Merge commit 'origin/master' into gallium-0.2
[mesa.git] / src / gallium / auxiliary / draw / draw_pt_vcache.c
index 6c17edba3401a9ebcae88490dc7e88c9566d6991..80d7200ca637cc17dea0ce279c03be2c4b75901c 100644 (file)
   *   Keith Whitwell <keith@tungstengraphics.com>
   */
 
-#include "pipe/p_util.h"
+#include "util/u_memory.h"
 #include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_pt.h"
 
 
-#define CACHE_MAX 32
-#define FETCH_MAX 128
+#define CACHE_MAX 256
+#define FETCH_MAX 256
 #define DRAW_MAX (16*1024)
 
 struct vcache_frontend {
@@ -52,15 +52,28 @@ struct vcache_frontend {
 
    unsigned draw_count;
    unsigned fetch_count;
+   unsigned fetch_max;
    
    struct draw_pt_middle_end *middle;
 
    unsigned input_prim;
    unsigned output_prim;
+
+   unsigned middle_prim;
+   unsigned opt;
 };
 
-static void vcache_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_flush( struct vcache_frontend *vcache )
 {
+   if (vcache->middle_prim != vcache->output_prim) {
+      vcache->middle_prim = vcache->output_prim;
+      vcache->middle->prepare( vcache->middle, 
+                               vcache->middle_prim, 
+                               vcache->opt, 
+                               &vcache->fetch_max );
+   }
+
    if (vcache->draw_count) {
       vcache->middle->run( vcache->middle,
                            vcache->fetch_elts,
@@ -74,7 +87,8 @@ static void vcache_flush( struct vcache_frontend *vcache )
    vcache->draw_count = 0;
 }
 
-static void vcache_check_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_check_flush( struct vcache_frontend *vcache )
 {
    if ( vcache->draw_count + 6 >= DRAW_MAX ||
         vcache->fetch_count + 4 >= FETCH_MAX )
@@ -84,9 +98,10 @@ static void vcache_check_flush( struct vcache_frontend *vcache )
 }
 
 
-static INLINE void vcache_elt( struct vcache_frontend *vcache,
-                               unsigned felt,
-                               ushort flags )
+static INLINE void 
+vcache_elt( struct vcache_frontend *vcache,
+            unsigned felt,
+            ushort flags )
 {
    unsigned idx = felt % CACHE_MAX;
 
@@ -103,11 +118,25 @@ static INLINE void vcache_elt( struct vcache_frontend *vcache,
 
 
                    
-static void vcache_triangle( struct vcache_frontend *vcache,
-                             ushort flags,
-                             unsigned i0,
-                             unsigned i1,
-                             unsigned i2 )
+static INLINE void 
+vcache_triangle( struct vcache_frontend *vcache,
+                 unsigned i0,
+                 unsigned i1,
+                 unsigned i2 )
+{
+   vcache_elt(vcache, i0, 0);
+   vcache_elt(vcache, i1, 0);
+   vcache_elt(vcache, i2, 0);
+   vcache_check_flush(vcache);
+}
+
+                         
+static INLINE void 
+vcache_triangle_flags( struct vcache_frontend *vcache,
+                       ushort flags,
+                       unsigned i0,
+                       unsigned i1,
+                       unsigned i2 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -115,10 +144,22 @@ static void vcache_triangle( struct vcache_frontend *vcache,
    vcache_check_flush(vcache);
 }
 
-static void vcache_line( struct vcache_frontend *vcache,
-                         ushort flags,
-                         unsigned i0,
-                         unsigned i1 )
+static INLINE void 
+vcache_line( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1 )
+{
+   vcache_elt(vcache, i0, 0);
+   vcache_elt(vcache, i1, 0);
+   vcache_check_flush(vcache);
+}
+
+
+static INLINE void 
+vcache_line_flags( struct vcache_frontend *vcache,
+                   ushort flags,
+                   unsigned i0,
+                   unsigned i1 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -126,107 +167,316 @@ static void vcache_line( struct vcache_frontend *vcache,
 }
 
 
-static void vcache_point( struct vcache_frontend *vcache,
-                          unsigned i0 )
+static INLINE void 
+vcache_point( struct vcache_frontend *vcache,
+              unsigned i0 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_check_flush(vcache);
 }
 
-static void vcache_quad( struct vcache_frontend *vcache,
-                         unsigned i0,
-                         unsigned i1,
-                         unsigned i2,
-                         unsigned i3 )
+static INLINE void 
+vcache_quad( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1,
+             unsigned i2,
+             unsigned i3 )
 {
-   vcache_triangle( vcache,
-                    ( DRAW_PIPE_RESET_STIPPLE |
-                      DRAW_PIPE_EDGE_FLAG_0 |
-                      DRAW_PIPE_EDGE_FLAG_2 ),
-                    i0, i1, i3 );
+   vcache_triangle( vcache, i0, i1, i3 );
+   vcache_triangle( vcache, i1, i2, i3 );
+}
 
-   vcache_triangle( vcache,
-                    ( DRAW_PIPE_EDGE_FLAG_0 |
-                      DRAW_PIPE_EDGE_FLAG_1 ),
-                    i1, i2, i3 );
+static INLINE void 
+vcache_ef_quad( struct vcache_frontend *vcache,
+                unsigned i0,
+                unsigned i1,
+                unsigned i2,
+                unsigned i3 )
+{
+   vcache_triangle_flags( vcache,
+                          ( DRAW_PIPE_RESET_STIPPLE |
+                            DRAW_PIPE_EDGE_FLAG_0 |
+                            DRAW_PIPE_EDGE_FLAG_2 ),
+                          i0, i1, i3 );
+
+   vcache_triangle_flags( vcache,
+                          ( DRAW_PIPE_EDGE_FLAG_0 |
+                            DRAW_PIPE_EDGE_FLAG_1 ),
+                          i1, i2, i3 );
 }
 
 /* At least for now, we're back to using a template include file for
  * this.  The two paths aren't too different though - it may be
  * possible to reunify them.
  */
-#define TRIANGLE(flags,i0,i1,i2)                \
-  vcache_triangle(vcache,                       \
-                  flags,                        \
-                  get_elt(elts,i0),             \
-                  get_elt(elts,i1),             \
-                  get_elt(elts,i2))
-
-#define QUAD(i0,i1,i2,i3)                       \
-  vcache_quad(vcache,                           \
-              get_elt(elts,i0),                 \
-              get_elt(elts,i1),                 \
-              get_elt(elts,i2),                 \
-              get_elt(elts,i3))
-
-#define LINE(flags,i0,i1)                       \
-  vcache_line(vcache,                           \
-              flags,                            \
-              get_elt(elts,i0),                 \
-              get_elt(elts,i1))
-
-#define POINT(i0)                               \
-  vcache_point(vcache,                          \
-               get_elt(elts,i0))
-
+#define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle_flags(vc,flags,i0,i1,i2)
+#define QUAD(vc,i0,i1,i2,i3)        vcache_ef_quad(vc,i0,i1,i2,i3)
+#define LINE(vc,flags,i0,i1)        vcache_line_flags(vc,flags,i0,i1)
+#define POINT(vc,i0)                vcache_point(vc,i0)
+#define FUNC vcache_run_extras
+#include "draw_pt_vcache_tmp.h"
+
+#define TRIANGLE(vc,flags,i0,i1,i2) vcache_triangle(vc,i0,i1,i2)
+#define QUAD(vc,i0,i1,i2,i3)        vcache_quad(vc,i0,i1,i2,i3)
+#define LINE(vc,flags,i0,i1)        vcache_line(vc,i0,i1)
+#define POINT(vc,i0)                vcache_point(vc,i0)
 #define FUNC vcache_run
-#define ARGS                                    \
-    struct draw_pt_front_end *frontend,         \
-    pt_elt_func get_elt,                        \
-    const void *elts
+#include "draw_pt_vcache_tmp.h"
+
+static INLINE void 
+rebase_uint_elts( const unsigned *src,
+                  unsigned count,
+                  int delta,
+                  ushort *dest )
+{
+   unsigned i;
+
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i] + delta);
+}
+
+static INLINE void 
+rebase_ushort_elts( const ushort *src,
+                    unsigned count,
+                    int delta,
+                                ushort *dest )
+{
+   unsigned i;
+
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i] + delta);
+}
 
-#define LOCAL_VARS                                                      \
-   struct vcache_frontend *vcache = (struct vcache_frontend *)frontend; \
-   struct draw_context *draw = vcache->draw;                            \
-   boolean flatfirst = (draw->rasterizer->flatshade &&                  \
-                        draw->rasterizer->flatshade_first);             \
-   unsigned prim = vcache->input_prim;                                  \
-   unsigned i, flags;
+static INLINE void 
+rebase_ubyte_elts( const ubyte *src,
+                   unsigned count,
+                   int delta,
+                   ushort *dest )
+{
+   unsigned i;
 
-#define FLUSH vcache_flush( vcache )
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i] + delta);
+}
 
-#include "draw_pt_decompose.h"
 
 
+static INLINE void 
+translate_uint_elts( const unsigned *src,
+                     unsigned count,
+                     ushort *dest )
+{
+   unsigned i;
 
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i]);
+}
 
+static INLINE void 
+translate_ushort_elts( const ushort *src,
+                       unsigned count,
+                       ushort *dest )
+{
+   unsigned i;
+
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i]);
+}
 
-static void vcache_prepare( struct draw_pt_front_end *frontend,
-                            unsigned prim,
-                            struct draw_pt_middle_end *middle,
-                           unsigned opt )
+static INLINE void 
+translate_ubyte_elts( const ubyte *src,
+                      unsigned count,
+                      ushort *dest )
+{
+   unsigned i;
+
+   for (i = 0; i < count; i++) 
+      dest[i] = (ushort)(src[i]);
+}
+
+
+
+
+#if 0
+static INLINE enum pipe_format 
+format_from_get_elt( pt_elt_func get_elt )
+{
+   switch (draw->pt.user.eltSize) {
+   case 1: return PIPE_FORMAT_R8_UNORM;
+   case 2: return PIPE_FORMAT_R16_UNORM;
+   case 4: return PIPE_FORMAT_R32_UNORM;
+   default: return PIPE_FORMAT_NONE;
+   }
+}
+#endif
+
+static INLINE void 
+vcache_check_run( struct draw_pt_front_end *frontend, 
+                  pt_elt_func get_elt,
+                  const void *elts,
+                  unsigned draw_count )
+{
+   struct vcache_frontend *vcache = (struct vcache_frontend *)frontend; 
+   struct draw_context *draw = vcache->draw;
+   unsigned min_index = draw->pt.user.min_index;
+   unsigned max_index = draw->pt.user.max_index;
+   unsigned index_size = draw->pt.user.eltSize;
+   unsigned fetch_count = max_index + 1 - min_index;
+   const ushort *transformed_elts;
+   ushort *storage = NULL;
+   boolean ok;
+
+
+   if (0) debug_printf("fetch_count %d fetch_max %d draw_count %d\n", fetch_count, 
+                       vcache->fetch_max,
+                       draw_count);
+      
+   if (max_index == 0xffffffff ||
+       fetch_count > draw_count) {
+      if (0) debug_printf("fail\n");
+      goto fail;
+   }
+      
+   if (vcache->middle_prim != vcache->input_prim) {
+      vcache->middle_prim = vcache->input_prim;
+      vcache->middle->prepare( vcache->middle, 
+                               vcache->middle_prim, 
+                               vcache->opt, 
+                               &vcache->fetch_max );
+   }
+
+
+   if (min_index == 0 &&
+       index_size == 2) 
+   {
+      transformed_elts = (const ushort *)elts;
+   }
+   else 
+   {
+      storage = MALLOC( draw_count * sizeof(ushort) );
+      if (!storage)
+         goto fail;
+      
+      if (min_index == 0) {
+         switch(index_size) {
+         case 1:
+            translate_ubyte_elts( (const ubyte *)elts,
+                                  draw_count,
+                                  storage );
+            break;
+
+         case 2:
+            translate_ushort_elts( (const ushort *)elts,
+                                   draw_count,
+                                   storage );
+            break;
+
+         case 4:
+            translate_uint_elts( (const uint *)elts,
+                                 draw_count,
+                                 storage );
+            break;
+
+         default:
+            assert(0);
+            return;
+         }
+      }
+      else {
+         switch(index_size) {
+         case 1:
+            rebase_ubyte_elts( (const ubyte *)elts,
+                                  draw_count,
+                                  0 - (int)min_index,
+                                  storage );
+            break;
+
+         case 2:
+            rebase_ushort_elts( (const ushort *)elts,
+                                   draw_count,
+                                   0 - (int)min_index,
+                                   storage );
+            break;
+
+         case 4:
+            rebase_uint_elts( (const uint *)elts,
+                                 draw_count,
+                                 0 - (int)min_index,
+                                 storage );
+            break;
+
+         default:
+            assert(0);
+            return;
+         }
+      }
+      transformed_elts = storage;
+   }
+
+   ok = vcache->middle->run_linear_elts( vcache->middle,
+                                         min_index, /* start */
+                                         fetch_count,
+                                         transformed_elts,
+                                         draw_count );
+   
+   FREE(storage);
+
+   if (ok)
+      return;
+
+   debug_printf("failed to execute atomic draw elts for %d/%d, splitting up\n",
+                fetch_count, draw_count);
+
+ fail:
+   vcache_run( frontend, get_elt, elts, draw_count );
+}
+
+
+
+
+static void 
+vcache_prepare( struct draw_pt_front_end *frontend,
+                unsigned prim,
+                struct draw_pt_middle_end *middle,
+                unsigned opt )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 
-   vcache->base.run = vcache_run;
+   if (opt & PT_PIPELINE)
+   {
+      vcache->base.run = vcache_run_extras;
+   }
+   else 
+   {
+      vcache->base.run = vcache_check_run;
+   }
+
    vcache->input_prim = prim;
    vcache->output_prim = draw_pt_reduced_prim(prim);
 
    vcache->middle = middle;
-   middle->prepare( middle, vcache->output_prim, opt );
+   vcache->opt = opt;
+
+   /* Have to run prepare here, but try and guess a good prim for
+    * doing so:
+    */
+   vcache->middle_prim = (opt & PT_PIPELINE) ? vcache->output_prim : vcache->input_prim;
+   middle->prepare( middle, vcache->middle_prim, opt, &vcache->fetch_max );
 }
 
 
 
 
-static void vcache_finish( struct draw_pt_front_end *frontend )
+static void 
+vcache_finish( struct draw_pt_front_end *frontend )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
    vcache->middle->finish( vcache->middle );
    vcache->middle = NULL;
 }
 
-static void vcache_destroy( struct draw_pt_front_end *frontend )
+static void 
+vcache_destroy( struct draw_pt_front_end *frontend )
 {
    FREE(frontend);
 }