i915: Accumulate the VB into a local buffer and subdata it in.

This lets GEM use pwrite, for an additional 4% or so speedup.

src/mesa/drivers/dri/i915/i830_vtbl.c

@@ -677,9 +677,6 @@ i830_new_batch(struct intel_context *intel)
    struct i830_context *i830 = i830_context(&intel->ctx);
    i830->state.emitted = 0;
 
-   /* Signal that we should put new vertices into a new vertex buffer. */
-   intel->prim.needs_new_vb = GL_TRUE;
-
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!intel->no_batch_wrap);
 }
@@ -722,4 +719,5 @@ i830InitVtbl(struct i830_context *i830)
    i830->intel.vtbl.render_prevalidate = i830_render_prevalidate;
    i830->intel.vtbl.assert_not_dirty = i830_assert_not_dirty;
    i830->intel.vtbl.note_unlock = i830_note_unlock; 
+   i830->intel.vtbl.finish_batch = intel_finish_vb;
 }

src/mesa/drivers/dri/i915/i915_vtbl.c

@@ -589,8 +589,6 @@ i915_new_batch(struct intel_context *intel)
     * difficulties associated with them (physical address requirements).
     */
    i915->state.emitted = 0;
-   /* Signal that we should put new vertices into a new vertex buffer. */
-   intel->prim.needs_new_vb = GL_TRUE;
 
    /* Check that we didn't just wrap our batchbuffer at a bad time. */
    assert(!intel->no_batch_wrap);
@@ -633,4 +631,5 @@ i915InitVtbl(struct i915_context *i915)
    i915->intel.vtbl.flush_cmd = i915_flush_cmd;
    i915->intel.vtbl.assert_not_dirty = i915_assert_not_dirty;
    i915->intel.vtbl.note_unlock = i915_note_unlock; 
+   i915->intel.vtbl.finish_batch = intel_finish_vb;
 }

src/mesa/drivers/dri/i915/intel_tris.c

@@ -77,31 +77,28 @@ uint32_t *intel_get_prim_space(struct intel_context *intel, unsigned int count)
 
    /* Check for space in the existing VB */
    if (intel->prim.vb_bo == NULL ||
-       intel->prim.needs_new_vb ||
        (intel->prim.current_offset +
 	count * intel->vertex_size * 4) > INTEL_VB_SIZE ||
        (intel->prim.count + count) >= (1 << 16)) {
       /* Flush existing prim if any */
       INTEL_FIREVERTICES(intel);
 
+      intel_finish_vb(intel);
+
       /* Start a new VB */
-      dri_bo_unreference(intel->prim.vb_bo);
+      if (intel->prim.vb == NULL)
+	 intel->prim.vb = malloc(INTEL_VB_SIZE);
       intel->prim.vb_bo = dri_bo_alloc(intel->bufmgr, "vb",
 				       INTEL_VB_SIZE, 4);
       intel->prim.start_offset = 0;
       intel->prim.current_offset = 0;
 
       dri_bufmgr_check_aperture_space(intel->prim.vb_bo);
-
-      intel->prim.needs_new_vb = GL_FALSE;
-
-      dri_bo_map(intel->prim.vb_bo, GL_TRUE);
    }
 
    intel->prim.flush = intel_flush_prim;
 
-   addr = (uint32_t *)((char *)intel->prim.vb_bo->virtual +
-		       intel->prim.current_offset);
+   addr = (uint32_t *)(intel->prim.vb + intel->prim.current_offset);
    intel->prim.current_offset += intel->vertex_size * 4 * count;
    intel->prim.count += count;
 
@@ -112,6 +109,7 @@ uint32_t *intel_get_prim_space(struct intel_context *intel, unsigned int count)
 void intel_flush_prim(struct intel_context *intel)
 {
    BATCH_LOCALS;
+   dri_bo *vb_bo;
 
    /* Must be called after an intel_start_prim. */
    assert(intel->prim.primitive != ~0);
@@ -119,9 +117,13 @@ void intel_flush_prim(struct intel_context *intel)
    if (intel->prim.count == 0)
       return;
 
-   intel_wait_flips(intel);
+   /* Keep a reference on the BO as it may get finished as we start the
+    * batch emit.
+    */
+   vb_bo = intel->prim.vb_bo;
+   dri_bo_reference(vb_bo);
 
-   dri_bo_unmap(intel->prim.vb_bo);
+   intel_wait_flips(intel);
 
    intel->vtbl.emit_state(intel);
 
@@ -147,7 +149,7 @@ void intel_flush_prim(struct intel_context *intel)
       OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
 		I1_LOAD_S(0) | I1_LOAD_S(1) | 1);
       assert((intel->prim.start_offset & !S0_VB_OFFSET_MASK) == 0);
-      OUT_RELOC(intel->prim.vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
+      OUT_RELOC(vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
 		intel->prim.start_offset);
       OUT_BATCH((intel->vertex_size << S1_VERTEX_WIDTH_SHIFT) |
 		(intel->vertex_size << S1_VERTEX_PITCH_SHIFT));
@@ -167,7 +169,7 @@ void intel_flush_prim(struct intel_context *intel)
 		I1_LOAD_S(0) | I1_LOAD_S(2) | 1);
       /* S0 */
       assert((intel->prim.start_offset & !S0_VB_OFFSET_MASK_830) == 0);
-      OUT_RELOC(intel->prim.vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
+      OUT_RELOC(vb_bo, I915_GEM_DOMAIN_VERTEX, 0,
 		intel->prim.start_offset |
 		(intel->vertex_size << S0_VB_PITCH_SHIFT_830) |
 		S0_VB_ENABLE_830);
@@ -193,17 +195,35 @@ void intel_flush_prim(struct intel_context *intel)
 
    intel->no_batch_wrap = GL_FALSE;
 
-   /* If we're going to keep using this VB for more primitives, map it
-    * again.
-    */
-   if (!intel->prim.needs_new_vb)
-      dri_bo_map(intel->prim.vb_bo, GL_TRUE);
-
    intel->prim.flush = NULL;
    intel->prim.start_offset = intel->prim.current_offset;
    if (!IS_9XX(intel->intelScreen->deviceID))
       intel->prim.start_offset = ALIGN(intel->prim.start_offset, 128);
    intel->prim.count = 0;
+
+   dri_bo_unreference(vb_bo);
+}
+
+/**
+ * Uploads the locally-accumulated VB into the buffer object.
+ *
+ * This avoids us thrashing the cachelines in and out as the buffer gets
+ * filled, dispatched, then reused as the hardware completes rendering from it,
+ * and also lets us clflush less if we dispatch with a partially-filled VB.
+ *
+ * This is called normally from get_space when we're finishing a BO, but also
+ * at batch flush time so that we don't try accessing the contents of a
+ * just-dispatched buffer.
+ */
+void intel_finish_vb(struct intel_context *intel)
+{
+   if (intel->prim.vb_bo == NULL)
+      return;
+
+   dri_bo_subdata(intel->prim.vb_bo, 0, intel->prim.start_offset,
+		  intel->prim.vb);
+   dri_bo_unreference(intel->prim.vb_bo);
+   intel->prim.vb_bo = NULL;
 }
 
 /***********************************************************************

src/mesa/drivers/dri/i915/intel_tris.h

@@ -30,7 +30,7 @@
 
 #include "mtypes.h"
 
-#define INTEL_VB_SIZE		(8 * 1024)
+#define INTEL_VB_SIZE		(32 * 1024)
 /** 3 dwords of state_immediate and 2 of 3dprim, in intel_flush_prim */
 #define INTEL_PRIM_EMIT_SIZE	(5 * 4)
 
@@ -49,5 +49,6 @@ extern void intelChooseRenderState(GLcontext * ctx);
 void intel_set_prim(struct intel_context *intel, uint32_t prim);
 GLuint *intel_get_prim_space(struct intel_context *intel, unsigned int count);
 void intel_flush_prim(struct intel_context *intel);
+void intel_finish_vb(struct intel_context *intel);
 
 #endif

src/mesa/drivers/dri/intel/intel_batchbuffer.c

@@ -250,6 +250,9 @@ _intel_batchbuffer_flush(struct intel_batchbuffer *batch, const char *file,
     * avoid that in the first place. */
    batch->ptr = batch->map;
 
+   if (intel->vtbl.finish_batch)
+      intel->vtbl.finish_batch(intel);
+
    /* TODO: Just pass the relocation list and dma buffer up to the
     * kernel.
     */

src/mesa/drivers/dri/intel/intel_context.c

@@ -727,6 +727,7 @@ intelDestroyContext(__DRIcontextPrivate * driContextPriv)
       intel->Fallback = 0;      /* don't call _swrast_Flush later */
 
       intel_batchbuffer_free(intel->batch);
+      free(intel->prim.vb);
 
       if (release_texture_heaps) {
          /* This share group is about to go away, free our private

src/mesa/drivers/dri/intel/intel_context.h

@@ -86,6 +86,7 @@ struct intel_context
    {
       void (*destroy) (struct intel_context * intel);
       void (*emit_state) (struct intel_context * intel);
+      void (*finish_batch) (struct intel_context * intel);
       void (*new_batch) (struct intel_context * intel);
       void (*emit_invarient_state) (struct intel_context * intel);
       void (*note_fence) (struct intel_context *intel, GLuint fence);
@@ -185,16 +186,10 @@ struct intel_context
       uint32_t primitive;	/**< Current hardware primitive type */
       void (*flush) (struct intel_context *);
       dri_bo *vb_bo;
+      uint8_t *vb;
       unsigned int start_offset; /**< Byte offset of primitive sequence */
       unsigned int current_offset; /**< Byte offset of next vertex */
       unsigned int count;	/**< Number of vertices in current primitive */
-      /**
-       * Signals when a new VB should be started, regardless of remaining
-       * space.
-       *
-       * Used to avoid rewriting a VB that's being rendered from.
-       */
-      GLboolean needs_new_vb;
    } prim;
 
    GLuint stats_wm;