v3d: emit 'Wait for transform feedback' commands when needed

The hardware can flush transform feedback writes before reads in the same
job by inserting this command.

This patch detects when the rendering state for the current draw call reads
resources that had been previously written by transform feedback in the
same job and inserts the 'Wait for transform feedback' command before
emitting the new draw.

v2 (Eric):
  - this was intended to look at job->tf_write_prscs for TF jobs.
  - clear job->tf_write_prscs after we emit the TF flush.
  - can skip flushes for fragment shader reads from TF.

v3 (Eric):
  - all resources in job->tf_write_prscs are resources written by TF so
   we don't need to check if they are bound to PIPE_BIND_STREAM_OUTPUT.
  - documented optimization opportunity for geometry stages.

Reviewed-by: Eric Anholt <eric@anholt.net>

diff --git a/src/gallium/drivers/v3d/v3dx_draw.c b/src/gallium/drivers/v3d/v3dx_draw.c
index fe07231d4744287187df0fbd99e87989e619340d..201813c69d6d2a164235ac9d3a75842107d4ced4 100644 (file)
--- a/src/gallium/drivers/v3d/v3dx_draw.c
+++ b/src/gallium/drivers/v3d/v3dx_draw.c
@@ -208,6 +208,116 @@ v3d_predraw_check_outputs(struct pipe_context *pctx)
         }
 }
 
+/**
+ * Checks if the state for the current draw reads a particular resource in
+ * in the given shader stage.
+ */
+static bool
+v3d_state_reads_resource(struct v3d_context *v3d,
+                         struct pipe_resource *prsc,
+                         enum pipe_shader_type s)
+{
+        struct v3d_resource *rsc = v3d_resource(prsc);
+
+        /* Vertex buffers */
+        if (s == PIPE_SHADER_VERTEX) {
+                foreach_bit(i, v3d->vertexbuf.enabled_mask) {
+                        struct pipe_vertex_buffer *vb = &v3d->vertexbuf.vb[i];
+                        if (!vb->buffer.resource)
+                                continue;
+
+                        struct v3d_resource *vb_rsc =
+                                v3d_resource(vb->buffer.resource);
+                        if (rsc->bo == vb_rsc->bo)
+                                return true;
+                }
+        }
+
+        /* Constant buffers */
+        foreach_bit(i, v3d->constbuf[s].enabled_mask) {
+                struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
+                if (!cb->buffer)
+                        continue;
+
+                struct v3d_resource *cb_rsc = v3d_resource(cb->buffer);
+                if (rsc->bo == cb_rsc->bo)
+                        return true;
+        }
+
+        /* Shader storage buffers */
+        foreach_bit(i, v3d->ssbo[s].enabled_mask) {
+                struct pipe_shader_buffer *sb = &v3d->ssbo[s].sb[i];
+                if (!sb->buffer)
+                        continue;
+
+                struct v3d_resource *sb_rsc = v3d_resource(sb->buffer);
+                if (rsc->bo == sb_rsc->bo)
+                        return true;
+        }
+
+        /* Textures  */
+        for (int i = 0; i < v3d->tex[s].num_textures; i++) {
+                struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
+                if (!pview)
+                        continue;
+
+                struct v3d_sampler_view *view = v3d_sampler_view(pview);
+                struct v3d_resource *v_rsc = v3d_resource(view->texture);
+                if (rsc->bo == v_rsc->bo)
+                        return true;
+        }
+
+        return false;
+}
+
+static void
+v3d_emit_wait_for_tf(struct v3d_job *job)
+{
+        /* XXX: we might be able to skip this in some cases, for now we
+         * always emit it.
+         */
+        cl_emit(&job->bcl, FLUSH_TRANSFORM_FEEDBACK_DATA, flush);
+
+        cl_emit(&job->bcl, WAIT_FOR_TRANSFORM_FEEDBACK, wait) {
+                /* XXX: Wait for all outstanding writes... maybe we can do
+                 * better in some cases.
+                 */
+                wait.block_count = 255;
+        }
+
+        /* We have just flushed all our outstanding TF work in this job so make
+         * sure we don't emit TF flushes again for any of it again.
+         */
+        _mesa_set_clear(job->tf_write_prscs, NULL);
+}
+
+static void
+v3d_emit_wait_for_tf_if_needed(struct v3d_context *v3d, struct v3d_job *job)
+{
+        if (!job->tf_enabled)
+            return;
+
+        set_foreach(job->tf_write_prscs, entry) {
+                struct pipe_resource *prsc = (struct pipe_resource *)entry->key;
+                for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
+                        /* Fragment shaders can only start executing after all
+                         * binning (and thus TF) is complete.
+                         *
+                         * XXX: For VS/GS/TES, if the binning shader does not
+                         * read the resource then we could also avoid emitting
+                         * the wait.
+                         */
+                        if (s == PIPE_SHADER_FRAGMENT)
+                            continue;
+
+                        if (v3d_state_reads_resource(v3d, prsc, s)) {
+                                v3d_emit_wait_for_tf(job);
+                                return;
+                        }
+                }
+        }
+}
+
 static void
 v3d_emit_gl_shader_state(struct v3d_context *v3d,
                          const struct pipe_draw_info *info)
@@ -596,6 +706,16 @@ v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
         v3d_update_compiled_shaders(v3d, info->mode);
         v3d_update_job_ez(v3d, job);
 
+        /* If this job was writing to transform feedback buffers before this
+         * draw and we are reading from them here, then we need to wait for TF
+         * to complete before we emit this draw.
+         *
+         * Notice this check needs to happen before we emit state for the
+         * current draw call, where we update job->tf_enabled, so we can ensure
+         * that we only check TF writes for prior draws.
+         */
+        v3d_emit_wait_for_tf_if_needed(v3d, job);
+
 #if V3D_VERSION >= 41
         v3d41_emit_state(pctx);
 #else