Metal: EEVEE Next: Optimize Virtual shadow maps for Apple Silicon

Opening WIP PR for optimization and implementation of EEVEE Next's
Virtual Shadow Maps for TBDRs. The core of these optimizations lie in
eliminating use of atomic shadow atlas writes and instead utilise tile
memory to perform depth accumulation as a secondary pass once
all geometry updates for a given shadow view have been updated.

This also allows use of fast on-tile depth testing/sorting, reducing
overdraw and redundant fragment operations, while also allowing
for tile indirection calculations to be offloaded into the vertex
shader to increase fragment storage efficiency and throughput.

Authored by Apple: Michael Parkin-White

@@ -573,6 +573,8 @@ set(GLSL_SRC
   engines/eevee_next/shaders/eevee_surf_forward_frag.glsl
   engines/eevee_next/shaders/eevee_surf_lib.glsl
   engines/eevee_next/shaders/eevee_surf_shadow_frag.glsl
+  engines/eevee_next/shaders/eevee_surf_shadow_accum_vert.glsl

@@ -160,0 +165,4 @@
+
+#ifdef WITH_METAL_BACKEND
+  /* Perform accumulation step for storing final tile depth results back to shadow atlas.  */
+  accum_ps_.init();

@@ -102,6 +102,9 @@ class ShadowPipeline {
   Instance &inst_;
 
   PassMain surface_ps_ = {"Shadow.Surface"};
+#ifdef WITH_METAL_BACKEND

@@ -1175,2 +1201,4 @@
       shadow_multi_view_.compute_procedural_bounds();
 
+#ifdef WITH_METAL_BACKEND
+      /* Specify load-store config right before render to ensure flags are retained. */

@@ -259,2 +259,3 @@
   /** Framebuffer with the atlas_tx attached. */
-  Framebuffer render_fb_;
+#ifdef WITH_METAL_BACKEND
+  Framebuffer render_fb_ = {"shadow_write_framebuffer"};

@@ -12,3 +12,3 @@
 #pragma BLENDER_REQUIRE(eevee_shadow_tilemap_lib.glsl)
 
-shared bool directional_range_changed;
+shared uint directional_range_changed;

@@ -69,3 +69,3 @@
   ivec2 tile_shifted = tile_co + tilemap.grid_shift;
   /* Ensure value is shifted into positive range to avoid modulo on negative. */
-  ivec2 tile_wrapped = ivec2((ivec2(SHADOW_TILEMAP_RES) + tile_shifted) % SHADOW_TILEMAP_RES);
+  ivec2 tile_wrapped = ivec2((ivec2(SHADOW_TILEMAP_RES * SHADOW_TILEMAP_PER_ROW) + tile_shifted) %

@@ -0,0 +74,4 @@
+
+  /** Interpolate output texel  */
+  /* Using bitwise ops is way faster than integer ops. */
+  const int page_shift = SHADOW_PAGE_LOD;

@@ -0,0 +94,4 @@
+  /* NOTE: To avoid redundant writes, we still enable the depth test and configure the accumulation
+   * pass quad depth such that only the fragments updated during the surface depth pass will run.
+   *
+   * We use `DRW_STATE_DEPTH_GREATER_EQUAL` such that we only run the fragment shader if the

@@ -159,0 +214,4 @@
+    inst_.sampling.bind_resources(&tbdr_page_store_ps_);
+    inst_.bind_uniform_data(&tbdr_page_store_ps_);
+    tbdr_page_store_ps_.draw_procedural(
+        GPU_PRIM_TRIS, (SHADOW_TILEMAP_RES) * (SHADOW_TILEMAP_RES)*SHADOW_VIEW_MAX, 6);

@@ -53,2 +55,4 @@
                                                                {0, 0, 0.5, 1}};
 
+/* Technique used for updating the virtual shadow map contents. */
+enum eShadowUpdateTechnique {

@@ -148,2 +158,4 @@
 }
 
+#ifdef GPU_METAL
+/* NOTE: Metal can support multiple shadow atlas formats. E.g. U32, F32. */

@@ -0,0 +1,64 @@
+

@@ -0,0 +3,4 @@
+ * Virtual Shadow map accumulation storage.
+ *
+ * On Apple silicon, we can use a three-pass method to perform virtual shadow map updates,
+ * leveraging efficient use of tile-based GPUs.Shadow updates rasterize geometry for each view in

@@ -0,0 +57,4 @@
+  shadow_atlas_img.texture->write(u_depth, ushort2(out_texel_xy), out_page_z);
+#  endif
+
+#  ifdef SHADOW_ATLAS_F32

@@ -21,6 +21,12 @@ shared int global_max;
 
 void main()
 {
+  uint index = gl_GlobalInvocationID.x;

@@ -99,3 +99,3 @@
     /* Issue one view if there is an update in the LOD. */
     if (all(equal(gl_LocalInvocationID, uvec3(0)))) {
-      bool lod_has_update = rect_min.x < rect_max.x;
+      bool lod_has_update = rect_min_x < rect_max_x;

@@ -665,6 +665,64 @@ void DrawMultiBuf::bind(RecordingState &state,
   GPU_debug_group_end();
 }
 
+void DrawMultiBuf::bind_no_comp(RecordingState &state,

@@ -18,2 +18,3 @@
 
-  /* Clear to FLT_MAX instead of 1 so the far plane doesn't cast shadows onto farther objects. */
+/* Clear to FLT_MAX instead of 1 so the far plane doesn't cast shadows onto farther objects. */
+#ifdef GPU_METAL

@@ -479,2 +479,4 @@
   Vector<FragOut> fragment_outputs_;
 
+  using FragTileIn = FragOut;
+  Vector<FragTileIn> fragment_tile_inputs_;

@@ -130,6 +130,8 @@ void MTLFrameBuffer::bind(bool enabled_srgb)
     Shader::set_framebuffer_srgb_target(enabled_srgb && srgb_);
   }
 
+  this->mark_dirty();

@@ -22,3 +22,3 @@
 #include "gpu_py_shader.h" /* own include */
 
-//#define USE_PYGPU_SHADER_INFO_IMAGE_METHOD
+// #define USE_PYGPU_SHADER_INFO_IMAGE_METHOD