From: Jason Ekstrand <jason.ekstrand@intel.com>
Date: Wed, 15 Jul 2015 18:09:02 +0000 (-0700)
Subject: vk/device: More documentation on surface state flushing
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=e375f722a697286b235d1699bb1bf73021215ed4;p=mesa.git

vk/device: More documentation on surface state flushing
---

diff --git a/src/vulkan/device.c b/src/vulkan/device.c
index 51d729a6f49..fd96457033e 100644
--- a/src/vulkan/device.c
+++ b/src/vulkan/device.c
@@ -3044,16 +3044,42 @@ anv_cmd_buffer_new_surface_state_bo(struct anv_cmd_buffer *cmd_buffer)
     */
    anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 
-   /* The sampler unit caches SURFACE_STATE and RENDER_SURFACE_STATE entries,
-    * and the data port uses the same cache. When changing the Surface State
-    * Base Address, we need to flush the texture cache so that it can pick up
-    * on the new SURFACE_STATE's. From the Broadwell PRM,
+   /* After re-setting the surface state base address, we have to do some
+    * cache flusing so that the sampler engine will pick up the new
+    * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
     * Shared Function > 3D Sampler > State > State Caching (page 96):
     *
-    *     Whenever the value of the Dynamic_State_Base_Addr,
-    *     Surface_State_Base_Addr are altered, the L1 state cache
-    *     must be invalidated to ensure the new surface or sampler state is
-    *     fetched from system memory.
+    *    Coherency with system memory in the state cache, like the texture
+    *    cache is handled partially by software. It is expected that the
+    *    command stream or shader will issue Cache Flush operation or
+    *    Cache_Flush sampler message to ensure that the L1 cache remains
+    *    coherent with system memory.
+    *
+    *    [...]
+    *
+    *    Whenever the value of the Dynamic_State_Base_Addr,
+    *    Surface_State_Base_Addr are altered, the L1 state cache must be
+    *    invalidated to ensure the new surface or sampler state is fetched
+    *    from system memory.
+    *
+    * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
+    * which, according the PIPE_CONTROL instruction documentation in the
+    * Broadwell PRM:
+    *
+    *    Setting this bit is independent of any other bit in this packet.
+    *    This bit controls the invalidation of the L1 and L2 state caches
+    *    at the top of the pipe i.e. at the parsing time.
+    *
+    * Unfortunately, experimentation seems to indicate that state cache
+    * invalidation through a PIPE_CONTROL does nothing whatsoever in
+    * regards to surface state and binding tables.  In stead, it seems that
+    * invalidating the texture cache is what is actually needed.
+    *
+    * XXX:  As far as we have been able to determine through
+    * experimentation, shows that flush the texture cache appears to be
+    * sufficient.  The theory here is that all of the sampling/rendering
+    * units cache the binding table in the texture cache.  However, we have
+    * yet to be able to actually confirm this.
     */
    anv_batch_emit(&cmd_buffer->batch, GEN8_PIPE_CONTROL,
                   .TextureCacheInvalidationEnable = true);