Implement optimal huffman encoding for (M)JPEG.

> seems to break
> make fate-vsynth1-mjpeg-444

Fixed.

11 files changed, 847 insertions, 91 deletions

diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 43a6add317..4765e9c96c 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -39,6 +39,7 @@ OBJS = allcodecs.o                                                      \
        mediacodec.o                                                     \
        mpeg12framerate.o                                                \
        options.o                                                        \
+       mjpegenc_huffman.o                                               \
        parser.o                                                         \
        profiles.o                                                       \
        qsv_api.o                                                        \
@@ -175,7 +176,8 @@ OBJS-$(CONFIG_AMRWB_DECODER)           += amrwbdec.o celp_filters.o   \
                                           celp_math.o acelp_filters.o \
                                           acelp_vectors.o             \
                                           acelp_pitch_delay.o
-OBJS-$(CONFIG_AMV_ENCODER)             += mjpegenc.o mjpegenc_common.o
+OBJS-$(CONFIG_AMV_ENCODER)             += mjpegenc.o mjpegenc_common.o \
+                                          mjpegenc_huffman.o
 OBJS-$(CONFIG_ANM_DECODER)             += anm.o
 OBJS-$(CONFIG_ANSI_DECODER)            += ansi.o cga_data.o
 OBJS-$(CONFIG_APE_DECODER)             += apedec.o
@@ -377,7 +379,8 @@ OBJS-$(CONFIG_METASOUND_DECODER)       += metasound.o metasound_data.o \
 OBJS-$(CONFIG_MICRODVD_DECODER)        += microdvddec.o ass.o
 OBJS-$(CONFIG_MIMIC_DECODER)           += mimic.o
 OBJS-$(CONFIG_MJPEG_DECODER)           += mjpegdec.o
-OBJS-$(CONFIG_MJPEG_ENCODER)           += mjpegenc.o mjpegenc_common.o
+OBJS-$(CONFIG_MJPEG_ENCODER)           += mjpegenc.o mjpegenc_common.o \
+                                          mjpegenc_huffman.o
 OBJS-$(CONFIG_MJPEGB_DECODER)          += mjpegbdec.o
 OBJS-$(CONFIG_MJPEG_VAAPI_ENCODER)     += vaapi_encode_mjpeg.o
 OBJS-$(CONFIG_MLP_DECODER)             += mlpdec.o mlpdsp.o
@@ -1014,6 +1017,7 @@ TESTPROGS = avpacket                                                    \
             jpeg2000dwt                                                 \
             mathops                                                    \
             options                                                     \
+            mjpegenc_huffman                                            \
             utils                                                       \
 
 TESTPROGS-$(CONFIG_CABAC)                 += cabac
diff --git a/libavcodec/mjpegenc.c b/libavcodec/mjpegenc.c
index 3d113770df..a4899c5d6e 100644
--- a/libavcodec/mjpegenc.c
+++ b/libavcodec/mjpegenc.c
@@ -39,39 +39,15 @@
 #include "mjpeg.h"
 #include "mjpegenc.h"
 
-static uint8_t uni_ac_vlc_len[64 * 64 * 2];
-static uint8_t uni_chroma_ac_vlc_len[64 * 64 * 2];
-
-static av_cold void init_uni_ac_vlc(const uint8_t huff_size_ac[256], uint8_t *uni_ac_vlc_len)
-{
-    int i;
-
-    for (i = 0; i < 128; i++) {
-        int level = i - 64;
-        int run;
-        if (!level)
-            continue;
-        for (run = 0; run < 64; run++) {
-            int len, code, nbits;
-            int alevel = FFABS(level);
-
-            len = (run >> 4) * huff_size_ac[0xf0];
-
-            nbits= av_log2_16bit(alevel) + 1;
-            code = ((15&run) << 4) | nbits;
-
-            len += huff_size_ac[code] + nbits;
-
-            uni_ac_vlc_len[UNI_AC_ENC_INDEX(run, i)] = len;
-            // We ignore EOB as its just a constant which does not change generally
-        }
-    }
-}
+// Don't know, but let's guess 16 bits per code
+#define MJPEG_HUFFMAN_EST_BITS_PER_CODE 16
 
 av_cold int ff_mjpeg_encode_init(MpegEncContext *s)
 {
     MJpegContext *m;
 
+    av_assert0(s->slice_context_count == 1);
+
     if (s->width > 65500 || s->height > 65500) {
         av_log(s, AV_LOG_ERROR, "JPEG does not support resolutions above 65500x65500\n");
         return AVERROR(EINVAL);
@@ -84,7 +60,9 @@ av_cold int ff_mjpeg_encode_init(MpegEncContext *s)
     s->min_qcoeff=-1023;
     s->max_qcoeff= 1023;
 
-    /* build all the huffman tables */
+    // Build default Huffman tables.
+    // These may be overwritten later with more optimal Huffman tables, but
+    // they are needed at least right now for some processes like trellis.
     ff_mjpeg_build_huffman_codes(m->huff_size_dc_luminance,
                                  m->huff_code_dc_luminance,
                                  avpriv_mjpeg_bits_dc_luminance,
@@ -102,12 +80,18 @@ av_cold int ff_mjpeg_encode_init(MpegEncContext *s)
                                  avpriv_mjpeg_bits_ac_chrominance,
                                  avpriv_mjpeg_val_ac_chrominance);
 
-    init_uni_ac_vlc(m->huff_size_ac_luminance,   uni_ac_vlc_len);
-    init_uni_ac_vlc(m->huff_size_ac_chrominance, uni_chroma_ac_vlc_len);
+    init_uni_ac_vlc(m->huff_size_ac_luminance,   m->uni_ac_vlc_len);
+    init_uni_ac_vlc(m->huff_size_ac_chrominance, m->uni_chroma_ac_vlc_len);
     s->intra_ac_vlc_length      =
-    s->intra_ac_vlc_last_length = uni_ac_vlc_len;
+    s->intra_ac_vlc_last_length = m->uni_ac_vlc_len;
     s->intra_chroma_ac_vlc_length      =
-    s->intra_chroma_ac_vlc_last_length = uni_chroma_ac_vlc_len;
+    s->intra_chroma_ac_vlc_last_length = m->uni_chroma_ac_vlc_len;
+
+    // Buffers start out empty.
+    m->huff_buffer = NULL;
+    m->huff_ncode = 0;
+    m->huff_capacity = 0;
+    m->error = 0;
 
     s->mjpeg_ctx = m;
     return 0;
@@ -115,71 +99,193 @@ av_cold int ff_mjpeg_encode_init(MpegEncContext *s)
 
 av_cold void ff_mjpeg_encode_close(MpegEncContext *s)
 {
+    av_freep(&s->mjpeg_ctx->huff_buffer);
     av_freep(&s->mjpeg_ctx);
 }
 
+/**
+ * Encodes and outputs the entire frame in the JPEG format.
+ *
+ * @param s The MpegEncContext.
+ */
+void ff_mjpeg_encode_picture_frame(MpegEncContext *s)
+{
+    int i, nbits, code, table_id;
+    MJpegContext *m = s->mjpeg_ctx;
+    uint8_t *huff_size[4] = {m->huff_size_dc_luminance,
+                             m->huff_size_dc_chrominance,
+                             m->huff_size_ac_luminance,
+                             m->huff_size_ac_chrominance};
+    uint16_t *huff_code[4] = {m->huff_code_dc_luminance,
+                              m->huff_code_dc_chrominance,
+                              m->huff_code_ac_luminance,
+                              m->huff_code_ac_chrominance};
+    size_t total_bits = 0;
+    size_t bytes_needed;
+
+    // Estimate the total size first
+    for (i = 0; i < m->huff_ncode; i++) {
+        table_id = m->huff_buffer[i].table_id;
+        code = m->huff_buffer[i].code;
+        nbits = code & 0xf;
+
+        total_bits += huff_size[table_id][code] + nbits;
+    }
+
+    bytes_needed = (total_bits + 7) / 8;
+    ff_mpv_reallocate_putbitbuffer(s, bytes_needed, bytes_needed);
+
+    for (i = 0; i < m->huff_ncode; i++) {
+        table_id = m->huff_buffer[i].table_id;
+        code = m->huff_buffer[i].code;
+        nbits = code & 0xf;
+
+        put_bits(&s->pb, huff_size[table_id][code], huff_code[table_id][code]);
+        if (nbits != 0) {
+            put_sbits(&s->pb, nbits, m->huff_buffer[i].mant);
+        }
+    }
+
+    m->huff_ncode = 0;
+}
+
+/**
+ * Add code and table_id to the JPEG buffer.
+ *
+ * @param s The MJpegContext which contains the JPEG buffer.
+ * @param table_id Which Huffman table the code belongs to.
+ * @param code The encoded exponent of the coefficients and the run-bits.
+ */
+static inline void ff_mjpeg_encode_code(MJpegContext *s, uint8_t table_id, int code)
+{
+    MJpegHuffmanCode *c = &s->huff_buffer[s->huff_ncode++];
+    av_assert0(s->huff_ncode < s->huff_capacity);
+    c->table_id = table_id;
+    c->code = code;
+}
+
+/**
+ * Add the coefficient's data to the JPEG buffer.
+ *
+ * @param s The MJpegContext which contains the JPEG buffer.
+ * @param table_id Which Huffman table the code belongs to.
+ * @param val The coefficient.
+ * @param run The run-bits.
+ */
+static void ff_mjpeg_encode_coef(MJpegContext *s, uint8_t table_id, int val, int run)
+{
+    int mant, code;
+
+    if (val == 0) {
+        av_assert0(run == 0);
+        ff_mjpeg_encode_code(s, table_id, 0);
+    } else {
+        mant = val;
+        if (val < 0) {
+            val = -val;
+            mant--;
+        }
+
+        code = (run << 4) | (av_log2_16bit(val) + 1);
+
+        s->huff_buffer[s->huff_ncode].mant = mant;
+        ff_mjpeg_encode_code(s, table_id, code);
+    }
+}
+
+/**
+ * Add the block's data into the JPEG buffer.
+ *
+ * @param s The MJpegEncContext that contains the JPEG buffer.
+ * @param block The block.
+ * @param n The block's index or number.
+ */
 static void encode_block(MpegEncContext *s, int16_t *block, int n)
 {
-    int mant, nbits, code, i, j;
-    int component, dc, run, last_index, val;
+    int i, j, table_id;
+    int component, dc, last_index, val, run;
     MJpegContext *m = s->mjpeg_ctx;
-    uint8_t *huff_size_ac;
-    uint16_t *huff_code_ac;
+
+    if (m->error) return;
+
+    av_assert0(m->huff_capacity >= m->huff_ncode + 64);
 
     /* DC coef */
     component = (n <= 3 ? 0 : (n&1) + 1);
+    table_id = (n <= 3 ? 0 : 1);
     dc = block[0]; /* overflow is impossible */
     val = dc - s->last_dc[component];
-    if (n < 4) {
-        ff_mjpeg_encode_dc(&s->pb, val, m->huff_size_dc_luminance, m->huff_code_dc_luminance);
-        huff_size_ac = m->huff_size_ac_luminance;
-        huff_code_ac = m->huff_code_ac_luminance;
-    } else {
-        ff_mjpeg_encode_dc(&s->pb, val, m->huff_size_dc_chrominance, m->huff_code_dc_chrominance);
-        huff_size_ac = m->huff_size_ac_chrominance;
-        huff_code_ac = m->huff_code_ac_chrominance;
-    }
+
+    ff_mjpeg_encode_coef(m, table_id, val, 0);
+
     s->last_dc[component] = dc;
 
     /* AC coefs */
 
     run = 0;
     last_index = s->block_last_index[n];
+    table_id |= 2;
+
     for(i=1;i<=last_index;i++) {
         j = s->intra_scantable.permutated[i];
         val = block[j];
+
         if (val == 0) {
             run++;
         } else {
             while (run >= 16) {
-                put_bits(&s->pb, huff_size_ac[0xf0], huff_code_ac[0xf0]);
+                ff_mjpeg_encode_code(m, table_id, 0xf0);
                 run -= 16;
             }
-            mant = val;
-            if (val < 0) {
-                val = -val;
-                mant--;
-            }
-
-            nbits= av_log2_16bit(val) + 1;
-            code = (run << 4) | nbits;
-
-            put_bits(&s->pb, huff_size_ac[code], huff_code_ac[code]);
-
-            put_sbits(&s->pb, nbits, mant);
+            ff_mjpeg_encode_coef(m, table_id, val, run);
             run = 0;
         }
     }
 
     /* output EOB only if not already 64 values */
     if (last_index < 63 || run != 0)
-        put_bits(&s->pb, huff_size_ac[0], huff_code_ac[0]);
+        ff_mjpeg_encode_code(m, table_id, 0);
 }
 
-void ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64])
+// Possibly reallocate the huffman code buffer, assuming blocks_per_mb.
+// Set s->mjpeg_ctx->error on ENOMEM.
+static void realloc_huffman(MpegEncContext *s, int blocks_per_mb)
 {
-    int i;
+    MJpegContext *m = s->mjpeg_ctx;
+    size_t num_mbs, num_blocks, num_codes;
+    MJpegHuffmanCode *new_buf;
+    if (m->error) return;
+    // Make sure we have enough space to hold this frame.
+    num_mbs = s->mb_width * s->mb_height;
+    num_blocks = num_mbs * blocks_per_mb;
+    av_assert0(m->huff_ncode <=
+               (s->mb_y * s->mb_width + s->mb_x) * blocks_per_mb * 64);
+    num_codes = num_blocks * 64;
+
+    new_buf = av_fast_realloc(m->huff_buffer, &m->huff_capacity,
+                              num_codes * sizeof(MJpegHuffmanCode));
+    if (!new_buf) {
+        m->error = AVERROR(ENOMEM);
+    } else {
+        m->huff_buffer = new_buf;
+    }
+}
+
+int ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64])
+{
+    int i, is_chroma_420;
+
+    // Number of bits used depends on future data.
+    // So, nothing that relies on encoding many times and taking the
+    // one with the fewest bits will work properly here.
+    if (s->i_tex_bits != MJPEG_HUFFMAN_EST_BITS_PER_CODE *
+        s->mjpeg_ctx->huff_ncode) {
+        av_log(NULL, AV_LOG_ERROR, "Unsupported encoding method\n");
+        return AVERROR(EINVAL);
+    }
+
     if (s->chroma_format == CHROMA_444) {
+        realloc_huffman(s, 12);
         encode_block(s, block[0], 0);
         encode_block(s, block[2], 2);
         encode_block(s, block[4], 4);
@@ -196,10 +302,12 @@ void ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64])
             encode_block(s, block[11], 11);
         }
     } else {
+        is_chroma_420 = (s->chroma_format == CHROMA_420);
+        realloc_huffman(s, 5 + (is_chroma_420 ? 1 : 3));
         for(i=0;i<5;i++) {
             encode_block(s, block[i], i);
         }
-        if (s->chroma_format == CHROMA_420) {
+        if (is_chroma_420) {
             encode_block(s, block[5], 5);
         } else {
             encode_block(s, block[6], 6);
@@ -207,8 +315,11 @@ void ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64])
             encode_block(s, block[7], 7);
         }
     }
+    if (s->mjpeg_ctx->error)
+        return s->mjpeg_ctx->error;
 
-    s->i_tex_bits += get_bits_diff(s);
+    s->i_tex_bits = MJPEG_HUFFMAN_EST_BITS_PER_CODE * s->mjpeg_ctx->huff_ncode;
+    return 0;
 }
 
 // maximum over s->mjpeg_vsample[i]
@@ -261,7 +372,9 @@ FF_MPV_COMMON_OPTS
     { "left",   NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 1 }, INT_MIN, INT_MAX, VE, "pred" },
     { "plane",  NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 2 }, INT_MIN, INT_MAX, VE, "pred" },
     { "median", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = 3 }, INT_MIN, INT_MAX, VE, "pred" },
-
+{ "huffman", "Huffman table strategy", OFFSET(huffman), AV_OPT_TYPE_INT, { .i64 = HUFFMAN_TABLE_DEFAULT }, 0, NB_HUFFMAN_TABLE_OPTION - 1, VE, "huffman" },
+    { "default", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = HUFFMAN_TABLE_DEFAULT }, INT_MIN, INT_MAX, VE, "huffman" },
+    { "optimal", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = HUFFMAN_TABLE_OPTIMAL }, INT_MIN, INT_MAX, VE, "huffman" },
 { NULL},
 };
 
diff --git a/libavcodec/mjpegenc.h b/libavcodec/mjpegenc.h
index 60cd56677f..577fa1fec6 100644
--- a/libavcodec/mjpegenc.h
+++ b/libavcodec/mjpegenc.h
@@ -39,18 +39,67 @@
 #include "mpegvideo.h"
 #include "put_bits.h"
 
+/**
+ * Buffer of JPEG frame data.
+ *
+ * Optimal Huffman table generation requires the frame data to be loaded into
+ * a buffer so that the tables can be computed.
+ * There are at most mb_width*mb_height*12*64 of these per frame.
+ */
+typedef struct MJpegHuffmanCode {
+    // 0=DC lum, 1=DC chrom, 2=AC lum, 3=AC chrom
+    uint8_t table_id; ///< The Huffman table id associated with the data.
+    uint8_t code;     ///< The exponent.
+    uint16_t mant;    ///< The mantissa.
+} MJpegHuffmanCode;
+
+/**
+ * Holds JPEG frame data and Huffman table data.
+ */
 typedef struct MJpegContext {
-    uint8_t huff_size_dc_luminance[12]; //FIXME use array [3] instead of lumi / chroma, for easier addressing
-    uint16_t huff_code_dc_luminance[12];
-    uint8_t huff_size_dc_chrominance[12];
-    uint16_t huff_code_dc_chrominance[12];
+    //FIXME use array [3] instead of lumi / chroma, for easier addressing
+    uint8_t huff_size_dc_luminance[12];     ///< DC luminance Huffman table size.
+    uint16_t huff_code_dc_luminance[12];    ///< DC luminance Huffman table codes.
+    uint8_t huff_size_dc_chrominance[12];   ///< DC chrominance Huffman table size.
+    uint16_t huff_code_dc_chrominance[12];  ///< DC chrominance Huffman table codes.
+
+    uint8_t huff_size_ac_luminance[256];    ///< AC luminance Huffman table size.
+    uint16_t huff_code_ac_luminance[256];   ///< AC luminance Huffman table codes.
+    uint8_t huff_size_ac_chrominance[256];  ///< AC chrominance Huffman table size.
+    uint16_t huff_code_ac_chrominance[256]; ///< AC chrominance Huffman table codes.
 
-    uint8_t huff_size_ac_luminance[256];
-    uint16_t huff_code_ac_luminance[256];
-    uint8_t huff_size_ac_chrominance[256];
-    uint16_t huff_code_ac_chrominance[256];
+    /** Storage for AC luminance VLC (in MpegEncContext) */
+    uint8_t uni_ac_vlc_len[64 * 64 * 2];
+    /** Storage for AC chrominance VLC (in MpegEncContext) */
+    uint8_t uni_chroma_ac_vlc_len[64 * 64 * 2];
+
+    // Default DC tables have exactly 12 values
+    uint8_t bits_dc_luminance[17];   ///< DC luminance Huffman bits.
+    uint8_t val_dc_luminance[12];    ///< DC luminance Huffman values.
+    uint8_t bits_dc_chrominance[17]; ///< DC chrominance Huffman bits.
+    uint8_t val_dc_chrominance[12];  ///< DC chrominance Huffman values.
+
+    // 8-bit JPEG has max 256 values
+    uint8_t bits_ac_luminance[17];   ///< AC luminance Huffman bits.
+    uint8_t val_ac_luminance[256];   ///< AC luminance Huffman values.
+    uint8_t bits_ac_chrominance[17]; ///< AC chrominance Huffman bits.
+    uint8_t val_ac_chrominance[256]; ///< AC chrominance Huffman values.
+
+    unsigned int huff_capacity;      ///< Size of the buffer, in entries.
+    size_t huff_ncode;               ///< Number of current entries in the buffer.
+    MJpegHuffmanCode *huff_buffer;   ///< Buffer for Huffman code values.
+    int error;                       ///< Error code.
 } MJpegContext;
 
+/**
+ * Enum for the Huffman encoding strategy.
+ */
+enum HuffmanTableOption {
+    HUFFMAN_TABLE_DEFAULT = 0, ///< Use the default Huffman tables.
+    HUFFMAN_TABLE_OPTIMAL = 1, ///< Compute and use optimal Huffman tables.
+    NB_HUFFMAN_TABLE_OPTION = 2
+};
+
 static inline void put_marker(PutBitContext *p, enum JpegMarker code)
 {
     put_bits(p, 8, 0xff);
@@ -58,7 +107,8 @@ static inline void put_marker(PutBitContext *p, enum JpegMarker code)
 }
 
 int  ff_mjpeg_encode_init(MpegEncContext *s);
+void ff_mjpeg_encode_picture_frame(MpegEncContext *s);
 void ff_mjpeg_encode_close(MpegEncContext *s);
-void ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64]);
+int ff_mjpeg_encode_mb(MpegEncContext *s, int16_t block[12][64]);
 
 #endif /* AVCODEC_MJPEGENC_H */
diff --git a/libavcodec/mjpegenc_common.c b/libavcodec/mjpegenc_common.c
index 7a6fe7468f..8ec3df9090 100644
--- a/libavcodec/mjpegenc_common.c
+++ b/libavcodec/mjpegenc_common.c
@@ -33,8 +33,35 @@
 #include "put_bits.h"
 #include "mjpegenc.h"
 #include "mjpegenc_common.h"
+#include "mjpegenc_huffman.h"
 #include "mjpeg.h"
 
+av_cold void init_uni_ac_vlc(const uint8_t huff_size_ac[256], uint8_t *uni_ac_vlc_len)
+{
+    int i;
+
+    for (i = 0; i < 128; i++) {
+        int level = i - 64;
+        int run;
+        if (!level)
+            continue;
+        for (run = 0; run < 64; run++) {
+            int len, code, nbits;
+            int alevel = FFABS(level);
+
+            len = (run >> 4) * huff_size_ac[0xf0];
+
+            nbits= av_log2_16bit(alevel) + 1;
+            code = ((15&run) << 4) | nbits;
+
+            len += huff_size_ac[code] + nbits;
+
+            uni_ac_vlc_len[UNI_AC_ENC_INDEX(run, i)] = len;
+            // We ignore EOB as its just a constant which does not change generally
+        }
+    }
+}
+
 /* table_class: 0 = DC coef, 1 = AC coefs */
 static int put_huffman_table(PutBitContext *p, int table_class, int table_id,
                              const uint8_t *bits_table, const uint8_t *value_table)
@@ -104,15 +131,30 @@ static void jpeg_table_header(AVCodecContext *avctx, PutBitContext *p,
     ptr = put_bits_ptr(p);
     put_bits(p, 16, 0); /* patched later */
     size = 2;
-    size += put_huffman_table(p, 0, 0, avpriv_mjpeg_bits_dc_luminance,
-                              avpriv_mjpeg_val_dc);
-    size += put_huffman_table(p, 0, 1, avpriv_mjpeg_bits_dc_chrominance,
-                              avpriv_mjpeg_val_dc);
-
-    size += put_huffman_table(p, 1, 0, avpriv_mjpeg_bits_ac_luminance,
-                              avpriv_mjpeg_val_ac_luminance);
-    size += put_huffman_table(p, 1, 1, avpriv_mjpeg_bits_ac_chrominance,
-                              avpriv_mjpeg_val_ac_chrominance);
+
+    // Only MJPEG can have a variable Huffman variable. All other
+    // formats use the default Huffman table.
+    if (s->out_format == FMT_MJPEG && s->huffman == HUFFMAN_TABLE_OPTIMAL) {
+        size += put_huffman_table(p, 0, 0, s->mjpeg_ctx->bits_dc_luminance,
+                                  s->mjpeg_ctx->val_dc_luminance);
+        size += put_huffman_table(p, 0, 1, s->mjpeg_ctx->bits_dc_chrominance,
+                                  s->mjpeg_ctx->val_dc_chrominance);
+
+        size += put_huffman_table(p, 1, 0, s->mjpeg_ctx->bits_ac_luminance,
+                                  s->mjpeg_ctx->val_ac_luminance);
+        size += put_huffman_table(p, 1, 1, s->mjpeg_ctx->bits_ac_chrominance,
+                                  s->mjpeg_ctx->val_ac_chrominance);
+    } else {
+        size += put_huffman_table(p, 0, 0, avpriv_mjpeg_bits_dc_luminance,
+                                  avpriv_mjpeg_val_dc);
+        size += put_huffman_table(p, 0, 1, avpriv_mjpeg_bits_dc_chrominance,
+                                  avpriv_mjpeg_val_dc);
+
+        size += put_huffman_table(p, 1, 0, avpriv_mjpeg_bits_ac_luminance,
+                                  avpriv_mjpeg_val_ac_luminance);
+        size += put_huffman_table(p, 1, 1, avpriv_mjpeg_bits_ac_chrominance,
+                                  avpriv_mjpeg_val_ac_chrominance);
+    }
     AV_WB16(ptr, size);
 }
 
@@ -372,14 +414,116 @@ void ff_mjpeg_escape_FF(PutBitContext *pb, int start)
     }
 }
 
+/**
+ * Builds all 4 optimal Huffman tables.
+ *
+ * Uses the data stored in the JPEG buffer to compute the tables.
+ * Stores the Huffman tables in the bits_* and val_* arrays in the MJpegContext.
+ *
+ * @param m MJpegContext containing the JPEG buffer.
+ */
+static void ff_mjpeg_build_optimal_huffman(MJpegContext *m)
+{
+    int i, ret, table_id, code;
+
+    MJpegEncHuffmanContext dc_luminance_ctx;
+    MJpegEncHuffmanContext dc_chrominance_ctx;
+    MJpegEncHuffmanContext ac_luminance_ctx;
+    MJpegEncHuffmanContext ac_chrominance_ctx;
+    MJpegEncHuffmanContext *ctx[4] = {&dc_luminance_ctx,
+                                      &dc_chrominance_ctx,
+                                      &ac_luminance_ctx,
+                                      &ac_chrominance_ctx};
+    for (i = 0; i < 4; i++) {
+        ff_mjpeg_encode_huffman_init(ctx[i]);
+    }
+    for (i = 0; i < m->huff_ncode; i++) {
+        table_id = m->huff_buffer[i].table_id;
+        code = m->huff_buffer[i].code;
+
+        ff_mjpeg_encode_huffman_increment(ctx[table_id], code);
+    }
+
+    ret = ff_mjpeg_encode_huffman_close(&dc_luminance_ctx,
+                                        m->bits_dc_luminance,
+                                        m->val_dc_luminance, 12);
+    av_assert0(!ret);
+    ret = ff_mjpeg_encode_huffman_close(&dc_chrominance_ctx,
+                                        m->bits_dc_chrominance,
+                                        m->val_dc_chrominance, 12);
+    av_assert0(!ret);
+    ret = ff_mjpeg_encode_huffman_close(&ac_luminance_ctx,
+                                        m->bits_ac_luminance,
+                                        m->val_ac_luminance, 256);
+    av_assert0(!ret);
+    ret = ff_mjpeg_encode_huffman_close(&ac_chrominance_ctx,
+                                        m->bits_ac_chrominance,
+                                        m->val_ac_chrominance, 256);
+    av_assert0(!ret);
+
+    ff_mjpeg_build_huffman_codes(m->huff_size_dc_luminance,
+                                 m->huff_code_dc_luminance,
+                                 m->bits_dc_luminance,
+                                 m->val_dc_luminance);
+    ff_mjpeg_build_huffman_codes(m->huff_size_dc_chrominance,
+                                 m->huff_code_dc_chrominance,
+                                 m->bits_dc_chrominance,
+                                 m->val_dc_chrominance);
+    ff_mjpeg_build_huffman_codes(m->huff_size_ac_luminance,
+                                 m->huff_code_ac_luminance,
+                                 m->bits_ac_luminance,
+                                 m->val_ac_luminance);
+    ff_mjpeg_build_huffman_codes(m->huff_size_ac_chrominance,
+                                 m->huff_code_ac_chrominance,
+                                 m->bits_ac_chrominance,
+                                 m->val_ac_chrominance);
+}
+
+/**
+ * Writes the complete JPEG frame.
+ *
+ * Header + values + stuffing.
+ *
+ * @param s The MpegEncContext.
+ * @return int Error code, 0 if successful.
+ */
 int ff_mjpeg_encode_stuffing(MpegEncContext *s)
 {
     int i;
     PutBitContext *pbc = &s->pb;
     int mb_y = s->mb_y - !s->mb_x;
+    int ret;
+    MJpegContext *m;
+
+    m = s->mjpeg_ctx;
+
+    if (m->error)
+        return m->error;
+
+    if (s->huffman == HUFFMAN_TABLE_OPTIMAL) {
+        ff_mjpeg_build_optimal_huffman(m);
+
+        // Replace the VLCs with the optimal ones.
+        // The default ones may be used for trellis during quantization.
+        init_uni_ac_vlc(m->huff_size_ac_luminance,   m->uni_ac_vlc_len);
+        init_uni_ac_vlc(m->huff_size_ac_chrominance, m->uni_chroma_ac_vlc_len);
+        s->intra_ac_vlc_length      =
+        s->intra_ac_vlc_last_length = m->uni_ac_vlc_len;
+        s->intra_chroma_ac_vlc_length      =
+        s->intra_chroma_ac_vlc_last_length = m->uni_chroma_ac_vlc_len;
+    }
+
+    ff_mjpeg_encode_picture_header(s->avctx, &s->pb, &s->intra_scantable,
+                                   s->pred, s->intra_matrix, s->chroma_intra_matrix);
+    ff_mjpeg_encode_picture_frame(s);
+    if (m->error < 0) {
+        ret = m->error;
+        return ret;
+    }
+
+    ret = ff_mpv_reallocate_putbitbuffer(s, put_bits_count(&s->pb) / 8 + 100,
+                                            put_bits_count(&s->pb) / 4 + 1000);
 
-    int ret = ff_mpv_reallocate_putbitbuffer(s, put_bits_count(&s->pb) / 8 + 100,
-                                                put_bits_count(&s->pb) / 4 + 1000);
     if (ret < 0) {
         av_log(s->avctx, AV_LOG_ERROR, "Buffer reallocation failed\n");
         goto fail;
diff --git a/libavcodec/mjpegenc_common.h b/libavcodec/mjpegenc_common.h
index 6e51ca04f8..7d760f8a32 100644
--- a/libavcodec/mjpegenc_common.h
+++ b/libavcodec/mjpegenc_common.h
@@ -40,4 +40,6 @@ void ff_mjpeg_init_hvsample(AVCodecContext *avctx, int hsample[4], int vsample[4
 void ff_mjpeg_encode_dc(PutBitContext *pb, int val,
                         uint8_t *huff_size, uint16_t *huff_code);
 
+av_cold void init_uni_ac_vlc(const uint8_t huff_size_ac[256], uint8_t *uni_ac_vlc_len);
+
 #endif /* AVCODEC_MJPEGENC_COMMON_H */
diff --git a/libavcodec/mjpegenc_huffman.c b/libavcodec/mjpegenc_huffman.c
new file mode 100644
index 0000000000..a5d7d5365b
--- /dev/null
+++ b/libavcodec/mjpegenc_huffman.c
@@ -0,0 +1,195 @@
+/*
+ * MJPEG encoder
+ * Copyright (c) 2016 William Ma, Ted Ying, Jerry Jiang
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <string.h>
+#include <stdint.h>
+#include <assert.h>
+#include <stdlib.h>
+#include "libavutil/common.h"
+#include "libavutil/error.h"
+#include "libavutil/qsort.h"
+#include "mjpegenc_huffman.h"
+
+/**
+ * Comparison function for two PTables by prob
+ *
+ * @param a First PTable to compare
+ * @param b Second PTable to compare
+ * @return < 0 for less than, 0 for equals, > 0 for greater than
+ */
+static int compare_by_prob(const void *a, const void *b)
+{
+    PTable a_val = *(PTable *) a;
+    PTable b_val = *(PTable *) b;
+    return a_val.prob - b_val.prob;
+}
+
+/**
+ * Comparison function for two HuffTables by length
+ *
+ * @param a First HuffTable to compare
+ * @param b Second HuffTable to compare
+ * @return < 0 for less than, 0 for equals, > 0 for greater than
+ */
+static int compare_by_length(const void *a, const void *b)
+{
+    HuffTable a_val = *(HuffTable *) a;
+    HuffTable b_val = *(HuffTable *) b;
+    return a_val.length - b_val.length;
+}
+
+/**
+ * Computes the length of the Huffman encoding for each distinct input value.
+ * Uses package merge algorithm as follows:
+ * 1. start with an empty list, lets call it list(0), set i = 0
+ * 2. add 1 entry to list(i) for each symbol we have and give each a score equal to the probability of the respective symbol
+ * 3. merge the 2 symbols of least score and put them in list(i+1), and remove them from list(i). The new score will be the sum of the 2 scores
+ * 4. if there is more than 1 symbol left in the current list(i), then goto 3
+ * 5. i++
+ * 6. if i < 16 goto 2
+ * 7. select the n-1 elements in the last list with the lowest score (n = the number of symbols)
+ * 8. the length of the huffman code for symbol s will be equal to the number of times the symbol occurs in the select elements
+ * Go to guru.multimedia.cx/small-tasks-for-ffmpeg/ for more details
+ *
+ * All probabilities should be positive integers. The output is sorted by code,
+ * not by length.
+ *
+ * @param prob_table input array of a PTable for each distinct input value
+ * @param distincts  output array of a HuffTable that will be populated by this function
+ * @param size       size of the prob_table array
+ * @param max_length max length of an encoding
+ */
+void ff_mjpegenc_huffman_compute_bits(PTable *prob_table, HuffTable *distincts, int size, int max_length)
+{
+    PackageMergerList list_a, list_b, *to = &list_a, *from = &list_b, *temp;
+
+    int times, i, j, k;
+
+    int nbits[257] = {0};
+
+    int min;
+
+    to->nitems = 0;
+    from->nitems = 0;
+    to->item_idx[0] = 0;
+    from->item_idx[0] = 0;
+    AV_QSORT(prob_table, size, PTable, compare_by_prob);
+
+    for (times = 0; times <= max_length; times++) {
+        to->nitems = 0;
+        to->item_idx[0] = 0;
+
+        j = 0;
+        k = 0;
+
+        if (times < max_length) {
+            i = 0;
+        }
+        while (i < size || j + 1 < from->nitems) {
+            to->nitems++;
+            to->item_idx[to->nitems] = to->item_idx[to->nitems - 1];
+            if (i < size &&
+                (j + 1 >= from->nitems ||
+                 prob_table[i].prob <
+                     from->probability[j] + from->probability[j + 1])) {
+                to->items[to->item_idx[to->nitems]++] = prob_table[i].value;
+                to->probability[to->nitems - 1] = prob_table[i].prob;
+                i++;
+            } else {
+                for (k = from->item_idx[j]; k < from->item_idx[j + 2]; k++) {
+                    to->items[to->item_idx[to->nitems]++] = from->items[k];
+                }
+                to->probability[to->nitems - 1] =
+                    from->probability[j] + from->probability[j + 1];
+                j += 2;
+            }
+        }
+        temp = to;
+        to = from;
+        from = temp;
+    }
+
+    min = (size - 1 < from->nitems) ? size - 1 : from->nitems;
+    for (i = 0; i < from->item_idx[min]; i++) {
+        nbits[from->items[i]]++;
+    }
+    // we don't want to return the 256 bit count (it was just in here to prevent
+    // all 1s encoding)
+    j = 0;
+    for (i = 0; i < 256; i++) {
+        if (nbits[i] > 0) {
+            distincts[j].code = i;
+            distincts[j].length = nbits[i];
+            j++;
+        }
+    }
+}
+
+void ff_mjpeg_encode_huffman_init(MJpegEncHuffmanContext *s)
+{
+    memset(s->val_count, 0, sizeof(s->val_count));
+}
+
+/**
+ * Produces a Huffman encoding with a given input
+ *
+ * @param s         input to encode
+ * @param bits      output array where the ith character represents how many input values have i length encoding
+ * @param val       output array of input values sorted by their encoded length
+ * @param max_nval  maximum number of distinct input values
+ * @return int      Return code, 0 if succeeded.
+ */
+int ff_mjpeg_encode_huffman_close(MJpegEncHuffmanContext *s, uint8_t bits[17],
+                                  uint8_t val[], int max_nval)
+{
+    int i, j;
+    int nval = 0;
+    PTable val_counts[257];
+    HuffTable distincts[256];
+
+    for (i = 0; i < 256; i++) {
+        if (s->val_count[i]) nval++;
+    }
+    if (nval > max_nval) {
+        return AVERROR(EINVAL);
+    }
+
+    j = 0;
+    for (i = 0; i < 256; i++) {
+        if (s->val_count[i]) {
+            val_counts[j].value = i;
+            val_counts[j].prob = s->val_count[i];
+            j++;
+        }
+    }
+    val_counts[j].value = 256;
+    val_counts[j].prob = 0;
+    ff_mjpegenc_huffman_compute_bits(val_counts, distincts, nval + 1, 16);
+    AV_QSORT(distincts, nval, HuffTable, compare_by_length);
+
+    memset(bits, 0, sizeof(bits[0]) * 17);
+    for (i = 0; i < nval; i++) {
+        val[i] = distincts[i].code;
+        bits[distincts[i].length]++;
+    }
+
+    return 0;
+}
diff --git a/libavcodec/mjpegenc_huffman.h b/libavcodec/mjpegenc_huffman.h
new file mode 100644
index 0000000000..ec6251b519
--- /dev/null
+++ b/libavcodec/mjpegenc_huffman.h
@@ -0,0 +1,74 @@
+/*
+ * MJPEG encoder
+ * Copyright (c) 2016 William Ma, Ted Ying, Jerry Jiang
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Huffman table generation for MJPEG encoder.
+ */
+
+#ifndef AVCODEC_MJPEGENC_HUFFMAN_H
+#define AVCODEC_MJPEGENC_HUFFMAN_H
+
+typedef struct MJpegEncHuffmanContext {
+    int val_count[256];
+} MJpegEncHuffmanContext;
+
+// Uses the package merge algorithm to compute the Huffman table.
+void ff_mjpeg_encode_huffman_init(MJpegEncHuffmanContext *s);
+static inline void ff_mjpeg_encode_huffman_increment(MJpegEncHuffmanContext *s,
+                                                     uint8_t val)
+{
+    s->val_count[val]++;
+}
+int ff_mjpeg_encode_huffman_close(MJpegEncHuffmanContext *s,
+                                  uint8_t bits[17], uint8_t val[],
+                                  int max_nval);
+
+
+/**
+ * Used to assign a occurrence count or "probability" to an input value
+ */
+typedef struct PTable {
+    int value;  ///< input value
+    int prob;   ///< number of occurences of this value in input
+} PTable;
+
+/**
+ * Used to store intermediate lists in the package merge algorithm
+ */
+typedef struct PackageMergerList {
+    int nitems;             ///< number of items in the list and probability      ex. 4
+    int item_idx[515];      ///< index range for each item in items                   0, 2, 5, 9, 13
+    int probability[514];   ///< probability of each item                             3, 8, 18, 46
+    int items[257 * 16];    ///< chain of all individual values that make up items    A, B, A, B, C, A, B, C, D, C, D, D, E
+} PackageMergerList;
+
+/**
+ * Used to store optimal huffman encoding results
+ */
+typedef struct HuffTable {
+    int code;       ///< code is the input value
+    int length;     ///< length of the encoding
+} HuffTable;
+
+void ff_mjpegenc_huffman_compute_bits(PTable *prob_table, HuffTable *distincts,
+                                      int size, int max_length);
+#endif /* AVCODEC_MJPEGENC_HUFFMAN_H */
diff --git a/libavcodec/mpegvideo.h b/libavcodec/mpegvideo.h
index c82fa3e1a1..85df19154f 100644
--- a/libavcodec/mpegvideo.h
+++ b/libavcodec/mpegvideo.h
@@ -422,6 +422,7 @@ typedef struct MpegEncContext {
     struct MJpegContext *mjpeg_ctx;
     int esc_pos;
     int pred;
+    int huffman;
 
     /* MSMPEG4 specific */
     int mv_table_index;
diff --git a/libavcodec/mpegvideo_enc.c b/libavcodec/mpegvideo_enc.c
index cdda73b654..b1d8dae4b2 100644
--- a/libavcodec/mpegvideo_enc.c
+++ b/libavcodec/mpegvideo_enc.c
@@ -266,7 +266,8 @@ static void mpv_encode_defaults(MpegEncContext *s)
     s->picture_in_gop_number = 0;
 }
 
-av_cold int ff_dct_encode_init(MpegEncContext *s) {
+av_cold int ff_dct_encode_init(MpegEncContext *s)
+{
     if (ARCH_X86)
         ff_dct_encode_init_x86(s);
 
@@ -642,6 +643,15 @@ FF_ENABLE_DEPRECATION_WARNINGS
         return -1;
     }
 
+    if ((s->mpv_flags & FF_MPV_FLAG_QP_RD) &&
+            (s->codec_id == AV_CODEC_ID_AMV ||
+             s->codec_id == AV_CODEC_ID_MJPEG)) {
+        // Used to produce garbage with MJPEG.
+        av_log(avctx, AV_LOG_ERROR,
+               "QP RD is no longer compatible with MJPEG or AMV\n");
+        return -1;
+    }
+
 #if FF_API_PRIVATE_OPT
 FF_DISABLE_DEPRECATION_WARNINGS
     if (avctx->scenechange_threshold)
@@ -3932,9 +3942,8 @@ static int encode_picture(MpegEncContext *s, int picture_number)
     s->last_bits= put_bits_count(&s->pb);
     switch(s->out_format) {
     case FMT_MJPEG:
-        if (CONFIG_MJPEG_ENCODER)
-            ff_mjpeg_encode_picture_header(s->avctx, &s->pb, &s->intra_scantable,
-                                           s->pred, s->intra_matrix, s->chroma_intra_matrix);
+        /* The MJPEG headers are printed after the initial encoding so that the
+         * optimal huffman encoding can be found. */
         break;
     case FMT_H261:
         if (CONFIG_H261_ENCODER)
diff --git a/libavcodec/tests/.gitignore b/libavcodec/tests/.gitignore
index 0ab029696d..f22ac821cb 100644
--- a/libavcodec/tests/.gitignore
+++ b/libavcodec/tests/.gitignore
@@ -10,6 +10,7 @@
 /imgconvert
 /jpeg2000dwt
 /mathops
+/mjpegenc_huffman
 /motion
 /options
 /rangecoder
diff --git a/libavcodec/tests/mjpegenc_huffman.c b/libavcodec/tests/mjpegenc_huffman.c
new file mode 100644
index 0000000000..4559cbe765
--- /dev/null
+++ b/libavcodec/tests/mjpegenc_huffman.c
@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) 2016 William Ma, Sofia Kim, Dustin Woo
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Optimal Huffman Encoding tests.
+ */
+
+#include "libavcodec/avcodec.h"
+#include <stdlib.h>
+#include "libavcodec/mjpegenc.h"
+#include "libavcodec/mjpegenc_huffman.h"
+#include "libavcodec/mjpegenc_common.h"
+#include "libavcodec/mpegvideo.h"
+
+// Validate the computed lengths satisfy the JPEG restrictions and is optimal.
+static int check_lengths(int L, int expected_length,
+                         const int *probs, int nprobs)
+{
+    HuffTable lengths[256];
+    PTable val_counts[256];
+    int actual_length = 0, i, j, k, prob, length;
+    int ret = 0;
+    double cantor_measure = 0;
+    assert(nprobs <= 256);
+
+    for (i = 0; i < nprobs; i++) {
+        val_counts[i] = (PTable){.value = i, .prob = probs[i]};
+    }
+
+    ff_mjpegenc_huffman_compute_bits(val_counts, lengths, nprobs, L);
+
+    for (i = 0; i < nprobs; i++) {
+        // Find the value's prob and length
+        for (j = 0; j < nprobs; j++)
+            if (val_counts[j].value == i) break;
+        for (k = 0; k < nprobs; k++)
+            if (lengths[k].code == i) break;
+        if (!(j < nprobs && k < nprobs)) return 1;
+        prob = val_counts[j].prob;
+        length = lengths[k].length;
+
+        if (prob) {
+            actual_length += prob * length;
+            cantor_measure += 1. / (1 << length);
+        }
+
+        if (length > L || length < 1) return 1;
+    }
+    // Check that the codes can be prefix-free.
+    if (cantor_measure > 1) ret = 1;
+    // Check that the total length is optimal
+    if (actual_length != expected_length) ret = 1;
+
+    if (ret == 1) {
+      fprintf(stderr,
+              "Cantor measure: %f\n"
+              "Actual length: %d\n"
+              "Expected length: %d\n",
+              cantor_measure, actual_length, expected_length);
+    }
+
+    return ret;
+}
+
+static const int probs_zeroes[] = {6, 6, 0, 0, 0};
+static const int probs_skewed[] = {2, 0, 0, 0, 0, 1, 0, 0, 20, 0, 2,
+    0, 10, 5, 1, 1, 9, 1, 1, 6, 0, 5, 0, 1, 0, 7, 6, 1, 1, 5, 0, 0, 0, 0,
+    11, 0, 0, 0, 51, 1, 0, 20, 0, 1, 0, 0, 0, 0, 6, 106, 1, 0, 1, 0, 2, 1,
+    16, 0, 0, 5, 0, 0, 0, 4, 3, 15, 4, 4, 0, 0, 0, 3, 0, 0, 1, 0, 3, 0, 3,
+    2, 2, 0, 0, 4, 3, 40, 1, 2, 0, 22, 0, 0, 0, 9, 0, 0, 0, 0, 1, 1, 0, 1,
+    6, 11, 4, 10, 28, 6, 1, 0, 0, 9, 9, 4, 0, 0, 0, 0, 8, 33844, 2, 0, 2,
+    1, 1, 5, 0, 0, 1, 9, 1, 0, 4, 14, 4, 0, 0, 3, 8, 0, 51, 9, 6, 1, 1, 2,
+    2, 3, 1, 5, 5, 29, 0, 0, 0, 0, 14, 29, 6, 4, 13, 12, 2, 3, 1, 0, 5, 4,
+    1, 1, 0, 0, 29, 1, 0, 0, 0, 0, 4, 0, 0, 1, 0, 1, 7, 0, 42, 0, 0, 0, 0,
+    0, 2, 0, 3, 9, 0, 0, 0, 2, 1, 0, 0, 6, 5, 6, 1, 2, 3, 0, 0, 0, 3, 0, 0,
+    28, 0, 2, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 23, 0, 0, 0, 0,
+    0, 21, 1, 0, 3, 24, 2, 0, 0, 7, 0, 0, 1, 5, 1, 2, 0, 5};
+static const int probs_sat[] = {74, 8, 14, 7, 9345, 40, 0, 2014, 2, 1,
+    115, 0, 2, 1, 194, 388, 20, 0, 0, 2, 1, 121, 1, 1583, 0, 16, 21, 2, 132,
+    2, 15, 9, 13, 1, 0, 2293, 2, 8, 5, 2, 30, 0, 0, 4, 54, 783, 4, 1, 2, 4,
+    0, 22, 93, 1, 143, 19, 0, 36, 32, 4, 6, 33, 3, 45, 0, 8, 1, 0, 18, 17, 1,
+    0, 1, 0, 0, 1, 1004, 38, 3, 8, 90, 23, 0, 2819, 3, 0, 970, 158, 9, 6, 4,
+    48, 4, 0, 1, 0, 0, 60, 3, 62, 0, 2, 2, 2, 279, 66, 16, 1, 20, 0, 7, 9,
+    32, 1411, 6, 3, 27, 1, 5, 49, 0, 0, 0, 0, 0, 2, 10, 1, 1, 2, 3, 801, 3,
+    25, 5, 1, 1, 0, 632, 0, 14, 18, 5, 8, 200, 4, 4, 22, 12, 0, 4, 1, 0, 2,
+    4, 9, 3, 16, 7, 2, 2, 213, 0, 2, 620, 39303, 0, 1, 0, 2, 1, 183781, 1,
+    0, 0, 0, 94, 7, 3, 4, 0, 4, 306, 43, 352, 76, 34, 13, 11, 0, 51, 1, 13,
+    19, 0, 26, 0, 7276, 4, 207, 31, 1, 2, 4, 6, 19, 8, 17, 4, 6, 0, 1085, 0,
+    0, 0, 3, 489, 36, 1, 0, 1, 9420, 294, 28, 0, 57, 5, 0, 9, 2, 0, 1, 2, 2,
+    0, 0, 9, 2, 29, 2, 2, 7, 0, 5, 490, 0, 7, 5, 0, 1, 8, 0, 0, 23255, 0, 1};
+
+// Test the example given on @see
+// http://guru.multimedia.cx/small-tasks-for-ffmpeg/
+int main(int argc, char **argv)
+{
+    int i, ret = 0;
+    // Probabilities of symbols 0..4
+    static PTable val_counts[] = {
+        {.value = 0, .prob = 1},
+        {.value = 1, .prob = 2},
+        {.value = 2, .prob = 5},
+        {.value = 3, .prob = 10},
+        {.value = 4, .prob = 21},
+    };
+    // Expected code lengths for each symbol
+    static const HuffTable expected[] = {
+        {.code = 0, .length = 3},
+        {.code = 1, .length = 3},
+        {.code = 2, .length = 3},
+        {.code = 3, .length = 3},
+        {.code = 4, .length = 1},
+    };
+    // Actual code lengths
+    HuffTable distincts[5];
+
+    // Build optimal huffman tree using an internal function, to allow for
+    // smaller-than-normal test cases. This mutates val_counts by sorting.
+    ff_mjpegenc_huffman_compute_bits(val_counts, distincts,
+                                     FF_ARRAY_ELEMS(distincts), 3);
+
+    for (i = 0; i < FF_ARRAY_ELEMS(distincts); i++) {
+        if (distincts[i].code != expected[i].code ||
+            distincts[i].length != expected[i].length) {
+            fprintf(stderr,
+                    "Built huffman does not equal expectations. "
+                    "Expected: code %d probability %d, "
+                    "Actual: code %d probability %d\n",
+                    expected[i].code, expected[i].length,
+                    distincts[i].code, distincts[i].length);
+            ret = 1;
+        }
+    }
+
+    // Check handling of zero probabilities
+    if (check_lengths(16, 18, probs_zeroes, FF_ARRAY_ELEMS(probs_zeroes)))
+        ret = 1;
+    // Check skewed distribution over 256 without saturated lengths
+    if (check_lengths(16, 41282, probs_skewed, FF_ARRAY_ELEMS(probs_skewed)))
+        ret = 1;
+    // Check skewed distribution over 256 with saturated lengths
+    if (check_lengths(16, 669904, probs_sat, FF_ARRAY_ELEMS(probs_sat)))
+        ret = 1;
+
+    return ret;
+}