atom based state upload - wippbo+blit+atoms

1 files changed, 58 insertions, 8 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_state_upload.c b/src/mesa/drivers/dri/i965/brw_state_upload.c
index a579781a85..d312f753c2 100644
--- a/src/mesa/drivers/dri/i965/brw_state_upload.c
+++ b/src/mesa/drivers/dri/i965/brw_state_upload.c
@@ -37,6 +37,9 @@
 #include "intel_batchbuffer.h"
 #include "intel_buffers.h"
 
+static uint64_t mesa_flag_to_atoms[64] = { 0 };
+static uint64_t brw_flag_to_atoms[64] = { 0 };
+
 static const struct brw_tracked_state *gen4_atoms[] =
 {
    &brw_vs_prog, /* must do before GS prog, state base address. */
@@ -373,6 +376,7 @@ void brw_init_state( struct brw_context *brw )
       num_atoms = ARRAY_SIZE(gen4_atoms);
    }
 
+   assert(num_atoms < 64);
    brw->atoms = atoms;
    brw->num_atoms = num_atoms;
 
@@ -382,6 +386,20 @@ void brw_init_state( struct brw_context *brw )
       atoms++;
    }
 
+   for (int i = 0; i < brw->num_atoms; i++) {
+      const struct brw_tracked_state *atom = brw->atoms[i];
+      int f;
+      for (GLuint flags = atom->dirty.mesa; flags; flags &= ~(1ull << f)) {
+         f = ffs(flags) - 1;
+         mesa_flag_to_atoms[f] |= 1ull << i;
+      }
+
+      for (uint64_t flags = atom->dirty.brw; flags; flags &= ~(1ull << f)) {
+         f = ffsll(flags) - 1;
+         brw_flag_to_atoms[f] |= 1ull << i;
+      }
+   }
+
    brw_upload_initial_gpu_state(brw);
 
    brw->state.dirty.mesa = ~0;
@@ -547,6 +565,24 @@ brw_print_dirty_count(struct dirty_bit_map *bit_map)
    }
 }
 
+static uint64_t
+get_dirty_atoms(struct brw_state_flags *state)
+{
+   int f;
+   uint64_t dirty_atoms = 0;
+   for (GLuint flags = state->mesa; flags; flags &= ~(1ull << f)) {
+      f = ffs(flags) - 1;
+      dirty_atoms |= mesa_flag_to_atoms[f];
+   }
+
+   for (uint64_t flags = state->brw; flags; flags &= ~(1ull << f)) {
+      f = ffsll(flags) - 1;
+      dirty_atoms |= brw_flag_to_atoms[f];
+   }
+
+   return dirty_atoms;
+}
+
 /***********************************************************************
  * Emit all state:
  */
@@ -624,14 +660,28 @@ void brw_upload_state(struct brw_context *brw)
 	 assert(!check_state(&examined, &generated));
 	 prev = *state;
       }
-   }
-   else {
-      for (i = 0; i < brw->num_atoms; i++) {
-	 const struct brw_tracked_state *atom = brw->atoms[i];
-
-	 if (check_state(state, &atom->dirty)) {
-	    atom->emit(brw);
-	 }
+   } else {
+      uint64_t dirty_atoms = get_dirty_atoms(state);
+      uint64_t emitted_atoms_mask = ~0ull;
+      struct brw_state_flags last_state = *state;
+      while (dirty_atoms & emitted_atoms_mask) {
+         int a = ffsll(dirty_atoms & emitted_atoms_mask) - 1;
+         brw->atoms[a]->emit(brw);
+         emitted_atoms_mask &= ~(1ull << a);
+         if (unlikely(state->mesa != last_state.mesa || state->brw != last_state.brw)) {
+            last_state.mesa ^= state->mesa;
+            last_state.brw ^= state->brw;
+            const uint64_t new_dirty_atoms = get_dirty_atoms(&last_state);
+
+            /* Make sure an older atom wasn't flagged as dirty. If it was,
+             * then it would indicate that the state atoms aren't ordered
+             * correctly.
+             */
+            assert(unlikely(INTEL_DEBUG && (new_dirty_atoms & ((1ull << (a + 1)) - 1) == 0)));
+
+            dirty_atoms |= get_dirty_atoms(&last_state) & ~((1ull << (a + 1)) - 1);
+            last_state = *state;
+         }
       }
    }