Add GPU command creation

8 files changed, 122 insertions, 8 deletions

diff --git a/source/blender/draw/CMakeLists.txt b/source/blender/draw/CMakeLists.txt
index 377baeb8361..d0735540068 100644
--- a/source/blender/draw/CMakeLists.txt
+++ b/source/blender/draw/CMakeLists.txt
@@ -487,6 +487,7 @@ set(GLSL_SRC
   intern/shaders/common_subdiv_vbo_sculpt_data_comp.glsl
   intern/shaders/common_view_clipping_lib.glsl
   intern/shaders/common_view_lib.glsl
+  intern/shaders/draw_command_generate_comp.glsl
   intern/shaders/draw_debug_draw_display_frag.glsl
   intern/shaders/draw_debug_draw_display_vert.glsl
   intern/shaders/draw_debug_info.hh
@@ -496,6 +497,7 @@ set(GLSL_SRC
   intern/shaders/draw_visibility_comp.glsl
 
   intern/draw_common_shader_shared.h
+  intern/draw_command_shared.hh
   intern/draw_shader_shared.h
   intern/draw_defines.h
 
diff --git a/source/blender/draw/intern/draw_command.hh b/source/blender/draw/intern/draw_command.hh
index 39c66d2ad6f..a73d35e48bf 100644
--- a/source/blender/draw/intern/draw_command.hh
+++ b/source/blender/draw/intern/draw_command.hh
@@ -529,16 +529,16 @@ class DrawMultiBuf {
 
   void bind(Vector<Header> &, Vector<Undetermined> &, ResourceIdBuf &resource_id_buf)
   {
-    uint prefix_sum = 0u;
+    uint instance_sum = 0u;
     for (DrawGroup &group : group_buf_) {
       /* Compute prefix sum of all instance of previous group. */
-      group.start = prefix_sum;
-      prefix_sum += group.len;
+      group.start = instance_sum;
+      instance_sum += group.len;
 
       int batch_inst_len;
       /* Now that GPUBatches are guaranteed to be finished, extract their parameters. */
       GPU_batch_draw_parameter_get(group.gpu_batch, &group.vertex_len, &batch_inst_len);
-      /* Tag group as using index draw (changes indirect drawcall structure). */
+      /* Tag group as using index draw (changes indirect draw call structure). */
       if (group.gpu_batch->elem != nullptr) {
         group.vertex_len = -group.vertex_len;
       }
@@ -552,8 +552,10 @@ class DrawMultiBuf {
     }
 
     group_buf_.push_update();
+    prototype_buf_.push_update();
     /* Allocate enough for the expansion pass. */
-    resource_id_buf.get_or_resize(prefix_sum);
+    resource_id_buf.get_or_resize(instance_sum);
+    command_buf_.get_or_resize(group_count_);
 
     // GPU_compute_dispatch(resource_id_expand_shader, n, 1, 1);
   }
diff --git a/source/blender/draw/intern/draw_command_shared.hh b/source/blender/draw/intern/draw_command_shared.hh
index f73e38cb8cc..afd3e06deb7 100644
--- a/source/blender/draw/intern/draw_command_shared.hh
+++ b/source/blender/draw/intern/draw_command_shared.hh
@@ -49,15 +49,15 @@ struct DrawGroup {
       /** GPUBatch values to be copied to DrawCommand after sorting (if not overriden). */
       int vertex_len; /** NOTE: Negative if using indexed draw. */
       /** Atomic counters used during command sorting. */
-      uint total_counter;
       uint front_facing_counter;
       uint back_facing_counter;
+      uint total_counter;
 #ifndef GPU_SHADER
     };
   };
 #endif
 };
-BLI_STATIC_ASSERT(sizeof(DrawGroup) == 32, "DrawGroup might not have the same size on GPU and CPU")
+BLI_STATIC_ASSERT_ALIGN(DrawGroup, 16)
 
 /**
  * Representation of a future draw call inside a DrawGroup. This #DrawPrototype is then
@@ -73,6 +73,7 @@ struct DrawPrototype {
   uint instance_len;
   uint _pad0;
 };
+BLI_STATIC_ASSERT_ALIGN(DrawPrototype, 16)
 
 /** \} */
 
diff --git a/source/blender/draw/intern/draw_defines.h b/source/blender/draw/intern/draw_defines.h
index a8731dedaf8..2769f5f71d1 100644
--- a/source/blender/draw/intern/draw_defines.h
+++ b/source/blender/draw/intern/draw_defines.h
@@ -18,6 +18,7 @@
 #define DRW_OBJ_INFOS_SLOT 5
 #define DRW_OBJ_ATTR_SLOT 4
 
+#define DRW_COMMAND_GROUP_SIZE 64
 #define DRW_FINALIZE_GROUP_SIZE 64
 /* Must be multiple of 32. Set to 64 for optimal scheduling on GCN hardware. */
-#define DRW_VISIBILITY_GROUP_SIZE 64
\ No newline at end of file
+#define DRW_VISIBILITY_GROUP_SIZE 64
diff --git a/source/blender/draw/intern/draw_shader.cc b/source/blender/draw/intern/draw_shader.cc
index 24ecd4efe00..960348b4a94 100644
--- a/source/blender/draw/intern/draw_shader.cc
+++ b/source/blender/draw/intern/draw_shader.cc
@@ -25,6 +25,7 @@ static struct {
   struct GPUShader *debug_draw_display_sh;
   struct GPUShader *draw_visibility_compute_sh;
   struct GPUShader *draw_resource_finalize_sh;
+  struct GPUShader *draw_command_generate_sh;
 } e_data = {{nullptr}};
 
 /* -------------------------------------------------------------------- */
@@ -143,6 +144,14 @@ GPUShader *DRW_shader_draw_resource_finalize_get()
   return e_data.draw_resource_finalize_sh;
 }
 
+GPUShader *DRW_shader_draw_command_generate_get()
+{
+  if (e_data.draw_command_generate_sh == nullptr) {
+    e_data.draw_command_generate_sh = GPU_shader_create_from_info_name("draw_command_generate");
+  }
+  return e_data.draw_command_generate_sh;
+}
+
 /** \} */
 
 void DRW_shaders_free()
@@ -154,4 +163,5 @@ void DRW_shaders_free()
   DRW_SHADER_FREE_SAFE(e_data.debug_draw_display_sh);
   DRW_SHADER_FREE_SAFE(e_data.draw_visibility_compute_sh);
   DRW_SHADER_FREE_SAFE(e_data.draw_resource_finalize_sh);
+  DRW_SHADER_FREE_SAFE(e_data.draw_command_generate_sh);
 }
diff --git a/source/blender/draw/intern/draw_shader.h b/source/blender/draw/intern/draw_shader.h
index 77c3410eb98..3b8c0425fa9 100644
--- a/source/blender/draw/intern/draw_shader.h
+++ b/source/blender/draw/intern/draw_shader.h
@@ -34,6 +34,7 @@ struct GPUShader *DRW_shader_debug_print_display_get(void);
 struct GPUShader *DRW_shader_debug_draw_display_get(void);
 struct GPUShader *DRW_shader_draw_visibility_compute_get(void);
 struct GPUShader *DRW_shader_draw_resource_finalize_get(void);
+struct GPUShader *DRW_shader_draw_command_generate_get(void);
 
 void DRW_shaders_free(void);
 
diff --git a/source/blender/draw/intern/shaders/draw_command_generate_comp.glsl b/source/blender/draw/intern/shaders/draw_command_generate_comp.glsl
new file mode 100644
index 00000000000..2c736996d60
--- /dev/null
+++ b/source/blender/draw/intern/shaders/draw_command_generate_comp.glsl
@@ -0,0 +1,84 @@
+
+/**
+ * Convert DrawPrototype into draw commands.
+ */
+
+#pragma BLENDER_REQUIRE(common_math_lib.glsl)
+
+#define atomicAddAndGet(dst, val) (atomicAdd(dst, val) + val)
+
+/* This is only called by the last thread executed over the group's prototype draws. */
+void write_draw_call(DrawGroup group, uint group_id)
+{
+  DrawCommand cmd;
+  cmd.vertex_len = abs(group.vertex_len);
+  cmd.vertex_first = 0;
+  if (group.vertex_len < 0) {
+    cmd.base_index = 0;
+    cmd.instance_first_indexed = group.start;
+  }
+  else {
+    cmd._instance_first_array = group.start;
+  }
+  /* Back-facing command. */
+  cmd.instance_len = group_buf[group_id].back_facing_counter;
+  command_buf[group_id * 2 + 0] = cmd;
+  /* Front-facing command. */
+  cmd.instance_len = group_buf[group_id].front_facing_counter;
+  command_buf[group_id * 2 + 1] = cmd;
+
+  /* Reset the counters for a next command gen dispatch. Avoids resending the whole data just
+   * for this purpose. Only the last thread will execute this so it is threadsafe.  */
+  group_buf[group_id].front_facing_counter = 0u;
+  group_buf[group_id].back_facing_counter = 0u;
+  group_buf[group_id].total_counter = 0u;
+}
+
+void main()
+{
+  uint proto_id = gl_GlobalInvocationID.x;
+  if (proto_id >= prototype_len) {
+    return;
+  }
+
+  DrawPrototype proto = prototype_buf[proto_id];
+  uint group_id = proto.group_id;
+  bool is_inverted = (proto.resource_handle & 0x80000000u) != 0;
+  uint resource_index = (proto.resource_handle & 0x7FFFFFFFu);
+
+  /* Visibility test result. */
+  bool is_visible = ((visibility_buf[resource_index / 32u] & (1u << (resource_index % 32u)))) != 0;
+
+  DrawGroup group = group_buf[group_id];
+
+  if (!is_visible || proto.instance_len == 0) {
+    /* Skip the draw but still count towards the completion. */
+    if (atomicAddAndGet(group_buf[group_id].total_counter, proto.instance_len) == group.len) {
+      write_draw_call(group, group_id);
+    }
+    return;
+  }
+
+  uint back_facing_len = group.len - group.front_facing_len;
+  uint front_facing_len = group.front_facing_len;
+  uint dst_index = group.start;
+  if (is_inverted) {
+    uint offset = atomicAdd(group_buf[group_id].back_facing_counter, proto.instance_len);
+    dst_index += offset;
+    if (atomicAddAndGet(group_buf[group_id].total_counter, proto.instance_len) == group.len) {
+      write_draw_call(group, group_id);
+    }
+  }
+  else {
+    uint offset = atomicAdd(group_buf[group_id].front_facing_counter, proto.instance_len);
+    dst_index += back_facing_len + offset;
+    if (atomicAddAndGet(group_buf[group_id].total_counter, proto.instance_len) == group.len) {
+      write_draw_call(group, group_id);
+    }
+  }
+
+  for (uint i = dst_index; i < dst_index + proto.instance_len; i++) {
+    /* Fill resource_id buffer for each instance of this draw */
+    resource_id_buf[i] = resource_index;
+  }
+}
\ No newline at end of file
diff --git a/source/blender/draw/intern/shaders/draw_view_info.hh b/source/blender/draw/intern/shaders/draw_view_info.hh
index 6149812a644..e2c6a806ac2 100644
--- a/source/blender/draw/intern/shaders/draw_view_info.hh
+++ b/source/blender/draw/intern/shaders/draw_view_info.hh
@@ -162,6 +162,19 @@ GPU_SHADER_CREATE_INFO(draw_visibility_compute)
     .compute_source("draw_visibility_comp.glsl")
     .additional_info("draw_view");
 
+GPU_SHADER_CREATE_INFO(draw_command_generate)
+    .do_static_compilation(true)
+    .typedef_source("draw_shader_shared.h")
+    .typedef_source("draw_command_shared.hh")
+    .local_group_size(DRW_COMMAND_GROUP_SIZE)
+    .storage_buf(0, Qualifier::READ_WRITE, "DrawGroup", "group_buf[]")
+    .storage_buf(1, Qualifier::READ, "DrawPrototype", "prototype_buf[]")
+    .storage_buf(2, Qualifier::WRITE, "DrawCommand", "command_buf[]")
+    .storage_buf(3, Qualifier::WRITE, "uint", "resource_id_buf[]")
+    .storage_buf(4, Qualifier::READ, "uint", "visibility_buf[]")
+    .push_constant(Type::UINT, "prototype_len")
+    .compute_source("draw_command_generate_comp.glsl");
+
 /** \} */
 
 /* -------------------------------------------------------------------- */