#include <xmmintrin.h>

int * pix32tbl;

void

_PsetL(int * ddscr, unsigned char * bkline)
{

  register int i,acc, *ptrl,*ptrh;

   __m128 xmm0,xmm1,xmm2,xmm3,xmm4,xmm5,xmm6,xmm7;


  if (bkline == 0)
  {

   xmm0 = _mm_xor_ps(xmm0,xmm0);

   for (i = 0; i < 32; i++)
    {
   _mm_stream_ps(ddscr+0, xmm0);
   _mm_stream_ps(ddscr+4, xmm0);
   _mm_stream_ps(ddscr+8, xmm0);
   _mm_stream_ps(ddscr+12,xmm0);
   _mm_stream_ps(ddscr+16,xmm0);
   _mm_stream_ps(ddscr+20,xmm0);
   _mm_stream_ps(ddscr+24,xmm0);
   _mm_stream_ps(ddscr+28,xmm0);

    ddscr += 32;
    }
    return;
  }

   for (i = 0; i < 32; i++) {

    acc = bkline[0];
    ptrl = &pix32tbl[acc*8];
    acc = bkline[1];
    ptrh = &pix32tbl[acc*8];
//
// загрузка
// 111111 222222 333333 444444
// 111111 222222 333333 444444
//
    xmm0 = _mm_load_ps(ptrl);
    xmm2 = _mm_load_ps(ptrl+4);
    xmm4 = _mm_load_ps(ptrh);
    xmm6 = _mm_load_ps(ptrh+4);

    xmm5  = xmm4;
    xmm7  = xmm6;
    xmm1  = xmm0;
    xmm3  = xmm2;
//
// преобразование
// 111111 111111 222222 222222
// 333333 333333 444444 444444
//
    xmm0 = _mm_shuffle_ps(xmm0,xmm0,0x50);
    xmm1 = _mm_shuffle_ps(xmm1,xmm1,0xfa);
    xmm2 = _mm_shuffle_ps(xmm2,xmm2,0x50);
    xmm3 = _mm_shuffle_ps(xmm3,xmm3,0xfa);
    xmm4 = _mm_shuffle_ps(xmm4,xmm4,0x50);
    xmm5 = _mm_shuffle_ps(xmm5,xmm5,0xfa);
    xmm6 = _mm_shuffle_ps(xmm6,xmm6,0x50);
    xmm7 = _mm_shuffle_ps(xmm7,xmm7,0xfa);
//
// вывод
//
   _mm_stream_ps(ddscr+0, xmm0);
   _mm_stream_ps(ddscr+4, xmm1);
   _mm_stream_ps(ddscr+8, xmm2);
   _mm_stream_ps(ddscr+12,xmm3);
   _mm_stream_ps(ddscr+16,xmm4);
   _mm_stream_ps(ddscr+20,xmm5);
   _mm_stream_ps(ddscr+24,xmm6);
   _mm_stream_ps(ddscr+28,xmm7);
//
    ddscr += 4*8;
    bkline+= 2;
  }
}

/*     */
void DrawScreen(int BKScr,int ScrollReg)
{
//   extern RECT WndRect;
   int *DDrawScr,y,ys;//,x;

// проверить доступность и восстановление экранов
  if ( IDirectDrawSurface_IsLost(lpDDSBack) != S_OK )
  {
    IDirectDrawSurface_Restore(lpDDSBack);
    IDirectDrawSurface_Restore(lpDDSPrimary);
  }

  if ( IDirectDrawSurface_Lock(lpDDSBack,NULL, &ddsd,
                               DDLOCK_SURFACEMEMORYPTR, 0) != S_OK) return;
       DDrawScr = (int*)ddsd.lpSurface;


       for (y = 0; y < 256; y++) {
            ys = (y-(ScrollReg-216)) & 0xFF ;

            _PsetL(DDrawScr+(ys*d_x),(unsigned char*)BKScr+y*64);
      // очистка неактивной части экрана в режиме "РП"
            if ( (y>63) && ((ScrollReg & 0x200) == 0) )
            _PsetL((int*)DDrawScr+(y*d_x),0);
       }

  if (IDirectDrawSurface_Unlock(lpDDSBack,NULL) != S_OK) return;
      //IDirectDraw_WaitForVerticalBlank(lpDD,1,0);
  if (IDirectDrawSurface_Blt(lpDDSPrimary, &WndRect,lpDDSBack,NULL, 0, NULL)
    == DDERR_SURFACELOST) IDirectDrawSurface_Restore(lpDDSPrimary);
}