R3/5xx EXA: Minimise number of draw primitives used for Composite operations.

This should reduce the kernel CS checker overhead, if nothing else.

I'll leave porting this to other chipset families to others who can test it.

2 files changed, 37 insertions, 12 deletions

diff --git a/src/radeon.h b/src/radeon.h
index db7b26b8..0322bf00 100644
--- a/src/radeon.h
+++ b/src/radeon.h
@@ -687,6 +687,9 @@ struct radeon_accel_state {
     /* Size of tiles ... set to 65536x65536 if not tiling in that direction */
     Bool              src_tile_width;
     Bool              src_tile_height;
+    uint32_t          *draw_header;
+    unsigned          vtx_count;
+    unsigned          num_vtx;
 
     Bool              vsync;
 
diff --git a/src/radeon_exa_render.c b/src/radeon_exa_render.c
index 6053eef2..c266fb71 100644
--- a/src/radeon_exa_render.c
+++ b/src/radeon_exa_render.c
@@ -2041,6 +2041,16 @@ static void FUNC_NAME(RadeonDoneComposite)(PixmapPtr pDst)
     ENTER_DRAW(0);
 
     if (IS_R300_3D || IS_R500_3D) {
+	if (info->accel_state->draw_header) {
+		info->accel_state->draw_header[0] = CP_PACKET3(R200_CP_PACKET3_3D_DRAW_IMMD_2,
+							       info->accel_state->num_vtx *
+							       info->accel_state->vtx_count);
+		info->accel_state->draw_header[1] = RADEON_CP_VC_CNTL_PRIM_TYPE_QUAD_LIST |
+		    RADEON_CP_VC_CNTL_PRIM_WALK_RING |
+		    (info->accel_state->num_vtx << RADEON_CP_VC_CNTL_NUM_SHIFT);
+		info->accel_state->draw_header = NULL;
+	}
+
 	BEGIN_ACCEL(3);
 	OUT_ACCEL_REG(R300_SC_CLIP_RULE, 0xAAAA);
 	OUT_ACCEL_REG(R300_RB3D_DSTCACHE_CTLSTAT, R300_RB3D_DC_FLUSH_ALL);
@@ -2126,8 +2136,10 @@ static void FUNC_NAME(RadeonCompositeTile)(ScrnInfoPtr pScrn,
     /* ErrorF("RadeonComposite (%d,%d) (%d,%d) (%d,%d) (%d,%d)\n",
        srcX, srcY, maskX, maskY,dstX, dstY, w, h); */
 
-#if defined(ACCEL_CP) && defined(XF86DRM_MODE)
-    if (info->cs && CS_FULL(info->cs)) {
+#if defined(ACCEL_CP)
+    if ((info->cs && CS_FULL(info->cs)) ||
+	(!info->cs && (info->cp->indirectBuffer->used + 4 * 32) >
+	 info->cp->indirectBuffer->total)) {
 	FUNC_NAME(RadeonDoneComposite)(info->accel_state->dst_pix);
 	radeon_cs_flush_indirect(pScrn);
 	info->accel_state->exa->PrepareComposite(info->accel_state->composite_op,
@@ -2202,12 +2214,26 @@ static void FUNC_NAME(RadeonCompositeTile)(ScrnInfoPtr pScrn,
 		 RADEON_CP_VC_CNTL_VTX_FMT_RADEON_MODE |
 		 (3 << RADEON_CP_VC_CNTL_NUM_SHIFT));
     } else if (IS_R300_3D || IS_R500_3D) {
-	BEGIN_RING(4 * vtx_count + 4);
-	OUT_RING(CP_PACKET3(R200_CP_PACKET3_3D_DRAW_IMMD_2,
-			    4 * vtx_count));
-	OUT_RING(RADEON_CP_VC_CNTL_PRIM_TYPE_QUAD_LIST |
-		 RADEON_CP_VC_CNTL_PRIM_WALK_RING |
-		 (4 << RADEON_CP_VC_CNTL_NUM_SHIFT));
+	if (!info->accel_state->draw_header) {
+	    BEGIN_RING(2);
+
+	    if (info->cs)
+		info->accel_state->draw_header = info->cs->packets + info->cs->cdw;
+	    else
+		info->accel_state->draw_header = __head;
+	    info->accel_state->num_vtx = 0;
+	    info->accel_state->vtx_count = vtx_count;
+
+	    OUT_RING(CP_PACKET3(R200_CP_PACKET3_3D_DRAW_IMMD_2,
+				4 * vtx_count));
+	    OUT_RING(RADEON_CP_VC_CNTL_PRIM_TYPE_QUAD_LIST |
+		     RADEON_CP_VC_CNTL_PRIM_WALK_RING |
+		     (4 << RADEON_CP_VC_CNTL_NUM_SHIFT));
+	    ADVANCE_RING();
+	}
+
+	info->accel_state->num_vtx += 4;
+	BEGIN_RING(4 * vtx_count);
     } else {
 	BEGIN_RING(3 * vtx_count + 2);
 	OUT_RING(CP_PACKET3(R200_CP_PACKET3_3D_DRAW_IMMD_2,
@@ -2267,10 +2293,6 @@ static void FUNC_NAME(RadeonCompositeTile)(ScrnInfoPtr pScrn,
 		xFixedToFloat(srcTopRight.x) / info->accel_state->texW[0],     xFixedToFloat(srcTopRight.y) / info->accel_state->texH[0]);
     }
 
-    if (IS_R300_3D || IS_R500_3D)
-	/* flushing is pipelined, free/finish is not */
-	OUT_ACCEL_REG(R300_RB3D_DSTCACHE_CTLSTAT, R300_DC_FLUSH_3D);
-
 #ifdef ACCEL_CP
     ADVANCE_RING();
 #else