anv: Only flush render target cache when detecting RT changes

We setup an empty render target when there are no color attachments,
which effectively makes it a different surface state. In most cases
the compiler will insert a null-rt bit in the extended descriptor
which means the RT isn't even accessed. But in some cases like
alpha-to-coverage output + depth/stencil write, we will access the
render target because using the null-rt will prevent alpha-to-coverage
from happening.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: 2bd304bc8f ("anv: Skip the RT flush when doing depth-only rendering.")
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/31196>

src/intel/vulkan/anv_private.h

@@ -3822,6 +3822,12 @@ struct anv_cmd_graphics_state {
    struct anv_attachment stencil_att;
    struct anv_state null_surface_state;
 
+   /* Bitfield of color attachments disabled by a pipeline (pointing a null
+    * surface state in the last emitted binding table for the fragment
+    * stage)
+    */
+   uint8_t disabled_color_atts;
+
    anv_cmd_dirty_mask_t dirty;
    uint32_t vb_dirty;
 

src/intel/vulkan/genX_cmd_buffer.c

@@ -5176,15 +5176,24 @@ void genX(CmdBeginRendering)(
    };
 
    const uint32_t color_att_count = pRenderingInfo->colorAttachmentCount;
+
    result = anv_cmd_buffer_init_attachments(cmd_buffer, color_att_count);
    if (result != VK_SUCCESS)
       return;
 
    genX(flush_pipeline_select_3d)(cmd_buffer);
 
+   UNUSED bool render_target_change = false;
    for (uint32_t i = 0; i < gfx->color_att_count; i++) {
-      if (pRenderingInfo->pColorAttachments[i].imageView == VK_NULL_HANDLE)
+      if (pRenderingInfo->pColorAttachments[i].imageView == VK_NULL_HANDLE) {
+         render_target_change |= gfx->color_att[i].iview != NULL;
+
+         gfx->color_att[i].vk_format = VK_FORMAT_UNDEFINED;
+         gfx->color_att[i].iview = NULL;
+         gfx->color_att[i].layout = VK_IMAGE_LAYOUT_UNDEFINED;
+         gfx->color_att[i].aux_usage = ISL_AUX_USAGE_NONE;
          continue;
+      }
 
       const VkRenderingAttachmentInfo *att =
          &pRenderingInfo->pColorAttachments[i];
@@ -5211,6 +5220,13 @@ void genX(CmdBeginRendering)(
                                  att->imageLayout,
                                  cmd_buffer->queue_family->queueFlags);
 
+      render_target_change |= gfx->color_att[i].iview != iview;
+
+      gfx->color_att[i].vk_format = iview->vk.format;
+      gfx->color_att[i].iview = iview;
+      gfx->color_att[i].layout = att->imageLayout;
+      gfx->color_att[i].aux_usage = aux_usage;
+
       union isl_color_value fast_clear_color = { .u32 = { 0, } };
 
       if (att->loadOp == VK_ATTACHMENT_LOAD_OP_CLEAR &&
@@ -5325,11 +5341,6 @@ void genX(CmdBeginRendering)(
          assert(att->imageLayout == initial_layout);
       }
 
-      gfx->color_att[i].vk_format = iview->vk.format;
-      gfx->color_att[i].iview = iview;
-      gfx->color_att[i].layout = att->imageLayout;
-      gfx->color_att[i].aux_usage = aux_usage;
-
       struct isl_view isl_view = iview->planes[0].isl;
       if (pRenderingInfo->viewMask) {
          assert(isl_view.array_len >= util_last_bit(pRenderingInfo->viewMask));
@@ -5606,14 +5617,7 @@ void genX(CmdBeginRendering)(
    gfx->dirty |= ANV_CMD_DIRTY_PIPELINE;
 
 #if GFX_VER >= 11
-   bool has_color_att = false;
+   if (render_target_change) {
-   for (uint32_t i = 0; i < gfx->color_att_count; i++) {
-      if (pRenderingInfo->pColorAttachments[i].imageView != VK_NULL_HANDLE) {
-         has_color_att = true;
-         break;
-      }
-   }
-   if (has_color_att) {
       /* The PIPE_CONTROL command description says:
       *
       *    "Whenever a Binding Table Index (BTI) used by a Render Target Message

src/intel/vulkan/genX_cmd_draw.c

@@ -687,6 +687,60 @@ genX(emit_ds)(struct anv_cmd_buffer *cmd_buffer)
 #endif
 }
 
+ALWAYS_INLINE static void
+cmd_buffer_maybe_flush_rt_writes(struct anv_cmd_buffer *cmd_buffer,
+                                 struct anv_graphics_pipeline *pipeline)
+{
+#if GFX_VER >= 11
+   if (!anv_pipeline_has_stage(pipeline, MESA_SHADER_FRAGMENT))
+      return;
+
+   const struct anv_shader_bin *shader =
+      pipeline->base.shaders[MESA_SHADER_FRAGMENT];
+   const struct anv_pipeline_bind_map *bind_map = &shader->bind_map;
+
+   unsigned s;
+   uint8_t disabled_mask = 0;
+   for (s = 0; s < bind_map->surface_count; s++) {
+      const struct anv_pipeline_binding *binding =
+         &bind_map->surface_to_descriptor[s];
+
+      if (binding->set != ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS)
+         break;
+
+      if (binding->index >= cmd_buffer->state.gfx.color_att_count)
+         disabled_mask |= BITFIELD_BIT(s);
+   }
+
+   const uint8_t diff_mask = (1u << s) - 1;
+
+   if ((cmd_buffer->state.gfx.disabled_color_atts & diff_mask) != disabled_mask) {
+      cmd_buffer->state.gfx.disabled_color_atts = disabled_mask |
+         (cmd_buffer->state.gfx.disabled_color_atts & ~diff_mask);
+
+      /* The PIPE_CONTROL command description says:
+       *
+       *    "Whenever a Binding Table Index (BTI) used by a Render Target Message
+       *     points to a different RENDER_SURFACE_STATE, SW must issue a Render
+       *     Target Cache Flush by enabling this bit. When render target flush
+       *     is set due to new association of BTI, PS Scoreboard Stall bit must
+       *     be set in this packet."
+       *
+       * Within a renderpass, the render target entries in the binding tables
+       * remain the same as what was setup at CmdBeginRendering() with one
+       * exception where have to setup a null render target because a fragment
+       * writes only depth/stencil yet the renderpass has been setup with at
+       * least one color attachment. This is because our render target messages
+       * in the shader always send the color.
+       */
+      anv_add_pending_pipe_bits(cmd_buffer,
+                                ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT |
+                                ANV_PIPE_STALL_AT_SCOREBOARD_BIT,
+                                "change RT due to shader outputs");
+   }
+#endif
+}
+
 ALWAYS_INLINE static void
 genX(cmd_buffer_flush_gfx_state)(struct anv_cmd_buffer *cmd_buffer)
 {
@@ -708,16 +762,18 @@ genX(cmd_buffer_flush_gfx_state)(struct anv_cmd_buffer *cmd_buffer)
 
    genX(flush_pipeline_select_3d)(cmd_buffer);
 
+   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE) {
       /* Wa_14015814527
        *
        * Apply task URB workaround when switching from task to primitive.
        */
-   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE) {
       if (anv_pipeline_is_primitive(pipeline)) {
          genX(apply_task_urb_workaround)(cmd_buffer);
       } else if (anv_pipeline_has_stage(pipeline, MESA_SHADER_TASK)) {
          cmd_buffer->state.gfx.used_task_shader = true;
       }
+
+      cmd_buffer_maybe_flush_rt_writes(cmd_buffer, pipeline);
    }
 
    /* Apply any pending pipeline flushes we may have.  We want to apply them