Simplfy clear() and swapbuffers() code.

Fix various mishandling of cliprects.
Allow multiple primitives to be emitted to a single dma buffer, which
was largely impossible previously.
Re-enable the fast unclipped render stage.

1 files changed, 324 insertions, 541 deletions

diff --git a/src/mesa/drivers/dri/unichrome/via_ioctl.c b/src/mesa/drivers/dri/unichrome/via_ioctl.c
index 431a26b2a9..55220f18f6 100644
--- a/src/mesa/drivers/dri/unichrome/via_ioctl.c
+++ b/src/mesa/drivers/dri/unichrome/via_ioctl.c
@@ -73,14 +73,127 @@ v * copy of this software and associated documentation files (the "Software"),
 #define VIA_BLIT_SET 0xFF
 
  
-#define DEPTH_SCALE ((1 << 16) - 1)
-
 void viaCheckDma(viaContextPtr vmesa, GLuint bytes)
 {
     VIA_FINISH_PRIM( vmesa );
     if (vmesa->dmaLow + bytes > VIA_DMA_HIGHWATER) {
-	viaFlushPrims(vmesa);
+	viaFlushDma(vmesa);
+    }
+}
+
+
+
+#define SetReg2DAGP(nReg, nData) do {		\
+    OUT_RING( ((nReg) >> 2) | 0xF0000000 );	\
+    OUT_RING( nData );				\
+} while (0)
+
+
+static void viaBlit(viaContextPtr vmesa, GLuint bpp,GLuint srcBase,
+		    GLuint srcPitch,GLuint dstBase,GLuint dstPitch,
+		    GLuint w,GLuint h,int xdir,int ydir, GLuint blitMode, 
+		    GLuint color, GLuint nMask ) 
+{
+
+    GLuint dwGEMode = 0, srcY=0, srcX, dstY=0, dstX;
+    GLuint cmd;
+    RING_VARS;
+
+    if (VIA_DEBUG)
+       fprintf(stderr, "%s bpp %d src %x/%x dst %x/%x w %d h %d dir %d,%d mode: %x color: 0x%08x mask 0x%08x\n",
+	       __FUNCTION__, bpp, srcBase, srcPitch, dstBase, dstPitch, w,h, xdir, ydir, blitMode, color, nMask);
+
+
+    if (!w || !h)
+        return;
+
+    srcX = srcBase & 31;
+    dstX = dstBase & 31;
+    switch (bpp) {
+    case 16:
+        dwGEMode |= VIA_GEM_16bpp;
+	srcX >>= 1;
+	dstX >>= 1;
+        break;
+    case 32:
+        dwGEMode |= VIA_GEM_32bpp;
+	srcX >>= 2;
+	dstX >>= 2;
+	break;
+    default:
+        dwGEMode |= VIA_GEM_8bpp;
+        break;
     }
+
+
+    cmd = 0; 
+
+    if (xdir < 0) {
+        cmd |= VIA_GEC_DECX;
+        srcX += (w - 1);
+        dstX += (w - 1);
+    }
+    if (ydir < 0) {
+        cmd |= VIA_GEC_DECY;
+        srcY += (h - 1);
+        dstY += (h - 1);
+    }
+
+    switch(blitMode) {
+    case VIA_BLIT_FILL:
+        BEGIN_RING((2 + 9) * 2);
+	SetReg2DAGP(VIA_REG_GEMODE, dwGEMode);
+	SetReg2DAGP( VIA_REG_FGCOLOR, color);
+	cmd |= VIA_GEC_BLT | VIA_GEC_FIXCOLOR_PAT | (VIA_BLIT_FILL << 24);
+	break;
+    case VIA_BLIT_COPY:
+        BEGIN_RING((2 + 9) * 2);
+	SetReg2DAGP(VIA_REG_GEMODE, dwGEMode);
+	SetReg2DAGP( VIA_REG_KEYCONTROL, 0x0);
+	cmd |= VIA_GEC_BLT | (VIA_BLIT_COPY << 24);
+    }	
+
+    SetReg2DAGP( 0x2C, nMask);
+    SetReg2DAGP( VIA_REG_SRCBASE, (srcBase & ~31) >> 3);
+    SetReg2DAGP( VIA_REG_DSTBASE, (dstBase & ~31) >> 3);
+    SetReg2DAGP( VIA_REG_PITCH, VIA_PITCH_ENABLE |
+	       (srcPitch >> 3) | (((dstPitch) >> 3) << 16));
+    SetReg2DAGP( VIA_REG_SRCPOS, ((srcY << 16) | srcX));
+    SetReg2DAGP( VIA_REG_DSTPOS, ((dstY << 16) | dstX));
+    SetReg2DAGP( VIA_REG_DIMENSION, (((h - 1) << 16) | (w - 1)));
+    SetReg2DAGP( VIA_REG_GECMD, cmd);
+    SetReg2DAGP( 0x2C, 0x00000000);
+    ADVANCE_RING();
+}
+
+static void viaFillBuffer(viaContextPtr vmesa,
+			  viaBuffer *buffer,
+			  drm_clip_rect_t *pbox,
+			  int nboxes,
+			  GLuint pixel,
+			  GLuint mask)
+{
+   GLuint bytePerPixel = buffer->bpp >> 3;
+   GLuint i;
+
+   for (i = 0; i < nboxes ; i++) {        
+      int x = pbox[i].x1 - vmesa->drawX;
+      int y = pbox[i].y1 - vmesa->drawY;
+      int w = pbox[i].x2 - pbox[i].x1;
+      int h = pbox[i].y2 - pbox[i].y1;
+
+      int offset = (buffer->orig + 
+		    y * buffer->pitch + 
+		    x * bytePerPixel);
+
+      viaBlit(vmesa,
+	      buffer->bpp, 
+	      offset, buffer->pitch,
+	      offset, buffer->pitch, 
+	      w, h,
+	      0, 0, 
+	      VIA_BLIT_FILL, pixel, mask); 
+   }
 }
 
 
@@ -133,73 +246,107 @@ static void viaClear(GLcontext *ctx, GLbitfield mask, GLboolean all,
    }
     
    if (flag) {
+      drm_clip_rect_t *boxes, *tmp_boxes = 0;
+      int nr = 0;
+
       LOCK_HARDWARE(vmesa);
+	    
       /* flip top to bottom */
       cy = dPriv->h - cy - ch;
-      cx += vmesa->drawX;
+      cx += vmesa->drawX + vmesa->drawXoff;
       cy += vmesa->drawY;
         
-      if (vmesa->numClipRects) {
-	 int nr = MIN2(i + VIA_NR_SAREA_CLIPRECTS, vmesa->numClipRects);
-	 drm_clip_rect_t *box = vmesa->pClipRects;
-	 drm_clip_rect_t *b = vmesa->sarea->boxes;
-	 int n = 0;
-	    
-	 if (!all) {
-	    for (; i < nr; i++) {
-	       GLint x = box[i].x1;
-	       GLint y = box[i].y1;
-	       GLint w = box[i].x2 - x;
-	       GLint h = box[i].y2 - y;
-
-	       if (x < cx) w -= cx - x, x = cx;
-	       if (y < cy) h -= cy - y, y = cy;
-	       if (x + w > cx + cw) w = cx + cw - x;
-	       if (y + h > cy + ch) h = cy + ch - y;
-	       if (w <= 0) continue;
-	       if (h <= 0) continue;
-
-	       b->x1 = x;
-	       b->y1 = y;
-	       b->x2 = x + w;
-	       b->y2 = y + h;
-	       b++;
-	       n++;
-	    }
-	 }
-	 else {
-	    for (; i < nr; i++) {
-	       *b++ = *(drm_clip_rect_t *)&box[i];
-	       n++;
-	    }
+      if (!all) {
+	 drm_clip_rect_t *b = vmesa->pClipRects;	 
+	 
+	 boxes = tmp_boxes = (drm_clip_rect_t *)malloc(vmesa->numClipRects * 
+						       sizeof(drm_clip_rect_t)); 
+	 if (!boxes) {
+	    UNLOCK_HARDWARE(vmesa);
+	    return;
 	 }
-	 vmesa->sarea->nbox = n;		
 
+	 for (; i < vmesa->numClipRects; i++) {
+	    GLint x = b[i].x1;
+	    GLint y = b[i].y1;
+	    GLint w = b[i].x2 - x;
+	    GLint h = b[i].y2 - y;
+
+	    if (x < cx) w -= cx - x, x = cx;
+	    if (y < cy) h -= cy - y, y = cy;
+	    if (x + w > cx + cw) w = cx + cw - x;
+	    if (y + h > cy + ch) h = cy + ch - y;
+	    if (w <= 0) continue;
+	    if (h <= 0) continue;
+
+	    boxes[nr].x1 = x;
+	    boxes[nr].y1 = y;
+	    boxes[nr].x2 = x + w;
+	    boxes[nr].y2 = y + h;
+	    nr++;
+	 }
+      }
+      else {
+	 boxes = vmesa->pClipRects;
+	 nr = vmesa->numClipRects;
       }
 	    
       if (flag & VIA_FRONT) {
-	 if (vmesa->drawType == GLX_PBUFFER_BIT)
-	    viaFillFrontPBuffer(vmesa);
-	 else
-	    viaFillFrontBuffer(vmesa);
+	 viaFillBuffer(vmesa, &vmesa->front, boxes, nr, vmesa->ClearColor, 0);
       } 
 		
       if (flag & VIA_BACK) {
-	 viaFillBackBuffer(vmesa);
+	 viaFillBuffer(vmesa, &vmesa->back, boxes, nr, vmesa->ClearColor, 0); /* FIXME: masks */
       }
 
       if (flag & VIA_DEPTH) {
-	 viaFillDepthBuffer(vmesa, clear_depth, clear_depth_mask);
+	 viaFillBuffer(vmesa, &vmesa->depth, boxes, nr, clear_depth, clear_depth_mask);
       }		
 
+      viaFlushDmaLocked(vmesa, VIA_NO_CLIPRECTS);
       UNLOCK_HARDWARE(vmesa);
+
+      if (tmp_boxes)
+	 free(tmp_boxes);
    }
    
    if (mask)
       _swrast_Clear(ctx, mask, all, cx, cy, cw, ch);
-   if (VIA_DEBUG) fprintf(stderr, "%s out\n", __FUNCTION__);    
 }
 
+
+
+
+static void viaDoSwapBuffers(viaContextPtr vmesa,
+			     drm_clip_rect_t *b,
+			     GLuint nbox)
+{    
+   GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
+   viaBuffer *front = &vmesa->front;
+   viaBuffer *back = &vmesa->back;
+   GLuint i;
+        
+   for (i = 0; i < nbox; i++, b++) {        
+      GLint x = b->x1 - vmesa->drawX;
+      GLint y = b->y1 - vmesa->drawY;
+      GLint w = b->x2 - b->x1;
+      GLint h = b->y2 - b->y1;
+	
+      GLuint src = back->orig + y * back->pitch + x * bytePerPixel;
+      GLuint dest = front->orig + y * front->pitch + x * bytePerPixel;
+
+      viaBlit(vmesa, 
+	      bytePerPixel << 3, 
+	      src, back->pitch,
+	      dest, front->pitch,
+	      w, h,
+	      0,0,VIA_BLIT_COPY, 0, 0); 
+   }
+}
+
+
+
+
 /*
  * Copy the back buffer to the front buffer. 
  */
@@ -207,29 +354,21 @@ void viaCopyBuffer(const __DRIdrawablePrivate *dPriv)
 {
    viaContextPtr vmesa = (viaContextPtr)dPriv->driContextPriv->driverPrivate;
    GLboolean missed_target;
-   int64_t ust;
 
    VIA_FLUSH_DMA(vmesa);
-
    driWaitForVBlank( dPriv, & vmesa->vbl_seq, vmesa->vblank_flags, & missed_target );
-
-   LOCK_HARDWARE(vmesa);
-   if (vmesa->drawType == GLX_PBUFFER_BIT) {
-      viaDoSwapPBuffers(vmesa);
-   }
-   else {
-      viaDoSwapBuffers(vmesa);
-   }
-   UNLOCK_HARDWARE(vmesa);
-
-   vmesa->swap_count++;
-   (*vmesa->get_ust)( & ust );
    if ( missed_target ) {
       vmesa->swap_missed_count++;
-      vmesa->swap_missed_ust = ust - vmesa->swap_ust;
+      vmesa->get_ust( &vmesa->swap_missed_ust );
    }
-    
-   vmesa->swap_ust = ust;
+   LOCK_HARDWARE(vmesa);
+
+   viaDoSwapBuffers(vmesa, dPriv->pClipRects, dPriv->numClipRects);
+   viaFlushDmaLocked(vmesa, VIA_NO_CLIPRECTS);
+
+   UNLOCK_HARDWARE(vmesa);
+   vmesa->swap_count++;
+   vmesa->get_ust( &vmesa->swap_ust );
 }
 
 /*
@@ -237,31 +376,17 @@ void viaCopyBuffer(const __DRIdrawablePrivate *dPriv)
  */
 void viaPageFlip(const __DRIdrawablePrivate *dPriv)
 {
-    /*=* John Sheng [2003.5.31] flip *=*/
     viaContextPtr vmesa = (viaContextPtr)dPriv->driContextPriv->driverPrivate;
-    GLcontext *ctx = vmesa->glCtx;
-    GLuint nBackBase;
     viaBuffer buffer_tmp;
     GLboolean missed_target;
-    int retcode;
-
-    if (VIA_DEBUG) fprintf(stderr, "%s - in\n", __FUNCTION__);        
-    assert(dPriv);
-    assert(dPriv->driContextPriv);
-    assert(dPriv->driContextPriv->driverPrivate);
 
 
     VIA_FLUSH_DMA(vmesa);
-
-    /* Now wait for the vblank:
-     */
-    retcode = driWaitForVBlank( dPriv, &vmesa->vbl_seq, 
-				vmesa->vblank_flags, &missed_target );
+    driWaitForVBlank( dPriv, &vmesa->vbl_seq, vmesa->vblank_flags, &missed_target );
     if ( missed_target ) {
        vmesa->swap_missed_count++;
-       (void) (*vmesa->get_ust)( &vmesa->swap_missed_ust );
+       vmesa->get_ust( &vmesa->swap_missed_ust );
     }
-
     LOCK_HARDWARE(vmesa);
 
     {
@@ -276,49 +401,36 @@ void viaPageFlip(const __DRIdrawablePrivate *dPriv)
     	    OUT_RING(0x0000000e);
 	    ADVANCE_RING();
 	}
-	nBackBase = (vmesa->back.offset );
 
 	BEGIN_RING(4);
 	OUT_RING( HALCYON_HEADER2 );
 	OUT_RING( 0x00fe0000 );
-	OUT_RING((HC_SubA_HFBBasL << 24) | (nBackBase & 0xFFFFF8) | 0x2);
+	OUT_RING((HC_SubA_HFBBasL << 24) | (vmesa->back.offset & 0xFFFFF8) | 0x2);
 	OUT_RING((HC_SubA_HFBDrawFirst << 24) |
-		 ((nBackBase & 0xFF000000) >> 24) | 0x0100);
+		 ((vmesa->back.offset & 0xFF000000) >> 24) | 0x0100);
 	ADVANCE_RING();
 
-	viaFlushPrimsLocked(vmesa);
     }
-    
-    UNLOCK_HARDWARE(vmesa);
 
+    viaFlushDmaLocked(vmesa, VIA_NO_CLIPRECTS);    
+    UNLOCK_HARDWARE(vmesa);
     vmesa->swap_count++;
-    (void) (*vmesa->get_ust)( &vmesa->swap_ust );
-   
+    vmesa->get_ust( &vmesa->swap_ust );
+
+    /* KW: FIXME: When buffers are freed, could free frontbuffer by
+     * accident:
+     */
     memcpy(&buffer_tmp, &vmesa->back, sizeof(viaBuffer));
     memcpy(&vmesa->back, &vmesa->front, sizeof(viaBuffer));
     memcpy(&vmesa->front, &buffer_tmp, sizeof(viaBuffer));
     
-    /* KW: BOGUS BOGUS BOGUS: The first time an app calls glDrawBuffer
-     * while pageflipping, this will blow up:  FIXME
-     */
     if(vmesa->currentPage) {
 	vmesa->currentPage = 0;
-	if (vmesa->glCtx->Color._DrawDestMask[0] == DD_BACK_LEFT_BIT) {
-		ctx->Driver.DrawBuffer(ctx, GL_BACK);
-	}
-	else {
-		ctx->Driver.DrawBuffer(ctx, GL_FRONT);
-	}
     }
     else {
 	vmesa->currentPage = 1;
-	if (vmesa->glCtx->Color._DrawDestMask[0] == DD_BACK_LEFT_BIT) {
-		ctx->Driver.DrawBuffer(ctx, GL_BACK);
-	}
-	else {
-		ctx->Driver.DrawBuffer(ctx, GL_FRONT);
-	}
     }
+
     if (VIA_DEBUG) fprintf(stderr, "%s - out\n", __FUNCTION__);        
 }
 
@@ -327,13 +439,13 @@ void viaPageFlip(const __DRIdrawablePrivate *dPriv)
 
 #define VIA_CMDBUF_MAX_LAG 50000
 
-static int fire_buffer(viaContextPtr vmesa, drm_via_flush_sys_t *buf)
+static int fire_buffer(viaContextPtr vmesa)
 {
    drmVIACommandBuffer bufI;
    int ret;
 
-   bufI.buf = (char *) (buf->index + buf->offset);
-   bufI.size = buf->size;
+   bufI.buf = (char *)vmesa->dma;
+   bufI.size = vmesa->dmaLow;
 
    if (vmesa->useAgp) {
       drmVIACmdBufSize bSiz;
@@ -387,74 +499,40 @@ static int fire_buffer(viaContextPtr vmesa, drm_via_flush_sys_t *buf)
  * into the head of the DMA buffer being flushed.  Fires the buffer
  * for each cliprect.
  */
-static int via_flush_sys(viaContextPtr vmesa, drm_via_flush_sys_t* buf) 
+static void via_emit_cliprect(viaContextPtr vmesa,
+			      drm_clip_rect_t *b) 
 {
-   GLuint *vb = (GLuint *)vmesa->dmaAddr;
+   viaBuffer *buffer = vmesa->drawBuffer;
+   GLuint *vb = (GLuint *)(vmesa->dma + vmesa->dmaCliprectAddr);
+
    GLuint format = (vmesa->viaScreen->bitsPerPixel == 0x20 
 		    ? HC_HDBFM_ARGB8888 
 		    : HC_HDBFM_RGB565);
 
-   if (vmesa->glCtx->Color._DrawDestMask[0] == DD_BACK_LEFT_BIT) {
-      GLuint offset = vmesa->back.offset;
-      GLuint pitch = vmesa->back.pitch;
+   GLuint pitch = buffer->pitch;
+   GLuint offset = buffer->orig;
 
-      vb[0] = HC_HEADER2;
-      vb[1] = HC_ParaType_NotTex << 16;
-	  
-      if (vmesa->driDrawable->w == 0 || vmesa->driDrawable->h == 0) {
-	 vb[2] = (HC_SubA_HClipTB << 24) | 0x0;
-	 vb[3] = (HC_SubA_HClipLR << 24) | 0x0;
-      }
-      else {
-	 vb[2] = (HC_SubA_HClipTB << 24) | vmesa->driDrawable->h;
-	 vb[3] = (HC_SubA_HClipLR << 24) | vmesa->driDrawable->w; 
-      }
-	
-      vb[4] = ((HC_SubA_HDBBasL << 24) | (offset & 0xFFFFFF));
-      vb[5] = ((HC_SubA_HDBBasH << 24) | ((offset & 0xFF000000)) >> 24); 
-      vb[6] = ((HC_SubA_HDBFM << 24) | HC_HDBLoc_Local | format | pitch); 
-      vb[7] = 0xcccccccc;
+   GLuint clipL = b->x1 - vmesa->drawX;
+   GLuint clipR = b->x2 - vmesa->drawX;
+   GLuint clipT = b->y1 - vmesa->drawY;
+   GLuint clipB = b->y2 - vmesa->drawY;
+	    
+   vb[0] = HC_HEADER2;
+   vb[1] = (HC_ParaType_NotTex << 16);
 
-      return fire_buffer( vmesa, buf );
+   if (vmesa->driDrawable->w == 0 || vmesa->driDrawable->h == 0) {
+      vb[2] = (HC_SubA_HClipTB << 24) | 0x0;
+      vb[3] = (HC_SubA_HClipLR << 24) | 0x0;
    }
    else {
-      GLuint i, ret;
-      drm_clip_rect_t *b = vmesa->sarea->boxes;
-		
-      for (i = 0; i < vmesa->sarea->nbox; i++, b++) {        
-	 GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
-	 GLuint pitch = vmesa->front.pitch;
-	 GLuint offset = (vmesa->viaScreen->fbOffset + (vmesa->drawY * pitch + vmesa->drawX * bytePerPixel)) & ~0x1f;
-
-	 GLuint clipL = b->x1 + vmesa->drawXoff;
-	 GLuint clipR = b->x2 + vmesa->drawXoff;
-	 GLuint clipT = b->y1;
-	 GLuint clipB = b->y2;
-	    
-	 vb[0] = HC_HEADER2;
-	 vb[1] = (HC_ParaType_NotTex << 16);
-
-	 if (vmesa->driDrawable->w == 0 || vmesa->driDrawable->h == 0) {
-	    vb[2] = (HC_SubA_HClipTB << 24) | 0x0;
-	    vb[3] = (HC_SubA_HClipLR << 24) | 0x0;
-	 }
-	 else {
-	    vb[2] = (HC_SubA_HClipTB << 24) | (clipT << 12) | clipB;
-	    vb[3] = (HC_SubA_HClipLR << 24) | (clipL << 12) | clipR;
-	 }
-	    
-	 vb[4] = ((HC_SubA_HDBBasL << 24) | (offset & 0xFFFFFF));
-	 vb[5] = ((HC_SubA_HDBBasH << 24) | ((offset & 0xFF000000)) >> 24); 
-	 vb[6] = ((HC_SubA_HDBFM << 24) | HC_HDBLoc_Local | format | pitch);            
-	 vb[7] = 0xcccccccc;
-
-	 ret = fire_buffer( vmesa, buf );
-	 if (ret)
-	    return ret;
-      }
+      vb[2] = (HC_SubA_HClipTB << 24) | (clipT << 12) | clipB;
+      vb[3] = (HC_SubA_HClipLR << 24) | (clipL << 12) | clipR;
    }
-
-   return 0;
+	    
+   vb[4] = ((HC_SubA_HDBBasL << 24) | (offset & 0xFFFFFF));
+   vb[5] = ((HC_SubA_HDBBasH << 24) | ((offset & 0xFF000000)) >> 24); 
+   vb[6] = ((HC_SubA_HDBFM << 24) | HC_HDBLoc_Local | format | pitch);            
+   vb[7] = 0xcccccccc;
 }
 
 
@@ -476,12 +554,23 @@ static int intersect_rect(drm_clip_rect_t *out,
     return 1;
 }
 
-void viaFlushPrimsLocked(viaContextPtr vmesa)
+static void dump_dma( viaContextPtr vmesa )
+{
+   GLuint i;
+   GLuint *data = (GLuint *)vmesa->dma;
+   for (i = 0; i < vmesa->dmaLow; i += 16) {
+      fprintf(stderr, "%04x:   ", i);
+      fprintf(stderr, "%08x  ", *data++);
+      fprintf(stderr, "%08x  ", *data++);
+      fprintf(stderr, "%08x  ", *data++);
+      fprintf(stderr, "%08x\n", *data++);
+   }
+   fprintf(stderr, "******************************************\n");
+}
+
+
+void viaFlushDmaLocked(viaContextPtr vmesa, GLuint flags)
 {
-   drm_clip_rect_t *pbox = (drm_clip_rect_t *)vmesa->pClipRects;
-   int nbox = vmesa->numClipRects;
-   drm_via_sarea_t *sarea = vmesa->sarea;
-   drm_via_flush_sys_t sysCmd;
    int i;
    RING_VARS;
 
@@ -491,7 +580,7 @@ void viaFlushPrimsLocked(viaContextPtr vmesa)
       abort();
    }
 
-   if (vmesa->dmaLow == DMA_OFFSET) {
+   if (vmesa->dmaLow == 0) {
       return;
    }
 
@@ -543,94 +632,80 @@ void viaFlushPrimsLocked(viaContextPtr vmesa)
       if (VIA_DEBUG)
 	 fprintf(stderr, "%s: unaligned value for vmesa->dmaLow: %x\n",
 		 __FUNCTION__, vmesa->dmaLow);
-/*       abort(); */
-/*       goto done; */
    }
 
-/*    assert((vmesa->dmaLow & 0x1f) == 0); */
 
-   sysCmd.offset = 0x0;
-   sysCmd.size = vmesa->dmaLow;
-   sysCmd.index = (GLuint)vmesa->dma;
-   sysCmd.discard = 0;
-    
-   if (!nbox) {
-      sysCmd.size = 0;		/* KW: FIXME bogus if we ever start emitting partial state */
-      sarea->nbox = 0;
-      sysCmd.discard = 1;
-      via_flush_sys(vmesa, &sysCmd);
-   }
-   else {
-      for (i = 0; i < nbox; ) {
-	 int nr = MIN2(i + VIA_NR_SAREA_CLIPRECTS, nbox);
-	 drm_clip_rect_t *b = sarea->boxes;
+   if (VIA_DEBUG)
+      dump_dma( vmesa );
 
+   if (flags & VIA_NO_CLIPRECTS) {
+      assert(vmesa->dmaCliprectAddr == 0);
+      fire_buffer( vmesa );
+   }
+   else if (!vmesa->dmaCliprectAddr) {
+      /* Contains only state.  Could just dump the packet?
+       */
+      if (0) fprintf(stderr, "no dmaCliprectAddr\n");
+      if (0) fire_buffer( vmesa );
+   }
+   else if (vmesa->numClipRects) {
+      int ret;
+      drm_clip_rect_t *pbox = vmesa->pClipRects;
+      
+      for (i = 0; i < vmesa->numClipRects; i++) {
 	 if (vmesa->glCtx->Scissor.Enabled) {
-	    sarea->nbox = 0;
-
-	    for (; i < nr; i++) {
-	       b->x1 = pbox[i].x1 - vmesa->drawX;
-	       b->y1 = pbox[i].y1 - vmesa->drawY;
-	       b->x2 = pbox[i].x2 - vmesa->drawX;
-	       b->y2 = pbox[i].y2 - vmesa->drawY;
-	       if (intersect_rect(b, b, &vmesa->scissorRect)) {
-		  sarea->nbox++;
-		  b++;
-	       }
-	    }
-	    if (!sarea->nbox) {
-	       if (nr < nbox) continue;
-	       sysCmd.size = 0;
-	    }
+	    drm_clip_rect_t b;
+	    if (!intersect_rect(&b, &pbox[i], &vmesa->scissorRect)) 
+	       continue;
+	    via_emit_cliprect(vmesa, &b);	 
 	 }
 	 else {
-	    sarea->nbox = nr - i;
-	    for (; i < nr; i++, b++) {
-	       b->x1 = pbox[i].x1 - vmesa->drawX;
-	       b->y1 = pbox[i].y1 - vmesa->drawY;
-	       b->x2 = pbox[i].x2 - vmesa->drawX;
-	       b->y2 = pbox[i].y2 - vmesa->drawY;
-	    }
+	    via_emit_cliprect(vmesa, &pbox[i]);
 	 }
-
-	 if (nr == nbox) {
-	    sysCmd.discard = 1;
-	 }
-
-	 via_flush_sys(vmesa, &sysCmd);
+	 
+	 ret = fire_buffer(vmesa);
+	 if (ret)
+	    goto done;
       }
+   } else {
+      UNLOCK_HARDWARE(vmesa);
+      sched_yield();
+      LOCK_HARDWARE(vmesa);
    }
 
-   if (VIA_DEBUG) {
-      GLuint i;
-      GLuint *data = (GLuint *)vmesa->dmaAddr;
-      for (i = 0; i < vmesa->dmaLow; i += 16) {
-	 fprintf(stderr, "%04x:   ", i);
-	 fprintf(stderr, "%08x  ", *data++);
-	 fprintf(stderr, "%08x  ", *data++);
-	 fprintf(stderr, "%08x  ", *data++);
-	 fprintf(stderr, "%08x\n", *data++);
-      }
-      fprintf(stderr, "******************************************\n");
-   }  
-
  done:
    /* Reset vmesa vars:
     */
-   vmesa->dmaLow = DMA_OFFSET;
-   vmesa->dmaAddr = (unsigned char *)vmesa->dma;
+   vmesa->dmaLow = 0;
+   vmesa->dmaCliprectAddr = 0;
+   viaValidateState(vmesa->glCtx);
 }
 
-void viaFlushPrims(viaContextPtr vmesa)
+void viaWrapPrimitive( viaContextPtr vmesa )
 {
-   if (VIA_DEBUG) fprintf(stderr, "%s in\n", __FUNCTION__);
+   if (VIA_DEBUG) fprintf(stderr, "%s\n", __FUNCTION__);
+   viaFinishPrimitive( vmesa );
+   
+   LOCK_HARDWARE(vmesa); 
+   viaFlushDmaLocked(vmesa, 0);
+   UNLOCK_HARDWARE(vmesa);
 
-   if (vmesa->dmaLow != DMA_OFFSET) {
-      LOCK_HARDWARE(vmesa); 
-      viaFlushPrimsLocked(vmesa);
-      UNLOCK_HARDWARE(vmesa);
+   viaRasterPrimitive( vmesa->glCtx,
+		       vmesa->renderPrimitive,
+		       vmesa->hwPrimitive );
+}
+
+void viaFlushDma(viaContextPtr vmesa)
+{
+   if (vmesa->dmaLow) {
+      if (vmesa->dmaLastPrim)
+	 viaWrapPrimitive(vmesa); 
+      else {      
+	 LOCK_HARDWARE(vmesa); 
+	 viaFlushDmaLocked(vmesa, 0);
+	 UNLOCK_HARDWARE(vmesa);
+      }
    }
-   if (VIA_DEBUG) fprintf(stderr, "%s in\n", __FUNCTION__);
 }
 
 static void viaFlush(GLcontext *ctx)
@@ -660,315 +735,23 @@ void viaInitIoctlFuncs(GLcontext *ctx)
 }
 
 
-#define SetReg2DAGP(nReg, nData) do {		\
-    OUT_RING( ((nReg) >> 2) | 0xF0000000 );	\
-    OUT_RING( nData );				\
-} while (0)
-
-
-static void viaBlit(viaContextPtr vmesa, GLuint bpp,GLuint srcBase,
-		    GLuint srcPitch,GLuint dstBase,GLuint dstPitch,
-		    GLuint w,GLuint h,int xdir,int ydir, GLuint blitMode, 
-		    GLuint color, GLuint nMask ) 
-{
-
-    GLuint dwGEMode = 0, srcY=0, srcX, dstY=0, dstX;
-    GLuint cmd;
-    RING_VARS;
-
-    if (VIA_DEBUG)
-       fprintf(stderr, "%s bpp %d src %x/%x dst %x/%x w %d h %d dir %d,%d mode: %x color: 0x%08x mask 0x%08x\n",
-	       __FUNCTION__, bpp, srcBase, srcPitch, dstBase, dstPitch, w,h, xdir, ydir, blitMode, color, nMask);
-
-
-    if (!w || !h)
-        return;
-
-    srcX = srcBase & 31;
-    dstX = dstBase & 31;
-    switch (bpp) {
-    case 16:
-        dwGEMode |= VIA_GEM_16bpp;
-	srcX >>= 1;
-	dstX >>= 1;
-        break;
-    case 32:
-        dwGEMode |= VIA_GEM_32bpp;
-	srcX >>= 2;
-	dstX >>= 2;
-	break;
-    default:
-        dwGEMode |= VIA_GEM_8bpp;
-        break;
-    }
-
-
-    cmd = 0; 
-
-    if (xdir < 0) {
-        cmd |= VIA_GEC_DECX;
-        srcX += (w - 1);
-        dstX += (w - 1);
-    }
-    if (ydir < 0) {
-        cmd |= VIA_GEC_DECY;
-        srcY += (h - 1);
-        dstY += (h - 1);
-    }
-
-    switch(blitMode) {
-    case VIA_BLIT_FILL:
-        BEGIN_RING((2 + 9) * 2);
-	SetReg2DAGP(VIA_REG_GEMODE, dwGEMode);
-	SetReg2DAGP( VIA_REG_FGCOLOR, color);
-	cmd |= VIA_GEC_BLT | VIA_GEC_FIXCOLOR_PAT | (VIA_BLIT_FILL << 24);
-	break;
-    case VIA_BLIT_COPY:
-        BEGIN_RING((2 + 9) * 2);
-	SetReg2DAGP(VIA_REG_GEMODE, dwGEMode);
-	SetReg2DAGP( VIA_REG_KEYCONTROL, 0x0);
-	cmd |= VIA_GEC_BLT | (VIA_BLIT_COPY << 24);
-    }	
-
-    SetReg2DAGP( 0x2C, nMask);
-    SetReg2DAGP( VIA_REG_SRCBASE, (srcBase & ~31) >> 3);
-    SetReg2DAGP( VIA_REG_DSTBASE, (dstBase & ~31) >> 3);
-    SetReg2DAGP( VIA_REG_PITCH, VIA_PITCH_ENABLE |
-	       (srcPitch >> 3) | (((dstPitch) >> 3) << 16));
-    SetReg2DAGP( VIA_REG_SRCPOS, ((srcY << 16) | srcX));
-    SetReg2DAGP( VIA_REG_DSTPOS, ((dstY << 16) | dstX));
-    SetReg2DAGP( VIA_REG_DIMENSION, (((h - 1) << 16) | (w - 1)));
-    SetReg2DAGP( VIA_REG_GECMD, cmd);
-    SetReg2DAGP( 0x2C, 0x00000000);
-    ADVANCE_RING();
-}
-
-void viaFillFrontBuffer(viaContextPtr vmesa)
-{
-    GLuint nDestBase, nDestPitch, nDestWidth, nDestHeight,i; 
-    drm_clip_rect_t *b = vmesa->sarea->boxes;
-    GLuint pixel = (GLuint)vmesa->ClearColor;
-    GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
-    nDestPitch = vmesa->front.pitch;
-
-    for (i = 0; i < vmesa->sarea->nbox ; i++) {        
-        nDestWidth = b->x2 - b->x1;
-	nDestHeight = b->y2 - b->y1;
-	nDestBase = vmesa->viaScreen->fbOffset + 
-	    (b->y1* nDestPitch + b->x1 * bytePerPixel);
-	viaBlit(vmesa,vmesa->viaScreen->bitsPerPixel, nDestBase, nDestPitch,
-		nDestBase , nDestPitch, nDestWidth, nDestHeight,
-		0,0,VIA_BLIT_FILL, pixel, 0x0); 
-	b++;
-    }
-
-    viaFlushPrimsLocked(vmesa);
-}
-
-void viaFillFrontPBuffer(viaContextPtr vmesa)
-{
-    GLuint nDestBase, nDestPitch, nDestWidth, nDestHeight, offset; 
-    GLuint pixel = (GLuint)vmesa->ClearColor;
-
-    offset = vmesa->front.offset;
-    if (VIA_DEBUG) fprintf(stderr, "Fill PFront offset = %08x\n", offset);
-    nDestBase = offset;
-    nDestPitch = vmesa->front.pitch;
-
-    nDestWidth = vmesa->driDrawable->w;
-    nDestHeight = vmesa->driDrawable->h;
-
-    viaBlit(vmesa,vmesa->viaScreen->bitsPerPixel, nDestBase, nDestPitch,
-	    nDestBase , nDestPitch, nDestWidth, nDestHeight,
-	    0,0,VIA_BLIT_FILL, pixel, 0x0); 
-	
-    viaFlushPrimsLocked(vmesa);
-}
-
-void viaFillBackBuffer(viaContextPtr vmesa)
-{
-    GLuint nDestBase, nDestPitch, nDestWidth, nDestHeight, offset; 
-    GLcontext *ctx = vmesa->glCtx;
-    GLuint pixel = (GLuint)vmesa->ClearColor;
-    GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
-
-    offset = vmesa->back.offset;
-    if (VIA_DEBUG) fprintf(stderr, "Fill Back offset = %08x\n", offset);
-    nDestBase = offset;
-    nDestPitch = vmesa->back.pitch;
-    
-    if (!ctx->Scissor.Enabled) {
-	nDestWidth = (vmesa->back.pitch / bytePerPixel);
-	nDestHeight = vmesa->driDrawable->h;
-	viaBlit(vmesa,vmesa->viaScreen->bitsPerPixel, nDestBase, nDestPitch,
-		nDestBase , nDestPitch, nDestWidth, nDestHeight,
-		0,0,VIA_BLIT_FILL, pixel, 0x0); 
-	
-    }
-    /*=* John Sheng [2003.7.18] texenv *=*/
-    else {
-	int i;
-	drm_clip_rect_t *b = vmesa->sarea->boxes;
-	for (i = 0; i < vmesa->sarea->nbox ; i++) {        
-	    nDestWidth = b->x2 - b->x1;
-	    nDestHeight = b->y2 - b->y1;
-	    nDestBase = offset + ((b->y1 - vmesa->drawY) * nDestPitch) + 
-		(b->x1 - vmesa->drawX + vmesa->drawXoff) * bytePerPixel;
-	    viaBlit(vmesa,vmesa->viaScreen->bitsPerPixel, nDestBase, nDestPitch,
-		    nDestBase , nDestPitch, nDestWidth, nDestHeight,
-		    0,0,VIA_BLIT_FILL, pixel, 0x0); 
-	    b++;
-	}
-    }
-    if (VIA_DEBUG) {
-	fprintf(stderr," width = %08x\n", nDestWidth);	
-	fprintf(stderr," height = %08x\n", nDestHeight);	
-    }	     
-}
-
-
-void viaFillDepthBuffer(viaContextPtr vmesa, GLuint pixel, GLuint mask)
-{
-    GLuint nDestBase, nDestPitch, nDestWidth, nDestHeight, offsetX, offset;
-
-    offset = vmesa->depth.offset;
-    if (VIA_DEBUG)
-       fprintf(stderr, "Fill Depth offset = %08x, pixel %x, mask %x\n", offset, pixel, mask);
-    nDestBase = offset;
-    nDestPitch = vmesa->depth.pitch;
-    
-    offsetX = vmesa->drawXoff;
-    nDestWidth = (vmesa->depth.pitch / vmesa->depth.bpp * 8) - offsetX;
-    nDestHeight = vmesa->driDrawable->h;
-
-    viaBlit(vmesa, vmesa->depth.bpp , nDestBase, nDestPitch,
-	    nDestBase , nDestPitch, nDestWidth, nDestHeight,
-	    0,0,VIA_BLIT_FILL, pixel, mask); 
-
-    
-    if (vmesa->glCtx->Color._DrawDestMask[0] == DD_BACK_LEFT_BIT) {
-	viaFlushPrimsLocked(vmesa);
-    }
-}
-
-void viaDoSwapBuffers(viaContextPtr vmesa)
-{    
-    GLuint nFrontPitch;
-    GLuint nBackPitch;
-    GLuint nFrontWidth, nFrontHeight;
-    GLuint nFrontBase, nBackBase;
-    drm_clip_rect_t *b = vmesa->pClipRects;
-    GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
-    GLuint i;
-    
-    nFrontPitch = vmesa->front.pitch;
-    nBackPitch = vmesa->back.pitch;
-    
-    for (i = 0; i < vmesa->numClipRects; i++) {        
-
-	/* Width, Height */
-        nFrontWidth = b->x2 - b->x1;
-	nFrontHeight = b->y2 - b->y1;
-	
-	nFrontBase = vmesa->viaScreen->fbOffset + (b->y1* nFrontPitch + 
-						   b->x1 * bytePerPixel);
-	nBackBase = vmesa->back.offset + ((b->y1 - vmesa->drawY) * nBackPitch) + 
-	    (b->x1 - vmesa->drawX + vmesa->drawXoff) * bytePerPixel;
-	
-	viaBlit(vmesa, bytePerPixel << 3 , nBackBase, nBackPitch,
-		nFrontBase , nFrontPitch, nFrontWidth, nFrontHeight,
-		0,0,VIA_BLIT_COPY, 0, 0); 
-	b++;
-    }
-
-    viaFlushPrimsLocked(vmesa);
-    if (VIA_DEBUG) fprintf(stderr, "Do Swap Buffer\n");
-}
-
-
-void viaDoSwapPBuffers(viaContextPtr vmesa)
-{    
-    GLuint nFrontPitch;
-    GLuint nBackPitch;
-    GLuint nFrontWidth, nFrontHeight;
-    GLuint nFrontBase, nBackBase;
-    GLuint nFrontOffsetX, nFrontOffsetY, nBackOffsetX, nBackOffsetY;
-    GLuint bytePerPixel = vmesa->viaScreen->bitsPerPixel >> 3;
-    GLuint EngStatus = *(vmesa->pnGEMode);
-    GLuint blitMode; 
-    RING_VARS;
-
-    switch(bytePerPixel) {
-    case 4:
-	blitMode = 0x300;
-	break;
-    case 2:
-	blitMode = 0x100;
-	break;
-    default:
-	blitMode = 0x000;
-	break;
-    }
-
-    nFrontPitch = vmesa->front.pitch;
-    nBackPitch = vmesa->back.pitch;
-    
-    /* Caculate Base */
-    nFrontBase = vmesa->front.offset;
-    nBackBase = vmesa->back.offset;
-    
-    /* Width, Height */
-    nFrontWidth = nFrontPitch / bytePerPixel;
-    nFrontHeight = nBackPitch / bytePerPixel;
-    
-    /* Offset */
-    nFrontOffsetX = 0;
-    nFrontOffsetY = 0;
-    nBackOffsetX = nFrontOffsetX;
-    nBackOffsetY = nFrontOffsetY;
-
-    BEGIN_RING(8 * 2);
-    /* Restore mode */
-    SetReg2DAGP(0x04, (EngStatus & 0xFFFFFCFF) | blitMode);
-    /* GEWD */
-    SetReg2DAGP(0x10, nFrontWidth | (nFrontHeight << 16));
-    /* GEDST */
-    SetReg2DAGP(0x0C, nFrontOffsetX | (nFrontOffsetY << 16));
-    /* GESRC */
-    SetReg2DAGP(0x08, nBackOffsetX | (nBackOffsetY << 16));
-    /* GEDSTBASE */
-    SetReg2DAGP(0x34, (nFrontBase >> 3));
-    /* GESCRBASE */
-    SetReg2DAGP(0x30, (nBackBase >> 3));
-    /* GEPITCH */
-    SetReg2DAGP(0x38, (((nFrontPitch >> 3) << 16) & 0x7FF0000) | 0x80000000 |
-                      ((nBackPitch >> 3) & 0x7FF));
-    /* BITBLT */
-    SetReg2DAGP(0x0, 0x1 | 0xCC000000);
-    ADVANCE_RING();
-
-    viaFlushPrimsLocked(vmesa);
-    if (VIA_DEBUG) fprintf(stderr, "Do Swap PBuffer\n");
-}
-
-
 
 
 GLuint *viaAllocDmaFunc(viaContextPtr vmesa, int bytes, const char *func, int line)
 {
-    if (vmesa->dmaLow + bytes > VIA_DMA_HIGHWATER) {
-	if (VIA_DEBUG) fprintf(stderr, "buffer overflow in check dma = %d + %d = %d\n", 
-			       vmesa->dmaLow, bytes, vmesa->dmaLow + bytes);
-	viaFlushPrims(vmesa);
-    }
+   if (vmesa->dmaLow + bytes > VIA_DMA_HIGHWATER) {
+      if (VIA_DEBUG) fprintf(stderr, "buffer overflow in check dma = %d + %d = %d\n", 
+			     vmesa->dmaLow, bytes, vmesa->dmaLow + bytes);
+      viaFlushDma(vmesa);
+   }
 
-    {
-        GLuint *start = (GLuint *)(vmesa->dmaAddr + vmesa->dmaLow);
-	vmesa->dmaLow += bytes;
-	if (VIA_DEBUG && (vmesa->dmaLow & 0x4))
-	   fprintf(stderr, "%s/%d: alloc 0x%x --> dmaLow 0x%x\n", func, line, bytes, vmesa->dmaLow);
-        return start;
-    }
+   {
+      GLuint *start = (GLuint *)(vmesa->dma + vmesa->dmaLow);
+      vmesa->dmaLow += bytes;
+      if (VIA_DEBUG && (vmesa->dmaLow & 0x4))
+	 fprintf(stderr, "%s/%d: alloc 0x%x --> dmaLow 0x%x\n", func, line, bytes, vmesa->dmaLow);
+      return start;
+   }
 }