source/blender/draw/CMakeLists.txt

@@ -605,6 +605,7 @@ set(GLSL_SRC
   engines/eevee_next/shaders/eevee_thickness_lib.glsl
   engines/eevee_next/shaders/eevee_transparency_lib.glsl
   engines/eevee_next/shaders/eevee_velocity_lib.glsl
+  engines/eevee_next/shaders/eevee_vertex_copy_comp.glsl
   engines/eevee_next/shaders/eevee_volume_integration_comp.glsl
   engines/eevee_next/shaders/eevee_volume_lib.glsl
   engines/eevee_next/shaders/eevee_volume_material_comp.glsl

source/blender/draw/engines/eevee_next/eevee_defines.hh

@@ -155,6 +155,9 @@
 #define VOLUME_INTEGRATION_GROUP_SIZE 8
 #define VOLUME_HIT_DEPTH_MAX 16
 
+/* Velocity. */
+#define VERTEX_COPY_GROUP_SIZE 64
+
 /* Resource bindings. */
 
 /* Textures. */

source/blender/draw/engines/eevee_next/eevee_shader.cc

@@ -274,6 +274,8 @@ const char *ShaderModule::static_shader_create_info_name_get(eShaderType shader_
       return "eevee_surfel_list_sort";
     case SURFEL_RAY:
       return "eevee_surfel_ray";
+    case VERTEX_COPY:
+      return "eevee_vertex_copy";
     case VOLUME_INTEGRATION:
       return "eevee_volume_integration";
     case VOLUME_OCCUPANCY_CONVERT:

source/blender/draw/engines/eevee_next/eevee_shader.hh

@@ -137,6 +137,8 @@ enum eShaderType {
   SURFEL_LIST_SORT,
   SURFEL_RAY,
 
+  VERTEX_COPY,
+
   VOLUME_INTEGRATION,
   VOLUME_OCCUPANCY_CONVERT,
   VOLUME_RESOLVE,

source/blender/draw/engines/eevee_next/eevee_velocity.cc

@@ -259,13 +259,34 @@ void VelocityModule::geometry_steps_fill()
    * `tot_len * sizeof(float4)` is greater than max SSBO size. */
   geometry_steps[step_]->resize(max_ii(16, dst_ofs));
 
+  PassSimple copy_ps("Velocity Copy Pass");
+  copy_ps.init();
+  copy_ps.state_set(DRW_STATE_NO_DRAW);
+  copy_ps.shader_set(inst_.shaders.static_shader_get(VERTEX_COPY));
+  copy_ps.bind_ssbo("out_buf", *geometry_steps[step_]);
+
   for (VelocityGeometryData &geom : geometry_map.values()) {
-    GPU_storagebuf_copy_sub_from_vertbuf(*geometry_steps[step_],
-                                         geom.pos_buf,
-                                         geom.ofs * sizeof(float4),
-                                         0,
-                                         geom.len * sizeof(float4));
+    const GPUVertFormat *format = GPU_vertbuf_get_format(geom.pos_buf);
+    if (format->stride == 16) {
+      GPU_storagebuf_copy_sub_from_vertbuf(*geometry_steps[step_],
+                                           geom.pos_buf,
+                                           geom.ofs * sizeof(float4),
+                                           0,
+                                           geom.len * sizeof(float4));
+    }
+    else {
+      BLI_assert(format->stride % 4 == 0);
+      copy_ps.bind_ssbo("in_buf", geom.pos_buf);
+      copy_ps.push_constant("start_offset", geom.ofs);
+      copy_ps.push_constant("vertex_stride", int(format->stride / 4));
+      copy_ps.push_constant("vertex_count", geom.len);
+      copy_ps.dispatch(int3(divide_ceil_u(geom.len, VERTEX_COPY_GROUP_SIZE), 1, 1));
+    }
   }
+
+  copy_ps.barrier(GPU_BARRIER_SHADER_STORAGE);
+  inst_.manager->submit(copy_ps);
+
   /* Copy back the #VelocityGeometryIndex into #VelocityObjectData which are
    * indexed using persistent keys (unlike geometries which are indexed by volatile ID). */
   for (VelocityObjectData &vel : velocity_map.values()) {

source/blender/draw/engines/eevee_next/shaders/eevee_vertex_copy_comp.glsl

@@ -0,0 +1,15 @@
+/* SPDX-FileCopyrightText: 2022-2023 Blender Authors
+ *
+ * SPDX-License-Identifier: GPL-2.0-or-later */
+
+void main()
+{
+  uint vertex = gl_GlobalInvocationID.x;
+  if (vertex >= vertex_count) {
+    return;
+  }
+  out_buf[start_offset + vertex] = vec4(in_buf[vertex * vertex_stride + 0],
+                                        in_buf[vertex * vertex_stride + 1],
+                                        in_buf[vertex * vertex_stride + 2],
+                                        1.0);
+}

source/blender/draw/engines/eevee_next/shaders/infos/eevee_velocity_info.hh

@@ -38,4 +38,14 @@ GPU_SHADER_CREATE_INFO(eevee_velocity_geom)
     .fragment_out(0, Type::VEC4, "out_velocity")
     .additional_info("eevee_velocity_camera");
 
+GPU_SHADER_CREATE_INFO(eevee_vertex_copy)
+    .compute_source("eevee_vertex_copy_comp.glsl")
+    .local_group_size(VERTEX_COPY_GROUP_SIZE)
+    .storage_buf(0, Qualifier::READ, "float", "in_buf[]")
+    .storage_buf(1, Qualifier::WRITE, "vec4", "out_buf[]")
+    .push_constant(Type::INT, "start_offset")
+    .push_constant(Type::INT, "vertex_stride")
+    .push_constant(Type::INT, "vertex_count")
+    .do_static_compilation(true);
+
 /** \} */