X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fllvmpipe%2Flp_setup_tri.c;h=4e2e17f77b1eb03b4a4a43c444f07fc8badb46bb;hb=392a2515c0967c395be098cac6a37f325dd66b90;hp=ac6264dc73ec58a09e773690050be4bb0f6c26bc;hpb=9d3a873f425fde83628b547c4b4638f39d78c1b9;p=mesa.git

diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index ac6264dc73e..4e2e17f77b1 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -34,6 +34,7 @@
 #include "lp_perf.h"
 #include "lp_setup_context.h"
 #include "lp_rast.h"
+#include "lp_state_fs.h"
 
 #define NUM_CHANNELS 4
 
@@ -146,32 +147,57 @@ setup_fragcoord_coef(struct lp_setup_context *setup,
                      unsigned slot,
                      const float (*v1)[4],
                      const float (*v2)[4],
-                     const float (*v3)[4])
+                     const float (*v3)[4],
+                     unsigned usage_mask)
 {
    /*X*/
-   tri->inputs.a0[slot][0] = 0.0;
-   tri->inputs.dadx[slot][0] = 1.0;
-   tri->inputs.dady[slot][0] = 0.0;
+   if (usage_mask & TGSI_WRITEMASK_X) {
+      tri->inputs.a0[slot][0] = 0.0;
+      tri->inputs.dadx[slot][0] = 1.0;
+      tri->inputs.dady[slot][0] = 0.0;
+   }
+
    /*Y*/
-   tri->inputs.a0[slot][1] = 0.0;
-   tri->inputs.dadx[slot][1] = 0.0;
-   tri->inputs.dady[slot][1] = 1.0;
+   if (usage_mask & TGSI_WRITEMASK_Y) {
+      tri->inputs.a0[slot][1] = 0.0;
+      tri->inputs.dadx[slot][1] = 0.0;
+      tri->inputs.dady[slot][1] = 1.0;
+   }
+
    /*Z*/
-   linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 2);
+   if (usage_mask & TGSI_WRITEMASK_Z) {
+      linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 2);
+   }
+
    /*W*/
-   linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 3);
+   if (usage_mask & TGSI_WRITEMASK_W) {
+      linear_coef(setup, tri, oneoverarea, slot, v1, v2, v3, 0, 3);
+   }
 }
 
 
+/**
+ * Setup the fragment input attribute with the front-facing value.
+ * \param frontface  is the triangle front facing?
+ */
 static void setup_facing_coef( struct lp_setup_context *setup,
                                struct lp_rast_triangle *tri,
                                unsigned slot,
-                               boolean frontface )
+                               boolean frontface,
+                               unsigned usage_mask)
 {
-   constant_coef( setup, tri, slot, 1.0f - frontface, 0 );
-   constant_coef( setup, tri, slot, 0.0f, 1 ); /* wasted */
-   constant_coef( setup, tri, slot, 0.0f, 2 ); /* wasted */
-   constant_coef( setup, tri, slot, 0.0f, 3 ); /* wasted */
+   /* convert TRUE to 1.0 and FALSE to -1.0 */
+   if (usage_mask & TGSI_WRITEMASK_X)
+      constant_coef( setup, tri, slot, 2.0f * frontface - 1.0f, 0 );
+
+   if (usage_mask & TGSI_WRITEMASK_Y)
+      constant_coef( setup, tri, slot, 0.0f, 1 ); /* wasted */
+
+   if (usage_mask & TGSI_WRITEMASK_Z)
+      constant_coef( setup, tri, slot, 0.0f, 2 ); /* wasted */
+
+   if (usage_mask & TGSI_WRITEMASK_W)
+      constant_coef( setup, tri, slot, 0.0f, 3 ); /* wasted */
 }
 
 
@@ -186,48 +212,65 @@ static void setup_tri_coefficients( struct lp_setup_context *setup,
 				    const float (*v3)[4],
 				    boolean frontface)
 {
+   unsigned fragcoord_usage_mask = TGSI_WRITEMASK_XYZ;
    unsigned slot;
 
-   /* The internal position input is in slot zero:
-    */
-   setup_fragcoord_coef(setup, tri, oneoverarea, 0, v1, v2, v3);
-
    /* setup interpolation for all the remaining attributes:
     */
    for (slot = 0; slot < setup->fs.nr_inputs; slot++) {
       unsigned vert_attr = setup->fs.input[slot].src_index;
+      unsigned usage_mask = setup->fs.input[slot].usage_mask;
       unsigned i;
 
       switch (setup->fs.input[slot].interp) {
       case LP_INTERP_CONSTANT:
-         for (i = 0; i < NUM_CHANNELS; i++)
-            constant_coef(setup, tri, slot+1, v3[vert_attr][i], i);
+         if (setup->flatshade_first) {
+            for (i = 0; i < NUM_CHANNELS; i++)
+               if (usage_mask & (1 << i))
+                  constant_coef(setup, tri, slot+1, v1[vert_attr][i], i);
+         }
+         else {
+            for (i = 0; i < NUM_CHANNELS; i++)
+               if (usage_mask & (1 << i))
+                  constant_coef(setup, tri, slot+1, v3[vert_attr][i], i);
+         }
          break;
 
       case LP_INTERP_LINEAR:
          for (i = 0; i < NUM_CHANNELS; i++)
-            linear_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
+            if (usage_mask & (1 << i))
+               linear_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
          break;
 
       case LP_INTERP_PERSPECTIVE:
          for (i = 0; i < NUM_CHANNELS; i++)
-            perspective_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
+            if (usage_mask & (1 << i))
+               perspective_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3, vert_attr, i);
+         fragcoord_usage_mask |= TGSI_WRITEMASK_W;
          break;
 
       case LP_INTERP_POSITION:
-         /* XXX: fix me - duplicates the values in slot zero.
+         /*
+          * The generated pixel interpolators will pick up the coeffs from
+          * slot 0, so all need to ensure that the usage mask is covers all
+          * usages.
           */
-         setup_fragcoord_coef(setup, tri, oneoverarea, slot+1, v1, v2, v3);
+         fragcoord_usage_mask |= usage_mask;
          break;
 
       case LP_INTERP_FACING:
-         setup_facing_coef(setup, tri, slot+1, frontface);
+         setup_facing_coef(setup, tri, slot+1, frontface, usage_mask);
          break;
 
       default:
          assert(0);
       }
    }
+
+   /* The internal position input is in slot zero:
+    */
+   setup_fragcoord_coef(setup, tri, oneoverarea, 0, v1, v2, v3,
+                        fragcoord_usage_mask);
 }
 
 
@@ -261,17 +304,45 @@ alloc_triangle(struct lp_scene *scene, unsigned nr_inputs, unsigned *tri_size)
 
    tri = lp_scene_alloc_aligned( scene, bytes, 16 );
 
-   inputs = (char *) (tri + 1);
-   tri->inputs.a0   = (float (*)[4]) inputs;
-   tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
-   tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
+   if (tri) {
+      inputs = (char *) (tri + 1);
+      tri->inputs.a0   = (float (*)[4]) inputs;
+      tri->inputs.dadx = (float (*)[4]) (inputs + input_array_sz);
+      tri->inputs.dady = (float (*)[4]) (inputs + 2 * input_array_sz);
 
-   *tri_size = bytes;
+      *tri_size = bytes;
+   }
 
    return tri;
 }
 
 
+/**
+ * Print triangle vertex attribs (for debug).
+ */
+static void
+print_triangle(struct lp_setup_context *setup,
+               const float (*v1)[4],
+               const float (*v2)[4],
+               const float (*v3)[4])
+{
+   uint i;
+
+   debug_printf("llvmpipe triangle\n");
+   for (i = 0; i < setup->fs.nr_inputs; i++) {
+      debug_printf("  v1[%d]:  %f %f %f %f\n", i,
+                   v1[i][0], v1[i][1], v1[i][2], v1[i][3]);
+   }
+   for (i = 0; i < setup->fs.nr_inputs; i++) {
+      debug_printf("  v2[%d]:  %f %f %f %f\n", i,
+                   v2[i][0], v2[i][1], v2[i][2], v2[i][3]);
+   }
+   for (i = 0; i < setup->fs.nr_inputs; i++) {
+      debug_printf("  v3[%d]:  %f %f %f %f\n", i,
+                   v3[i][0], v3[i][1], v3[i][2], v3[i][3]);
+   }
+}
+
 
 /**
  * Do basic setup for triangle rasterization and determine which
@@ -300,7 +371,12 @@ do_triangle_ccw(struct lp_setup_context *setup,
    int minx, maxx, miny, maxy;
    unsigned tri_bytes;
 
+   if (0)
+      print_triangle(setup, v1, v2, v3);
+
    tri = alloc_triangle(scene, setup->fs.nr_inputs, &tri_bytes);
+   if (!tri)
+      return;
 
 #ifdef DEBUG
    tri->v[0][0] = v1[0][0];
@@ -361,6 +437,8 @@ do_triangle_ccw(struct lp_setup_context *setup,
     */
    setup_tri_coefficients( setup, tri, oneoverarea, v1, v2, v3, frontfacing );
 
+   tri->inputs.facing = frontfacing ? 1.0F : -1.0F;
+
    /* half-edge constants, will be interated over the whole render target.
     */
    tri->c1 = tri->dy12 * x1 - tri->dx12 * y1;
@@ -533,7 +611,8 @@ do_triangle_ccw(struct lp_setup_context *setup,
                /* triangle covers the whole tile- shade whole tile */
                LP_COUNT(nr_fully_covered_64);
 	       in = TRUE;
-	       if(setup->fs.current.opaque) {
+	       if (setup->fs.current.variant->opaque &&
+	           !setup->fb.zsbuf) {
 	          lp_scene_bin_reset( scene, x, y );
 	          lp_scene_bin_command( scene, x, y,
 	                                lp_rast_set_state,
@@ -570,6 +649,9 @@ do_triangle_ccw(struct lp_setup_context *setup,
 }
 
 
+/**
+ * Draw triangle if it's CW, cull otherwise.
+ */
 static void triangle_cw( struct lp_setup_context *setup,
 			 const float (*v0)[4],
 			 const float (*v1)[4],
@@ -579,6 +661,9 @@ static void triangle_cw( struct lp_setup_context *setup,
 }
 
 
+/**
+ * Draw triangle if it's CCW, cull otherwise.
+ */
 static void triangle_ccw( struct lp_setup_context *setup,
 			 const float (*v0)[4],
 			 const float (*v1)[4],
@@ -588,6 +673,10 @@ static void triangle_ccw( struct lp_setup_context *setup,
 }
 
 
+
+/**
+ * Draw triangle whether it's CW or CCW.
+ */
 static void triangle_both( struct lp_setup_context *setup,
 			   const float (*v0)[4],
 			   const float (*v1)[4],
@@ -619,14 +708,14 @@ void
 lp_setup_choose_triangle( struct lp_setup_context *setup )
 {
    switch (setup->cullmode) {
-   case PIPE_WINDING_NONE:
+   case PIPE_FACE_NONE:
       setup->triangle = triangle_both;
       break;
-   case PIPE_WINDING_CCW:
-      setup->triangle = triangle_cw;
+   case PIPE_FACE_BACK:
+      setup->triangle = setup->ccw_is_frontface ? triangle_ccw : triangle_cw;
       break;
-   case PIPE_WINDING_CW:
-      setup->triangle = triangle_ccw;
+   case PIPE_FACE_FRONT:
+      setup->triangle = setup->ccw_is_frontface ? triangle_cw : triangle_ccw;
       break;
    default:
       setup->triangle = triangle_nop;