Mesa (main): turnip: Move binning decisions from FB usage time to FB creation time.

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 6c37b4ded1006850c4b92799c9e0c6f9823eb45f
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=6c37b4ded1006850c4b92799c9e0c6f9823eb45f

Author: Emma Anholt <emma at anholt.net>
Date:   Mon May 16 13:18:06 2022 -0700

turnip: Move binning decisions from FB usage time to FB creation time.

This is mostly about helping me understand which choices are constant for the object as opposed to runtime decisions.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16826>

---

 src/freedreno/vulkan/tu_cmd_buffer.c | 34 ++++++++--------------------------
 src/freedreno/vulkan/tu_private.h    |  6 ++++++
 src/freedreno/vulkan/tu_util.c       | 32 +++++++++++++++++++++++++++++++-
 3 files changed, 45 insertions(+), 27 deletions(-)

diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index 5fd8c16ce7c..17ae859f621 100644
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -556,41 +556,23 @@ tu_cs_emit_draw_state(struct tu_cs *cs, uint32_t id, struct tu_draw_state state)
    tu_cs_emit_qw(cs, state.iova);
 }
 
-static bool
-is_hw_binning_possible(const struct tu_framebuffer *fb)
-{
-   /* Similar to older gens, # of tiles per pipe cannot be more than 32.
-    * But there are no hangs with 16 or more tiles per pipe in either
-    * X or Y direction, so that limit does not seem to apply.
-    */
-   uint32_t tiles_per_pipe = fb->pipe0.width * fb->pipe0.height;
-   return tiles_per_pipe <= 32;
-}
-
 static bool
 use_hw_binning(struct tu_cmd_buffer *cmd)
 {
    const struct tu_framebuffer *fb = cmd->state.framebuffer;
 
-   /* XFB commands are emitted for BINNING || SYSMEM, which makes it incompatible
-    * with non-hw binning GMEM rendering. this is required because some of the
-    * XFB commands need to only be executed once
+   /* XFB commands are emitted for BINNING || SYSMEM, which makes it
+    * incompatible with non-hw binning GMEM rendering. this is required because
+    * some of the XFB commands need to only be executed once.
+    * use_sysmem_rendering() should have made sure we only ended up here if no
+    * XFB was used.
     */
    if (cmd->state.xfb_used) {
-      assert(is_hw_binning_possible(fb));
+      assert(fb->binning_possible);
       return true;
    }
 
-   if (!is_hw_binning_possible(fb))
-      return false;
-
-   if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
-      return false;
-
-   if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN))
-      return true;
-
-   return (fb->tile_count.width * fb->tile_count.height) > 2;
+   return fb->binning;
 }
 
 static bool
@@ -619,7 +601,7 @@ use_sysmem_rendering(struct tu_cmd_buffer *cmd,
       return true;
 
    /* XFB is incompatible with non-hw binning GMEM rendering, see use_hw_binning */
-   if (cmd->state.xfb_used && !is_hw_binning_possible(cmd->state.framebuffer))
+   if (cmd->state.xfb_used && !cmd->state.framebuffer->binning_possible)
       return true;
 
    if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_GMEM))
diff --git a/src/freedreno/vulkan/tu_private.h b/src/freedreno/vulkan/tu_private.h
index 2c9f66b71e5..2e8110c93b2 100644
--- a/src/freedreno/vulkan/tu_private.h
+++ b/src/freedreno/vulkan/tu_private.h
@@ -1816,6 +1816,12 @@ struct tu_framebuffer
    /* number of VSC pipes */
    VkExtent2D pipe_count;
 
+   /* Whether binning should be used for gmem rendering using this framebuffer. */
+   bool binning;
+
+   /* Whether binning could be used for gmem rendering using this framebuffer. */
+   bool binning_possible;
+
    /* pipe register values */
    uint32_t pipe_config[MAX_VSC_PIPES];
    uint32_t pipe_sizes[MAX_VSC_PIPES];
diff --git a/src/freedreno/vulkan/tu_util.c b/src/freedreno/vulkan/tu_util.c
index 2d54af74e61..5be9100702b 100644
--- a/src/freedreno/vulkan/tu_util.c
+++ b/src/freedreno/vulkan/tu_util.c
@@ -208,6 +208,35 @@ tu_tiling_config_update_pipes(struct tu_framebuffer *fb,
           sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
 }
 
+static bool
+is_hw_binning_possible(const struct tu_framebuffer *fb)
+{
+   /* Similar to older gens, # of tiles per pipe cannot be more than 32.
+    * But there are no hangs with 16 or more tiles per pipe in either
+    * X or Y direction, so that limit does not seem to apply.
+    */
+   uint32_t tiles_per_pipe = fb->pipe0.width * fb->pipe0.height;
+   return tiles_per_pipe <= 32;
+}
+
+static void
+tu_tiling_config_update_binning(struct tu_framebuffer *fb, const struct tu_device *device)
+{
+   fb->binning_possible = is_hw_binning_possible(fb);
+
+   if (fb->binning_possible) {
+      fb->binning = (fb->tile_count.width * fb->tile_count.height) > 2;
+
+      if (unlikely(device->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN))
+         fb->binning = true;
+      if (unlikely(device->physical_device->instance->debug_flags &
+                   TU_DEBUG_NOBIN))
+         fb->binning = false;
+   } else {
+      fb->binning = false;
+   }
+}
+
 void
 tu_framebuffer_tiling_config(struct tu_framebuffer *fb,
                              const struct tu_device *device,
@@ -216,6 +245,7 @@ tu_framebuffer_tiling_config(struct tu_framebuffer *fb,
    tu_tiling_config_update_tile_layout(fb, device, pass);
    tu_tiling_config_update_pipe_layout(fb, device);
    tu_tiling_config_update_pipes(fb, device);
+   tu_tiling_config_update_binning(fb, device);
 }
 
 void
@@ -268,4 +298,4 @@ tu_dbg_log_gmem_load_store_skips(struct tu_device *device)
    last_total_stores = current_total_stores;
 
    pthread_mutex_unlock(&device->submit_mutex);
-}
\ No newline at end of file
+}