Merge pull request #1440 from rouault/rate_alloc_speedup

Significant speed-up rate allocation by rate/distoratio ratio

src/bin/jp2/opj_compress.c

@@ -846,7 +846,7 @@ static int parse_cmdline_encoder(int argc, char **argv,
 
         /* ----------------------------------------------------- */
 
-        case 'q': {         /* add fixed_quality */
+        case 'q': {         /* layer allocation by distortion ratio (PSNR) */
             char *s = opj_optarg;
             while (sscanf(s, "%f", &parameters->tcp_distoratio[parameters->tcp_numlayers])
                     == 1) {
@@ -866,7 +866,7 @@ static int parse_cmdline_encoder(int argc, char **argv,
         /* dda */
         /* ----------------------------------------------------- */
 
-        case 'f': {         /* mod fixed_quality (before : -q) */
+        case 'f': {         /* layer allocation by fixed layer */
             int *row = NULL, *col = NULL;
             OPJ_UINT32 numlayers = 0, numresolution = 0, matrix_width = 0;
 
@@ -1812,7 +1812,7 @@ static int parse_cmdline_encoder(int argc, char **argv,
                   parameters->cp_fixed_quality))) {
         fprintf(stderr, "[ERROR] options -r -q and -f cannot be used together !!\n");
         return 1;
-    }               /* mod fixed_quality */
+    }
 
 
     /* if no rate entered, lossless by default */

src/lib/openjp2/bio.c

@@ -43,12 +43,6 @@
 /** @name Local static functions */
 /*@{*/
 
-/**
-Write a bit
-@param bio BIO handle
-@param b Bit to write (0 or 1)
-*/
-static void opj_bio_putbit(opj_bio_t *bio, OPJ_UINT32 b);
 /**
 Read a bit
 @param bio BIO handle
@@ -100,16 +94,6 @@ static OPJ_BOOL opj_bio_bytein(opj_bio_t *bio)
     return OPJ_TRUE;
 }
 
-static void opj_bio_putbit(opj_bio_t *bio, OPJ_UINT32 b)
-{
-    if (bio->ct == 0) {
-        opj_bio_byteout(
-            bio); /* MSD: why not check the return value of this function ? */
-    }
-    bio->ct--;
-    bio->buf |= b << bio->ct;
-}
-
 static OPJ_UINT32 opj_bio_getbit(opj_bio_t *bio)
 {
     if (bio->ct == 0) {
@@ -162,6 +146,16 @@ void opj_bio_init_dec(opj_bio_t *bio, OPJ_BYTE *bp, OPJ_UINT32 len)
     bio->ct = 0;
 }
 
+void opj_bio_putbit(opj_bio_t *bio, OPJ_UINT32 b)
+{
+    if (bio->ct == 0) {
+        opj_bio_byteout(
+            bio); /* MSD: why not check the return value of this function ? */
+    }
+    bio->ct--;
+    bio->buf |= b << bio->ct;
+}
+
 void opj_bio_write(opj_bio_t *bio, OPJ_UINT32 v, OPJ_UINT32 n)
 {
     OPJ_INT32 i;

src/lib/openjp2/bio.h

@@ -106,6 +106,14 @@ Write bits
 @param n Number of bits to write
 */
 void opj_bio_write(opj_bio_t *bio, OPJ_UINT32 v, OPJ_UINT32 n);
+
+/**
+Write a bit
+@param bio BIO handle
+@param b Bit to write (0 or 1)
+*/
+void opj_bio_putbit(opj_bio_t *bio, OPJ_UINT32 b);
+
 /**
 Read bits
 @param bio BIO handle

src/lib/openjp2/j2k.c

@@ -7666,6 +7666,27 @@ OPJ_BOOL opj_j2k_setup_encoder(opj_j2k_t *p_j2k,
         return OPJ_FALSE;
     }
 
+    if (parameters->cp_fixed_alloc) {
+        if (parameters->cp_matrice == NULL) {
+            opj_event_msg(p_manager, EVT_ERROR,
+                          "cp_fixed_alloc set, but cp_matrice missing\n");
+            return OPJ_FALSE;
+        }
+
+        if (parameters->tcp_numlayers > J2K_TCD_MATRIX_MAX_LAYER_COUNT) {
+            opj_event_msg(p_manager, EVT_ERROR,
+                          "tcp_numlayers when cp_fixed_alloc set should not exceed %d\n",
+                          J2K_TCD_MATRIX_MAX_LAYER_COUNT);
+            return OPJ_FALSE;
+        }
+        if (parameters->numresolution > J2K_TCD_MATRIX_MAX_RESOLUTION_COUNT) {
+            opj_event_msg(p_manager, EVT_ERROR,
+                          "numresolution when cp_fixed_alloc set should not exceed %d\n",
+                          J2K_TCD_MATRIX_MAX_RESOLUTION_COUNT);
+            return OPJ_FALSE;
+        }
+    }
+
     p_j2k->m_specific_param.m_encoder.m_nb_comps = image->numcomps;
 
     /* keep a link to cp so that we can destroy it later in j2k_destroy_compress */
@@ -7885,15 +7906,17 @@ OPJ_BOOL opj_j2k_setup_encoder(opj_j2k_t *p_j2k,
     cp->m_specific_param.m_enc.m_max_comp_size = (OPJ_UINT32)
             parameters->max_comp_size;
     cp->rsiz = parameters->rsiz;
-    cp->m_specific_param.m_enc.m_disto_alloc = (OPJ_UINT32)
+    if (parameters->cp_fixed_alloc) {
-            parameters->cp_disto_alloc & 1u;
+        cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy = FIXED_LAYER;
-    cp->m_specific_param.m_enc.m_fixed_alloc = (OPJ_UINT32)
+    } else if (parameters->cp_fixed_quality) {
-            parameters->cp_fixed_alloc & 1u;
+        cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy =
-    cp->m_specific_param.m_enc.m_fixed_quality = (OPJ_UINT32)
+            FIXED_DISTORTION_RATIO;
-            parameters->cp_fixed_quality & 1u;
+    } else {
+        cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy =
+            RATE_DISTORTION_RATIO;
+    }
 
-    /* mod fixed_quality */
+    if (parameters->cp_fixed_alloc) {
-    if (parameters->cp_fixed_alloc && parameters->cp_matrice) {
         size_t array_size = (size_t)parameters->tcp_numlayers *
                             (size_t)parameters->numresolution * 3 * sizeof(OPJ_INT32);
         cp->m_specific_param.m_enc.m_matrice = (OPJ_INT32 *) opj_malloc(array_size);
@@ -8051,22 +8074,25 @@ OPJ_BOOL opj_j2k_setup_encoder(opj_j2k_t *p_j2k,
 
     for (tileno = 0; tileno < cp->tw * cp->th; tileno++) {
         opj_tcp_t *tcp = &cp->tcps[tileno];
+        const OPJ_BOOL fixed_distoratio =
+            cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
+            FIXED_DISTORTION_RATIO;
         tcp->numlayers = (OPJ_UINT32)parameters->tcp_numlayers;
 
         for (j = 0; j < tcp->numlayers; j++) {
             if (OPJ_IS_CINEMA(cp->rsiz) || OPJ_IS_IMF(cp->rsiz)) {
-                if (cp->m_specific_param.m_enc.m_fixed_quality) {
+                if (fixed_distoratio) {
                     tcp->distoratio[j] = parameters->tcp_distoratio[j];
                 }
                 tcp->rates[j] = parameters->tcp_rates[j];
             } else {
-                if (cp->m_specific_param.m_enc.m_fixed_quality) {       /* add fixed_quality */
+                if (fixed_distoratio) {
                     tcp->distoratio[j] = parameters->tcp_distoratio[j];
                 } else {
                     tcp->rates[j] = parameters->tcp_rates[j];
                 }
             }
-            if (!cp->m_specific_param.m_enc.m_fixed_quality &&
+            if (!fixed_distoratio &&
                     tcp->rates[j] <= 1.0) {
                 tcp->rates[j] = 0.0;    /* force lossless */
             }

src/lib/openjp2/j2k.h

@@ -113,6 +113,9 @@ The functions in J2K.C have for goal to read/write the several parts of the code
 
 #define J2K_MAX_POCS    32      /**< Maximum number of POCs */
 
+#define J2K_TCD_MATRIX_MAX_LAYER_COUNT 10
+#define J2K_TCD_MATRIX_MAX_RESOLUTION_COUNT 10
+
 /* ----------------------------------------------------------------------- */
 
 /**
@@ -272,7 +275,7 @@ typedef struct opj_tcp {
     OPJ_UINT32 ppt_data_size;
     /** size of ppt_data*/
     OPJ_UINT32 ppt_len;
-    /** add fixed_quality */
+    /** PSNR values */
     OPJ_FLOAT32 distoratio[100];
     /** tile-component coding parameters */
     opj_tccp_t *tccps;
@@ -314,6 +317,14 @@ typedef struct opj_tcp {
 } opj_tcp_t;
 
 
+/**
+Rate allocation strategy
+*/
+typedef enum {
+    RATE_DISTORTION_RATIO = 0,    /** allocation by rate/distortion */
+    FIXED_DISTORTION_RATIO = 1,   /** allocation by fixed distortion ratio (PSNR) (fixed quality) */
+    FIXED_LAYER = 2,              /** allocation by fixed layer (number of passes per layer / resolution / subband) */
+} J2K_QUALITY_LAYER_ALLOCATION_STRATEGY;
 
 
 typedef struct opj_encoding_param {
@@ -325,12 +336,8 @@ typedef struct opj_encoding_param {
     OPJ_INT32 *m_matrice;
     /** Flag determining tile part generation*/
     OPJ_BYTE m_tp_flag;
-    /** allocation by rate/distortion */
+    /** Quality layer allocation strategy */
-    OPJ_BITFIELD m_disto_alloc : 1;
+    J2K_QUALITY_LAYER_ALLOCATION_STRATEGY m_quality_layer_alloc_strategy;
-    /** allocation by fixed layer */
-    OPJ_BITFIELD m_fixed_alloc : 1;
-    /** add fixed_quality */
-    OPJ_BITFIELD m_fixed_quality : 1;
     /** Enabling Tile part generation*/
     OPJ_BITFIELD m_tp_on : 1;
 }

src/lib/openjp2/openjpeg.h

@@ -405,7 +405,7 @@ typedef struct opj_cparameters {
     int cp_disto_alloc;
     /** allocation by fixed layer */
     int cp_fixed_alloc;
-    /** add fixed_quality */
+    /** allocation by fixed quality (PSNR) */
     int cp_fixed_quality;
     /** fixed layer */
     int *cp_matrice;
@@ -829,9 +829,9 @@ typedef struct opj_tile_info {
     int pdy[33];
     /** information concerning packets inside tile */
     opj_packet_info_t *packet;
-    /** add fixed_quality */
+    /** number of pixels of the tile */
     int numpix;
-    /** add fixed_quality */
+    /** distortion of the tile */
     double distotile;
     /** number of markers */
     int marknum;

src/lib/openjp2/t1.c

@@ -1410,7 +1410,6 @@ static void opj_t1_dec_clnpass(
 }
 
 
-/** mod fixed_quality */
 static OPJ_FLOAT64 opj_t1_getwmsedec(
     OPJ_INT32 nmsedec,
     OPJ_UINT32 compno,
@@ -2313,7 +2312,7 @@ OPJ_BOOL opj_t1_encode_cblks(opj_tcd_t* tcd,
     OPJ_UINT32 compno, resno, bandno, precno, cblkno;
     opj_mutex_t* mutex = opj_mutex_create();
 
-    tile->distotile = 0;        /* fixed_quality */
+    tile->distotile = 0;
 
     for (compno = 0; compno < tile->numcomps; ++compno) {
         opj_tcd_tilecomp_t* tilec = &tile->comps[compno];
@@ -2401,7 +2400,6 @@ static int opj_t1_enc_is_term_pass(opj_tcd_cblk_enc_t* cblk,
 }
 
 
-/** mod fixed_quality */
 static OPJ_FLOAT64 opj_t1_encode_cblk(opj_t1_t *t1,
                                       opj_tcd_cblk_enc_t* cblk,
                                       OPJ_UINT32 orient,
@@ -2505,7 +2503,6 @@ static OPJ_FLOAT64 opj_t1_encode_cblk(opj_t1_t *t1,
             break;
         }
 
-        /* fixed_quality */
         tempwmsedec = opj_t1_getwmsedec(nmsedec, compno, level, orient, bpno, qmfbid,
                                         stepsize, numcomps, mct_norms, mct_numcomps) ;
         cumwmsedec += tempwmsedec;

src/lib/openjp2/t2.c

@@ -167,9 +167,9 @@ static OPJ_BOOL opj_t2_init_seg(opj_tcd_cblk_dec_t* cblk,
 static void opj_t2_putcommacode(opj_bio_t *bio, OPJ_INT32 n)
 {
     while (--n >= 0) {
-        opj_bio_write(bio, 1, 1);
+        opj_bio_putbit(bio, 1);
     }
-    opj_bio_write(bio, 0, 1);
+    opj_bio_putbit(bio, 0);
 }
 
 static OPJ_UINT32 opj_t2_getcommacode(opj_bio_t *bio)
@@ -184,7 +184,7 @@ static OPJ_UINT32 opj_t2_getcommacode(opj_bio_t *bio)
 static void opj_t2_putnumpasses(opj_bio_t *bio, OPJ_UINT32 n)
 {
     if (n == 1) {
-        opj_bio_write(bio, 0, 1);
+        opj_bio_putbit(bio, 0);
     } else if (n == 2) {
         opj_bio_write(bio, 2, 2);
     } else if (n <= 5) {
@@ -801,7 +801,7 @@ static OPJ_BOOL opj_t2_encode_packet(OPJ_UINT32 tileno,
         }
     }
 #endif
-    opj_bio_write(bio, packet_empty ? 0 : 1, 1);           /* Empty header bit */
+    opj_bio_putbit(bio, packet_empty ? 0 : 1);           /* Empty header bit */
 
     /* Writing Packet header */
     band = res->bands;
@@ -849,7 +849,7 @@ static OPJ_BOOL opj_t2_encode_packet(OPJ_UINT32 tileno,
             if (!cblk->numpasses) {
                 opj_tgt_encode(bio, prc->incltree, cblkno, (OPJ_INT32)(layno + 1));
             } else {
-                opj_bio_write(bio, layer->numpasses != 0, 1);
+                opj_bio_putbit(bio, layer->numpasses != 0);
             }
 
             /* if cblk not included, go to the next cblk  */
@@ -978,7 +978,9 @@ static OPJ_BOOL opj_t2_encode_packet(OPJ_UINT32 tileno,
                 return OPJ_FALSE;
             }
 
-            memcpy(c, layer->data, layer->len);
+            if (p_t2_mode == FINAL_PASS) {
+                memcpy(c, layer->data, layer->len);
+            }
             cblk->numpasses += layer->numpasses;
             c += layer->len;
             length -= layer->len;

src/lib/openjp2/tcd.c

@@ -42,6 +42,8 @@
 #include "opj_includes.h"
 #include "opj_common.h"
 
+// #define DEBUG_RATE_ALLOC
+
 /* ----------------------------------------------------------------------- */
 
 /* TODO MSD: */
@@ -143,6 +145,9 @@ static OPJ_BOOL opj_tcd_code_block_enc_allocate_data(opj_tcd_cblk_enc_t *
  */
 static void opj_tcd_code_block_enc_deallocate(opj_tcd_precinct_t * p_precinct);
 
+static
+void opj_tcd_makelayer_fixed(opj_tcd_t *tcd, OPJ_UINT32 layno,
+                             OPJ_UINT32 final);
 
 /**
 Free the memory allocated for encoding
@@ -224,6 +229,7 @@ opj_tcd_t* opj_tcd_create(OPJ_BOOL p_is_decoder)
 
 /* ----------------------------------------------------------------------- */
 
+static
 void opj_tcd_rateallocate_fixed(opj_tcd_t *tcd)
 {
     OPJ_UINT32 layno;
@@ -234,17 +240,23 @@ void opj_tcd_rateallocate_fixed(opj_tcd_t *tcd)
 }
 
 
-void opj_tcd_makelayer(opj_tcd_t *tcd,
+/* ----------------------------------------------------------------------- */
-                       OPJ_UINT32 layno,
+
-                       OPJ_FLOAT64 thresh,
+/** Returns OPJ_TRUE if the layer allocation is unchanged w.r.t to the previous
-                       OPJ_UINT32 final)
+ * invokation with a different threshold */
+static
+OPJ_BOOL opj_tcd_makelayer(opj_tcd_t *tcd,
+                           OPJ_UINT32 layno,
+                           OPJ_FLOAT64 thresh,
+                           OPJ_UINT32 final)
 {
     OPJ_UINT32 compno, resno, bandno, precno, cblkno;
     OPJ_UINT32 passno;
 
     opj_tcd_tile_t *tcd_tile = tcd->tcd_image->tiles;
+    OPJ_BOOL layer_allocation_is_same = OPJ_TRUE;
 
-    tcd_tile->distolayer[layno] = 0;        /* fixed_quality */
+    tcd_tile->distolayer[layno] = 0;
 
     for (compno = 0; compno < tcd_tile->numcomps; compno++) {
         opj_tcd_tilecomp_t *tilec = &tcd_tile->comps[compno];
@@ -304,7 +316,10 @@ void opj_tcd_makelayer(opj_tcd_t *tcd,
                             }
                         }
 
-                        layer->numpasses = n - cblk->numpassesinlayers;
+                        if (layer->numpasses != n - cblk->numpassesinlayers) {
+                            layer_allocation_is_same = OPJ_FALSE;
+                            layer->numpasses = n - cblk->numpassesinlayers;
+                        }
 
                         if (!layer->numpasses) {
                             layer->disto = 0;
@@ -323,7 +338,7 @@ void opj_tcd_makelayer(opj_tcd_t *tcd,
                                            cblk->passes[cblk->numpassesinlayers - 1].distortiondec;
                         }
 
-                        tcd_tile->distolayer[layno] += layer->disto;    /* fixed_quality */
+                        tcd_tile->distolayer[layno] += layer->disto;
 
                         if (final) {
                             cblk->numpassesinlayers = n;
@@ -333,14 +348,17 @@ void opj_tcd_makelayer(opj_tcd_t *tcd,
             }
         }
     }
+    return layer_allocation_is_same;
 }
 
+/** For m_quality_layer_alloc_strategy == FIXED_LAYER */
+static
 void opj_tcd_makelayer_fixed(opj_tcd_t *tcd, OPJ_UINT32 layno,
                              OPJ_UINT32 final)
 {
     OPJ_UINT32 compno, resno, bandno, precno, cblkno;
     OPJ_INT32 value;                        /*, matrice[tcd_tcp->numlayers][tcd_tile->comps[0].numresolutions][3]; */
-    OPJ_INT32 matrice[10][10][3];
+    OPJ_INT32 matrice[J2K_TCD_MATRIX_MAX_LAYER_COUNT][J2K_TCD_MATRIX_MAX_RESOLUTION_COUNT][3];
     OPJ_UINT32 i, j, k;
 
     opj_cp_t *cp = tcd->cp;
@@ -440,6 +458,11 @@ void opj_tcd_makelayer_fixed(opj_tcd_t *tcd, OPJ_UINT32 layno,
     }
 }
 
+/** Rate allocation for the following methods:
+ * - allocation by rate/distortio (m_quality_layer_alloc_strategy == RATE_DISTORTION_RATIO)
+ * - allocation by fixed quality  (m_quality_layer_alloc_strategy == FIXED_DISTORTION_RATIO)
+ */
+static
 OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
                               OPJ_BYTE *dest,
                               OPJ_UINT32 * p_data_written,
@@ -450,8 +473,8 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
     OPJ_UINT32 compno, resno, bandno, precno, cblkno, layno;
     OPJ_UINT32 passno;
     OPJ_FLOAT64 min, max;
-    OPJ_FLOAT64 cumdisto[100];      /* fixed_quality */
+    OPJ_FLOAT64 cumdisto[100];
-    const OPJ_FLOAT64 K = 1;                /* 1.1; fixed_quality */
+    const OPJ_FLOAT64 K = 1;
     OPJ_FLOAT64 maxSE = 0;
 
     opj_cp_t *cp = tcd->cp;
@@ -461,7 +484,7 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
     min = DBL_MAX;
     max = 0;
 
-    tcd_tile->numpix = 0;           /* fixed_quality */
+    tcd_tile->numpix = 0;
 
     for (compno = 0; compno < tcd_tile->numcomps; compno++) {
         opj_tcd_tilecomp_t *tilec = &tcd_tile->comps[compno];
@@ -511,9 +534,12 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
                             }
                         } /* passno */
 
-                        /* fixed_quality */
+                        {
-                        tcd_tile->numpix += ((cblk->x1 - cblk->x0) * (cblk->y1 - cblk->y0));
+                            const OPJ_SIZE_T cblk_pix_count = (OPJ_SIZE_T)((cblk->x1 - cblk->x0) *
-                        tilec->numpix += ((cblk->x1 - cblk->x0) * (cblk->y1 - cblk->y0));
+                                                              (cblk->y1 - cblk->y0));
+                            tcd_tile->numpix += cblk_pix_count;
+                            tilec->numpix += cblk_pix_count;
+                        }
                     } /* cbklno */
                 } /* precno */
             } /* bandno */
@@ -527,8 +553,8 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
     /* index file */
     if (cstr_info) {
         opj_tile_info_t *tile_info = &cstr_info->tile[tcd->tcd_tileno];
-        tile_info->numpix = tcd_tile->numpix;
+        tile_info->numpix = (int)tcd_tile->numpix;
-        tile_info->distotile = tcd_tile->distotile;
+        tile_info->distotile = (int)tcd_tile->distotile;
         tile_info->thresh = (OPJ_FLOAT64 *) opj_malloc(tcd_tcp->numlayers * sizeof(
                                 OPJ_FLOAT64));
         if (!tile_info->thresh) {
@@ -545,35 +571,54 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
         OPJ_FLOAT64 goodthresh = 0;
         OPJ_FLOAT64 stable_thresh = 0;
         OPJ_UINT32 i;
-        OPJ_FLOAT64 distotarget;                /* fixed_quality */
+        OPJ_FLOAT64 distotarget;
 
-        /* fixed_quality */
         distotarget = tcd_tile->distotile - ((K * maxSE) / pow((OPJ_FLOAT32)10,
                                              tcd_tcp->distoratio[layno] / 10));
 
         /* Don't try to find an optimal threshold but rather take everything not included yet, if
-          -r xx,yy,zz,0   (disto_alloc == 1 and rates == 0)
+          -r xx,yy,zz,0   (m_quality_layer_alloc_strategy == RATE_DISTORTION_RATIO and rates == NULL)
-          -q xx,yy,zz,0   (fixed_quality == 1 and distoratio == 0)
+          -q xx,yy,zz,0   (m_quality_layer_alloc_strategy == FIXED_DISTORTION_RATIO and distoratio == NULL)
           ==> possible to have some lossy layers and the last layer for sure lossless */
-        if (((cp->m_specific_param.m_enc.m_disto_alloc == 1) &&
+        if (((cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
+                RATE_DISTORTION_RATIO) &&
                 (tcd_tcp->rates[layno] > 0.0f)) ||
-                ((cp->m_specific_param.m_enc.m_fixed_quality == 1) &&
+                ((cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
+                  FIXED_DISTORTION_RATIO) &&
                  (tcd_tcp->distoratio[layno] > 0.0))) {
             opj_t2_t*t2 = opj_t2_create(tcd->image, cp);
             OPJ_FLOAT64 thresh = 0;
+            OPJ_BOOL last_layer_allocation_ok = OPJ_FALSE;
 
             if (t2 == 00) {
                 return OPJ_FALSE;
             }
 
             for (i = 0; i < 128; ++i) {
-                OPJ_FLOAT64 distoachieved = 0;  /* fixed_quality */
+                OPJ_FLOAT64 distoachieved = 0;
+                OPJ_BOOL layer_allocation_is_same;
 
-                thresh = (lo + hi) / 2;
+                OPJ_FLOAT64 new_thresh = (lo + hi) / 2;
+                /* Stop iterating when the threshold has stabilized enough */
+                /* 0.5 * 1e-5 is somewhat arbitrary, but has been selected */
+                /* so that this doesn't change the results of the regression */
+                /* test suite. */
+                if (fabs(new_thresh - thresh) <= 0.5 * 1e-5 * thresh) {
+                    break;
+                }
+                thresh = new_thresh;
+#ifdef DEBUG_RATE_ALLOC
+                opj_event_msg(p_manager, EVT_INFO, "layno=%u, iter=%u, thresh=%g",
+                              layno, i, new_thresh);
+#endif
 
-                opj_tcd_makelayer(tcd, layno, thresh, 0);
+                layer_allocation_is_same = opj_tcd_makelayer(tcd, layno, thresh, 0) && i != 0;
-
+#ifdef DEBUG_RATE_ALLOC
-                if (cp->m_specific_param.m_enc.m_fixed_quality) {       /* fixed_quality */
+                opj_event_msg(p_manager, EVT_INFO, "--> layer_allocation_is_same = %d",
+                              layer_allocation_is_same);
+#endif
+                if (cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
+                        FIXED_DISTORTION_RATIO) {
                     if (OPJ_IS_CINEMA(cp->rsiz) || OPJ_IS_IMF(cp->rsiz)) {
                         if (! opj_t2_encode_packets(t2, tcd->tcd_tileno, tcd_tile, layno + 1, dest,
                                                     p_data_written, maxlen, cstr_info, NULL, tcd->cur_tp_num, tcd->tp_pos,
@@ -605,17 +650,41 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
                         }
                         lo = thresh;
                     }
-                } else {
+                } else { /* Disto/rate based optimization */
-                    if (! opj_t2_encode_packets(t2, tcd->tcd_tileno, tcd_tile, layno + 1, dest,
+                    /* Check if the layer allocation done by opj_tcd_makelayer()
-                                                p_data_written, maxlen, cstr_info, NULL, tcd->cur_tp_num, tcd->tp_pos,
+                     * is compatible of the maximum rate allocation. If not,
-                                                tcd->cur_pino,
+                     * retry with a higher threshold.
-                                                THRESH_CALC, p_manager)) {
+                     * If OK, try with a lower threshold.
-                        /* TODO: what to do with l ??? seek / tell ??? */
+                     * Call opj_t2_encode_packets() only if opj_tcd_makelayer()
-                        /* opj_event_msg(tcd->cinfo, EVT_INFO, "rate alloc: len=%d, max=%d\n", l, maxlen); */
+                     * has resulted in different truncation points since its last
+                     * call. */
+                    if ((layer_allocation_is_same && !last_layer_allocation_ok) ||
+                            (!layer_allocation_is_same &&
+                             ! opj_t2_encode_packets(t2, tcd->tcd_tileno, tcd_tile, layno + 1, dest,
+                                                     p_data_written, maxlen, cstr_info, NULL, tcd->cur_tp_num, tcd->tp_pos,
+                                                     tcd->cur_pino,
+                                                     THRESH_CALC, p_manager))) {
+
+#ifdef DEBUG_RATE_ALLOC
+                        if (!layer_allocation_is_same) {
+                            opj_event_msg(p_manager, EVT_INFO,
+                                          "--> check rate alloc failed (> maxlen=%u)\n", maxlen);
+                        }
+#endif
+                        last_layer_allocation_ok = OPJ_FALSE;
                         lo = thresh;
                         continue;
                     }
 
+#ifdef DEBUG_RATE_ALLOC
+                    if (!layer_allocation_is_same) {
+                        opj_event_msg(p_manager, EVT_INFO,
+                                      "--> check rate alloc success (len=%u <= maxlen=%u)\n", *p_data_written,
+                                      maxlen);
+                    }
+#endif
+
+                    last_layer_allocation_ok = OPJ_TRUE;
                     hi = thresh;
                     stable_thresh = thresh;
                 }
@@ -635,7 +704,6 @@ OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
 
         opj_tcd_makelayer(tcd, layno, goodthresh, 1);
 
-        /* fixed_quality */
         cumdisto[layno] = (layno == 0) ? tcd_tile->distolayer[0] :
                           (cumdisto[layno - 1] + tcd_tile->distolayer[layno]);
     }
@@ -2599,10 +2667,10 @@ static OPJ_BOOL opj_tcd_rate_allocate_encode(opj_tcd_t *p_tcd,
         p_cstr_info->index_write = 0;
     }
 
-    if (l_cp->m_specific_param.m_enc.m_disto_alloc ||
+    if (l_cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
-            l_cp->m_specific_param.m_enc.m_fixed_quality)  {
+            RATE_DISTORTION_RATIO ||
-        /* fixed_quality */
+            l_cp->m_specific_param.m_enc.m_quality_layer_alloc_strategy ==
-        /* Normal Rate/distortion allocation */
+            FIXED_DISTORTION_RATIO)  {
         if (! opj_tcd_rateallocate(p_tcd, p_dest_data, &l_nb_written, p_max_dest_size,
                                    p_cstr_info, p_manager)) {
             return OPJ_FALSE;

src/lib/openjp2/tcd.h

@@ -222,8 +222,8 @@ typedef struct opj_tcd_tilecomp {
     OPJ_UINT32 win_x1;
     OPJ_UINT32 win_y1;
 
-    /* add fixed_quality */
+    /* number of pixels */
-    OPJ_INT32 numpix;
+    OPJ_SIZE_T numpix;
 } opj_tcd_tilecomp_t;
 
 
@@ -235,9 +235,9 @@ typedef struct opj_tcd_tile {
     OPJ_INT32 x0, y0, x1, y1;
     OPJ_UINT32 numcomps;            /* number of components in tile */
     opj_tcd_tilecomp_t *comps;  /* Components information */
-    OPJ_INT32 numpix;               /* add fixed_quality */
+    OPJ_SIZE_T numpix;               /* number of pixels */
-    OPJ_FLOAT64 distotile;          /* add fixed_quality */
+    OPJ_FLOAT64 distotile;          /* distortion of the tile */
-    OPJ_FLOAT64 distolayer[100];    /* add fixed_quality */
+    OPJ_FLOAT64 distolayer[100];    /* distortion per layer */
     OPJ_UINT32 packno;              /* packet number */
 } opj_tcd_tile_t;
 
@@ -369,23 +369,6 @@ OPJ_BOOL opj_tcd_init(opj_tcd_t *p_tcd,
 OPJ_BOOL opj_tcd_init_decode_tile(opj_tcd_t *p_tcd, OPJ_UINT32 p_tile_no,
                                   opj_event_mgr_t* p_manager);
 
-void opj_tcd_makelayer_fixed(opj_tcd_t *tcd, OPJ_UINT32 layno,
-                             OPJ_UINT32 final);
-
-void opj_tcd_rateallocate_fixed(opj_tcd_t *tcd);
-
-void opj_tcd_makelayer(opj_tcd_t *tcd,
-                       OPJ_UINT32 layno,
-                       OPJ_FLOAT64 thresh,
-                       OPJ_UINT32 final);
-
-OPJ_BOOL opj_tcd_rateallocate(opj_tcd_t *tcd,
-                              OPJ_BYTE *dest,
-                              OPJ_UINT32 * p_data_written,
-                              OPJ_UINT32 len,
-                              opj_codestream_info_t *cstr_info,
-                              opj_event_mgr_t *p_manager);
-
 /**
  * Gets the maximum tile size that will be taken by the tile once decoded.
  */

src/lib/openjp2/tgt.c

@@ -287,12 +287,12 @@ void opj_tgt_encode(opj_bio_t *bio, opj_tgt_tree_t *tree, OPJ_UINT32 leafno,
         while (low < threshold) {
             if (low >= node->value) {
                 if (!node->known) {
-                    opj_bio_write(bio, 1, 1);
+                    opj_bio_putbit(bio, 1);
                     node->known = 1;
                 }
                 break;
             }
-            opj_bio_write(bio, 0, 1);
+            opj_bio_putbit(bio, 0);
             ++low;
         }