i965: Reimplement ARB_transform_feedback2 on Haswell and later.

My old implementation accumulated <start, end> pairs in a buffer,
and eventually processed that data on the CPU.  This meant flushing
the batchbuffer and waiting for it to completely execute before we
could map it, resulting in really long stalls.  We could also run out
of space in the buffer, and have to do this early.

Instead, we can use Haswell's MI_MATH command to do the (end - start)
subtraction, as well as the multiplication by 2 or 3 to convert from
the number of primitives written to the number of vertices written.
We still need to CS stall to read the counters, but otherwise everything
is completely pipelined - there's no CPU<->GPU synchronization required.
It also uses only 80 bytes in the buffer, no matter what.

Improves performance in Manhattan on Skylake GT3e at 800x600 by
6.1086% +/- 0.954166% (n=9).  At 1920x1080, improves performance
by 2.82103% +/- 0.148596% (n=84).

v2: Fix number of primitives -> number of vertices calculation for
    GL_TRIANGLES (I was multiplying by 4 instead of 3.)  Caught by
    Jordan Justen.

Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jordan Justen <[email protected]>

1 files changed, 30 insertions, 8 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_draw.c b/src/mesa/drivers/dri/i965/brw_draw.c
index afa8a4e9eae..9d034cfdb33 100644
--- a/src/mesa/drivers/dri/i965/brw_draw.c
+++ b/src/mesa/drivers/dri/i965/brw_draw.c
@@ -153,7 +153,9 @@ trim(GLenum prim, GLuint length)
 static void
 brw_emit_prim(struct brw_context *brw,
               const struct _mesa_prim *prim,
-              uint32_t hw_prim)
+              uint32_t hw_prim,
+              struct brw_transform_feedback_object *xfb_obj,
+              unsigned stream)
 {
    int verts_per_instance;
    int vertex_access_type;
@@ -185,7 +187,7 @@ brw_emit_prim(struct brw_context *brw,
       verts_per_instance = prim->count;
 
    /* If nothing to emit, just return. */
-   if (verts_per_instance == 0 && !prim->is_indirect)
+   if (verts_per_instance == 0 && !prim->is_indirect && !xfb_obj)
       return;
 
    /* If we're set to always flush, do it before and after the primitive emit.
@@ -197,7 +199,25 @@ brw_emit_prim(struct brw_context *brw,
       brw_emit_mi_flush(brw);
 
    /* If indirect, emit a bunch of loads from the indirect BO. */
-   if (prim->is_indirect) {
+   if (xfb_obj) {
+      indirect_flag = GEN7_3DPRIM_INDIRECT_PARAMETER_ENABLE;
+
+      brw_load_register_mem(brw, GEN7_3DPRIM_VERTEX_COUNT,
+                            xfb_obj->prim_count_bo,
+                            I915_GEM_DOMAIN_VERTEX, 0,
+                            stream * sizeof(uint32_t));
+      BEGIN_BATCH(9);
+      OUT_BATCH(MI_LOAD_REGISTER_IMM | (9 - 2));
+      OUT_BATCH(GEN7_3DPRIM_INSTANCE_COUNT);
+      OUT_BATCH(prim->num_instances);
+      OUT_BATCH(GEN7_3DPRIM_START_VERTEX);
+      OUT_BATCH(0);
+      OUT_BATCH(GEN7_3DPRIM_BASE_VERTEX);
+      OUT_BATCH(0);
+      OUT_BATCH(GEN7_3DPRIM_START_INSTANCE);
+      OUT_BATCH(0);
+      ADVANCE_BATCH();
+   } else if (prim->is_indirect) {
       struct gl_buffer_object *indirect_buffer = brw->ctx.DrawIndirectBuffer;
       drm_intel_bo *bo = intel_bufferobj_buffer(brw,
             intel_buffer_object(indirect_buffer),
@@ -382,6 +402,8 @@ brw_try_draw_prims(struct gl_context *ctx,
                    const struct _mesa_index_buffer *ib,
                    GLuint min_index,
                    GLuint max_index,
+                   struct brw_transform_feedback_object *xfb_obj,
+                   unsigned stream,
                    struct gl_buffer_object *indirect)
 {
    struct brw_context *brw = brw_context(ctx);
@@ -531,7 +553,7 @@ retry:
 	 brw_upload_render_state(brw);
       }
 
-      brw_emit_prim(brw, &prims[i], brw->primitive);
+      brw_emit_prim(brw, &prims[i], brw->primitive, xfb_obj, stream);
 
       brw->no_batch_wrap = false;
 
@@ -573,14 +595,14 @@ brw_draw_prims(struct gl_context *ctx,
                GLboolean index_bounds_valid,
                GLuint min_index,
                GLuint max_index,
-               struct gl_transform_feedback_object *unused_tfb_object,
+               struct gl_transform_feedback_object *gl_xfb_obj,
                unsigned stream,
                struct gl_buffer_object *indirect)
 {
    struct brw_context *brw = brw_context(ctx);
    const struct gl_client_array **arrays = ctx->Array._DrawArrays;
-
-   assert(unused_tfb_object == NULL);
+   struct brw_transform_feedback_object *xfb_obj =
+      (struct brw_transform_feedback_object *) gl_xfb_obj;
 
    if (!brw_check_conditional_render(brw))
       return;
@@ -619,7 +641,7 @@ brw_draw_prims(struct gl_context *ctx,
     * to it.
     */
    brw_try_draw_prims(ctx, arrays, prims, nr_prims, ib, min_index, max_index,
-                      indirect);
+                      xfb_obj, stream, indirect);
 }
 
 void