r600: add support for draw_elements_base_vertexr600-basevtx

use VTX_BASE_VTX_LOC for offset, last time using INDEX_OFFSET was
probably a wrong register for this

3 files changed, 36 insertions, 26 deletions

diff --git a/src/mesa/drivers/dri/r600/r600_context.c b/src/mesa/drivers/dri/r600/r600_context.c
index 86e1c1589d..389b0412ba 100644
--- a/src/mesa/drivers/dri/r600/r600_context.c
+++ b/src/mesa/drivers/dri/r600/r600_context.c
@@ -73,6 +73,7 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 #define need_GL_VERSION_2_0
 #define need_GL_VERSION_2_1
+#define need_GL_ARB_draw_elements_base_vertex
 #define need_GL_ARB_occlusion_query
 #define need_GL_ARB_point_parameters
 #define need_GL_ARB_vertex_program
@@ -141,6 +142,7 @@ static const struct dri_extension card_extensions[] = {
   {"GL_NV_vertex_program",		GL_NV_vertex_program_functions},
   {"GL_SGIS_generate_mipmap",		NULL},
   {"GL_ARB_pixel_buffer_object",        NULL},
+  {"GL_ARB_draw_elements_base_vertex",	GL_ARB_draw_elements_base_vertex_functions },
   {NULL,				NULL}
   /* *INDENT-ON* */
 };
diff --git a/src/mesa/drivers/dri/r600/r700_chip.c b/src/mesa/drivers/dri/r600/r700_chip.c
index dc4998265a..bf8063391a 100644
--- a/src/mesa/drivers/dri/r600/r700_chip.c
+++ b/src/mesa/drivers/dri/r600/r700_chip.c
@@ -265,17 +265,6 @@ static void r700SendVTXState(GLcontext *ctx, struct radeon_state_atom *atom)
     if (context->radeon.tcl.aos_count == 0)
 	    return;
 
-    BEGIN_BATCH_NO_AUTOSTATE(6);
-    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
-    R600_OUT_BATCH(mmSQ_VTX_BASE_VTX_LOC - ASIC_CTL_CONST_BASE_INDEX);
-    R600_OUT_BATCH(0);
-
-    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
-    R600_OUT_BATCH(mmSQ_VTX_START_INST_LOC - ASIC_CTL_CONST_BASE_INDEX);
-    R600_OUT_BATCH(0);
-    END_BATCH();
-    COMMIT_BATCH();
-
     for(i=0; i<VERT_ATTRIB_MAX; i++) {
 	    if(vp->mesa_program->Base.InputsRead & (1 << i))
 	    {
@@ -1480,9 +1469,6 @@ static int check_vtx(GLcontext *ctx, struct radeon_state_atom *atom)
 	context_t *context = R700_CONTEXT(ctx);
 	int count = context->radeon.tcl.aos_count * 18;
 
-	if (count)
-		count += 6;
-
 	radeon_print(RADEON_STATE, RADEON_TRACE, "%s %d\n", __func__, count);
 	return count;
 }
@@ -1605,7 +1591,7 @@ void r600InitAtoms(context_t *context)
 	ALLOC_STATE(ps, always, 24, r700SendPSState);
 	ALLOC_STATE(vs_consts, vs_consts, (2 + (R700_MAX_DX9_CONSTS * 4)), r700SendVSConsts);
 	ALLOC_STATE(ps_consts, ps_consts, (2 + (R700_MAX_DX9_CONSTS * 4)), r700SendPSConsts);
-	ALLOC_STATE(vtx, vtx, (6 + (VERT_ATTRIB_MAX * 18)), r700SendVTXState);
+	ALLOC_STATE(vtx, vtx, (VERT_ATTRIB_MAX * 18), r700SendVTXState);
 	ALLOC_STATE(tx, tx, (R700_TEXTURE_NUMBERUNITS * 20), r700SendTexState);
 	ALLOC_STATE(tx_smplr, tx, (R700_TEXTURE_NUMBERUNITS * 5), r700SendTexSamplerState);
 	ALLOC_STATE(tx_brdr_clr, tx, (R700_TEXTURE_NUMBERUNITS * 6), r700SendTexBorderColorState);
diff --git a/src/mesa/drivers/dri/r600/r700_render.c b/src/mesa/drivers/dri/r600/r700_render.c
index ba55f38e05..43f32d628b 100644
--- a/src/mesa/drivers/dri/r600/r700_render.c
+++ b/src/mesa/drivers/dri/r600/r700_render.c
@@ -244,7 +244,8 @@ static int r700NumVerts(int num_verts, int prim)
 	return num_verts - verts_off;
 }
 
-static void r700RunRenderPrimitive(GLcontext * ctx, int start, int end, int prim)
+static void r700RunRenderPrimitive(GLcontext * ctx, int start, int end,
+				   int prim, GLint basevertex)
 {
     context_t *context = R700_CONTEXT(ctx);
     BATCH_LOCALS(&context->radeon);
@@ -282,6 +283,7 @@ static void r700RunRenderPrimitive(GLcontext * ctx, int start, int end, int prim
     total_emit =   3  /* VGT_PRIMITIVE_TYPE */
 	         + 2  /* VGT_INDEX_TYPE */
 	         + 2  /* NUM_INSTANCES */
+		 + 6  /* VTX_BASE_VTX_LOC + VTX_START_INST_LOC */
 	         + 5 + 2; /* DRAW_INDEX */
 
     BEGIN_BATCH_NO_AUTOSTATE(total_emit);
@@ -294,6 +296,14 @@ static void r700RunRenderPrimitive(GLcontext * ctx, int start, int end, int prim
     // num instances
     R600_OUT_BATCH(CP_PACKET3(R600_IT_NUM_INSTANCES, 0));
     R600_OUT_BATCH(1);
+    /* offset */
+    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
+    R600_OUT_BATCH(mmSQ_VTX_BASE_VTX_LOC - ASIC_CTL_CONST_BASE_INDEX);
+    R600_OUT_BATCH(basevertex);
+    /* inst offset - 0 */
+    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
+    R600_OUT_BATCH(mmSQ_VTX_START_INST_LOC - ASIC_CTL_CONST_BASE_INDEX);
+    R600_OUT_BATCH(0);
     // draw packet
     R600_OUT_BATCH(CP_PACKET3(R600_IT_DRAW_INDEX, 3));
     R600_OUT_BATCH(context->ind_buf.bo_offset);
@@ -364,6 +374,7 @@ static void r700RunRenderPrimitiveImmediate(GLcontext * ctx, int start, int end,
     total_emit +=   3 /* VGT_PRIMITIVE_TYPE */
 	          + 2 /* VGT_INDEX_TYPE */
 	          + 2 /* NUM_INSTANCES */
+		  + 6 /* VTX_BASE_VTX_LOC + VTX_START_INST_LOC */
 	          + 3; /* DRAW */
 
     BEGIN_BATCH_NO_AUTOSTATE(total_emit);
@@ -376,6 +387,14 @@ static void r700RunRenderPrimitiveImmediate(GLcontext * ctx, int start, int end,
     // num instances
     R600_OUT_BATCH(CP_PACKET3(R600_IT_NUM_INSTANCES, 0));
     R600_OUT_BATCH(1);
+    /* offset */
+    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
+    R600_OUT_BATCH(mmSQ_VTX_BASE_VTX_LOC - ASIC_CTL_CONST_BASE_INDEX);
+    R600_OUT_BATCH(0);
+    /* inst offset - 0 */
+    R600_OUT_BATCH(CP_PACKET3(R600_IT_SET_CTL_CONST, 1));
+    R600_OUT_BATCH(mmSQ_VTX_START_INST_LOC - ASIC_CTL_CONST_BASE_INDEX);
+    R600_OUT_BATCH(0);
     // draw packet
     if(start == 0)
     {
@@ -433,16 +452,16 @@ static GLuint r700PredictRenderSize(GLcontext* ctx,
 
     dwords = PRE_EMIT_STATE_BUFSZ;
     if (ib)
-	    dwords += nr_prims * 14;
+	    dwords += nr_prims * 20;
     else {
 	    for (i = 0; i < nr_prims; ++i)
 	    {
 		    if (prim[i].start == 0)
-			    dwords += 10;
+			    dwords += 16;
 		    else if (prim[i].count > 0xffff)
-			    dwords += prim[i].count + 10;
+			    dwords += prim[i].count + 16;
 		    else
-			    dwords += ((prim[i].count + 1) / 2) + 10;
+			    dwords += ((prim[i].count + 1) / 2) + 16;
 	    }
     }
 
@@ -923,7 +942,8 @@ static GLboolean r700TryDrawPrims(GLcontext *ctx,
 		    r700RunRenderPrimitive(ctx,
 					   prim[i].start,
 					   prim[i].start + prim[i].count,
-					   prim[i].mode);
+					   prim[i].mode,
+					   prim[i].basevertex);
 	    else
 		    r700RunRenderPrimitiveImmediate(ctx,
 						    prim[i].start,
@@ -975,15 +995,17 @@ static void r700DrawPrims(GLcontext *ctx,
 	/* This check should get folded into just the places that
 	 * min/max index are really needed.
 	 */
-	if (!index_bounds_valid) {
-		vbo_get_minmax_index(ctx, prim, ib, &min_index, &max_index);
-	}
 
-	if (min_index) {
+	if (!vbo_all_varyings_in_vbos(arrays)) {
+	    if (!index_bounds_valid)
+		vbo_get_minmax_index(ctx, prim, ib, &min_index, &max_index);
+	    /* do we want to rebase, minimizes the 
+	     * amount of data to upload? */
+	    if (min_index) {
 		vbo_rebase_prims( ctx, arrays, prim, nr_prims, ib, min_index, max_index, r700DrawPrims );
 		return;
+	    }
 	}
-
 	/* Make an attempt at drawing */
 	retval = r700TryDrawPrims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);