/* nut.c - A simple NUT encoder with the following properties:
 *
 *  - Each frame is an uncompressed 8-bit RGB (specifically `(msb)2B
 *    3G 3R(lsb)`, known as the 4CC "BGR\x08") framebuffer.
 *
 *  - Each pixel in the framebuffer is non-square; it is twice as tall
 *    as it is wide.
 *
 *  - The resolution may change mid-stream, between the values
 *    640x480, 720x480, and 720x576.  This is the display-resolution;
 *    because of non-square pixels, the framebuffer for each would be
 *    640x240, 720x240, and 720x288, respectively.
 *
 *  - VFR (Variable Frame Rate) - Each frame has its own timestamp.
 */

#include <assert.h>  /* for static_assert() */
#include <stdbool.h> /* for bool, true, false */
#include <stdint.h>  /* for uint{n}_t, UINT64_MAX */
#include <stdio.h>   /* for perror() */
#include <stdlib.h>  /* for realloc(), free() */
#include <string.h>  /* for memcpy(), memset() */
#include <unistd.h>  /* for write() */
#include <fcntl.h>   /* for open() */

/* Generic ********************************************************************/

#include "common.h"

/* NUT ************************************************************************/

/* Magic values.  *****************************************/
uint8_t nut_file_id_string[]      = "nut/multimedia container"; /* including NUL terminator */
uint8_t nut_startcode_main[]      = {'N', 'M', 0x7A, 0x56, 0x1F, 0x5F, 0x04, 0xAD };
uint8_t nut_startcode_stream[]    = {'N', 'S', 0x11, 0x40, 0x5B, 0xF2, 0xF9, 0xDB };
uint8_t nut_startcode_syncpoint[] = {'N', 'K', 0xE4, 0xAD, 0xEE, 0xCA, 0x45, 0x69 };
uint8_t nut_startcode_index[]     = {'N', 'X', 0xDD, 0x67, 0x2F, 0x23, 0xE6, 0x4E };
uint8_t nut_startcode_info[]      = {'N', 'I', 0xAB, 0x68, 0xB5, 0x96, 0xBA, 0x78 };

#define NUT_FRAMEFLAG_KEY         (1<< 0)
#define NUT_FRAMEFLAG_EOR         (1<< 1)
#define _NUT_FRAMEFLAG_2          (1<< 2)
#define NUT_FRAMEFLAG_CODED_PTS   (1<< 3)
#define NUT_FRAMEFLAG_STREAM_ID   (1<< 4)
#define NUT_FRAMEFLAG_SIZE_MSB    (1<< 5)
#define NUT_FRAMEFLAG_CHECKSUM    (1<< 6)
#define NUT_FRAMEFLAG_RESERVED    (1<< 7)
#define NUT_FRAMEFLAG_SM_DATA     (1<< 8)
#define _NUT_FRAMEFLAG_9          (1<< 9)
#define NUT_FRAMEFLAG_HEADER_IDX  (1<<10)
#define NUT_FRAMEFLAG_MATCH_TIME  (1<<11)
#define NUT_FRAMEFLAG_CODED       (1<<12)
#define NUT_FRAMEFLAG_INVALID     (1<<13)

#define NUT_MAINFLAG_BROADCAST_MODE (1<< 0)
#define NUT_MAINFLAG_PIPE_MODE      (1<< 1)

#define NUT_UNKNOWN_MATCH_TIME (1-(UINT64_C(1)<<62))

enum nut_stream_class {
	NUT_STREAMCLASS_VIDEO     = 0,
	NUT_STREAMCLASS_AUDIO     = 1,
	NUT_STREAMCLASS_SUBTITLES = 2,
	NUT_STREAMCLASS_USERDATA  = 3,
};

#define NUT_STREAMFLAG_FIXED_FPS (1<< 1)

enum nut_colorspace {
	NUT_COLORSPACE_UNKNOWN         =  0,
	NUT_COLORSPACE_ITU642_TRUNC    =  1,
	NUT_COLORSPACE_ITU709_TRUNC    =  2,
	NUT_COLORSPACE_ITU642_FULL     = 17,
	NUT_COLORSPACE_ITU709_FILL     = 18,
};

/* Basic fixed-length types.  *****************************/

void nut_append_u8(struct buf *buf, uint8_t val) {
	assert(buf);

	append(buf, &val, 1);
}

void nut_append_u32(struct buf *buf, uint32_t val) {
	assert(buf);

	append_u32be(buf, val);
}

/* Basic variable-length types.  **************************/

void nut_append_vu(struct buf *buf, uint64_t val) {
	assert(buf);

	uint8_t dat[10] = {
		0x80|((val>>(7*9))&0x7F),
		0x80|((val>>(7*8))&0x7F),
		0x80|((val>>(7*7))&0x7F),
		0x80|((val>>(7*6))&0x7F),
		0x80|((val>>(7*5))&0x7F),
		0x80|((val>>(7*4))&0x7F),
		0x80|((val>>(7*3))&0x7F),
		0x80|((val>>(7*2))&0x7F),
		0x80|((val>>(7*1))&0x7F),
		0x00|((val>>(7*0))&0x7F),
	};
	uint8_t skip = 0;
	while (dat[skip] == 0x80)
		skip++;
	append(buf, &dat[skip], sizeof(dat)-skip);
}

void nut_append_vs(struct buf *buf, int64_t val) {
	assert(buf);
	assert((val < 0 ? -val : val) <= UINT64_MAX>>1);

	uint64_t temp;
	if (val > 0)
		temp = val << 1;
	else
		temp = ((-val) << 1) | 1;
	temp--;
	nut_append_vu(buf, temp);
}

void nut_append_vb(struct buf *buf, void *dat, size_t len) {
	assert(buf);
	assert(len == 0 || dat);

	nut_append_vu(buf, len);
	append(buf, dat, len);
}

#define nut_append_vb_str(buf, str) nut_append_vb(buf, str, sizeof(str)-1)

/* Mid-level.  ********************************************/

uint32_t nut_crc32(uint8_t *dat, int len){
	assert(len == 0 || dat);

	/* "Generator polynomial is 0x104C11DB7."  */
	static const uint32_t table[16] = {
		0x00000000, 0x04C11DB7, 0x09823B6E, 0x0D4326D9,
		0x130476DC, 0x17C56B6B, 0x1A864DB2, 0x1E475005,
		0x2608EDB8, 0x22C9F00F, 0x2F8AD6D6, 0x2B4BCB61,
		0x350C9B64, 0x31CD86D3, 0x3C8EA00A, 0x384FBDBD,
	};

	/* "Starting value is zero."  */
	uint32_t crc = 0;

	while (len--) {
		crc ^= *dat++ << 24;
		crc = (crc<<4) ^ table[crc>>28];
		crc = (crc<<4) ^ table[crc>>28];
	}
	return crc;
}

void nut_append_packet(struct buf *buf, uint8_t *startcode, void *data, size_t len) {
	assert(buf);
	assert(startcode);
	assert(len == 0 || data);

	size_t packet_beg = buf->len;
	append(buf, startcode, 8);
	nut_append_vu(buf, len+4);
	if (len+4 > 4096)
		nut_append_u32(buf, nut_crc32(&buf->dat[packet_beg], buf->len - packet_beg));
	size_t data_beg = buf->len;
	append(buf, data, len);
	nut_append_u32(buf, nut_crc32(&buf->dat[data_beg], buf->len - data_beg));
}

/* High-level (specific to the properties we want) ****************************/

enum app_res {
	APP_RES_INVALID = -1,
	APP_RES_640_480 = 0,
	APP_RES_720_480 = 1,
	APP_RES_720_576 = 2,
	APP_RES_NUM = 3,
};
#define APP_RES_MIN APP_RES_640_480
#define APP_RES_MAX APP_RES_720_576

size_t app_res_w(enum app_res res) {
	switch (res) {
	case APP_RES_640_480: return 640;
	case APP_RES_720_480: return 720;
	case APP_RES_720_576: return 720;
	default: assert_notreached("invalid res");
	}
}
size_t app_res_h(enum app_res res) {
	switch (res) {
	case APP_RES_640_480: return 480/2;
	case APP_RES_720_480: return 480/2;
	case APP_RES_720_576: return 576/2;
	default: assert_notreached("invalid res");
	}
}
size_t app_res_fb_size(enum app_res res) {
	return app_res_w(res)*app_res_h(res);
}

#define APP_FRAME_SIZE_MSB_MUL 0x2000 /* must be less than 0x4000 */
#define APP_FRAME_MAX_HEADER   16

struct app_state {
	enum app_res res;
	uint64_t len;
	uint64_t last_syncpoint;
};

void app_append_intro(struct buf *out, enum app_res res) {
	struct buf pkt = {0};
	unsigned int frame_code = 0;
#define INC_FRAME_CODE() do { frame_code++; if (frame_code == 'N') frame_code++; } while(0)

	append(out, nut_file_id_string, sizeof(nut_file_id_string));

#define BOGUS(n) n
	/* main_header ********************************************************/
	pkt.len = 0;

	/* head *******************************************/
	nut_append_vu(&pkt, 3);                      /*! version */
	nut_append_vu(&pkt, 1);                      /*! stream_count */
	nut_append_vu(&pkt,                          /*! max_distance */
	              APP_FRAME_MAX_HEADER +
	              app_res_fb_size(res));

	/* time bases *************************************/
	nut_append_vu(&pkt, 1);                      /*! time_base_count */

	/* time_base[0] = 1ns */
	nut_append_vu(&pkt, 1);                      /*! numerator */
	nut_append_vu(&pkt, 1000000000ULL);          /*! denominator */

	/* frame codes ************************************/
	/* "A muxer SHOULD mark [frame codes] 0x00 and 0xFF as invalid" */

	/* frame_code=0 (invalid) */
	nut_append_vu(&pkt, NUT_FRAMEFLAG_INVALID);  /*! flags */
	nut_append_vu(&pkt, 2);                      /*! field_count */
	nut_append_vu(&pkt, BOGUS(0));               /*! 0: fields.pts */
	nut_append_vu(&pkt, 1);                      /*! 1: fields.size_msb_nul */
	INC_FRAME_CODE();

	/* frame_code=1 (res) */
	nut_append_vu(&pkt,                          /*! flags */
	              NUT_FRAMEFLAG_KEY       |  /* Because they're full framebuffers, all frames are keyframes. */
	              NUT_FRAMEFLAG_SIZE_MSB  |  /* 640*480/2 > 16384 (the max val of data_size_lsb). */
	              NUT_FRAMEFLAG_CODED_PTS |  /* framerate is unknown, each frame must have a timestamp. */
	              NUT_FRAMEFLAG_CHECKSUM  ); /* framerate is unknown, guard against exceeding max_pts_distance. */
	nut_append_vu(&pkt, 6);                      /*! field_count */
	nut_append_vu(&pkt, BOGUS(0));               /*! 0: fields.pts */
	nut_append_vu(&pkt, APP_FRAME_SIZE_MSB_MUL); /*! 1: fields.size_msb_nul */
	nut_append_vu(&pkt, 0);                      /*! 2: fields.stream */
	nut_append_vu(&pkt,                          /*! 3: fields.size_lsb */
	              app_res_fb_size(res) % APP_FRAME_SIZE_MSB_MUL);
	nut_append_vu(&pkt, 0);                      /*! 4: fields.reserved */
	nut_append_vu(&pkt, 1);                      /*! 5: fields.count */
	INC_FRAME_CODE();

	/* frame_code=N-255 (invalid) */
	nut_append_vu(&pkt, NUT_FRAMEFLAG_INVALID);  /*! flags */
	nut_append_vu(&pkt, 2);                      /*! field_count */
	nut_append_vu(&pkt, BOGUS(0));               /*! 0: fields.pts */
	nut_append_vu(&pkt,                          /*! 1: fields.size_msb_nul */
	              256-frame_code-(frame_code < 'N' ? 1 : 0));

	/* tail *******************************************/
	nut_append_vu(&pkt, 0);                      /*! header_count_minus_1 */

	nut_append_packet(out, nut_startcode_main, pkt.dat, pkt.len);

	/* stream_header ******************************************************/
	pkt.len = 0;

	nut_append_vu(&pkt, 0);                      /*! stream_id */
	nut_append_vu(&pkt, NUT_STREAMCLASS_VIDEO);  /*! stream_class */
	nut_append_vb(&pkt, "BGR\x08", 4);           /*! fourcc */
	nut_append_vu(&pkt, 0);                      /*! time_base_id */
	nut_append_vu(&pkt, 0);                      /*! msb_pts_shift */
	nut_append_vu(&pkt, BOGUS(0));               /*! max_pts_distance (all frames have a checksum)  */
	nut_append_vu(&pkt, 0);                      /*! decode_delay */
	nut_append_vu(&pkt, 0);                      /*! stream_flags */
	nut_append_vb(&pkt, NULL, 0);                /*! codec_specific_data */
	nut_append_vu(&pkt, app_res_w(res));         /*! width */
	nut_append_vu(&pkt, app_res_h(res));         /*! height */
	nut_append_vu(&pkt, 1);                      /*! sample_width */
	nut_append_vu(&pkt, 2);                      /*! sample_height */
	nut_append_vu(&pkt, NUT_COLORSPACE_UNKNOWN); /*! colorspace_type */

	nut_append_packet(out, nut_startcode_stream, pkt.dat, pkt.len);

	/* flush **************************************************************/
#undef BOGUS

	free(pkt.dat);
}

bool app_write_frame(int fd, struct app_state *state, uint64_t time_ns, enum app_res res, void *framebuffer) {
	assert(framebuffer);
	assert(0 <= res && res < APP_RES_NUM);

	struct buf out = {0};
	struct buf pkt = {0};

	if (res != state->res) {
		app_append_intro(&out, res);
		state->res = res;
	}

	/* syncpoint packet ***************************************************/
	pkt.len = 0;
	uint64_t syncpoint_beg = state->len + out.len;

	nut_append_vu(&pkt, time_ns);             /*! global_key_pts */
	nut_append_vu(&pkt, state->last_syncpoint /*! back_ptr_div16 */
	              ? (syncpoint_beg - state->last_syncpoint)/16
	              : 0);

	nut_append_packet(&out, nut_startcode_syncpoint, pkt.dat, pkt.len);

	/* frame header ( 1+10+1+4 = 16 bytes) */
	uint64_t frame_beg = out.len;
	nut_append_u8(&out, 1);                                                    /*! frame_code    (1 byte) */
	nut_append_vu(&out, time_ns);                                              /*! coded_pts     (<=10 bytes) */
	nut_append_vu(&out, app_res_fb_size(res) / APP_FRAME_SIZE_MSB_MUL);        /*! data_size_msb (1 byte) */
	nut_append_u32(&out, nut_crc32(&out.dat[frame_beg], out.len - frame_beg)); /*! checksum      (4 bytes) */
	assert(out.len - frame_beg <= APP_FRAME_MAX_HEADER);

	/* flush */
	bool err = xwrite(fd, out.dat, out.len);
	free(out.dat);
	free(pkt.dat);
	if (err)
		return true;

	/* frame data */
	if (xwrite(fd, framebuffer, app_res_fb_size(res)))
		return true;

	/* return */
	state->res = res;
	state->len += out.len + app_res_fb_size(res);
	state->last_syncpoint = syncpoint_beg;
	return false;
}

/* Demo application ***********************************************************/

int main() {
	struct app_state state = {
		.res = APP_RES_INVALID,
	};

	uint8_t framebuffer[app_res_fb_size(APP_RES_MAX)];

#define SCALE 10
	for (int i = 0; i < 10; i++) {
		enum app_res res = (i/2)%APP_RES_NUM;
		memset(framebuffer, 0b00011000, sizeof(framebuffer));
		for (int y = 0; y < 4*SCALE; y++)
			for (int x = 0; x < 8*SCALE; x++)
				framebuffer[(y*app_res_w(res))+x] = font[i][((y/SCALE)*8)+(x/SCALE)] == ' ' ? 0b11000000 : 0b00000011;
		if (i%2 == 0) {
			char name[] = {'o', 'u', 't', '-', '0'+(i/2), '.', 'n', 'u', 't', '\0'};
			int fd = open(name, O_WRONLY|O_TRUNC|O_CREAT, 0666);
			dup2(fd, 1);
			close(fd);
			state.len = 0;
			state.last_syncpoint = 0;
		}
		if (app_write_frame(1, &state, ((uint64_t)i)*1000000000ULL, res, framebuffer))
			return 1;
	}
	return 0;
}