#include <stdlib.h>
#include <string.h>
#include <spu_intrinsics.h>
#include "DrawSpan.h"
#include "polygon_pack.h"
#include "texture.h"
#include "viewer_types.h"
#include "Func.h"

#define SPAN_PACK_LOAD 0
#define TEX_LOAD 1
#define FB_STORE 2

//#define PROFILE

#if defined(PROFILE)
#  include <spu_mfcio.h>
#endif

SchedDefineTask(DrawSpan);

DrawSpan::~DrawSpan(void)
{
    smanager->dma_wait(FB_STORE);
    free((void*)((int)linebuf*doneWrite));
}

/**
 * テクスチャは、TEXTURE_SPLIT_PIXEL^2 のブロックに分割する
 *
 * +---+---+---+---+---+---+
 * | 0 | 1 | 2 | 3 | 4 | 5 |
 * +---+---+---+---+---+---+
 * |   |   |   |   |   |11 |
 * +---+---+---+---+---+---+
 * |   |   |   |   |   |17 |
 * +---+---+---+---+---+---+
 * |   |   |   |   |   |23 |
 * +---+---+---+---+---+---+
 * |   |   |   |   |   |29 |
 * +---+---+---+---+---+---+
 * |   |   |   |   |   |35 |
 * +---+---+---+---+---+---+
 *
 * 一辺を TEXTURE_SPLIT とする
 * 各ブロックの数字がブロックIDとなる。
 */

/**
 * テクスチャの座標から、
 * テクスチャのどのブロックかを求める
 *
 * @param[in] tx X coordinates of texture
 * @param[in] tx Y coordinates of texture
 * @param[in] twidth  Width of texture
 * @return block ID
 */
int
DrawSpan::getTexBlock(int tx, int ty, int twidth)
{
     int blockX, blockY;

     blockX = tx / TEXTURE_SPLIT_PIXEL;
     blockY = ty / TEXTURE_SPLIT_PIXEL;

     return blockX + (twidth/TEXTURE_SPLIT_PIXEL)*blockY;
}

/**
 * block ID と、テクスチャの TOP address から
 * (tx,ty) で使われるテクスチャの Tile addres を求める
 *
 * @param[in] tx X coordinates of texture
 * @param[in] tx Y coordinates of texture
 * @param[in] tw Width of texture
 * @param[in] tex_addr_top (tx,ty) で使うテクスチャの先頭address
 * @return block ID
 */
uint32*
DrawSpan::getTile(int tx, int ty, int tw, uint32 *tex_addr_top)
{
    int block = getTexBlock(tx, ty, tw);
    return tex_addr_top + block*TEXTURE_BLOCK_SIZE;
}

/**
 * FrameBuffer に書き込む rgb の領域初期化
 *
 * @param width  Width of Buffer
 * @param height Height of Buffer
 * @param rgb    Initial value of RGB at Buffer
 * @return Buffer
 */
int*
DrawSpan::linebuf_init(int width, int height, int rgb)
{
    int *buf = (int*)smanager->allocate(sizeof(int)*width*height);

    for (int i = 0; i < width*height; i++) {
	buf[i] = rgb;
    }

    return buf;
}

/**
 * Z-Buffer の初期化
 *
 * @param width  Width of Z-Buffer
 * @param height Height of Z-Buffer
 * @return Z-Buffer
 */
float*
DrawSpan::zRow_init(int width, int height)
{
    float *buf = (float*)smanager->allocate(sizeof(float)*width*height);
    float def = 65535.0f;

#if 1
    for (int i = 0; i < width*height; i++) {
	buf[i] = def;
    }
#else 
    vector float init = {0.0f, 0.0f, 0.0f, 0.0f};
    vector float defi = {def, def, def, def};

    for (int i = 0; i < width*height; i += 4) {
	vector float *out = (vector float *)&buf[i];

	*out = spu_add(init, defi);
    }
#endif

    return buf;
}


/**
 * Span が使う Texture Tile があるか
 *
 * @retval != NULL 存在する
 * @retval NULL    存在しない
 */
TilePtr
DrawSpan::isAvailableTile(uint32 *addr)
{
    return hash->get(addr);
}

void
DrawSpan::set_rgb(uint32 *addr)
{
    TilePtr tile;

    tile = tileList->nextTile();
    /**
     * FIFO なので、もし前のが残っていれば削除
     */
    hash->remove(tile->texture_addr);
    
    tile->texture_addr = addr;
    
    hash->put(tile->texture_addr, tile);
    
    smanager->dma_load(tile->pixel, (uint32)addr,
		       sizeof(uint32)*TEXTURE_BLOCK_SIZE, TEX_LOAD);
}

void
DrawSpan::set_rgbs(uint32 *addr, uint32 *max_addr)
{
    uint32 start = (uint32)addr;
    uint32 end = (uint32)max_addr;
    int length = (int)end-start;
    int diff = sizeof(int)*TEXTURE_BLOCK_SIZE;
    int max_tile = 16;

    for (int i = 0, j = 0; i <= length && j < max_tile; i += diff, j++) {
	set_rgb((uint32*)(start + i));
    }
}

uint32
DrawSpan::get_rgb(int tx, int ty, uint32 *addr)
{
    TilePtr tile;

    tile = hash->get(addr);
    return tile->pixel[(TEXTURE_SPLIT_PIXEL)*ty+tx];
}

/**
 * DrawSpan の再起動 (DrawSpanRenew 生成)
 *
 * @param[in] spack 現在処理している SpanPack
 * @param[in] cur_span_x span->length_x != 1 の時の Span の処理で
 *                       どこまで進んでいるか
 */
void
DrawSpan::reboot(SpanPackPtr spack, int cur_span_x)
{
    DrawSpanArgPtr args =
	(DrawSpanArgPtr)smanager->allocate(sizeof(DrawSpanArg));
    TaskPtr renew_task = smanager->create_task(TASK_DRAW_SPAN2);

    // 数が多いので構造体で渡す
    args->display      = smanager->get_param(0);
    args->screen_width = smanager->get_param(1);
    args->rangex_start = smanager->get_param(2);
    args->rangex_end   = smanager->get_param(3);
    args->rangey       = smanager->get_param(4);
    renew_task->add_param((int)args);

    /**
     * SpanPack は続きから開始するので、
     * 現在の状態をコピーしておく。
     * spack は rbuf から取得してる可能性があり
     * rbuf はシステムが自動的に free() するため
     * アドレスだけ渡すのはNG
     */ 
    SpanPackPtr curr = (SpanPackPtr)smanager->allocate(sizeof(SpanPack));
    memcpy(curr, spack, sizeof(SpanPack));
    renew_task->add_param((int)curr);
    renew_task->add_param(cur_span_x);

    // linebuf と zRow も引き継がせる
    renew_task->add_param((int)linebuf);
    renew_task->add_param((int)zRow);

    /**
     * 再起動したタスクを待つ
     */ 
    smanager->wait_task(renew_task);

    // next_spack は free() するので wait する
    smanager->dma_wait(SPAN_PACK_LOAD);
}

void
DrawSpan::writebuffer(unsigned int display, int buf_width, int height,
		      int screen_width)
{
    for (int i = 0; i < height; i++) {
	smanager->dma_store(&linebuf[i*buf_width],
			    display + (sizeof(int)*screen_width*i),
			    sizeof(int)*buf_width, FB_STORE);
    }

    doneWrite = 1;
}

#define Prof(st,cur) (((st)-(cur))/79800000.0f * 1000.0f)

int
DrawSpan::run(void *rbuf, void *wbuf)
{
    SpanPackPtr spack = (SpanPackPtr)smanager->get_input(0);
    SpanPackPtr next_spack = (SpanPackPtr)smanager->allocate(sizeof(SpanPack));
    SpanPackPtr free_spack = next_spack; // next_spack の free() 用
    Span *span;

    uint32 display   = smanager->get_param(0);
    int screen_width = smanager->get_param(1);
    int rangex_start = smanager->get_param(2);
    int rangex_end   = smanager->get_param(3); 

    // このタスクが担当する x の範囲
    int rangex = rangex_end - rangex_start + 1;

    // y の範囲
    int rangey = smanager->get_param(4);
    
    hash = (TileHashPtr)smanager->global_get(GLOBAL_TEXTURE_HASH);
    tileList = (TileListPtr)smanager->global_get(GLOBAL_TILE_LIST);

    zRow = zRow_init(rangex, rangey);
    linebuf = linebuf_init(rangex, rangey, 0xffffffff);

    doneWrite = 0;

    int tileNum = 0;
    int flag = 0;
    
#if defined(PROFILE)
    uint32 profile = spu_read_decrementer();
    printf("%d %u DrawSpan start\n", smanager->get_cpuid(), profile);
#endif
    do {
	/**
	 * SpanPack->next が存在する場合、
	 * 現在の SpanPack を処理してる間に
	 * 次の SpanPack の DMA 転送を行う
	 */
	if (spack->next != NULL) {
	    smanager->dma_load(next_spack, (uint32)spack->next,
			       sizeof(SpanPack), SPAN_PACK_LOAD);
	} else {
	    next_spack = NULL;
	}

	for (int t = 0; t < spack->info.size; t++) {	  
	    span = &spack->span[t];

	    uint32 rgb = 0x0000ff00;
	    float tex1 = span->tex_x1;
	    float tex2 = span->tex_x2;
	    float tey1 = span->tex_y1;
	    float tey2 = span->tex_y2;

	    /**
	     * Span が持つ 1 pixel 毎の
	     * テクスチャの座標
	     */
	    int tex_xpos;
	    int tex_ypos;

	    /**
	     * (tex_xpos, tex_ypos) の、ブロック内(上の図参照)での座標と
	     * そのブロックのアドレス(MainMemory)
	     */
	    int tex_localx;
	    int tex_localy;
	    uint32 *tex_addr;

	    int x = span->x;
	    int y = span->y;
	    int x_len = span->length_x;
	    float z = span->start_z;
	    float zpos = span->end_z;

	    // 座標が [0 .. split_screen_w-1] に入るように x,y を -1
	    int localx = getLocalX(x-1);
	    int localy = getLocalY(y-1);

	    if (x_len == 1) {
		if (x < rangex_start || rangex_end < x) {
		    continue;
		}

		flag = 1;
		tex_xpos = (int)((span->tex_width-1) * tex1);
		tex_ypos = (int)((span->tex_height-1) * tey1);

		if (zpos < zRow[localx + (rangex*localy)]) {
		    tex_addr = getTile(tex_xpos, tex_ypos,
				       span->tex_width, span->tex_addr);
		    tex_localx = tex_xpos % TEXTURE_SPLIT_PIXEL;
		    tex_localy = tex_ypos % TEXTURE_SPLIT_PIXEL;
		    
		    if (!isAvailableTile(tex_addr)) {
#  if defined(PROFILE)
			printf("%d %u start load tile\n",
			       smanager->get_cpuid(), spu_read_decrementer());
#  endif
			set_rgb(tex_addr);
			smanager->dma_wait(TEX_LOAD);
#  if defined(PROFILE)
			printf("%d %u end load tile\n",
			       smanager->get_cpuid(), spu_read_decrementer());
#  endif
			tileNum++;
		    }

		    rgb = get_rgb(tex_localx, tex_localy, tex_addr);

		    zRow[localx + (rangex*localy)] = zpos;
		    linebuf[localx + (rangex*localy)] = rgb;
		}
	    } else {
		int js = (x < rangex_start) ? rangex_start - x : 0;
		int je = (x + x_len > rangex_end) ? rangex_end - x : x_len;
		float tex_x, tex_y, tex_z;

		for (int j = js; j <= je; j++) {
		    flag = 1;
		    localx = getLocalX(x-1+j);

		    tex_z = z*(x_len-1-j)/(x_len-1) + zpos*j/(x_len-1);

		    tex_x = tex1*(x_len-1-j)/(x_len-1) + tex2*j/(x_len-1);
		    tex_y = tey1*(x_len-1-j)/(x_len-1) + tey2*j/(x_len-1);
		    if (tex_x > 1) tex_x = 1;
		    if (tex_x < 0) tex_x = 0;
		    if (tex_y > 1) tex_y = 1;
		    if (tex_y < 0) tex_y = 0;
		    tex_xpos = (int)((span->tex_width-1) * tex_x);
		    tex_ypos = (int)((span->tex_height-1) * tex_y);
		    
		    if (tex_z < zRow[localx + (rangex*localy)]) {
			tex_addr = getTile(tex_xpos, tex_ypos,
					   span->tex_width, span->tex_addr);
			tex_localx = tex_xpos % TEXTURE_SPLIT_PIXEL;
			tex_localy = tex_ypos % TEXTURE_SPLIT_PIXEL;

			/**
			 * Tile が無い場合、一旦タスクはここで中断し、
			 * Tile をロードするタスクを走らせた後に再起動する
			 */
			if (!isAvailableTile(tex_addr)) {
			    spack->info.start = t;
#if 0
			    set_rgbs(tex_addr,
				     getTile(span->tex_width-1, tex_ypos,
					     span->tex_width, span->tex_addr));
			    //smanager->dma_wait(TEX_LOAD);
			    reboot(spack, j);
			    goto FINISH;
#else

#  if defined(PROFILE)
			    printf("%d %u start load tile\n",
				   smanager->get_cpuid(), 
				   spu_read_decrementer());
#  endif

			    set_rgb(tex_addr);
			    smanager->dma_wait(TEX_LOAD);
			    tileNum++;

#  if defined(PROFILE)
			    printf("%d %u end load tile\n",
				   smanager->get_cpuid(),
				   spu_read_decrementer());
#  endif
#endif
			}
			
			rgb = get_rgb(tex_localx, tex_localy, tex_addr);

			zRow[localx + (rangex*localy)] = tex_z;
			linebuf[localx + (rangex*localy)] = rgb;
		    }
		}
	    }
	}

	for (int t = 0; t < spack->info.size; t++) {	  
	    span = &spack->span[t];

	    uint32 rgb = 0x0000ff00;
	    float tex1 = span->tex_x1;
	    float tex2 = span->tex_x2;
	    float tey1 = span->tex_y1;
	    float tey2 = span->tex_y2;

	    /**
	     * Span が持つ 1 pixel 毎の
	     * テクスチャの座標
	     */
	    int tex_xpos;
	    int tex_ypos;

	    /**
	     * (tex_xpos, tex_ypos) の、ブロック内(上の図参照)での座標と
	     * そのブロックのアドレス(MainMemory)
	     */
	    int tex_localx;
	    int tex_localy;
	    uint32 *tex_addr;

	    int x = span->x;
	    int y = span->y;
	    int x_len = span->length_x;
	    float z = span->start_z;
	    float zpos = span->end_z;

	    // 座標が [0 .. split_screen_w-1] に入るように x,y を -1
	    int localx = getLocalX(x-1);
	    int localy = getLocalY(y-1);

	    if (x_len == 1) {
		if (x < rangex_start || rangex_end < x) {
		    continue;
		}

		flag = 1;
		tex_xpos = (int)((span->tex_width-1) * tex1);
		tex_ypos = (int)((span->tex_height-1) * tey1);

		if (zpos < zRow[localx + (rangex*localy)]) {
		    tex_addr = getTile(tex_xpos, tex_ypos,
				       span->tex_width, span->tex_addr);
		    tex_localx = tex_xpos % TEXTURE_SPLIT_PIXEL;
		    tex_localy = tex_ypos % TEXTURE_SPLIT_PIXEL;
		    
		    if (!isAvailableTile(tex_addr)) {
#  if defined(PROFILE)
			printf("%d %u start load tile\n",
			       smanager->get_cpuid(), spu_read_decrementer());
#  endif
			set_rgb(tex_addr);
			smanager->dma_wait(TEX_LOAD);
#  if defined(PROFILE)
			printf("%d %u end load tile\n",
			       smanager->get_cpuid(), spu_read_decrementer());
#  endif
			tileNum++;
		    }

		    rgb = get_rgb(tex_localx, tex_localy, tex_addr);

		    zRow[localx + (rangex*localy)] = zpos;
		    linebuf[localx + (rangex*localy)] = rgb;
		}
	    } else {
		int js = (x < rangex_start) ? rangex_start - x : 0;
		int je = (x + x_len > rangex_end) ? rangex_end - x : x_len;
		float tex_x, tex_y, tex_z;

		for (int j = js; j <= je; j++) {
		    flag = 1;
		    localx = getLocalX(x-1+j);

		    tex_z = z*(x_len-1-j)/(x_len-1) + zpos*j/(x_len-1);

		    tex_x = tex1*(x_len-1-j)/(x_len-1) + tex2*j/(x_len-1);
		    tex_y = tey1*(x_len-1-j)/(x_len-1) + tey2*j/(x_len-1);
		    if (tex_x > 1) tex_x = 1;
		    if (tex_x < 0) tex_x = 0;
		    if (tex_y > 1) tex_y = 1;
		    if (tex_y < 0) tex_y = 0;
		    tex_xpos = (int)((span->tex_width-1) * tex_x);
		    tex_ypos = (int)((span->tex_height-1) * tex_y);
		    
		    if (tex_z < zRow[localx + (rangex*localy)]) {
			tex_addr = getTile(tex_xpos, tex_ypos,
					   span->tex_width, span->tex_addr);
			tex_localx = tex_xpos % TEXTURE_SPLIT_PIXEL;
			tex_localy = tex_ypos % TEXTURE_SPLIT_PIXEL;

			/**
			 * Tile が無い場合、一旦タスクはここで中断し、
			 * Tile をロードするタスクを走らせた後に再起動する
			 */
			if (!isAvailableTile(tex_addr)) {
			    spack->info.start = t;
#if 0
			    set_rgbs(tex_addr,
				     getTile(span->tex_width-1, tex_ypos,
					     span->tex_width, span->tex_addr));
			    //smanager->dma_wait(TEX_LOAD);
			    reboot(spack, j);
			    goto FINISH;
#else

#  if defined(PROFILE)
			    printf("%d %u start load tile\n",
				   smanager->get_cpuid(), 
				   spu_read_decrementer());
#  endif

			    set_rgb(tex_addr);
			    smanager->dma_wait(TEX_LOAD);
			    tileNum++;

#  if defined(PROFILE)
			    printf("%d %u end load tile\n",
				   smanager->get_cpuid(),
				   spu_read_decrementer());
#  endif
#endif
			}
			
			rgb = get_rgb(tex_localx, tex_localy, tex_addr);

			zRow[localx + (rangex*localy)] = tex_z;
			linebuf[localx + (rangex*localy)] = rgb;
		    }
		}
	    }
	}


	smanager->dma_wait(SPAN_PACK_LOAD);

	SpanPackPtr tmp_spack = spack;
	spack = next_spack;
	next_spack = tmp_spack;
    } while (spack);

    writebuffer(display, rangex, rangey, screen_width);

    // linebuf は、writebuffer() の dma_store を wait する
    // DrawSpan::~DrawSpan() 内で free する。
    //free(linebuf);
    free(zRow);

FINISH:
    /**
     * goto FINISH; の時は reboot なので
     * linebuf, zRow は free() しない
     */
    free(free_spack);
    
#if defined(PROFILE)
    printf("%d %u DrawSpan end\n",
	   smanager->get_cpuid(), spu_read_decrementer());
    if (flag) {
	printf("%d %d tileNum\n", smanager->get_cpuid(), tileNum);
    }
#endif

    return 0;
}