st/mesa: rewrite the primitive restart fallback code

Previously we were mapping/unmapping the index buffer each time we
found the restart index in the buffer.  This is bad when the restart
index is frequently used.  Now just map the index buffer once, scan
it to produce a list of sub-primitives, unmap the buffer, then draw
the sub-primitives.

Also, clean up the logic of testing for indexed primitives and calling
handle_fallback_primitive_restart().  Don't call it for non-indexed
primitives.

v2: per Jose, only map the relevant part of the index buffer with
pipe_buffer_map_range()

Reviewed-by: José Fonseca <jfonseca@vmware.com>

1 files changed, 145 insertions, 101 deletions

diff --git a/src/mesa/state_tracker/st_draw.c b/src/mesa/state_tracker/st_draw.c
index cb518e1afb3..05a71d35c23 100644
--- a/src/mesa/state_tracker/st_draw.c
+++ b/src/mesa/state_tracker/st_draw.c
@@ -648,45 +648,108 @@ check_uniforms(struct gl_context *ctx)
    }
 }
 
-/** Helper code for primitive restart fallback */
-#define DO_DRAW(pipe, cur_start, cur_count) \
-   do { \
-      info.start = cur_start; \
-      info.count = cur_count; \
-      if (u_trim_pipe_prim(info.mode, &info.count)) { \
-         if (transfer) \
-            pipe_buffer_unmap(pipe, transfer); \
-         pipe->draw_vbo(pipe, &info); \
-         if (transfer) { \
-            ptr = pipe_buffer_map(pipe, ibuffer->buffer, PIPE_TRANSFER_READ, &transfer); \
-            assert(ptr != NULL); \
-            ptr = ADD_POINTERS(ptr, ibuffer->offset); \
-         } \
-      } \
-   } while(0)
-      
-/** More helper code for primitive restart fallback */
-#define PRIM_RESTART_LOOP(elements) \
-   do { \
-      for (i = start; i < end; i++) { \
-         if (elements[i] == info.restart_index) { \
-            if (cur_count > 0) { \
-               /* draw elts up to prev pos */ \
-               DO_DRAW(pipe, cur_start, cur_count); \
-            } \
-            /* begin new prim at next elt */ \
-            cur_start = i + 1; \
-            cur_count = 0; \
-         } \
-         else { \
-            cur_count++; \
+
+/*
+ * Notes on primitive restart:
+ * The code below is used when the gallium driver does not support primitive
+ * restart itself.  We map the index buffer, find the restart indexes, unmap
+ * the index buffer then draw the sub-primitives delineated by the restarts.
+ * A couple possible optimizations:
+ * 1. Save the list of sub-primitive (start, count) values in a list attached
+ *    to the index buffer for re-use in subsequent draws.  The list would be
+ *    invalidated when the contents of the buffer changed.
+ * 2. If drawing triangle strips or quad strips, create a new index buffer
+ *    that uses duplicated vertices to render the disjoint strips as one
+ *    long strip.  We'd have to be careful to avoid using too much memory
+ *    for this.
+ * Finally, some apps might perform better if they don't use primitive restart
+ * at all rather than this fallback path.  Set MESA_EXTENSION_OVERRIDE to
+ * "-GL_NV_primitive_restart" to test that.
+ */
+
+
+struct sub_primitive
+{
+   unsigned start, count;
+};
+
+
+/**
+ * Scan the elements array to find restart indexes.  Return a list
+ * of primitive (start,count) pairs to indicate how to draw the sub-
+ * primitives delineated by the restart index.
+ */
+static struct sub_primitive *
+find_sub_primitives(const void *elements, unsigned element_size,
+                    unsigned start, unsigned end, unsigned restart_index,
+                    unsigned *num_sub_prims)
+{
+   const unsigned max_prims = end - start;
+   struct sub_primitive *sub_prims;
+   unsigned i, cur_start, cur_count, num;
+
+   sub_prims = (struct sub_primitive *)
+      malloc(max_prims * sizeof(struct sub_primitive));
+
+   if (!sub_prims) {
+      *num_sub_prims = 0;
+      return NULL;
+   }
+
+   cur_start = start;
+   cur_count = 0;
+   num = 0;
+
+#define SCAN_ELEMENTS(TYPE) \
+   for (i = start; i < end; i++) { \
+      if (((const TYPE *) elements)[i] == restart_index) { \
+         if (cur_count > 0) { \
+            assert(num < max_prims); \
+            sub_prims[num].start = cur_start; \
+            sub_prims[num].count = cur_count; \
+            num++; \
          } \
+         cur_start = i + 1; \
+         cur_count = 0; \
       } \
-      if (cur_count > 0) { \
-         DO_DRAW(pipe, cur_start, cur_count); \
+      else { \
+         cur_count++; \
       } \
-   } while (0)
+   } \
+   if (cur_count > 0) { \
+      assert(num < max_prims); \
+      sub_prims[num].start = cur_start; \
+      sub_prims[num].count = cur_count; \
+      num++; \
+   }
+
+   switch (element_size) {
+   case 1:
+      SCAN_ELEMENTS(ubyte);
+      break;
+   case 2:
+      SCAN_ELEMENTS(ushort);
+      break;
+   case 4:
+      SCAN_ELEMENTS(uint);
+      break;
+   default:
+      assert(0 && "bad index_size in find_sub_primitives()");
+   }
+
+#undef SCAN_ELEMENTS
+
+   *num_sub_prims = num;
+
+   return sub_prims;
+}
+
 
+/**
+ * For gallium drivers that don't support the primitive restart
+ * feature, handle it here by breaking up the indexed primitive into
+ * sub-primitives.
+ */
 static void
 handle_fallback_primitive_restart(struct pipe_context *pipe,
                                   const struct _mesa_index_buffer *ib,
@@ -695,78 +758,61 @@ handle_fallback_primitive_restart(struct pipe_context *pipe,
 {
    const unsigned start = orig_info->start;
    const unsigned count = orig_info->count;
-   const unsigned end = start + count;
    struct pipe_draw_info info = *orig_info;
    struct pipe_transfer *transfer = NULL;
-   unsigned instance, i, cur_start, cur_count;
-   const void *ptr;
+   unsigned instance, i;
+   const void *ptr = NULL;
+   struct sub_primitive *sub_prims;
+   unsigned num_sub_prims;
 
-   info.primitive_restart = FALSE;
-
-   if (!info.indexed) {
-      /* Splitting the draw arrays call is handled by the VBO module */
-      if (u_trim_pipe_prim(info.mode, &info.count))
-         pipe->draw_vbo(pipe, &info);
+   assert(info.indexed);
+   assert(ibuffer->buffer);
+   assert(ib);
 
+   if (!ibuffer->buffer || !ib)
       return;
-   }
 
-   /* info.indexed == TRUE */
-   assert(ibuffer);
-   assert(ibuffer->buffer);
+   info.primitive_restart = FALSE;
+   info.instance_count = 1;
 
-   if (ib) {
-      struct gl_buffer_object *bufobj = ib->obj;
-      if (bufobj && bufobj->Name) {
-         ptr = NULL;
-      }
-      else {
-         ptr = ib->ptr;
-      }
-   } else {
-      ptr = NULL;
+   if (ib->obj && _mesa_is_bufferobj(ib->obj)) {
+      ptr = pipe_buffer_map_range(pipe, ibuffer->buffer,
+                                  start * ibuffer->index_size, /* start */
+                                  count * ibuffer->index_size, /* length */
+                                  PIPE_TRANSFER_READ, &transfer);
+   }
+   else {
+      ptr = ib->ptr;
    }
 
    if (!ptr)
-      ptr = pipe_buffer_map(pipe, ibuffer->buffer, PIPE_TRANSFER_READ, &transfer);
+      return;
 
-   if (!ptr)
-     return;
    ptr = ADD_POINTERS(ptr, ibuffer->offset);
 
-   /* Need to loop over instances as well to preserve draw order */
+   sub_prims = find_sub_primitives(ptr, ibuffer->index_size,
+                                   0, count, orig_info->restart_index,
+                                   &num_sub_prims);
+
+   if (transfer)
+      pipe_buffer_unmap(pipe, transfer);
+
+   /* Now draw the sub primitives.
+    * Need to loop over instances as well to preserve draw order.
+    */
    for (instance = 0; instance < orig_info->instance_count; instance++) {
       info.start_instance = instance + orig_info->start_instance;
-      info.instance_count = 1;
-      cur_start = start;
-      cur_count = 0;
-
-      switch (ibuffer->index_size) {
-      case 1:
-         {
-            const ubyte *elt_ub = (const ubyte *)ptr; 
-            PRIM_RESTART_LOOP(elt_ub);
-         }
-         break;
-      case 2:
-         {
-            const ushort *elt_us = (const ushort *)ptr;
-            PRIM_RESTART_LOOP(elt_us);
-         }
-         break;
-      case 4:
-         {
-            const uint *elt_ui = (const uint *)ptr;
-            PRIM_RESTART_LOOP(elt_ui);
+      for (i = 0; i < num_sub_prims; i++) {
+         info.start = sub_prims[i].start;
+         info.count = sub_prims[i].count;
+         if (u_trim_pipe_prim(info.mode, &info.count)) {
+            pipe->draw_vbo(pipe, &info);
          }
-         break;
-      default:
-         assert(0 && "bad index_size in handle_fallback_primitive_restart()");
       }
    }
 
-   if (transfer)
-      pipe_buffer_unmap(pipe, transfer);
+   if (sub_prims)
+      free(sub_prims);
 }
 
 
@@ -978,10 +1024,13 @@ st_draw_vbo(struct gl_context *ctx,
          info.min_index = min_index;
          info.max_index = max_index;
       }
-   }
 
-   info.primitive_restart = ctx->Array.PrimitiveRestart;
-   info.restart_index = ctx->Array.RestartIndex;
+      /* The VBO module handles restart for the non-indexed GLDrawArrays
+       * so we only set these fields for indexed drawing:
+       */
+      info.primitive_restart = ctx->Array.PrimitiveRestart;
+      info.restart_index = ctx->Array.RestartIndex;
+   }
 
    /* do actual drawing */
    for (i = 0; i < nr_prims; i++) {
@@ -996,19 +1045,14 @@ st_draw_vbo(struct gl_context *ctx,
       }
 
       if (info.primitive_restart) {
-         /*
-          * Handle primitive restart for drivers that doesn't support it.
-          *
-          * The VBO module handles restart inside of draw_arrays for us,
-          * but we should still remove the primitive_restart flag on the
-          * info struct, the fallback function does this for us. Just
-          * remove the flag for all drivers in this case as well.
-          */
-         if (st->sw_primitive_restart || !info.indexed)
+         if (st->sw_primitive_restart) {
+            /* Handle primitive restart for drivers that doesn't support it */
             handle_fallback_primitive_restart(pipe, ib, &ibuffer, &info);
-         else
+         }
+         else {
             /* don't trim, restarts might be inside index list */
             pipe->draw_vbo(pipe, &info);
+         }
       }
       else if (u_trim_pipe_prim(info.mode, &info.count))
          pipe->draw_vbo(pipe, &info);