#include <SDL.h>
#include "viewer.h"
#include "viewer_types.h"
#include "SceneGraph.h"
#include "SceneGraphRoot.h"
#include "scene_graph_pack.h"
#include "matrix_calc.h"
#include "Func.h"
#include "error.h"
#include "TaskManager.h"
#include <wchar.h>
#include "Pad.h"
#include "Application.h"
#include "lindaapi.h"
#include "SchedTask.h"

/* measure for FPS (Frame Per Second) */
static int start_time;
static int this_time;
int frames;
// static void post2speRunLoop(void *viewer);
//static void post2runMove(void *viewer);
//static void post2exchange_sgroot(void *viewer);
//static void post2speRunLoop(void *viewer);

//SceneGraphRootPtr sgroot_2;

/* Data Pack sent to Other CPUs (ex. SPE) */

RenderingData r[2];

int  ppi, spi = 0;

Viewer::Viewer(TaskManager *m, ViewerDevice *vd, int b, int w, int h, int _num)
{
    spe_num = _num;
    manager = m;

    dev = vd;
    pixels = dev->video_init(manager, b, w, h);

    width = dev->width;
    height = dev->height;
    bpp = dev->bpp;

}

int
Viewer::get_ticks(void)
{
    int time;
    time = SDL_GetTicks();
    return time;
}

bool
Viewer::quit_check(void)
{
    SDL_Event event;

    while(SDL_PollEvent(&event)) {
        if (event.type==SDL_QUIT) {
            return true;
        }
    }

    Uint8 *keys=SDL_GetKeyState(NULL);

    if (keys[SDLK_q] == SDL_PRESSED) {
        return true;
    }

    return false;
}

void
Viewer::quit(void)
{
    SDL_Quit();
}


void
Viewer::run_init(TaskManager *manager, Application *app)
{
    this->manager = manager;

    if (spe_num == 0) spe_num = 1;

    sgroot = new SceneGraphRoot(this->width, this->height, manager);

    int light_num = 4;
    int size = sizeof(float)*4*light_num; //xyz+alfa(4) * light_num(4)
    int light_size = size / sizeof(float);

    light_xyz = (float *)manager->allocate(size);
	
    for (int i = 0; i < light_size ; i++) {
      light_xyz[i] = 0.0f;
    }


    for(int i = 0; i < spe_num; i++) {
        HTaskPtr data_load = manager->create_task(DataAllocate);
	data_load->set_param(0,(memaddr)3);                         // num of allocate block
	data_load->set_param(1,(memaddr)(sizeof(float)*4*light_num)); // 1st allocate size
	data_load->set_param(2,(memaddr)Light);                     // 1st id
	data_load->set_param(3,(memaddr)(light_num * sizeof(int)));   // 2nd size
	data_load->set_param(4,(memaddr)LightSwitch);               // 2nd id
	data_load->set_param(5,(memaddr)16);                        // 3rd size
	data_load->set_param(6,(memaddr)LightSysSwitch);            // 3rd id
	data_load->set_cpu((CPU_TYPE)((int)SPE_0 + i));
	data_load->spawn();
    }

    light_switch = (int*)manager->allocate(size);

    for (int i = 0; i < light_num; i++) {
        light_switch[i] = 0;
    }

    size = 16; // LightSysSwitch は 4byte. 残り 12byte は DMA転送の為のパディング

    light_sysswitch = (int*)manager->allocate(size);

    for (unsigned int i = 0; i < size / sizeof(int); i++) {
      light_sysswitch[i] = 0;
    }

    start_time = get_ticks();
    this_time  = 0;
    frames     = 0;
    pp_sum_num = 0;

    this->app = app;
    MainLoop *mainloop = app->init(this, this->width, this->height);
    mainloop->mainLoop();
}


HTaskPtr
Viewer::initLoop()
{
    HTaskPtr task_next;
    HTaskPtr task_tex;

    /*

      こうかな

      manager->createMemList(sizeof(SpanPack), POOL_SPANPACK);


     */



    for(int i=0;i<2;i++) {
	r[i].ppack  = (PolygonPack*)manager->allocate(sizeof(PolygonPack));
	r[i].ppack->next = 0;  

	r[i].spackList_length = (this->height + split_screen_h - 1) / split_screen_h;
	r[i].spackList = (SpanPack*)manager->allocate(sizeof(SpanPack)*r[i].spackList_length);
	// printf("spackList %0lx height %d\n",(unsigned long)r[i].spackList, this->height);

	/**
	 * SPU に送る address list は 16 バイト倍数でないといけない。
	 * spackList_length*sizeof(SpanPack*) が 16 バイト倍数になるような
	 * length_align を求めている。はみ出した部分は使われない
	 * (ex) spackList_length が 13 の場合
	 *   spackList_length_align = 16;
	 *     実際に送るデータは64バイトになるのでOK
	 *     14,15,16 の部分は何も入らない。
	 */
	r[i].spackList_length_align = (r[i].spackList_length + 3)&(~3);

	/* 各 SPU が持つ、SpanPack の address list */
	r[i].spackList_ptr =
	    (SpanPack**)manager->allocate(sizeof(SpanPack*)*r[i].spackList_length_align);

	for (int j = 0; j < r[i].spackList_length; j++) {
	    r[i].spackList_ptr[j] = &r[i].spackList[j];
	}

	for (int j = 1; j <= r[i].spackList_length; j++) {
	    r[i].spackList[j-1].init(j*split_screen_h);
	}
    }

    task_next = manager->create_task(RUN_LOOP_TASK,0,0,0,0);
    task_next->set_param(0, (void*)this);
   
    // ここは、Iterator を用意するべきだよね
    for (int j = 0; j < spe_num; j++) {
        task_tex = manager->create_task(AllocateSegment,0,0,0,0);
        task_tex->set_cpu((CPU_TYPE)((int)SPE_0 + j));
        task_next->wait_for(task_tex);
	task_tex->spawn();	
    }


    task_next->spawn();

    return 0;
}


void
Viewer::getKey()
{
    Pad *pad = sgroot->getController();
    if (pad->right.isHold()) {
		keyPtr->right = HOLD;
    } else if (pad->right.isPush()) {
		keyPtr->right = PUSH;
    } else {
		keyPtr->right = NONE;
    }
    
    if (pad->left.isHold()) {
		keyPtr->left = HOLD;
    } else if (pad->left.isPush()) {
		keyPtr->left = PUSH;
    } else {
		keyPtr->left = NONE;
    }

    if (pad->up.isHold()) {
		keyPtr->up = HOLD;
    } else if (pad->up.isPush()) {
		keyPtr->up = PUSH;
    } else {
		keyPtr->up = NONE;
    }

    if (pad->down.isHold()) {
		keyPtr->down = HOLD;
    } else if (pad->down.isPush()) {
		keyPtr->down = PUSH;
    } else {
		keyPtr->down = NONE;
    }

    if (pad->circle.isHold()) {
		keyPtr->circle = HOLD;
    } else if (pad->circle.isPush()) {
		keyPtr->circle = PUSH;
    } else {
		keyPtr->circle = NONE;
    }
}


void
Viewer::mainLoop()
{
    if (pixels) {
	initLoop();
    } else {
	HTaskPtr task_next = manager->create_task(EXEC_ONLY_TASK, 0, 0, 0, 0);
	task_next->set_param(0, (void*)this);
	
	task_next->spawn();
    }
}

bool
Viewer::main_exec(HTaskPtr task_next)
{
    psx_sync_n();
    
    task_next = app->application_task(task_next, this);
    dev->clear_screen();
    
    bool quit_flg;
    quit_flg = quit_check();
    if (quit_flg == true) {
        this_time = get_ticks();
        run_finish();
        return false;
    }

    dev->clean_pixels();
    pixels = dev->flip_screen(pixels);

    sgroot->updateControllerState();
    if (app->app_loop(this)) {
	//TaskArray を使うか使わないか
	if (sgroot->gtask_array != NULL) {
	    sgroot->create_task_array();
	    sgroot->allExecute(width, height);
	    sgroot->task_array_finish();
	} else {
	    sgroot->allExecute(width, height);	
	}
    } else {
	sgroot->treeApply(width, height);	
    }

    light_xyz_stock = sgroot->getLightVector();


    light_switch_stock = sgroot->getLightSwitch();
    light_sysswitch_stock = sgroot->getLightSysSwitch();

    return true;
}

void
Viewer::run_loop(HTaskPtr task_next)
{
    if (main_exec(task_next)) {
        //rendering(task_next);
        task_next->spawn();
    }
}

SchedDefineTask1(EXEC_ONLY_TASK,exec_only_task);

static int
exec_only_task(SchedTask *smanager, void *rbuf, void *wbuf)
{

  Viewer *viewer = (Viewer*)smanager->get_param(0);
  HTaskPtr task_next = smanager->create_task(EXEC_ONLY_TASK, 0, 0, 0, 0);
  task_next->set_param(0, (void*)viewer);

  if (viewer->main_exec(task_next)) {
      task_next->spawn();
  }
  return 0;
}

SchedDefineTask1(RUN_LOOP_TASK,run_loop_task);

static int
run_loop_task(SchedTask *smanager, void *rbuf, void *wbuf)
{

  Viewer *viewer = (Viewer*)smanager->get_param(0);
  HTaskPtr task_next = smanager->create_task(CREATE_PP_TASK, 0, 0, 0, 0);
  task_next->set_param(0, (void*)viewer);

  viewer->run_loop(task_next);

  return 0;
}



void
Viewer::run_collision()
{
}

SchedDefineTask1(CREATE_PP_TASK, create_pp_task);

static int
create_pp_task(SchedTask* smanager, void* rbuf, void* wbuf)
{

  Viewer *viewer = (Viewer*)smanager->get_param(0);
  HTaskPtr task_next = smanager->create_task(CREATE_SP_TASK, 0, 0, 0, 0);
  task_next->set_param(0, (void*)viewer);

  viewer->create_pp(task_next);

  return 0;

}

void 
Viewer::create_pp(HTaskPtr task_next) {

    rendering_pp(task_next, sgroot);
    
    // Barrier 同期
    // run_draw() を呼ぶ post2runDraw
    task_next->spawn(); // create_sp_task

}


SchedDefineTask1(CREATE_SP_TASK, create_sp_task);

static int
create_sp_task(SchedTask* smanager, void* rbuf, void* wbuf)
{

  Viewer *viewer = (Viewer*)smanager->get_param(0);
  HTaskPtr task_next = smanager->create_task(DRAW_TASK, 0, 0, 0, 0);
  task_next->set_param(0, (void*)viewer);

  viewer->create_sp(task_next);

  return 0;

}

void 
Viewer::create_sp(HTaskPtr task_next) {

    rendering_sp(task_next, sgroot);
    
    // Barrier 同期
    // run_draw() を呼ぶ post2runDraw
    task_next->spawn(); // draw_task

}
     

SchedDefineTask1(DRAW_TASK, draw_task);

static int
draw_task(SchedTask* smanager, void* rbuf, void* wbuf)
{

  Viewer* viewer = (Viewer*)smanager->get_param(0);
  HTaskPtr task_next = smanager->create_task(RUN_LOOP_TASK, 0, 0, 0, 0); 
  task_next->set_param(0, (void*)viewer);
  viewer->run_draw(task_next);
  
  return 0;

}

void
Viewer::run_draw(HTaskPtr task_next) // 引数に post2runLoop を入れるようにする
{
    common_draw(task_next);
   
    task_next->spawn(); // run_loop_task
    // TASK_DRAW_SPAN が全て終わったら DUMMY_TASK が Viewer::run_loop() を呼ぶ    

    frames++;
}


void
Viewer::rendering_pp(HTaskPtr task_next, SceneGraphRoot *sgroot)
{


    /*

      sg->pp をもっと細かく分けるか？
      
      DataSegment seg1;

      (ObjectPolygonPtr)seg1->in = (ObjectPolygonPtr)makeSegment(pp_sum_num);
      (SpanPackPtr)seg1->out = (SpanPackPtr)makeSegment(pp_sum_num);

      HTaskPtr create_pp = manager->create_task(CreatePolygon,seg1);
      create_pp->spawn();

      DataSegment seg;

     */
    


    HTaskPtr game_task_array = 0;

    /* GameTask の処理の終了を待ってからポリゴンを作る */
    if (sgroot->gtask_array != NULL) {
	game_task_array = sgroot->gtask_array->get_task_array();
    }

    PolygonPackPtr out_pp = r[ppi].ppack;
    out_pp->init();

    CameraPtr camera = sgroot->getCamera();
    
    //多分このsg_remove_listであってる?。チェック対象かも
    for (SceneGraphPtr t = sgroot->sg_remove_list; t != NULL; t = t->next) {
      if (t->size > 0) {
        pp_sum_num += t->pp_num;
        for (int i = 0; i < t->pp_num; i++) {

	  HTaskPtr create_pp = manager->create_task(CreatePolygonFromSceneGraph);
	  
	  create_pp->add_inData(&t->pp[i], sizeof(PolygonPack));
	  create_pp->add_inData(t->matrix, sizeof(float)*16);
	  create_pp->add_inData(camera->m_screen, sizeof(float)*16);
	  create_pp->add_inData(t->texture_info, sizeof(texture_list));
	  
	  if ( (unsigned long)t->matrix % 16) {
	    printf("marix not aligned\n");
	  }
	  
	  if ((unsigned long)t->texture_info % 16) {
	    printf("texture_info not aligned\n");
	  }
	  
	  create_pp->add_outData(out_pp, sizeof(PolygonPack));
	  
	  if (game_task_array != NULL) {
	    create_pp->wait_for(game_task_array);
	  }
	  
	  PolygonPackPtr tmp_pp = (PolygonPackPtr)manager->allocate(sizeof(PolygonPack));

	  tmp_pp->init();
	  create_pp->set_param(0, (memaddr)tmp_pp);
	  out_pp = tmp_pp;
	  
	  //create_pp->set_cpu(SPE_ANY);
	  create_pp->set_cpu(CPU_PPE);
	  task_next->wait_for(create_pp);
	  create_pp->spawn();
	} 
      }
    }
}


void
Viewer::rendering_sp(HTaskPtr task_next, SceneGraphRoot *sgroot)
{

#ifdef CreateSpanDivi

    /*

      CreateSpan -> SortSpan -> DrawSpan
                 |           |
                 v           v
                Data        Data

      Bestなデータ構造ってどういうのだろう。
      今のデータ構造は「良い」とは言えないだろうなぁ。
      間に wait_for もいまいち見たいだから、Taskで挟むのがいいか

      CreateSpan -> BridgeTask -> SortSpan -> BridgeTask -> DrawSpan
      
      みたいな感じ？ この wait for は見づらいしね。んで上の図みたな依存性がみれたり、
      実行のガントチャートが見れたりすると嬉しいって話は前からある。

      以下妄想

      CreateSpan(SegID in) {

        SpanPack out = makeSegment();
        PolygonPack pp = getSegment(in);

        --処理--

        return SortTask(out);

      }

      SortSpan(SegID in) {

        SpanPack out = makeSegment();
        SpanPack sp = getSegment(in):

        --処理--

        return DrawSpan(out);

      }

      DrawSpan() {

       

        return Start(); //Startは同期するん？

      }

      main() {

        --処理--

        PolygonPack out = makeSegment(3); //個数なのか？

        out[0] = pp[0];
        out[1] = pp[1];
        out[2] = pp[2];

        return CreateSpan(out); //3スレッドできるのか？
      
      }


      下みたいにはもう書きたくないか・・・
      うーんだいぶ書きなおすか。

     */
 

    PolygonPackPtr pp =  r[ppi].ppack;
    SpanPackPtr* spackList =  (SpanPackPtr*)manager->allocate(sizeof(SpanPackPtr)*pp_sum_num);

    for (PolygonPackPtr t = pp; t->next != NULL; t = t->next) {

        int span_num = t->info.span_num;
        int spack_num = (span_num + MAX_SIZE_SPAN -1) / MAX_SIZE_SPAN; 

        SpanPackPtr spack =  (SpanPackPtr)manager->allocate(sizeof(SpanPack)*spack_num);
        spackList[i] = spack;
        
        for (int i = 0; i < spack_num; i++) {
            HTaskPtr task_create_sp = manager->create_task(CreateSpan);
            task_create_sp->set_inData(0, t, sizeof(PolygonPack));
            task_create_sp->set_outData(0, &spack[i], sizeof(SpanPack));

            task_create_sp->spawn();
            task_next->wait_for(task_create_sp);
        }

    }

    HTaskPtr sort_span = (SortSpanPtr)manager->allocate(sizeof(SortSpan));

    /*まだ途中*/

#else

    int  range_base = spe_num;

    // 切り上げのつもり
    int range = (r[spi].spackList_length + range_base - 1) / range_base;

    for (int i = 0; i < range_base; i++) {
        int index_start = range*i;
        int index_end = (index_start + range >= r[spi].spackList_length)
            ? r[spi].spackList_length : index_start + range;
	int starty = index_start*split_screen_h + 1;
        int endy = index_end*split_screen_h;
	if (starty<=0) starty = 1;
	if (endy>height) endy = height;

	HTaskPtr task_create_sp = manager->create_task(CreateSpan);

        task_create_sp->set_param(0,index_start);
	
        /**
         * ex. screen_height が 480, spenum が 6 の場合、各SPEのy担当範囲
         *   [  1.. 80] [ 81..160] [161..240]
         *   [241..320] [321..400] [401..480]
         *
         * ex. screen_height が 1080, spenum が 5 の場合、
         *   [  1..216] [217..432] [433..648]
         *   [649..864] [865..1080]
         */

        task_create_sp->set_param(1,starty);
        task_create_sp->set_param(2,endy);

       	task_create_sp->add_inData(r[ppi].ppack, sizeof(PolygonPack));
        task_create_sp->add_inData(r[spi].spackList_ptr,
                                   sizeof(SpanPack*)*r[spi].spackList_length_align);
        task_create_sp->add_inData(&r[spi].spackList[index_start], sizeof(SpanPack));

	task_next->wait_for(task_create_sp);

	task_create_sp->set_cpu(SPE_ANY);
        task_create_sp->spawn();
    }

#endif

}

void
Viewer::common_draw(HTaskPtr task_next)
{

    //task_next = manager->create_task(Dummy);
    //task_next->set_post(post2runLoop, (void*)this);

    //Light info update
                                                                   
    int light_num = 4;
    int size = sizeof(float)*4*light_num; //xyz+alpha(4) * light_num(4) 
    int light_size = size / sizeof(float);

    for (int i = 0; i < light_size; i++) {
        light_xyz[i] = light_xyz_stock[i]; 
    }

    for (int i = 0; i < light_num; i++) {
        light_switch[i] = light_switch_stock[i];
    }

    light_sysswitch[0] = light_sysswitch_stock;

#if 1

    HTask *data_update_wait = 0;
    for (int i = 0; i < spe_num; i++) {
	data_update_wait = manager->create_task(DataUpdate);
	data_update_wait->set_param(0,3);
	data_update_wait->set_param(1,Light);              // GlobalSet ID base
	data_update_wait->set_inData(0,light_xyz,size);                           // Light
	data_update_wait->set_inData(1,light_switch,light_num * sizeof(int));     // LightSwitch = Light+1
	data_update_wait->set_inData(2,light_sysswitch,16);                       // LightSysSwitch = Light+2
        data_update_wait->set_cpu((CPU_TYPE)(SPE_0+i));
	data_update_wait->spawn();
	}

#else

    HTask *data_update_wait = manager->create_task(Dummy);
    for (int i = 0; i < spe_num; i++) {
      
	HTaskPtr data_update = manager->create_task(DataUpdate);
	data_update->set_param(0,3);
	data_update->set_param(1,Light);              // GlobalSet ID base
	data_update->set_inData(0,light_xyz,size);                           // Light
	data_update->set_inData(1,light_switch,light_num * sizeof(int));     // LightSwitch = Light+1
	data_update->set_inData(2,light_sysswitch,16);                       // LightSysSwitch = Light+2
        data_update->set_cpu((CPU_TYPE)(SPE_0+i));
        data_update->wait_for(data_update_wait);
	data_update->spawn();
	}

#endif

    ppi ^= 1; 
    r[ppi].ppack->clear();

    for (int i = 0; i < r[spi].spackList_length; i++) {
        SpanPack *spack = &r[spi].spackList[i];
        int startx = 1;
        int endx = split_screen_w;

        int starty = spack->info.y_top - split_screen_h + 1;
        //int endy = spack->info.y_top;
        int rangey = (starty + split_screen_h - 1 > this->height)
            ? this->height - starty + 1 : split_screen_h;

#ifdef USE_TASKARRAY

        int array_task_num = (this->width + split_screen_w - 1) / split_screen_w;
        HTaskPtr task_draw_array = manager->create_task_array(DrawSpan, array_task_num, 6, 1, rangey);
        Task *task_draw = 0;
        
	  while (startx < this->width) {
	    
	    // Draw SpanPack
	    
	    task_draw = task_draw_array->next_task_array(DrawSpan,task_draw);
	    task_draw->set_param(0,(memaddr)&pixels[(startx-1) + this->width*(starty-1)]);
	    task_draw->set_param(1,this->width);
	    task_draw->set_param(2,startx);
	    task_draw->set_param(3,endx);
	    task_draw->set_param(4,rangey);
	    task_draw->set_param(5,spack->info.size);
	    
	    task_draw->set_inData(0,spack, sizeof(SpanPack));
	    
	    for (int i = 0; i < rangey; i++) {
	      task_draw->set_outData(i,
				     &pixels[(startx-1) + this->width*(starty-1 + i) ],
				     (endx-startx+1)*sizeof(int));
	    }
	    
	    startx += split_screen_w;
	    endx += split_screen_w;
	    
	    if (endx > this->width) {
	      endx = this->width;
	    }
	    
	  }	
	  
	  task_draw_array->spawn_task_array(task_draw->next());
	  task_draw_array->set_cpu(SPE_ANY);
	  //task_draw_array->wait_for(data_update_wait);
#ifndef USE_PIPELINE
	  task_next->wait_for(task_draw_array);
#endif
	  task_draw_array->spawn();
#else

       HTaskPtr task_draw;

        while (startx < this->width) {
	  if (spack->info.size > 0 || mem_flag == 1) {
                // Draw SpanPack
                task_draw = manager->create_task(DrawSpan);

                task_draw->set_param(0,
                    (memaddr)&pixels[(startx-1) + this->width*(starty-1)]);
                task_draw->set_param(1,this->width);
		task_draw->set_param(2,startx);
		task_draw->set_param(3,endx);
		task_draw->set_param(4,rangey);
		task_draw->set_param(5,spack->info.size);

                task_draw->add_inData(spack, sizeof(SpanPack));

		for (int i = 0; i < rangey; i++) {
		    task_draw->add_outData(
			&pixels[(startx-1) + this->width*(starty-1 + i) ],
			(endx-startx+1)*sizeof(int));
		}

	} else {
		// 7.7.3 SL1 Data Cache Range Set to Zero コマンド
		//  を使って、DMAでclearするべき... ということは、
		// それもSPEでやる方が良い?
		  memset(&pixels[(startx-1)+this->width*(starty-1)],
		                0, (this->width)*sizeof(int)*rangey);
		  break;
	}

	    task_draw->set_cpu(SPE_ANY);
#ifndef USE_PIPELINE
            task_next->wait_for(task_draw);
#endif
            task_draw->spawn();

            startx += split_screen_w;
            endx += split_screen_w;

            if (endx > this->width) {
                endx = this->width;
            }
        }
#endif

    }

    //data_update_wait->spawn();

    spi ^= 1; 
    for (int i = 1; i <= r[spi].spackList_length; i++) {
        r[spi].spackList[i-1].reinit(i*split_screen_h);
    }

    if (profile) {
	if (frames % 50 == 49) {
	  manager->show_profile();
           this_time = get_ticks();
           if (this_time != start_time) {
              printf("\n%f FPS\n", ((((float)frames)*1000.0)/(this_time-start_time)));
              start_time = this_time; frames = 0;
           }
        }
    }
}

void
Viewer::run_finish()
{
    dev->free_device();
    if (this_time != start_time) {
        printf("%f FPS\n", (((float)frames)/(this_time-start_time))*1000.0);
    }
    if (profile) {
	    manager->show_profile();
    }

    delete sgroot;
//    delete sgroot_2;
    quit();
}

/* end */