llvmpipe: fixed-point rasterization

4 files changed, 147 insertions, 130 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index f371b709df3..97250071199 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -39,7 +39,10 @@
 struct lp_rasterizer;
 struct pipe_screen;
 
-#define TILESIZE 64
+#define FIXED_ORDER 4
+#define FIXED_ONE (1<<FIXED_ORDER)
+#define TILE_ORDER 6
+#define TILESIZE (1<<TILE_ORDER)
 
 
 struct lp_rast_state {
@@ -84,29 +87,29 @@ struct lp_rast_triangle {
    int maxy;
 
    /* one-pixel sized trivial accept offsets for each plane */
-   float ei1;                   
-   float ei2;
-   float ei3;
+   int ei1;                   
+   int ei2;
+   int ei3;
 
    /* one-pixel sized trivial reject offsets for each plane */
-   float eo1;                   
-   float eo2;
-   float eo3;
+   int eo1;                   
+   int eo2;
+   int eo3;
 
    /* y deltas for vertex pairs */
-   float dy12;
-   float dy23;
-   float dy31;
+   int dy12;
+   int dy23;
+   int dy31;
 
    /* x deltas for vertex pairs */
-   float dx12;
-   float dx23;
-   float dx31;
+   int dx12;
+   int dx23;
+   int dx31;
 
    /* edge function values at minx,miny ?? */
-   float c1;
-   float c2;
-   float c3;
+   int c1;
+   int c2;
+   int c3;
 
    /* XXX: this is only used inside lp_setup_tri.c, don't really
     * need it here:
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index f096972d638..f9a8be20c70 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -66,15 +66,15 @@ static void block_full( struct lp_rasterizer *rast,
 static INLINE unsigned
 do_quad( const struct lp_rast_triangle *tri,
 	 int x, int y,
-	 float c1, float c2, float c3 )
+	 int c1, int c2, int c3 )
 {
-   float xstep1 = -tri->dy12;
-   float xstep2 = -tri->dy23;
-   float xstep3 = -tri->dy31;
+   const int xstep1 = -tri->dy12 * FIXED_ONE;
+   const int xstep2 = -tri->dy23 * FIXED_ONE;
+   const int xstep3 = -tri->dy31 * FIXED_ONE;
 
-   float ystep1 = tri->dx12;
-   float ystep2 = tri->dx23;
-   float ystep3 = tri->dx31;
+   const int ystep1 = tri->dx12 * FIXED_ONE;
+   const int ystep2 = tri->dx23 * FIXED_ONE;
+   const int ystep3 = tri->dx31 * FIXED_ONE;
 
    unsigned mask = 0;
 
@@ -108,26 +108,26 @@ static void
 do_block( struct lp_rasterizer *rast,
           const struct lp_rast_triangle *tri,
           int x, int y,
-          float c1,
-          float c2,
-          float c3 )
+          int c1,
+          int c2,
+          int c3 )
 {
-   const int step = 2;
+   const int step = 2 * FIXED_ONE;
 
-   float xstep1 = -step * tri->dy12;
-   float xstep2 = -step * tri->dy23;
-   float xstep3 = -step * tri->dy31;
+   const int xstep1 = -step * tri->dy12;
+   const int xstep2 = -step * tri->dy23;
+   const int xstep3 = -step * tri->dy31;
 
-   float ystep1 = step * tri->dx12;
-   float ystep2 = step * tri->dx23;
-   float ystep3 = step * tri->dx31;
+   const int ystep1 = step * tri->dx12;
+   const int ystep2 = step * tri->dx23;
+   const int ystep3 = step * tri->dx31;
 
    int ix, iy;
 
    for (iy = 0; iy < BLOCKSIZE; iy += 2) {
-      float cx1 = c1;
-      float cx2 = c2;
-      float cx3 = c3;
+      int cx1 = c1;
+      int cx2 = c2;
+      int cx3 = c3;
 
       unsigned masks[4] = {0, 0, 0, 0};
 
@@ -160,23 +160,23 @@ void lp_rast_triangle( struct lp_rasterizer *rast,
 {
    const struct lp_rast_triangle *tri = arg.triangle;
 
-   const int step = BLOCKSIZE;
+   const int step = BLOCKSIZE * FIXED_ONE;
 
-   float ei1 = tri->ei1 * step;
-   float ei2 = tri->ei2 * step;
-   float ei3 = tri->ei3 * step;
+   int ei1 = tri->ei1 * step;
+   int ei2 = tri->ei2 * step;
+   int ei3 = tri->ei3 * step;
 
-   float eo1 = tri->eo1 * step;
-   float eo2 = tri->eo2 * step;
-   float eo3 = tri->eo3 * step;
+   int eo1 = tri->eo1 * step;
+   int eo2 = tri->eo2 * step;
+   int eo3 = tri->eo3 * step;
 
-   float xstep1 = -step * tri->dy12;
-   float xstep2 = -step * tri->dy23;
-   float xstep3 = -step * tri->dy31;
+   int xstep1 = -step * tri->dy12;
+   int xstep2 = -step * tri->dy23;
+   int xstep3 = -step * tri->dy31;
 
-   float ystep1 = step * tri->dx12;
-   float ystep2 = step * tri->dx23;
-   float ystep3 = step * tri->dx31;
+   int ystep1 = step * tri->dx12;
+   int ystep2 = step * tri->dx23;
+   int ystep3 = step * tri->dx31;
 
    /* Clamp to tile dimensions:
     */
@@ -186,8 +186,8 @@ void lp_rast_triangle( struct lp_rasterizer *rast,
    int maxy = MIN2(tri->maxy, rast->y + TILE_SIZE);
 
    int x, y;
-   float x0, y0;
-   float c1, c2, c3;
+   int x0, y0;
+   int c1, c2, c3;
 
    debug_printf("%s\n", __FUNCTION__);
 
@@ -196,23 +196,23 @@ void lp_rast_triangle( struct lp_rasterizer *rast,
       return;
    }
 
-   minx &= ~(step-1);
-   miny &= ~(step-1);
+   minx &= ~(BLOCKSIZE-1);
+   miny &= ~(BLOCKSIZE-1);
 
-   x0 = (float)minx;
-   y0 = (float)miny;
+   x0 = minx << FIXED_ORDER;
+   y0 = miny << FIXED_ORDER;
 
    c1 = tri->c1 + tri->dx12 * y0 - tri->dy12 * x0;
    c2 = tri->c2 + tri->dx23 * y0 - tri->dy23 * x0;
    c3 = tri->c3 + tri->dx31 * y0 - tri->dy31 * x0;
 
-   for (y = miny; y < maxy; y += step)
+   for (y = miny; y < maxy; y += BLOCKSIZE)
    {
-      float cx1 = c1;
-      float cx2 = c2;
-      float cx3 = c3;
+      int cx1 = c1;
+      int cx2 = c2;
+      int cx3 = c3;
 
-      for (x = minx; x < maxx; x += step)
+      for (x = minx; x < maxx; x += BLOCKSIZE)
       {
          if (cx1 + eo1 < 0 ||
              cx2 + eo2 < 0 ||
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_context.h b/src/gallium/drivers/llvmpipe/lp_setup_context.h
index bcd3b9b7aa3..d91ffc7c20a 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup_context.h
@@ -169,6 +169,15 @@ static INLINE void *get_data( struct data_block_list *list,
    }
 }
 
+/* Put back data if we decide not to use it, eg. culled triangles.
+ */
+static INLINE void putback_data( struct data_block_list *list,
+                                 unsigned size)
+{
+   list->tail->used -= size;
+}
+
+
 static INLINE void *get_data_aligned( struct data_block_list *list,
                                       unsigned size,
                                       unsigned alignment )
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index 89b2b4eb37c..44386a225d1 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -223,10 +223,9 @@ static void setup_tri_coefficients( struct setup_context *setup,
 
 /* XXX: do this by add/subtracting a large floating point number:
  */
-static inline float subpixel_snap( float a )
+static inline int subpixel_snap( float a )
 {
-   int i = a * 16;
-   return (float)i * (1.0/16);
+   return util_iround(FIXED_ONE * a);
 }
 
 
@@ -256,23 +255,18 @@ do_triangle_ccw(struct setup_context *setup,
 		const float (*v3)[4],
 		boolean frontfacing )
 {
-   const int rt_width = setup->fb.width;
-   const int rt_height = setup->fb.height;
 
-   const float y1 = subpixel_snap(v1[0][1]);
-   const float y2 = subpixel_snap(v2[0][1]);
-   const float y3 = subpixel_snap(v3[0][1]);
+   const int y1 = subpixel_snap(v1[0][1]);
+   const int y2 = subpixel_snap(v2[0][1]);
+   const int y3 = subpixel_snap(v3[0][1]);
 
-   const float x1 = subpixel_snap(v1[0][0]);
-   const float x2 = subpixel_snap(v2[0][0]);
-   const float x3 = subpixel_snap(v3[0][0]);
+   const int x1 = subpixel_snap(v1[0][0]);
+   const int x2 = subpixel_snap(v2[0][0]);
+   const int x3 = subpixel_snap(v3[0][0]);
    
    struct lp_rast_triangle *tri = get_data( &setup->data, sizeof *tri );
    float area;
    int minx, maxx, miny, maxy;
-   float c1, c2, c3;
-
-   tri->inputs.state = setup->fs.stored;
 
    tri->dx12 = x1 - x2;
    tri->dx23 = x2 - x3;
@@ -285,35 +279,32 @@ do_triangle_ccw(struct setup_context *setup,
    area = (tri->dx12 * tri->dy31 - 
 	   tri->dx31 * tri->dy12);
 
-   /* Cull non-ccw and zero-sized triangles.
+   /* Cull non-ccw and zero-sized triangles. 
+    *
+    * XXX: subject to overflow??
     */
-   if (area <= 0 || util_is_inf_or_nan(area))
+   if (area <= 0) {
+      putback_data( &setup->data, sizeof *tri );
       return;
+   }
 
    // Bounding rectangle
-   minx = util_iround(MIN3(x1, x2, x3) - .5);
-   maxx = util_iround(MAX3(x1, x2, x3) + .5);
-   miny = util_iround(MIN3(y1, y2, y3) - .5);
-   maxy = util_iround(MAX3(y1, y2, y3) + .5);
+   tri->minx = (MIN3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
+   tri->maxx = (MAX3(x1, x2, x3) + 0xf) >> FIXED_ORDER;
+   tri->miny = (MIN3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
+   tri->maxy = (MAX3(y1, y2, y3) + 0xf) >> FIXED_ORDER;
    
-   /* Clamp to framebuffer (or tile) dimensions:
-    */
-   miny = MAX2(0, miny);
-   minx = MAX2(0, minx);
-   maxy = MIN2(rt_height, maxy);
-   maxx = MIN2(rt_width, maxx);
-
-   if (miny == maxy || minx == maxx)
+   if (tri->miny == tri->maxy || 
+       tri->minx == tri->maxx) {
+      putback_data( &setup->data, sizeof *tri );
       return;
+   }
 
-   tri->miny = miny;
-   tri->minx = minx;
-   tri->maxy = maxy;
-   tri->maxx = maxx;
+   tri->inputs.state = setup->fs.stored;
 
-   /* The only divide in this code.  Is it really needed?
+   /* 
     */
-   tri->oneoverarea = 1.0f / area;
+   tri->oneoverarea = ((float)FIXED_ONE) / (float)area;
 
    /* Setup parameter interpolants:
     */
@@ -328,9 +319,9 @@ do_triangle_ccw(struct setup_context *setup,
 
    /* correct for top-left fill convention:
     */
-   if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1 += 1.0/16.0f;
-   if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2 += 1.0/16.0f;
-   if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3 += 1.0/16.0f;
+   if (tri->dy12 < 0 || (tri->dy12 == 0 && tri->dx12 > 0)) tri->c1++;
+   if (tri->dy23 < 0 || (tri->dy23 == 0 && tri->dx23 > 0)) tri->c2++;
+   if (tri->dy31 < 0 || (tri->dy31 == 0 && tri->dx31 > 0)) tri->c3++;
 
    /* find trivial reject offsets for each edge for a single-pixel
     * sized block.  These will be scaled up at each recursive level to
@@ -355,17 +346,10 @@ do_triangle_ccw(struct setup_context *setup,
    tri->ei2 = tri->dx23 - tri->dy23 - tri->eo2;
    tri->ei3 = tri->dx31 - tri->dy31 - tri->eo3;
 
-   minx &= ~(TILESIZE-1);		/* aligned blocks */
-   miny &= ~(TILESIZE-1);		/* aligned blocks */
-
-   c1 = tri->c1 + tri->dx12 * miny - tri->dy12 * minx;
-   c2 = tri->c2 + tri->dx23 * miny - tri->dy23 * minx;
-   c3 = tri->c3 + tri->dx31 * miny - tri->dy31 * minx;
-
-   minx /= TILESIZE;
-   miny /= TILESIZE;
-   maxx /= TILESIZE;
-   maxy /= TILESIZE;
+   minx = tri->minx / TILESIZE;
+   miny = tri->miny / TILESIZE;
+   maxx = tri->maxx / TILESIZE;
+   maxy = tri->maxy / TILESIZE;
 
    /* Convert to tile coordinates:
     */
@@ -378,23 +362,31 @@ do_triangle_ccw(struct setup_context *setup,
    }
    else 
    {
-      const int step = TILESIZE;
-
-      float ei1 = tri->ei1 * step;
-      float ei2 = tri->ei2 * step;
-      float ei3 = tri->ei3 * step;
-
-      float eo1 = tri->eo1 * step;
-      float eo2 = tri->eo2 * step;
-      float eo3 = tri->eo3 * step;
-
-      float xstep1 = -step * tri->dy12;
-      float xstep2 = -step * tri->dy23;
-      float xstep3 = -step * tri->dy31;
-
-      float ystep1 = step * tri->dx12;
-      float ystep2 = step * tri->dx23;
-      float ystep3 = step * tri->dx31;
+      int c1 = (tri->c1 + 
+                tri->dx12 * miny * TILESIZE * FIXED_ONE - 
+                tri->dy12 * minx * TILESIZE * FIXED_ONE);
+      int c2 = (tri->c2 + 
+                tri->dx23 * miny * TILESIZE * FIXED_ONE -
+                tri->dy23 * minx * TILESIZE * FIXED_ONE);
+      int c3 = (tri->c3 +
+                tri->dx31 * miny * TILESIZE * FIXED_ONE -
+                tri->dy31 * minx * TILESIZE * FIXED_ONE);
+
+      int ei1 = tri->ei1 << (FIXED_ORDER + TILE_ORDER);
+      int ei2 = tri->ei2 << (FIXED_ORDER + TILE_ORDER);
+      int ei3 = tri->ei3 << (FIXED_ORDER + TILE_ORDER);
+
+      int eo1 = tri->eo1 << (FIXED_ORDER + TILE_ORDER);
+      int eo2 = tri->eo2 << (FIXED_ORDER + TILE_ORDER);
+      int eo3 = tri->eo3 << (FIXED_ORDER + TILE_ORDER);
+
+      int xstep1 = -(tri->dy12 << (FIXED_ORDER + TILE_ORDER));
+      int xstep2 = -(tri->dy23 << (FIXED_ORDER + TILE_ORDER));
+      int xstep3 = -(tri->dy31 << (FIXED_ORDER + TILE_ORDER));
+
+      int ystep1 = tri->dx12 << (FIXED_ORDER + TILE_ORDER);
+      int ystep2 = tri->dx23 << (FIXED_ORDER + TILE_ORDER);
+      int ystep3 = tri->dx31 << (FIXED_ORDER + TILE_ORDER);
       int x, y;
 
 
@@ -406,12 +398,25 @@ do_triangle_ccw(struct setup_context *setup,
        */
       for (y = miny; y <= maxy; y++)
       {
-	 float cx1 = c1;
-	 float cx2 = c2;
-	 float cx3 = c3;
+	 int cx1 = c1;
+	 int cx2 = c2;
+	 int cx3 = c3;
 
 	 for (x = minx; x <= maxx; x++)
 	 {
+            assert(cx1 == 
+                   tri->c1 + 
+                   tri->dx12 * y * TILESIZE * FIXED_ONE - 
+                   tri->dy12 * x * TILESIZE * FIXED_ONE);
+            assert(cx2 ==
+                   tri->c2 + 
+                   tri->dx23 * y * TILESIZE * FIXED_ONE - 
+                   tri->dy23 * x * TILESIZE * FIXED_ONE);
+            assert(cx3 == 
+                   tri->c3 +
+                   tri->dx31 * y * TILESIZE * FIXED_ONE -
+                   tri->dy31 * x * TILESIZE * FIXED_ONE);
+
 	    if (cx1 + eo1 < 0 || 
 		cx2 + eo2 < 0 ||
 		cx3 + eo3 < 0) 
@@ -427,9 +432,9 @@ do_triangle_ccw(struct setup_context *setup,
                             lp_rast_arg_inputs(&tri->inputs) );
 	    }
 	    else 
-	    {
+	    { 
                /* shade partial tile */
-	       bin_command( &setup->tile[x][y], 
+               bin_command( &setup->tile[x][y], 
                             lp_rast_triangle, 
                             lp_rast_arg_triangle(tri) );
 	    }