Prelocate the batch and state buffers, avoid most SWZ relocs.

4 files changed, 48 insertions, 1 deletions

diff --git a/src/mesa/drivers/dri/i915tex/i915_differencer.c b/src/mesa/drivers/dri/i915tex/i915_differencer.c
index 90c404fc4b..d95526ad48 100644
--- a/src/mesa/drivers/dri/i915tex/i915_differencer.c
+++ b/src/mesa/drivers/dri/i915tex/i915_differencer.c
@@ -138,12 +138,18 @@ static void emit_indirect( struct intel_context *intel,
       for (i = 0; i < I915_MAX_CACHE; i++) {
 	 if (dirty & (1<<i)) {
 	    if (i != I915_CACHE_DYNAMIC) {
+#if DO_PRELOC
+	       EMIT_DWORD( ptr, ((intel->batch->state_offset + state->offsets[i]) | 
+				 flag | 
+				 SIS0_BUFFER_VALID) );
+#else
 	       EMIT_RELOC( intel, 
 			   ptr,
 			   intel->batch->state_buffer, 
 			   intel->batch->state_memflags,
 			   DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE,
 			   ( state->offsets[i] | flag | SIS0_BUFFER_VALID ) );
+#endif
 
 	       /* No state size dword for dynamic state:
 		*/
@@ -156,13 +162,18 @@ static void emit_indirect( struct intel_context *intel,
 	    }
 	    else {
 	       assert( (state->offsets[i] & 4095) >= 4 );
-
+#if DO_PRELOC
+	       EMIT_DWORD( ptr, ((intel->batch->state_offset + state->offsets[i] - 4) | 
+				 flag | 
+				 SIS0_BUFFER_VALID) );
+#else
 	       EMIT_RELOC( intel, 
 			   ptr,
 			   intel->batch->state_buffer, 
 			   intel->batch->state_memflags,
 			   DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE,
 			   ( (state->offsets[i] - 4) | flag | SIS0_BUFFER_VALID ) );
+#endif
 
 	       assert(intel->batch->segment_start_offset[1] <= state->offsets[i]);
 	       assert(intel->batch->segment_finish_offset[1] >= state->offsets[i]);
diff --git a/src/mesa/drivers/dri/i915tex/intel_batchbuffer.c b/src/mesa/drivers/dri/i915tex/intel_batchbuffer.c
index 1c36fc12f2..c4fdefda74 100644
--- a/src/mesa/drivers/dri/i915tex/intel_batchbuffer.c
+++ b/src/mesa/drivers/dri/i915tex/intel_batchbuffer.c
@@ -205,6 +205,29 @@ intel_batchbuffer_reset(struct intel_batchbuffer *batch)
       batch->state_map = driBOMap(batch->state_buffer, DRM_BO_FLAG_WRITE, 0);
    }
 
+
+   if (DO_PRELOC) {
+      GLboolean was_locked = batch->intel->locked;
+      if (!was_locked)
+	 LOCK_HARDWARE(batch->intel);
+
+      driBOValidate(batch->intel->driFd, batch->buffer,
+                    DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE | DRM_BO_FLAG_NO_MOVE,
+                    DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE | DRM_BO_FLAG_NO_MOVE, 
+		    DRM_BO_HINT_DONT_FENCE);
+
+      driBOValidate(batch->intel->driFd, batch->state_buffer,
+                    DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE | DRM_BO_FLAG_NO_MOVE,
+                    DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE | DRM_BO_FLAG_NO_MOVE, 
+		    DRM_BO_HINT_DONT_FENCE);
+
+      if (!was_locked)
+	 UNLOCK_HARDWARE(batch->intel);
+
+      batch->state_offset = driBOOffset(batch->state_buffer);
+      batch->batch_offset = driBOOffset(batch->buffer);
+   }
+
    batch->segment_finish_offset[0] = batch->segment_start_offset[0];
    batch->segment_finish_offset[1] = batch->segment_start_offset[1];
    batch->segment_finish_offset[2] = batch->segment_start_offset[2];
diff --git a/src/mesa/drivers/dri/i915tex/intel_batchbuffer.h b/src/mesa/drivers/dri/i915tex/intel_batchbuffer.h
index 06973d82d3..7ea1c3dc86 100644
--- a/src/mesa/drivers/dri/i915tex/intel_batchbuffer.h
+++ b/src/mesa/drivers/dri/i915tex/intel_batchbuffer.h
@@ -14,6 +14,11 @@ struct intel_context;
 #define INTEL_BATCH_CLIPRECTS        0x2
 #define INTEL_BATCH_HWZ              0x4 
 
+/* Prelocations - pre-assign an offset to the batch and state buffers
+ * to eliminate most of the relocation overhead on zone rendering.
+ */
+#define DO_PRELOC 1
+
 struct buffer_reloc
 {
    GLuint slot;
@@ -41,6 +46,9 @@ struct intel_batchbuffer
    GLuint state_memtype;
    GLuint state_memflags;
 
+   GLuint batch_offset;		/* DO_PRELOC only */
+   GLuint state_offset;		/* DO_PRELOC only */
+
    struct _DriBufferList *list;
    GLuint list_count;
    GLubyte *map;
diff --git a/src/mesa/drivers/dri/i915tex/intel_swz.h b/src/mesa/drivers/dri/i915tex/intel_swz.h
index 9d3308fce6..6bc8fb3bbf 100644
--- a/src/mesa/drivers/dri/i915tex/intel_swz.h
+++ b/src/mesa/drivers/dri/i915tex/intel_swz.h
@@ -261,6 +261,10 @@ static INLINE void zone_begin_batch( struct swz_render *swz,
 
    /* Both out and newptr are assumed to be in the batchbuffer!!
     */
+#if DO_PRELOC
+   *out++ = (swz->intel->batch->batch_offset + 
+	     newptr - swz->intel->batch->map);
+#else
    intel_batchbuffer_set_reloc( swz->intel->batch,
 				SEGMENT_IMMEDIATE,
 				((GLubyte *)out) - swz->intel->batch->map,
@@ -268,6 +272,7 @@ static INLINE void zone_begin_batch( struct swz_render *swz,
 				DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE,
 				DRM_BO_MASK_MEM | DRM_BO_FLAG_EXE,			 
 				newptr - swz->intel->batch->map );
+#endif
 }