r128_accel.c

x.org上有关ati系列显卡最新驱动

   x11perf -seg1              2960000.0/sec   2990000.0/sec
   x11perf -copyplane100         4400.0/sec      6700.0/sec
   x11perf -putimagexy100         138.0/sec       191.0/sec

*/
static void R128SetupForScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
							   int fg, int bg,
							   int rop,
							   unsigned int
							   planemask)
{
    R128InfoPtr   info      = R128PTR(pScrn);
    unsigned char *R128MMIO = info->MMIO;

    R128WaitForFifo(pScrn, 4);
#if X_BYTE_ORDER == X_LITTLE_ENDIAN
    OUTREG(R128_DP_GUI_MASTER_CNTL, (info->dp_gui_master_cntl
				     | R128_GMC_DST_CLIPPING
				     | R128_GMC_BRUSH_NONE
				     | (bg == -1
					? R128_GMC_SRC_DATATYPE_MONO_FG_LA
					: R128_GMC_SRC_DATATYPE_MONO_FG_BG)
				     | R128_ROP[rop].rop
				     | R128_GMC_BYTE_LSB_TO_MSB
				     | R128_DP_SRC_SOURCE_HOST_DATA));
#else	/* X_BYTE_ORDER == X_BIG_ENDIAN */
    OUTREG(R128_DP_GUI_MASTER_CNTL, (info->dp_gui_master_cntl
				     | R128_GMC_DST_CLIPPING
				     | R128_GMC_BRUSH_NONE
				     | (bg == -1
					? R128_GMC_SRC_DATATYPE_MONO_FG_LA
					: R128_GMC_SRC_DATATYPE_MONO_FG_BG)
				     | R128_ROP[rop].rop
				     | R128_DP_SRC_SOURCE_HOST_DATA));
#endif
    OUTREG(R128_DP_WRITE_MASK,      planemask);
    OUTREG(R128_DP_SRC_FRGD_CLR,    fg);
    OUTREG(R128_DP_SRC_BKGD_CLR,    bg);
}

/* Subsequent XAA indirect CPU-to-screen color expansion.  This is only
   called once for each rectangle. */
static void R128SubsequentScanlineCPUToScreenColorExpandFill(ScrnInfoPtr pScrn,
							     int x, int y,
							     int w, int h,
							     int skipleft)
{
    R128InfoPtr   info      = R128PTR(pScrn);
    unsigned char *R128MMIO = info->MMIO;
    int x1clip = x+skipleft;
    int x2clip = x+w;

    info->scanline_h      = h;
    info->scanline_words  = (w + 31) >> 5;

#if 0
    /* Seems as though the Rage128's doesn't like blitting directly
     * as we must be overwriting something too quickly, therefore we
     * render to the buffer first and then blit */
    if ((info->scanline_words * h) <= 9) {
	/* Turn on direct for less than 9 dword colour expansion */
	info->scratch_buffer[0]
	    = (unsigned char *)(ADDRREG(R128_HOST_DATA_LAST)
				- (info->scanline_words - 1));
	info->scanline_direct = 1;
    } else
#endif
    {
	/* Use indirect for anything else */
	info->scratch_buffer[0] = info->scratch_save;
	info->scanline_direct   = 0;
    }

    if (pScrn->bitsPerPixel == 24) {
	x1clip *= 3;
	x2clip *= 3;
    }

    R128WaitForFifo(pScrn, 4 + (info->scanline_direct ?
					(info->scanline_words * h) : 0) );
    OUTREG(R128_SC_TOP_LEFT,     (y << 16)       | (x1clip & 0xffff));
    OUTREG(R128_SC_BOTTOM_RIGHT, ((y+h-1) << 16) | ((x2clip-1) & 0xffff));
    OUTREG(R128_DST_Y_X,         (y << 16)       | (x & 0xffff));
    /* Have to pad the width here and use clipping engine */
    OUTREG(R128_DST_HEIGHT_WIDTH, (h << 16)      | ((w + 31) & ~31));
}

/* Subsequent XAA indirect CPU-to-screen color expansion.  This is called
   once for each scanline. */
static void R128SubsequentColorExpandScanline(ScrnInfoPtr pScrn, int bufno)
{
    R128InfoPtr     info      = R128PTR(pScrn);
    unsigned char   *R128MMIO = info->MMIO;
    CARD32          *p        = (pointer)info->scratch_buffer[bufno];
    int             i;
    int             left      = info->scanline_words;
    volatile CARD32 *d;

    if (info->scanline_direct) return;
    --info->scanline_h;
    while (left) {
        write_mem_barrier();
	if (left <= 8) {
	  /* Last scanline - finish write to DATA_LAST */
	  if (info->scanline_h == 0) {
	    R128WaitForFifo(pScrn, left);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA_LAST) - (left - 1); left; --left)
		*d++ = *p++;
	    return;
	  } else {
	    R128WaitForFifo(pScrn, left);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA7) - (left - 1); left; --left)
		*d++ = *p++;
	  }
	} else {
	    R128WaitForFifo(pScrn, 8);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA0), i = 0; i < 8; i++)
		*d++ = *p++;
	    left -= 8;
	}
    }
}

/* Setup for XAA indirect image write.

   1024x768@76Hz 8bpp
                             Without             With
   x11perf -putimage10   37500.0/sec      39300.0/sec
   x11perf -putimage100   2150.0/sec       1170.0/sec
   x11perf -putimage500    108.0/sec         49.8/sec
 */
static void R128SetupForScanlineImageWrite(ScrnInfoPtr pScrn,
					   int rop,
					   unsigned int planemask,
					   int trans_color,
					   int bpp,
					   int depth)
{
    R128InfoPtr   info      = R128PTR(pScrn);
    unsigned char *R128MMIO = info->MMIO;

    info->scanline_bpp = bpp;

    R128WaitForFifo(pScrn, 2);
    OUTREG(R128_DP_GUI_MASTER_CNTL, (info->dp_gui_master_cntl
				     | R128_GMC_DST_CLIPPING
				     | R128_GMC_BRUSH_1X8_COLOR
				     | R128_GMC_SRC_DATATYPE_COLOR
				     | R128_ROP[rop].rop
				     | R128_GMC_BYTE_LSB_TO_MSB
				     | R128_DP_SRC_SOURCE_HOST_DATA));
    OUTREG(R128_DP_WRITE_MASK,      planemask);

    if (trans_color != -1) {
				/* Set up for transparency */
	R128WaitForFifo(pScrn, 3);
	OUTREG(R128_CLR_CMP_CLR_SRC, trans_color);
	OUTREG(R128_CLR_CMP_MASK,    R128_CLR_CMP_MSK);
	OUTREG(R128_CLR_CMP_CNTL,    (R128_SRC_CMP_NEQ_COLOR
				      | R128_CLR_CMP_SRC_SOURCE));
    }
}

/* Subsequent XAA indirect image write. This is only called once for each
   rectangle. */
static void R128SubsequentScanlineImageWriteRect(ScrnInfoPtr pScrn,
						 int x, int y,
						 int w, int h,
						 int skipleft)
{
    R128InfoPtr   info      = R128PTR(pScrn);
    unsigned char *R128MMIO = info->MMIO;
    int x1clip = x+skipleft;
    int x2clip = x+w;

    int shift = 0; /* 32bpp */

    if (pScrn->bitsPerPixel == 8) shift = 3;
    else if (pScrn->bitsPerPixel == 16) shift = 1;

    info->scanline_h      = h;
    info->scanline_words  = (w * info->scanline_bpp + 31) >> 5;

#if 0
    /* Seeing as the CPUToScreen doesn't like this, I've done this
     * here too, as it uses pretty much the same path. */
    if ((info->scanline_words * h) <= 9) {
	/* Turn on direct for less than 9 dword colour expansion */
	info->scratch_buffer[0]
	    = (unsigned char *)(ADDRREG(R128_HOST_DATA_LAST)
				- (info->scanline_words - 1));
	info->scanline_direct = 1;
    } else
#endif
    {
	/* Use indirect for anything else */
	info->scratch_buffer[0] = info->scratch_save;
	info->scanline_direct   = 0;
    }

    if (pScrn->bitsPerPixel == 24) {
	x1clip *= 3;
	x2clip *= 3;
    }

    R128WaitForFifo(pScrn, 4 + (info->scanline_direct ?
					(info->scanline_words * h) : 0) );
    OUTREG(R128_SC_TOP_LEFT,      (y << 16)       | (x1clip & 0xffff));
    OUTREG(R128_SC_BOTTOM_RIGHT,  ((y+h-1) << 16) | ((x2clip-1) & 0xffff));
    OUTREG(R128_DST_Y_X,          (y << 16)       | (x & 0xffff));
    /* Have to pad the width here and use clipping engine */
    OUTREG(R128_DST_HEIGHT_WIDTH, (h << 16)       | ((w + shift) & ~shift));
}

/* Subsequent XAA indirect iamge write.  This is called once for each
   scanline. */
static void R128SubsequentImageWriteScanline(ScrnInfoPtr pScrn, int bufno)
{
    R128InfoPtr     info      = R128PTR(pScrn);
    unsigned char   *R128MMIO = info->MMIO;
    CARD32          *p        = (pointer)info->scratch_buffer[bufno];
    int             i;
    int             left      = info->scanline_words;
    volatile CARD32 *d;

    if (info->scanline_direct) return;
    --info->scanline_h;
    while (left) {
        write_mem_barrier();
	if (left <= 8) {
	  /* Last scanline - finish write to DATA_LAST */
	  if (info->scanline_h == 0) {
	    R128WaitForFifo(pScrn, left);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA_LAST) - (left - 1); left; --left)
		*d++ = *p++;
	    return;
	  } else {
	    R128WaitForFifo(pScrn, left);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA7) - (left - 1); left; --left)
		*d++ = *p++;
	  }
	} else {
	    R128WaitForFifo(pScrn, 8);
				/* Unrolling doesn't improve performance */
	    for (d = ADDRREG(R128_HOST_DATA0), i = 0; i < 8; i++)
		*d++ = *p++;
	    left -= 8;
	}
    }
}

/* Initialize the acceleration hardware. */
void R128EngineInit(ScrnInfoPtr pScrn)
{
    R128InfoPtr   info      = R128PTR(pScrn);
    unsigned char *R128MMIO = info->MMIO;

    R128TRACE(("EngineInit (%d/%d)\n", info->CurrentLayout.pixel_code, info->CurrentLayout.bitsPerPixel));

    OUTREG(R128_SCALE_3D_CNTL, 0);
    R128EngineReset(pScrn);

    switch (info->CurrentLayout.pixel_code) {
    case 8:  info->datatype = 2; break;
    case 15: info->datatype = 3; break;
    case 16: info->datatype = 4; break;
    case 24: info->datatype = 5; break;
    case 32: info->datatype = 6; break;
    default:
	R128TRACE(("Unknown depth/bpp = %d/%d (code = %d)\n",
		   info->CurrentLayout.depth, info->CurrentLayout.bitsPerPixel,
		   info->CurrentLayout.pixel_code));
    }
    info->pitch = (info->CurrentLayout.displayWidth / 8) * (info->CurrentLayout.pixel_bytes == 3 ? 3 : 1);

    R128TRACE(("Pitch for acceleration = %d\n", info->pitch));

    R128WaitForFifo(pScrn, 2);
    OUTREG(R128_DEFAULT_OFFSET, pScrn->fbOffset);
    OUTREG(R128_DEFAULT_PITCH,  info->pitch);

    R128WaitForFifo(pScrn, 4);
    OUTREG(R128_AUX_SC_CNTL,             0);
    OUTREG(R128_DEFAULT_SC_BOTTOM_RIGHT, (R128_DEFAULT_SC_RIGHT_MAX
					  | R128_DEFAULT_SC_BOTTOM_MAX));
    OUTREG(R128_SC_TOP_LEFT,             0);
    OUTREG(R128_SC_BOTTOM_RIGHT,         (R128_DEFAULT_SC_RIGHT_MAX
					  | R128_DEFAULT_SC_BOTTOM_MAX));

    info->dp_gui_master_cntl = ((info->datatype << R128_GMC_DST_DATATYPE_SHIFT)
				| R128_GMC_CLR_CMP_CNTL_DIS
				| R128_GMC_AUX_CLIP_DIS);
    R128WaitForFifo(pScrn, 1);
    OUTREG(R128_DP_GUI_MASTER_CNTL, (info->dp_gui_master_cntl
				     | R128_GMC_BRUSH_SOLID_COLOR
				     | R128_GMC_SRC_DATATYPE_COLOR));

    R128WaitForFifo(pScrn, 8);
    OUTREG(R128_DST_BRES_ERR,      0);
    OUTREG(R128_DST_BRES_INC,      0);
    OUTREG(R128_DST_BRES_DEC,      0);
    OUTREG(R128_DP_BRUSH_FRGD_CLR, 0xffffffff);
    OUTREG(R128_DP_BRUSH_BKGD_CLR, 0x00000000);
    OUTREG(R128_DP_SRC_FRGD_CLR,   0xffffffff);
    OUTREG(R128_DP_SRC_BKGD_CLR,   0x00000000);
    OUTREG(R128_DP_WRITE_MASK,     0xffffffff);

    R128WaitForFifo(pScrn, 1);

#if X_BYTE_ORDER == X_BIG_ENDIAN
    /* FIXME: this is a kludge for texture uploads in the 3D driver. Look at
     * how the radeon driver handles HOST_DATA_SWAP if you want to implement
     * CCE ImageWrite acceleration or anything needing this bit */
#ifdef XF86DRI
    if (info->directRenderingEnabled)
	OUTREGP(R128_DP_DATATYPE, 0, ~R128_HOST_BIG_ENDIAN_EN);
    else
#endif
	OUTREGP(R128_DP_DATATYPE,
		R128_HOST_BIG_ENDIAN_EN, ~R128_HOST_BIG_ENDIAN_EN);
#else /* X_LITTLE_ENDIAN */
    OUTREGP(R128_DP_DATATYPE, 0, ~R128_HOST_BIG_ENDIAN_EN);
#endif

#ifdef XF86DRI
    info->sc_left         = 0x00000000;
    info->sc_right        = R128_DEFAULT_SC_RIGHT_MAX;
    info->sc_top          = 0x00000000;
    info->sc_bottom       = R128_DEFAULT_SC_BOTTOM_MAX;

    info->re_top_left     = 0x00000000;
    info->re_width_height = ((0x7ff << R128_RE_WIDTH_SHIFT) |
			     (0x7ff << R128_RE_HEIGHT_SHIFT));

    info->aux_sc_cntl     = 0x00000000;
#endif

    R128WaitForIdle(pScrn);
}

#ifdef XF86DRI

/* Setup for XAA SolidFill. */
static void R128CCESetupForSolidFill(ScrnInfoPtr pScrn,
				     int color, int rop,
				     unsigned int planemask)
{
    R128InfoPtr   info = R128PTR(pScrn);
    RING_LOCALS;

    R128CCE_REFRESH( pScrn, info );

    BEGIN_RING( 8 );

    OUT_RING_REG( R128_DP_GUI_MASTER_CNTL,
		  (info->dp_gui_master_cntl
		   | R128_GMC_BRUSH_SOLID_COLOR
		   | R128_GMC_SRC_DATATYPE_COLOR
		   | R128_ROP[rop].pattern) );

    OUT_RING_REG( R128_DP_BRUSH_FRGD_CLR,  color );
    OUT_RING_REG( R128_DP_WRITE_MASK,	   planemask );
    OUT_RING_REG( R128_DP_CNTL,		   (R128_DST_X_LEFT_TO_RIGHT |
					    R128_DST_Y_TOP_TO_BOTTOM));
    ADVANCE_RING();
}

/* Subsequent XAA SolidFillRect.

   Tests: xtest CH06/fllrctngl, xterm
*/
static void R128CCESubsequentSolidFillRect(ScrnInfoPtr pScrn,