: Ia59ffac13176c85b809ddd6340d37b54ee9487ea

README

@@ -140,6 +140,8 @@ options:
   -f <int> ............... filter strength (0=off..100)
   -sharpness <int> ....... filter sharpness (0:most .. 7:least sharp)
   -strong ................ use strong filter instead of simple.
+  -partition_limit <int> . limit quality to fit the 512k limit on
+                           the first partition (0=no degradation ... 100=full)
   -alpha_comp <int> ...... set the transparency-compression
   -noalpha ............... discard any transparency information.
   -pass <int> ............ analysis pass number (1..10)

examples/cwebp.c

@@ -676,6 +676,9 @@ static void HelpLong(void) {
   printf("  -sharpness <int> ....... "
          "filter sharpness (0:most .. 7:least sharp)\n");
   printf("  -strong ................ use strong filter instead of simple.\n");
+  printf("  -partition_limit <int> . limit quality to fit the 512k limit on\n");
+  printf("                           "
+         "the first partition (0=no degradation ... 100=full)\n");
   printf("  -alpha_comp <int> ...... set the transparency-compression\n");
   printf("  -noalpha ............... discard any transparency information.\n");
   printf("  -pass <int> ............ analysis pass number (1..10)\n");
@@ -712,10 +715,15 @@ static const char* const kErrorMessages[] = {
   "BITSTREAM_OUT_OF_MEMORY: Out of memory re-allocating byte buffer",
   "NULL_PARAMETER: NULL parameter passed to function",
   "INVALID_CONFIGURATION: configuration is invalid",
-  "BAD_DIMENSION: Bad picture dimension",
-  "PARTITION0_OVERFLOW: Partition #0 is too big to fit 512k",
-  "PARTITION_OVERFLOW: Partition is too big to fir 16M",
-  "BAD_WRITE: Picture writer returned an error"
+  "BAD_DIMENSION: Bad picture dimension. Maximum width and height "
+  "allowed is 16383 pixels.",
+  "PARTITION0_OVERFLOW: Partition #0 is too big to fit 512k.\n"
+  "To reduce the size of this partition, try using less segments "
+  "with the -segments option, and eventually reduce the number of "
+  "header bits using -partition_limit. More details are available "
+  "in the manual (`man cwebp`)",
+  "PARTITION_OVERFLOW: Partition is too big to fit 16M",
+  "BAD_WRITE: Picture writer returned an I/O error"
 };
 
 //-----------------------------------------------------------------------------
@@ -789,6 +797,8 @@ int main(int argc, const char *argv[]) {
       config.preprocessing = strtol(argv[++c], NULL, 0);
     } else if (!strcmp(argv[c], "-segments") && c < argc - 1) {
       config.segments = strtol(argv[++c], NULL, 0);
+    } else if (!strcmp(argv[c], "-partition_limit") && c < argc - 1) {
+      config.partition_limit = strtol(argv[++c], NULL, 0);
     } else if (!strcmp(argv[c], "-alpha_comp") && c < argc - 1) {
       config.alpha_compression = strtol(argv[++c], NULL, 0);
     } else if (!strcmp(argv[c], "-noalpha")) {

man/cwebp.1

@@ -1,5 +1,5 @@
 .\"                                      Hey, EMACS: -*- nroff -*-
-.TH CWEBP 1 "June  20, 2011"
+.TH CWEBP 1 "August 23, 2011"
 .SH NAME
 cwebp \- compress an image file to a WebP file
 .SH SYNOPSIS
@@ -86,6 +86,25 @@ used thanks to the \fB\-f\fP option). Strong filtering is off by default.
 Change the number of partitions to use during the segmentation of the
 sns algorithm. Segments should be in range 1 to 4. Default value is 4.
 .TP
+.B \-partition_limit int
+Degrade quality by limiting the number of bits used by some macroblocks.
+Range is 0 (no degradation, the default) to 100 (full degradation).
+Useful values are usually around 30-70 for moderately large images.
+In the VP8 format, the so-called control partition has a limit of 512k and
+is used to store the following information: whether the macroblock is skipped,
+which segment it belongs to, whether it is coded as intra 4x4 or intra 16x16
+mode, and finally the prediction modes to use for each of the sub-blocks.
+For a very large image, 512k only leaves room to few bits per 16x16 macroblock.
+The absolute minimum is 4 bits per macroblock. Skip, segment, and mode
+information can use up almost all these 4 bits (although the case is unlikely),
+which is problematic for very large images. The partition_limit factor controls
+how frequently the most bit-costly mode (intra 4x4) will be used. This is
+useful in case the 512k limit is reached and the following message is displayed:
+\fIError code: 6 (PARTITION0_OVERFLOW: Partition #0 is too big to fit 512k)\fP.
+If using \fB-partition_limit\fP is not enough to meet the 512k constraint, one
+should use less segments in order to save more header bits per macroblock.
+See the \fB-segments\fP option.
+.TP
 .B \-size int
 Specify a target size (in bytes) to try and reach for the compressed output.
 Compressor will make several pass of partial encoding in order to get as

src/enc/config.c

@@ -42,6 +42,7 @@ int WebPConfigInitInternal(WebPConfig* const config,
   config->preprocessing = 0;
   config->autofilter = 0;
   config->alpha_compression = 0;
+  config->partition_limit = 0;
 
   // TODO(skal): tune.
   switch (preset) {
@@ -106,6 +107,8 @@ int WebPValidateConfig(const WebPConfig* const config) {
     return 0;
   if (config->partitions < 0 || config->partitions > 3)
     return 0;
+  if (config->partition_limit < 0 || config->partition_limit > 100)
+    return 0;
   if (config->alpha_compression < 0)
     return 0;
   return 1;

src/enc/quant.c

@@ -757,8 +757,13 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
   const int tlambda = dqm->tlambda_;
   const uint8_t* const src0 = it->yuv_in_ + Y_OFF;
   uint8_t* const best_blocks = it->yuv_out2_ + Y_OFF;
+  int total_header_bits = 0;
   VP8ModeScore rd_best;
 
+  if (enc->max_i4_header_bits_ == 0) {
+    return 0;
+  }
+
   InitScore(&rd_best);
   rd_best.score = 211;  // '211' is the value of VP8BitCost(0, 145)
   VP8IteratorStartI4(it);
@@ -799,7 +804,9 @@ static int PickBestIntra4(VP8EncIterator* const it, VP8ModeScore* const rd) {
     }
     SetRDScore(dqm->lambda_mode_, &rd_i4);
     AddScore(&rd_best, &rd_i4);
-    if (rd_best.score >= rd->score) {
+    total_header_bits += mode_costs[best_mode];
+    if (rd_best.score >= rd->score ||
+        total_header_bits > enc->max_i4_header_bits_) {
       return 0;
     }
     // Copy selected samples if not in the right place already.

src/enc/vp8enci.h

@@ -359,8 +359,9 @@ struct VP8Encoder {
   int      block_count_[3];
 
   // quality/speed settings
-  int method_;             // 0=fastest, 6=best/slowest.
-  int rd_opt_level_;       // Deduced from method_.
+  int method_;              // 0=fastest, 6=best/slowest.
+  int rd_opt_level_;        // Deduced from method_.
+  int max_i4_header_bits_;  // partition #0 safeness factor
 
   // Memory
   VP8MBInfo* mb_info_;   // contextual macroblock infos (mb_w_ + 1)

src/enc/webpenc.c

@@ -112,11 +112,15 @@ static void ResetBoundaryPredictions(VP8Encoder* const enc) {
 
 static void MapConfigToTools(VP8Encoder* const enc) {
   const int method = enc->config_->method;
+  const int limit = 100 - enc->config_->partition_limit;
   enc->method_ = method;
   enc->rd_opt_level_ = (method >= 6) ? 3
                      : (method >= 5) ? 2
                      : (method >= 3) ? 1
                      : 0;
+  enc->max_i4_header_bits_ =
+      256 * 16 * 16 *                 // upper bound: up to 16bit per 4x4 block
+      (limit * limit) / (100 * 100);  // ... modulated with a quadratic curve.
 }
 
 // Memory scaling with dimensions:

src/webp/encode.h

@@ -69,6 +69,8 @@ typedef struct {
   int preprocessing;     // preprocessing filter (0=none, 1=segment-smooth)
   int partitions;        // log2(number of token partitions) in [0..3]
                          // Default is set to 0 for easier progressive decoding.
+  int partition_limit;   // quality degradation allowed to fit the 512k limit on
+                         // prediction modes coding (0=no degradation, 100=full)
   int alpha_compression;  // Algorithm for optimizing the alpha plane (0 = none)
 } WebPConfig;