tiny_jpeg.h

/**
 * tiny_jpeg.h
 *
 * Tiny JPEG Encoder
 *  - Sergio Gonzalez
 *
 * This is a readable and simple single-header JPEG encoder.
 *
 * Features
 *  - Implements Baseline DCT JPEG compression.
 *  - No dynamic allocations.
 *
 * This library is coded in the spirit of the stb libraries and mostly follows
 * the stb guidelines.
 *
 * It is written in C99. And depends on the C standard library.
 * Works with C++11
 *
 *
 * ==== Thanks ====
 *
 *  AssociationSirius (Bug reports)
 *  Bernard van Gastel (Thread-safe defaults, BSD compilation)
 *
 *
 * ==== License ====
 *
 * This software is in the public domain. Where that dedication is not
 * recognized, you are granted a perpetual, irrevocable license to copy and
 * modify this file as you see fit.
 *
 */

// ============================================================
// Usage
// ============================================================
// Include "tiny_jpeg.h" to and use the public interface defined below.
//
// You *must* do:
//
//      #define TJE_IMPLEMENTATION
//      #include "tiny_jpeg.h"
//
// in exactly one of your C files to actually compile the implementation.


// Here is an example program that loads a bmp with stb_image and writes it
// with Tiny JPEG

/*

#define STB_IMAGE_IMPLEMENTATION
#include "stb_image.h"


#define TJE_IMPLEMENTATION
#include "tiny_jpeg.h"


int main()
{
    int width, height, num_components;
    unsigned char* data = stbi_load("in.bmp", &width, &height, &num_components, 0);
    if ( !data ) {
        puts("Could not find file");
        return EXIT_FAILURE;
    }

    if ( !tje_encode_to_file("out.jpg", width, height, num_components, data) ) {
        fprintf(stderr, "Could not write JPEG\n");
        return EXIT_FAILURE;
    }

    return EXIT_SUCCESS;
}

*/



#ifdef __cplusplus
extern "C"
{
#endif

#if defined(__GNUC__) || defined(__clang__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wmissing-field-initializers"  // We use {0}, which will zero-out the struct.
#pragma GCC diagnostic ignored "-Wmissing-braces"
#pragma GCC diagnostic ignored "-Wpadded"
#endif

// ============================================================
// Public interface:
// ============================================================

#ifndef TJE_HEADER_GUARD
#define TJE_HEADER_GUARD

// - tje_encode_to_file -
//
// Usage:
//  Takes bitmap data and writes a JPEG-encoded image to disk.
//
//  PARAMETERS
//      dest_path:          filename to which we will write. e.g. "out.jpg"
//      width, height:      image size in pixels
//      num_components:     3 is RGB. 4 is RGBA. Those are the only supported values
//      src_data:           pointer to the pixel data.
//
//  RETURN:
//      0 on error. 1 on success.

int tje_encode_to_file(const char* dest_path,
                       const int width,
                       const int height,
                       const int num_components,
                       const unsigned char* src_data);

// - tje_encode_to_file_at_quality -
//
// Usage:
//  Takes bitmap data and writes a JPEG-encoded image to disk.
//
//  PARAMETERS
//      dest_path:          filename to which we will write. e.g. "out.jpg"
//      quality:            3: Highest. Compression varies wildly (between 1/3 and 1/20).
//                          2: Very good quality. About 1/2 the size of 3.
//                          1: Noticeable. About 1/6 the size of 3, or 1/3 the size of 2.
//      width, height:      image size in pixels
//      num_components:     3 is RGB. 4 is RGBA. Those are the only supported values
//      src_data:           pointer to the pixel data.
//
//  RETURN:
//      0 on error. 1 on success.

int tje_encode_to_file_at_quality(const char* dest_path,
                                  const int quality,
                                  const int width,
                                  const int height,
                                  const int num_components,
                                  const unsigned char* src_data);

// - tje_encode_with_func -
//
// Usage
//  Same as tje_encode_to_file_at_quality, but it takes a callback that knows
//  how to handle (or ignore) `context`. The callback receives an array `data`
//  of `size` bytes, which can be written directly to a file. There is no need
//  to free the data.

typedef void tje_write_func(void* context, void* data, int size);

int tje_encode_with_func(tje_write_func* func,
                         void* context,
                         const int quality,
                         const int width,
                         const int height,
                         const int num_components,
                         const unsigned char* src_data);

#endif // TJE_HEADER_GUARD


#ifdef __cplusplus
}  // extern C
#endif


// Implementation: In exactly one of the source files of your application,
// define TJE_IMPLEMENTATION and include tiny_jpeg.h

// ============================================================
// Internal
// ============================================================
#ifdef TJE_IMPLEMENTATION


#define tjei_min(a, b) ((a) < b) ? (a) : (b);
#define tjei_max(a, b) ((a) < b) ? (b) : (a);


#if defined(_MSC_VER)
#define TJEI_FORCE_INLINE __forceinline
// #define TJEI_FORCE_INLINE __declspec(noinline)  // For profiling
#else
#define TJEI_FORCE_INLINE static // TODO: equivalent for gcc & clang
#endif

// Only use zero for debugging and/or inspection.
#define TJE_USE_FAST_DCT 1

// C std lib
#include <assert.h>
#include <inttypes.h>
#include <math.h>   // floorf, ceilf
#ifdef OFFTARGET
#include <stdio.h>  // FILE, puts
#endif
#include <string.h> // memcpy


#define TJEI_BUFFER_SIZE 1024

#ifdef _WIN32

#include <windows.h>
#ifndef snprintf
#define snprintf sprintf_s
#endif
// Not quite the same but it works for us. If I am not mistaken, it differs
// only in the return value.

#endif

#ifndef NDEBUG

#ifdef _WIN32
#define tje_log(msg) OutputDebugStringA(msg)
#elif defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)
#define tje_log(msg) qsee_log(QSEE_LOG_MSG_ERROR, msg)
#else
#warning "need a tje_log definition for your platform for debugging purposes (not needed if compiling with NDEBUG)"
#endif

#else  // NDEBUG
#define tje_log(msg)
#endif  // NDEBUG


typedef struct
{
    void*           context;
    tje_write_func* func;
} TJEWriteContext;

typedef struct
{
    // Huffman data.
    uint8_t         ehuffsize[4][257];
    uint16_t        ehuffcode[4][256];
    uint8_t const * ht_bits[4];
    uint8_t const * ht_vals[4];

    // Cuantization tables.
    uint8_t         qt_luma[64];
    uint8_t         qt_chroma[64];

    // fwrite by default. User-defined when using tje_encode_with_func.
    TJEWriteContext write_context;

    // Buffered output. Big performance win when using the usual stdlib implementations.
    size_t          output_buffer_count;
    uint8_t         output_buffer[TJEI_BUFFER_SIZE];
} TJEState;

// ============================================================
// Table definitions.
//
// The spec defines tjei_default reasonably good quantization matrices and huffman
// specification tables.
//
//
// Instead of hard-coding the final huffman table, we only hard-code the table
// spec suggested by the specification, and then derive the full table from
// there.  This is only for didactic purposes but it might be useful if there
// ever is the case that we need to swap huffman tables from various sources.
// ============================================================


// K.1 - suggested luminance QT
static const uint8_t tjei_default_qt_luma_from_spec[] =
{
   16,11,10,16, 24, 40, 51, 61,
   12,12,14,19, 26, 58, 60, 55,
   14,13,16,24, 40, 57, 69, 56,
   14,17,22,29, 51, 87, 80, 62,
   18,22,37,56, 68,109,103, 77,
   24,35,55,64, 81,104,113, 92,
   49,64,78,87,103,121,120,101,
   72,92,95,98,112,100,103, 99,
};

// Unused
#if 0
static const uint8_t tjei_default_qt_chroma_from_spec[] =
{
    // K.1 - suggested chrominance QT
   17,18,24,47,99,99,99,99,
   18,21,26,66,99,99,99,99,
   24,26,56,99,99,99,99,99,
   47,66,99,99,99,99,99,99,
   99,99,99,99,99,99,99,99,
   99,99,99,99,99,99,99,99,
   99,99,99,99,99,99,99,99,
   99,99,99,99,99,99,99,99,
};
#endif

static const uint8_t tjei_default_qt_chroma_from_paper[] =
{
    // Example QT from JPEG paper
    16,  12, 14,  14, 18, 24,  49,  72,
    11,  10, 16,  24, 40, 51,  61,  12,
    13,  17, 22,  35, 64, 92,  14,  16,
    22,  37, 55,  78, 95, 19,  24,  29,
    56,  64, 87,  98, 26, 40,  51,  68,
    81, 103, 112, 58, 57, 87,  109, 104,
    121,100, 60,  69, 80, 103, 113, 120,
    103, 55, 56,  62, 77, 92,  101, 99,
};

// == Procedure to 'deflate' the huffman tree: JPEG spec, C.2

// Number of 16 bit values for every code length. (K.3.3.1)
static const uint8_t tjei_default_ht_luma_dc_len[16] =
{
    0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0
};
// values
static const uint8_t tjei_default_ht_luma_dc[12] =
{
    0,1,2,3,4,5,6,7,8,9,10,11
};

// Number of 16 bit values for every code length. (K.3.3.1)
static const uint8_t tjei_default_ht_chroma_dc_len[16] =
{
    0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0
};
// values
static const uint8_t tjei_default_ht_chroma_dc[12] =
{
    0,1,2,3,4,5,6,7,8,9,10,11
};

// Same as above, but AC coefficients.
static const uint8_t tjei_default_ht_luma_ac_len[16] =
{
    0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d
};
static const uint8_t tjei_default_ht_luma_ac[] =
{
    0x01, 0x02, 0x03, 0x00, 0x04, 0x11, 0x05, 0x12, 0x21, 0x31, 0x41, 0x06, 0x13, 0x51, 0x61, 0x07,
    0x22, 0x71, 0x14, 0x32, 0x81, 0x91, 0xA1, 0x08, 0x23, 0x42, 0xB1, 0xC1, 0x15, 0x52, 0xD1, 0xF0,
    0x24, 0x33, 0x62, 0x72, 0x82, 0x09, 0x0A, 0x16, 0x17, 0x18, 0x19, 0x1A, 0x25, 0x26, 0x27, 0x28,
    0x29, 0x2A, 0x34, 0x35, 0x36, 0x37, 0x38, 0x39, 0x3A, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x49,
    0x4A, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5A, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69,
    0x6A, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79, 0x7A, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89,
    0x8A, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98, 0x99, 0x9A, 0xA2, 0xA3, 0xA4, 0xA5, 0xA6, 0xA7,
    0xA8, 0xA9, 0xAA, 0xB2, 0xB3, 0xB4, 0xB5, 0xB6, 0xB7, 0xB8, 0xB9, 0xBA, 0xC2, 0xC3, 0xC4, 0xC5,
    0xC6, 0xC7, 0xC8, 0xC9, 0xCA, 0xD2, 0xD3, 0xD4, 0xD5, 0xD6, 0xD7, 0xD8, 0xD9, 0xDA, 0xE1, 0xE2,
    0xE3, 0xE4, 0xE5, 0xE6, 0xE7, 0xE8, 0xE9, 0xEA, 0xF1, 0xF2, 0xF3, 0xF4, 0xF5, 0xF6, 0xF7, 0xF8,
    0xF9, 0xFA
};

static const uint8_t tjei_default_ht_chroma_ac_len[16] =
{
    0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77
};
static const uint8_t tjei_default_ht_chroma_ac[] =
{
    0x00, 0x01, 0x02, 0x03, 0x11, 0x04, 0x05, 0x21, 0x31, 0x06, 0x12, 0x41, 0x51, 0x07, 0x61, 0x71,
    0x13, 0x22, 0x32, 0x81, 0x08, 0x14, 0x42, 0x91, 0xA1, 0xB1, 0xC1, 0x09, 0x23, 0x33, 0x52, 0xF0,
    0x15, 0x62, 0x72, 0xD1, 0x0A, 0x16, 0x24, 0x34, 0xE1, 0x25, 0xF1, 0x17, 0x18, 0x19, 0x1A, 0x26,
    0x27, 0x28, 0x29, 0x2A, 0x35, 0x36, 0x37, 0x38, 0x39, 0x3A, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48,
    0x49, 0x4A, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5A, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68,
    0x69, 0x6A, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79, 0x7A, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
    0x88, 0x89, 0x8A, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x98, 0x99, 0x9A, 0xA2, 0xA3, 0xA4, 0xA5,
    0xA6, 0xA7, 0xA8, 0xA9, 0xAA, 0xB2, 0xB3, 0xB4, 0xB5, 0xB6, 0xB7, 0xB8, 0xB9, 0xBA, 0xC2, 0xC3,
    0xC4, 0xC5, 0xC6, 0xC7, 0xC8, 0xC9, 0xCA, 0xD2, 0xD3, 0xD4, 0xD5, 0xD6, 0xD7, 0xD8, 0xD9, 0xDA,
    0xE2, 0xE3, 0xE4, 0xE5, 0xE6, 0xE7, 0xE8, 0xE9, 0xEA, 0xF2, 0xF3, 0xF4, 0xF5, 0xF6, 0xF7, 0xF8,
    0xF9, 0xFA
};


// ============================================================
// Code
// ============================================================

// Zig-zag order:
static const uint8_t tjei_zig_zag[64] =
{
    0,   1,  5,  6, 14, 15, 27, 28,
    2,   4,  7, 13, 16, 26, 29, 42,
    3,   8, 12, 17, 25, 30, 41, 43,
    9,  11, 18, 24, 31, 40, 44, 53,
    10, 19, 23, 32, 39, 45, 52, 54,
    20, 22, 33, 38, 46, 51, 55, 60,
    21, 34, 37, 47, 50, 56, 59, 61,
    35, 36, 48, 49, 57, 58, 62, 63,
};

// Memory order as big endian. 0xhilo -> 0xlohi which looks as 0xhilo in memory.
static uint16_t tjei_be_word(const uint16_t le_word)
{
    uint16_t lo = (le_word & 0x00ff);
    uint16_t hi = ((le_word & 0xff00) >> 8);
    return (uint16_t)((lo << 8) | hi);
}

// ============================================================
// The following structs exist only for code clarity, debugability, and
// readability. They are used when writing to disk, but it is useful to have
// 1-packed-structs to document how the format works, and to inspect memory
// while developing.
// ============================================================

static const uint8_t tjeik_jfif_id[] = "JFIF";
static const uint8_t tjeik_com_str[] = "Created by Tiny JPEG Encoder";

// TODO: Get rid of packed structs!
#pragma pack(push)
#pragma pack(1)
typedef struct
{
    uint16_t SOI;
    // JFIF header.
    uint16_t APP0;
    uint16_t jfif_len;
    uint8_t  jfif_id[5];
    uint16_t version;
    uint8_t  units;
    uint16_t x_density;
    uint16_t y_density;
    uint8_t  x_thumb;
    uint8_t  y_thumb;
} TJEJPEGHeader;

typedef struct
{
    uint16_t com;
    uint16_t com_len;
    char     com_str[sizeof(tjeik_com_str) - 1];
} TJEJPEGComment;

// Helper struct for TJEFrameHeader (below).
typedef struct
{
    uint8_t  component_id;
    uint8_t  sampling_factors;    // most significant 4 bits: horizontal. 4 LSB: vertical (A.1.1)
    uint8_t  qt;                  // Quantization table selector.
} TJEComponentSpec;

typedef struct
{
    uint16_t         SOF;
    uint16_t         len;                   // 8 + 3 * frame.num_components
    uint8_t          precision;             // Sample precision (bits per sample).
    uint16_t         height;
    uint16_t         width;
    uint8_t          num_components;        // For this implementation, will be equal to 3.
    TJEComponentSpec component_spec[3];
} TJEFrameHeader;

typedef struct
{
    uint8_t component_id;                 // Just as with TJEComponentSpec
    uint8_t dc_ac;                        // (dc|ac)
} TJEFrameComponentSpec;

typedef struct
{
    uint16_t              SOS;
    uint16_t              len;
    uint8_t               num_components;  // 3.
    TJEFrameComponentSpec component_spec[3];
    uint8_t               first;  // 0
    uint8_t               last;  // 63
    uint8_t               ah_al;  // o
} TJEScanHeader;
#pragma pack(pop)


static void tjei_write(TJEState* state, const void* data, size_t num_bytes, size_t num_elements)
{
    size_t to_write = num_bytes * num_elements;

    // Cap to the buffer available size and copy memory.
    size_t capped_count = tjei_min(to_write, TJEI_BUFFER_SIZE - 1 - state->output_buffer_count);

    memcpy(state->output_buffer + state->output_buffer_count, data, capped_count);
    state->output_buffer_count += capped_count;

    assert (state->output_buffer_count <= TJEI_BUFFER_SIZE - 1);

    // Flush the buffer.
    if ( state->output_buffer_count == TJEI_BUFFER_SIZE - 1 ) {
        state->write_context.func(state->write_context.context, state->output_buffer, (int)state->output_buffer_count);
        state->output_buffer_count = 0;
    }

    // Recursively calling ourselves with the rest of the buffer.
    if (capped_count < to_write) {
        tjei_write(state, (uint8_t*)data+capped_count, to_write - capped_count, 1);
    }
}

static void tjei_write_DQT(TJEState* state, const uint8_t* matrix, uint8_t id)
{
    uint16_t DQT = tjei_be_word(0xffdb);
    tjei_write(state, &DQT, sizeof(uint16_t), 1);
    uint16_t len = tjei_be_word(0x0043); // 2(len) + 1(id) + 64(matrix) = 67 = 0x43
    tjei_write(state, &len, sizeof(uint16_t), 1);
    assert(id < 4);
    uint8_t precision_and_id = id;  // 0x0000 8 bits | 0x00id
    tjei_write(state, &precision_and_id, sizeof(uint8_t), 1);
    // Write matrix
    tjei_write(state, matrix, 64*sizeof(uint8_t), 1);
}

typedef enum
{
    TJEI_DC = 0,
    TJEI_AC = 1
} TJEHuffmanTableClass;

static void tjei_write_DHT(TJEState* state,
                           uint8_t const * matrix_len,
                           uint8_t const * matrix_val,
                           TJEHuffmanTableClass ht_class,
                           uint8_t id)
{
    int num_values = 0;
    for ( int i = 0; i < 16; ++i ) {
        num_values += matrix_len[i];
    }
    assert(num_values <= 0xffff);

    uint16_t DHT = tjei_be_word(0xffc4);
    // 2(len) + 1(Tc|th) + 16 (num lengths) + ?? (num values)
    uint16_t len = tjei_be_word(2 + 1 + 16 + (uint16_t)num_values);
    assert(id < 4);
    uint8_t tc_th = (uint8_t)((((uint8_t)ht_class) << 4) | id);

    tjei_write(state, &DHT, sizeof(uint16_t), 1);
    tjei_write(state, &len, sizeof(uint16_t), 1);
    tjei_write(state, &tc_th, sizeof(uint8_t), 1);
    tjei_write(state, matrix_len, sizeof(uint8_t), 16);
    tjei_write(state, matrix_val, sizeof(uint8_t), (size_t)num_values);
}
// ============================================================
//  Huffman deflation code.
// ============================================================

// Returns all code sizes from the BITS specification (JPEG C.3)
static uint8_t* tjei_huff_get_code_lengths(uint8_t huffsize[/*256*/], uint8_t const * bits)
{
    int k = 0;
    for ( int i = 0; i < 16; ++i ) {
        for ( int j = 0; j < bits[i]; ++j ) {
            huffsize[k++] = (uint8_t)(i + 1);
        }
        huffsize[k] = 0;
    }
    return huffsize;
}

// Fills out the prefixes for each code.
static uint16_t* tjei_huff_get_codes(uint16_t codes[], uint8_t* huffsize, int64_t count)
{
    uint16_t code = 0;
    int k = 0;
    uint8_t sz = huffsize[0];
    for(;;) {
        do {
            assert(k < count);
            codes[k++] = code++;
        } while (huffsize[k] == sz);
        if (huffsize[k] == 0) {
            return codes;
        }
        do {
            code = (uint16_t)(code << 1);
            ++sz;
        } while( huffsize[k] != sz );
    }
}

static void tjei_huff_get_extended(uint8_t* out_ehuffsize,
                                   uint16_t* out_ehuffcode,
                                   uint8_t const * huffval,
                                   uint8_t* huffsize,
                                   uint16_t* huffcode, int64_t count)
{
    int k = 0;
    do {
        uint8_t val = huffval[k];
        out_ehuffcode[val] = huffcode[k];
        out_ehuffsize[val] = huffsize[k];
        k++;
    } while ( k < count );
}
// ============================================================

// Returns:
//  out[1] : number of bits
//  out[0] : bits
TJEI_FORCE_INLINE void tjei_calculate_variable_length_int(int value, uint16_t out[2])
{
    int abs_val = value;
    if ( value < 0 ) {
        abs_val = -abs_val;
        --value;
    }
    out[1] = 1;
    while( abs_val >>= 1 ) {
        ++out[1];
    }
    out[0] = (uint16_t)(value & ((1 << out[1]) - 1));
}

// Write bits to file.
TJEI_FORCE_INLINE void tjei_write_bits(TJEState* state,
                                       uint32_t* bitbuffer, uint32_t* location,
                                       uint16_t num_bits, uint16_t bits)
{
    //   v-- location
    //  [                     ]   <-- bit buffer
    // 32                     0
    //
    // This call pushes to the bitbuffer and saves the location. Data is pushed
    // from most significant to less significant.
    // When we can write a full byte, we write a byte and shift.

    // Push the stack.
    uint32_t nloc = *location + num_bits;
    *bitbuffer |= (uint32_t)(bits << (32 - nloc));
    *location = nloc;
    while ( *location >= 8 ) {
        // Grab the most significant byte.
        uint8_t c = (uint8_t)((*bitbuffer) >> 24);
        // Write it to file.
        tjei_write(state, &c, 1, 1);
        if ( c == 0xff )  {
            // Special case: tell JPEG this is not a marker.
            char z = 0;
            tjei_write(state, &z, 1, 1);
        }
        // Pop the stack.
        *bitbuffer <<= 8;
        *location -= 8;
    }
}

// DCT implementation by Thomas G. Lane.
// Obtained through NVIDIA
//  http://developer.download.nvidia.com/SDK/9.5/Samples/vidimaging_samples.html#gpgpu_dct
//
// QUOTE:
//  This implementation is based on Arai, Agui, and Nakajima's algorithm for
//  scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
//  Japanese, but the algorithm is described in the Pennebaker & Mitchell
//  JPEG textbook (see REFERENCES section in file README).  The following code
//  is based directly on figure 4-8 in P&M.
//
static void tjei_fdct (float * data)
{
    float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    float tmp10, tmp11, tmp12, tmp13;
    float z1, z2, z3, z4, z5, z11, z13;
    float *dataptr;
    int ctr;

    /* Pass 1: process rows. */

    dataptr = data;
    for ( ctr = 7; ctr >= 0; ctr-- ) {
        tmp0 = dataptr[0] + dataptr[7];
        tmp7 = dataptr[0] - dataptr[7];
        tmp1 = dataptr[1] + dataptr[6];
        tmp6 = dataptr[1] - dataptr[6];
        tmp2 = dataptr[2] + dataptr[5];
        tmp5 = dataptr[2] - dataptr[5];
        tmp3 = dataptr[3] + dataptr[4];
        tmp4 = dataptr[3] - dataptr[4];

        /* Even part */

        tmp10 = tmp0 + tmp3;    /* phase 2 */
        tmp13 = tmp0 - tmp3;
        tmp11 = tmp1 + tmp2;
        tmp12 = tmp1 - tmp2;

        dataptr[0] = tmp10 + tmp11; /* phase 3 */
        dataptr[4] = tmp10 - tmp11;

        z1 = (tmp12 + tmp13) * ((float) 0.707106781); /* c4 */
        dataptr[2] = tmp13 + z1;    /* phase 5 */
        dataptr[6] = tmp13 - z1;

        /* Odd part */

        tmp10 = tmp4 + tmp5;    /* phase 2 */
        tmp11 = tmp5 + tmp6;
        tmp12 = tmp6 + tmp7;

        /* The rotator is modified from fig 4-8 to avoid extra negations. */
        z5 = (tmp10 - tmp12) * ((float) 0.382683433); /* c6 */
        z2 = ((float) 0.541196100) * tmp10 + z5; /* c2-c6 */
        z4 = ((float) 1.306562965) * tmp12 + z5; /* c2+c6 */
        z3 = tmp11 * ((float) 0.707106781); /* c4 */

        z11 = tmp7 + z3;        /* phase 5 */
        z13 = tmp7 - z3;

        dataptr[5] = z13 + z2;  /* phase 6 */
        dataptr[3] = z13 - z2;
        dataptr[1] = z11 + z4;
        dataptr[7] = z11 - z4;

        dataptr += 8;     /* advance pointer to next row */
    }

    /* Pass 2: process columns. */

    dataptr = data;
    for ( ctr = 8-1; ctr >= 0; ctr-- ) {
        tmp0 = dataptr[8*0] + dataptr[8*7];
        tmp7 = dataptr[8*0] - dataptr[8*7];
        tmp1 = dataptr[8*1] + dataptr[8*6];
        tmp6 = dataptr[8*1] - dataptr[8*6];
        tmp2 = dataptr[8*2] + dataptr[8*5];
        tmp5 = dataptr[8*2] - dataptr[8*5];
        tmp3 = dataptr[8*3] + dataptr[8*4];
        tmp4 = dataptr[8*3] - dataptr[8*4];

        /* Even part */

        tmp10 = tmp0 + tmp3;    /* phase 2 */
        tmp13 = tmp0 - tmp3;
        tmp11 = tmp1 + tmp2;
        tmp12 = tmp1 - tmp2;

        dataptr[8*0] = tmp10 + tmp11; /* phase 3 */
        dataptr[8*4] = tmp10 - tmp11;

        z1 = (tmp12 + tmp13) * ((float) 0.707106781); /* c4 */
        dataptr[8*2] = tmp13 + z1; /* phase 5 */
        dataptr[8*6] = tmp13 - z1;

        /* Odd part */

        tmp10 = tmp4 + tmp5;    /* phase 2 */
        tmp11 = tmp5 + tmp6;
        tmp12 = tmp6 + tmp7;

        /* The rotator is modified from fig 4-8 to avoid extra negations. */
        z5 = (tmp10 - tmp12) * ((float) 0.382683433); /* c6 */
        z2 = ((float) 0.541196100) * tmp10 + z5; /* c2-c6 */
        z4 = ((float) 1.306562965) * tmp12 + z5; /* c2+c6 */
        z3 = tmp11 * ((float) 0.707106781); /* c4 */

        z11 = tmp7 + z3;        /* phase 5 */
        z13 = tmp7 - z3;

        dataptr[8*5] = z13 + z2; /* phase 6 */
        dataptr[8*3] = z13 - z2;
        dataptr[8*1] = z11 + z4;
        dataptr[8*7] = z11 - z4;

        dataptr++;          /* advance pointer to next column */
    }
}
#if !TJE_USE_FAST_DCT
static float slow_fdct(int u, int v, float* data)
{
#define kPI 3.14159265f
    float res = 0.0f;
    float cu = (u == 0) ? 0.70710678118654f : 1;
    float cv = (v == 0) ? 0.70710678118654f : 1;
    for ( int y = 0; y < 8; ++y ) {
        for ( int x = 0; x < 8; ++x ) {
            res += (data[y * 8 + x]) *
                    cosf(((2.0f * x + 1.0f) * u * kPI) / 16.0f) *
                    cosf(((2.0f * y + 1.0f) * v * kPI) / 16.0f);
        }
    }
    res *= 0.25f * cu * cv;
    return res;
#undef kPI
}
#endif

#define ABS(x) ((x) < 0 ? -(x) : (x))

static void tjei_encode_and_write_MCU(TJEState* state,
                                      float* mcu,
#if TJE_USE_FAST_DCT
                                      float* qt,  // Pre-processed quantization matrix.
#else
                                      uint8_t* qt,
#endif
                                      uint8_t* huff_dc_len, uint16_t* huff_dc_code, // Huffman tables
                                      uint8_t* huff_ac_len, uint16_t* huff_ac_code,
                                      int* pred,  // Previous DC coefficient
                                      uint32_t* bitbuffer,  // Bitstack.
                                      uint32_t* location)
{
    int du[64];  // Data unit in zig-zag order

    float dct_mcu[64];
    memcpy(dct_mcu, mcu, 64 * sizeof(float));

#if TJE_USE_FAST_DCT
    tjei_fdct(dct_mcu);
    for ( int i = 0; i < 64; ++i ) {
        float fval = dct_mcu[i];
        fval *= qt[i];
#if 0
        fval = (fval > 0) ? floorf(fval + 0.5f) : ceilf(fval - 0.5f);
#else
        fval = floorf(fval + 1024 + 0.5f);
        fval -= 1024;
#endif
        int val = (int)fval;
        du[tjei_zig_zag[i]] = val;
    }
#else
    for ( int v = 0; v < 8; ++v ) {
        for ( int u = 0; u < 8; ++u ) {
            dct_mcu[v * 8 + u] = slow_fdct(u, v, mcu);
        }
    }
    for ( int i = 0; i < 64; ++i ) {
        float fval = dct_mcu[i] / (qt[i]);
        int val = (int)((fval > 0) ? floorf(fval + 0.5f) : ceilf(fval - 0.5f));
        du[tjei_zig_zag[i]] = val;
    }
#endif

    uint16_t vli[2];

    // Encode DC coefficient.
    int diff = du[0] - *pred;
    *pred = du[0];
    if ( diff != 0 ) {
        tjei_calculate_variable_length_int(diff, vli);
        // Write number of bits with Huffman coding
        tjei_write_bits(state, bitbuffer, location, huff_dc_len[vli[1]], huff_dc_code[vli[1]]);
        // Write the bits.
        tjei_write_bits(state, bitbuffer, location, vli[1], vli[0]);
    } else {
        tjei_write_bits(state, bitbuffer, location, huff_dc_len[0], huff_dc_code[0]);
    }

    // ==== Encode AC coefficients ====

    int last_non_zero_i = 0;
    // Find the last non-zero element.
    for ( int i = 63; i > 0; --i ) {
        if (du[i] != 0) {
            last_non_zero_i = i;
            break;
        }
    }

    for ( int i = 1; i <= last_non_zero_i; ++i ) {
        // If zero, increase count. If >=15, encode (FF,00)
        int zero_count = 0;
        while ( du[i] == 0 ) {
            ++zero_count;
            ++i;
            if (zero_count == 16) {
                // encode (ff,00) == 0xf0
                tjei_write_bits(state, bitbuffer, location, huff_ac_len[0xf0], huff_ac_code[0xf0]);
                zero_count = 0;
            }
        }
        tjei_calculate_variable_length_int(du[i], vli);

        assert(zero_count < 0x10);
        assert(vli[1] <= 10);

        uint16_t sym1 = (uint16_t)((uint16_t)zero_count << 4) | vli[1];

        assert(huff_ac_len[sym1] != 0);

        // Write symbol 1  --- (RUNLENGTH, SIZE)
        tjei_write_bits(state, bitbuffer, location, huff_ac_len[sym1], huff_ac_code[sym1]);
        // Write symbol 2  --- (AMPLITUDE)
        tjei_write_bits(state, bitbuffer, location, vli[1], vli[0]);
    }

    if (last_non_zero_i != 63) {
        // write EOB HUFF(00,00)
        tjei_write_bits(state, bitbuffer, location, huff_ac_len[0], huff_ac_code[0]);
    }
    return;
}

enum {
    TJEI_LUMA_DC,
    TJEI_LUMA_AC,
    TJEI_CHROMA_DC,
    TJEI_CHROMA_AC,
};

#if TJE_USE_FAST_DCT
struct TJEProcessedQT
{
    float chroma[64];
    float luma[64];
};
#endif

// Set up huffman tables in state.
static void tjei_huff_expand(TJEState* state)
{
    assert(state);

    state->ht_bits[TJEI_LUMA_DC]   = tjei_default_ht_luma_dc_len;
    state->ht_bits[TJEI_LUMA_AC]   = tjei_default_ht_luma_ac_len;
    state->ht_bits[TJEI_CHROMA_DC] = tjei_default_ht_chroma_dc_len;
    state->ht_bits[TJEI_CHROMA_AC] = tjei_default_ht_chroma_ac_len;

    state->ht_vals[TJEI_LUMA_DC]   = tjei_default_ht_luma_dc;
    state->ht_vals[TJEI_LUMA_AC]   = tjei_default_ht_luma_ac;
    state->ht_vals[TJEI_CHROMA_DC] = tjei_default_ht_chroma_dc;
    state->ht_vals[TJEI_CHROMA_AC] = tjei_default_ht_chroma_ac;

    // How many codes in total for each of LUMA_(DC|AC) and CHROMA_(DC|AC)
    int32_t spec_tables_len[4] = { 0 };

    for ( int i = 0; i < 4; ++i ) {
        for ( int k = 0; k < 16; ++k ) {
            spec_tables_len[i] += state->ht_bits[i][k];
        }
    }

    // Fill out the extended tables..
    uint8_t huffsize[4][257];
    uint16_t huffcode[4][256];
    for ( int i = 0; i < 4; ++i ) {
        assert (256 >= spec_tables_len[i]);
        tjei_huff_get_code_lengths(huffsize[i], state->ht_bits[i]);
        tjei_huff_get_codes(huffcode[i], huffsize[i], spec_tables_len[i]);
    }
    for ( int i = 0; i < 4; ++i ) {
        int64_t count = spec_tables_len[i];
        tjei_huff_get_extended(state->ehuffsize[i],
                               state->ehuffcode[i],
                               state->ht_vals[i],
                               &huffsize[i][0],
                               &huffcode[i][0], count);
    }
}

#ifdef TJE_USE_ARMADILLO
#include <armadillo>
#endif
static int tjei_encode_main(TJEState* state,
                            const unsigned char* src_data,
                            const int width,
                            const int height,
                            const int src_num_components)
{
    if (src_num_components != 3 && src_num_components != 4) {
        return 0;
    }

    if (width > 0xffff || height > 0xffff) {
        return 0;
    }

#if TJE_USE_FAST_DCT
    struct TJEProcessedQT pqt;
    // Again, taken from classic japanese implementation.
    //
    /* For float AA&N IDCT method, divisors are equal to quantization
     * coefficients scaled by scalefactor[row]*scalefactor[col], where
     *   scalefactor[0] = 1
     *   scalefactor[k] = cos(k*PI/16) * sqrt(2)    for k=1..7
     * We apply a further scale factor of 8.
     * What's actually stored is 1/divisor so that the inner loop can
     * use a multiplication rather than a division.
     */
    static const float aan_scales[] = {
        1.0f, 1.387039845f, 1.306562965f, 1.175875602f,
        1.0f, 0.785694958f, 0.541196100f, 0.275899379f
    };

    // build (de)quantization tables
    for(int y=0; y<8; y++) {
        for(int x=0; x<8; x++) {
            int i = y*8 + x;
            pqt.luma[y*8+x] = 1.0f / (8 * aan_scales[x] * aan_scales[y] * state->qt_luma[tjei_zig_zag[i]]);
            pqt.chroma[y*8+x] = 1.0f / (8 * aan_scales[x] * aan_scales[y] * state->qt_chroma[tjei_zig_zag[i]]);
        }
    }
#endif

    { // Write header
        TJEJPEGHeader header;
        // JFIF header.
        header.SOI = tjei_be_word(0xffd8);  // Sequential DCT
        header.APP0 = tjei_be_word(0xffe0);

        uint16_t jfif_len = sizeof(TJEJPEGHeader) - 4 /*SOI & APP0 markers*/;
        header.jfif_len = tjei_be_word(jfif_len);
        memcpy(header.jfif_id, (void*)tjeik_jfif_id, 5);
        header.version = tjei_be_word(0x0102);
        header.units = 0x01;  // Dots-per-inch
        header.x_density = tjei_be_word(0x0060);  // 96 DPI
        header.y_density = tjei_be_word(0x0060);  // 96 DPI
        header.x_thumb = 0;
        header.y_thumb = 0;
        tjei_write(state, &header, sizeof(TJEJPEGHeader), 1);
    }
    {  // Write comment
        TJEJPEGComment com;
        uint16_t com_len = 2 + sizeof(tjeik_com_str) - 1;
        // Comment
        com.com = tjei_be_word(0xfffe);
        com.com_len = tjei_be_word(com_len);
        memcpy(com.com_str, (void*)tjeik_com_str, sizeof(tjeik_com_str)-1);
        tjei_write(state, &com, sizeof(TJEJPEGComment), 1);
    }

    // Write quantization tables.
    tjei_write_DQT(state, state->qt_luma, 0x00);
    tjei_write_DQT(state, state->qt_chroma, 0x01);

    {  // Write the frame marker.
        TJEFrameHeader header;
        header.SOF = tjei_be_word(0xffc0);
        header.len = tjei_be_word(8 + 3 * 3);
        header.precision = 8;
        assert(width <= 0xffff);
        assert(height <= 0xffff);
        header.width = tjei_be_word((uint16_t)width);
        header.height = tjei_be_word((uint16_t)height);
        header.num_components = 3;
        uint8_t tables[3] = {
            0,  // Luma component gets luma table (see tjei_write_DQT call above.)
            1,  // Chroma component gets chroma table
            1,  // Chroma component gets chroma table
        };
        for (int i = 0; i < 3; ++i) {
            TJEComponentSpec spec;
            spec.component_id = (uint8_t)(i + 1);  // No particular reason. Just 1, 2, 3.
            spec.sampling_factors = (uint8_t)0x11;
            spec.qt = tables[i];

            header.component_spec[i] = spec;
        }
        // Write to file.
        tjei_write(state, &header, sizeof(TJEFrameHeader), 1);
    }

    tjei_write_DHT(state, state->ht_bits[TJEI_LUMA_DC],   state->ht_vals[TJEI_LUMA_DC], TJEI_DC, 0);
    tjei_write_DHT(state, state->ht_bits[TJEI_LUMA_AC],   state->ht_vals[TJEI_LUMA_AC], TJEI_AC, 0);
    tjei_write_DHT(state, state->ht_bits[TJEI_CHROMA_DC], state->ht_vals[TJEI_CHROMA_DC], TJEI_DC, 1);
    tjei_write_DHT(state, state->ht_bits[TJEI_CHROMA_AC], state->ht_vals[TJEI_CHROMA_AC], TJEI_AC, 1);

    // Write start of scan
    {
        TJEScanHeader header;
        header.SOS = tjei_be_word(0xffda);
        header.len = tjei_be_word((uint16_t)(6 + (sizeof(TJEFrameComponentSpec) * 3)));
        header.num_components = 3;

        uint8_t tables[3] = {
            0x00,
            0x11,
            0x11,
        };
        for (int i = 0; i < 3; ++i) {
            TJEFrameComponentSpec cs;
            // Must be equal to component_id from frame header above.
            cs.component_id = (uint8_t)(i + 1);
            cs.dc_ac = (uint8_t)tables[i];

            header.component_spec[i] = cs;
        }
        header.first = 0;
        header.last  = 63;
        header.ah_al = 0;
        tjei_write(state, &header, sizeof(TJEScanHeader), 1);

    }
    // Write compressed data.

#ifdef TJE_USE_ARMADILLO
    arma::fmat::fixed<64,3> du;
    float* du_y = du.memptr();
    float* du_b = du_y + 64;
    float* du_r = du_b + 64;
#else
    float du_y[64];
    float du_b[64];
    float du_r[64];
#endif

    // Set diff to 0.
    int pred_y = 0;
    int pred_b = 0;
    int pred_r = 0;

    // Bit stack
    uint32_t bitbuffer = 0;
    uint32_t location = 0;

#ifdef USE_RGB_888
    const arma::Mat<uint8_t> pixel ((uint8_t*) src_data, src_num_components, width*height, false, true);
#elif !defined(TJE_USE_DIRECT_YUV_DATA)
    // Convert NV12 format to RGB
    arma::Mat<uint8_t> image_rgb(3,width*height);

    const arma::fmat::fixed<3,2> VU2RGB = {
        {1.370705f, 0.f},
        {-0.698001f, -0.337633f},
        {0.f, 1.732446f}
    };

    const arma::fmat::fixed<3,3> RGB2YUV = {
        {0.299f   ,0.587f    ,0.114f },
        {-0.1687f ,-0.3313f   ,0.5f   },
        {0.5f     ,-0.4187f   ,-0.0813f}
        };
    const arma::fcolvec::fixed<3> offset = {-128.f, 0.f, 0.f};
#endif

typedef uint16_t uchar2;
#ifndef OFFTARGET
// Android YUV420SP has padding size of 32 rows
const size_t padding = width * 32;
#else
// OpenCV.cvtColor assumes zero padding size
const size_t padding = 0;
#endif

#ifdef TJE_USE_ARMADILLO
    // Map from camera_buffer
    const arma::Mat<uint8_t> Y((uint8_t*) src_data, width, height, false, true);
    const arma::Mat<uint16_t> UV((uchar2*) (src_data + padding + width*height), width/2, height/2, false, true);
#else
    const uint8_t* Y = src_data;
    const uchar2* UV = (const uchar2*) (src_data + width * height + padding);
#endif

    for ( int y = 0; y < height; y += 8 ) {
        for ( int x = 0; x < width; x += 8 ) {
            // Block loop: ====
            for ( int off_y = 0; off_y < 8; ++off_y ) {
                for ( int off_x = 0; off_x < 8; ++off_x ) {
                    int col = x + off_x;
                    int row = y + off_y;

                    // Use boundary pixels if the coordinate is outside image boundary
                    if(row >= height)
                        row = height - 1;
                    if(col >= width)
                        col = width - 1;

#ifdef USE_RGB_888
                    du.row(off_y*8 + off_x) = arma::trans (RGB2YUV * pixel.col(row*width + col) + offset);
#elif defined(TJE_USE_DIRECT_YUV_DATA) && defined(TJE_USE_ARMADILLO)
                    du(off_y*8 + off_x, 0) = static_cast<float>(Y(col, row)) - 128.f;
                    du(off_y*8 + off_x, 1) = static_cast<float>(UV(col/2, row/2) & 0xff) - 128.f;
                    du(off_y*8 + off_x, 2) = static_cast<float>(UV(col/2, row/2) >> 8) - 128.f;
#elif defined(TJE_USE_DIRECT_YUV_DATA) && !defined(TJE_USE_ARMADILLO)
                    const size_t UV_index = (row/2) * (width/2) + col/2;
                    du_y[off_y*8 + off_x] = (float) (Y[row*width + col]) - 128.f;
                    du_b[off_y*8 + off_x] = (float) (UV[UV_index] & 0xff) - 128.f;
                    du_r[off_y*8 + off_x] = (float) (UV[UV_index] >> 8) - 128.f;
#else
                    const arma::Col<uint8_t> vu((uint8_t*) &UV(col/2, row/2), 2, false, true);

                    du.row(off_y*8 + off_x) = arma::trans (
                            RGB2YUV * (
                                VU2RGB * (arma::conv_to<arma::fcolvec>::from(vu) - 128.f) + Y(col,row))
                             + offset );
#endif
                }
            }

            tjei_encode_and_write_MCU(state, du_y,
#if TJE_USE_FAST_DCT
                                     pqt.luma,
#else
                                     state->qt_luma,
#endif
                                     state->ehuffsize[TJEI_LUMA_DC], state->ehuffcode[TJEI_LUMA_DC],
                                     state->ehuffsize[TJEI_LUMA_AC], state->ehuffcode[TJEI_LUMA_AC],
                                     &pred_y, &bitbuffer, &location);
            tjei_encode_and_write_MCU(state, du_b,
#if TJE_USE_FAST_DCT
                                     pqt.chroma,
#else
                                     state->qt_chroma,
#endif
                                     state->ehuffsize[TJEI_CHROMA_DC], state->ehuffcode[TJEI_CHROMA_DC],
                                     state->ehuffsize[TJEI_CHROMA_AC], state->ehuffcode[TJEI_CHROMA_AC],
                                     &pred_b, &bitbuffer, &location);
            tjei_encode_and_write_MCU(state, du_r,
#if TJE_USE_FAST_DCT
                                     pqt.chroma,
#else
                                     state->qt_chroma,
#endif
                                     state->ehuffsize[TJEI_CHROMA_DC], state->ehuffcode[TJEI_CHROMA_DC],
                                     state->ehuffsize[TJEI_CHROMA_AC], state->ehuffcode[TJEI_CHROMA_AC],
                                     &pred_r, &bitbuffer, &location);


        }
    }

    // Finish the image.
    { // Flush
        if (location > 0 && location < 8) {
            tjei_write_bits(state, &bitbuffer, &location, (uint16_t)(8 - location), 0);
        }
    }
    uint16_t EOI = tjei_be_word(0xffd9);
    tjei_write(state, &EOI, sizeof(uint16_t), 1);

    if (state->output_buffer_count) {
        state->write_context.func(state->write_context.context, state->output_buffer, (int)state->output_buffer_count);
        state->output_buffer_count = 0;
    }

    return 1;
}

int tje_encode_to_file(const char* dest_path,
                       const int width,
                       const int height,
                       const int num_components,
                       const unsigned char* src_data)
{
    int res = tje_encode_to_file_at_quality(dest_path, 3, width, height, num_components, src_data);
    return res;
}

#ifdef OFFTARGET
static void tjei_stdlib_func(void* context, void* data, int size)
{
    FILE* fd = (FILE*)context;
    fwrite(data, size, 1, fd);
}
#endif

// Define public interface.
int tje_encode_to_file_at_quality(const char* dest_path,
                                  const int quality,
                                  const int width,
                                  const int height,
                                  const int num_components,
                                  const unsigned char* src_data)
{
#ifdef OFFTARGET
    FILE* fd = fopen(dest_path, "wb");
#else
    void* fd = NULL;
#endif
    if (!fd) {
        tje_log("Could not open file for writing.");
        return 0;
    }

#ifdef OFFTARGET
    int result = tje_encode_with_func(tjei_stdlib_func, fd,
                                      quality, width, height, num_components, src_data);

    result |= 0 == fclose(fd);
#else
    int result = 0;
#endif
    return result;
}

int tje_encode_with_func(tje_write_func* func,
                         void* context,
                         const int quality,
                         const int width,
                         const int height,
                         const int num_components,
                         const unsigned char* src_data)
{
    if (quality < 1 || quality > 3) {
        tje_log("[ERROR] -- Valid 'quality' values are 1 (lowest), 2, or 3 (highest)\n");
        return 0;
    }

    TJEState state = { 0 };

    uint8_t qt_factor = 1;
    switch(quality) {
    case 3:
        for ( int i = 0; i < 64; ++i ) {
            state.qt_luma[i]   = 1;
            state.qt_chroma[i] = 1;
        }
        break;
    case 2:
        qt_factor = 10;
        // don't break. fall through.
    case 1:
        for ( int i = 0; i < 64; ++i ) {
            state.qt_luma[i]   = tjei_default_qt_luma_from_spec[i] / qt_factor;
            if (state.qt_luma[i] == 0) {
                state.qt_luma[i] = 1;
            }
            state.qt_chroma[i] = tjei_default_qt_chroma_from_paper[i] / qt_factor;
            if (state.qt_chroma[i] == 0) {
                state.qt_chroma[i] = 1;
            }
        }
        break;
    default:
        assert(!"invalid code path");
        break;
    }

    TJEWriteContext wc = { 0 };

    wc.context = context;
    wc.func = func;

    state.write_context = wc;


    tjei_huff_expand(&state);

    int result = tjei_encode_main(&state, src_data, width, height, num_components);

    return result;
}
// ============================================================
#endif // TJE_IMPLEMENTATION
// ============================================================
//
#if defined(__GNUC__) || defined(__clang__)
#pragma GCC diagnostic pop
#endif