[Mesa-dev] [PATCH 02/12] i965/rbc: Set aux surface unconditionally for sampling engine

[Topi Pohjolainen]

Once mcs buffer gets allocated without delay for lossless
compression (same as we do for msaa), one gets regression in:

GL45-CTS.texture_barrier_ARB.same-texel-rw

Setting the auxiliary surface for both sampling engine and data
port seems to fix this. I haven't found any hardware documentation
backing this though.

Signed-off-by: Topi Pohjolainen <topi.pohjolainen at intel.com>
---
 src/mesa/drivers/dri/i965/brw_wm_surface_state.c | 46 ++++++++++++++++++++++--
 1 file changed, 43 insertions(+), 3 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
index bfd973b..5cd27ef 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
@@ -72,6 +72,48 @@ static const struct surface_state_info surface_state_infos[] = {
    [9] = {16, 64, 8,  10, SKL_MOCS_WB,  SKL_MOCS_PTE},
 };
 
+static bool
+brw_needs_aux_surface(const struct brw_context *brw,
+                      const struct intel_mipmap_tree *mt,
+                      const struct isl_view *view)
+{
+   if (!mt->mcs_mt)
+      return false;
+
+   if (view->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT)
+      return true;
+
+   const bool is_lossless_compressed =
+      intel_miptree_is_lossless_compressed(brw, mt);
+   const bool view_format_lossless_compressed =
+       isl_format_supports_lossless_compression(brw->intelScreen->devinfo,
+                                                view->format);
+
+   /* If the underlying surface is compressed but it is sampled using a
+    * format that the sampling engine doesn't support as compressed, there
+    * is no alternative but to treat the surface as non-compressed.
+    */
+   if (is_lossless_compressed && !view_format_lossless_compressed) {
+      /* Logic elsewhere needs to take care to resolve the color buffer prior
+       * to sampling it as non-compressed.
+       */
+      assert(mt->fast_clear_state == INTEL_FAST_CLEAR_STATE_RESOLVED);
+      return false;
+   }
+
+   /* In practise it looks that setting the same lossless compressed surface
+    * to be sampled without auxiliary surface and to be written with auxiliary
+    * surface confuses the hardware.
+    * Currently there is no direct way of knowing if render pass will use the
+    * sane surface as render target and therefore simply use the auxiliary
+    * surface unconditionally here.
+    */
+   if (is_lossless_compressed)
+      return true;
+
+   return mt->fast_clear_state != INTEL_FAST_CLEAR_STATE_RESOLVED;
+}
+
 static void
 brw_emit_surface_state(struct brw_context *brw,
                        struct intel_mipmap_tree *mt,
@@ -135,9 +177,7 @@ brw_emit_surface_state(struct brw_context *brw,
    struct isl_surf *aux_surf = NULL, aux_surf_s;
    uint64_t aux_offset = 0;
    enum isl_aux_usage aux_usage = ISL_AUX_USAGE_NONE;
-   if (mt->mcs_mt &&
-       ((view.usage & ISL_SURF_USAGE_RENDER_TARGET_BIT) ||
-        mt->fast_clear_state != INTEL_FAST_CLEAR_STATE_RESOLVED)) {
+   if (brw_needs_aux_surface(brw, mt, &view)) {
       intel_miptree_get_aux_isl_surf(brw, mt, &aux_surf_s, &aux_usage);
       aux_surf = &aux_surf_s;
       assert(mt->mcs_mt->offset == 0);
-- 
2.5.5