More synchronization fixes (#68)

- Fix write-after-write for animation-transforms in basic rasterization-path - Fix a read-after-write in gpu-cull - Fix write-after-write in mesh_compute module
crocdialer · Sep 14, 2024 · d1b69cf · d1b69cf
1 parent 7f5b388
commit d1b69cf
Show file tree

Hide file tree

Showing 4 changed files with 29 additions and 6 deletions.
diff --git a/src/Buffer.cpp b/src/Buffer.cpp
@@ -9,7 +9,6 @@ void barrier(VkCommandBuffer command_buffer, const VkBuffer *buffers, uint32_t n
 {
     if(buffers && num_buffers)
     {
-
         std::vector<VkBufferMemoryBarrier2> barriers;
 
         for(uint32_t i = 0; i < num_buffers; ++i)
@@ -29,7 +28,6 @@ void barrier(VkCommandBuffer command_buffer, const VkBuffer *buffers, uint32_t n
                 barriers.push_back(barrier);
             }
         }
-
         VkDependencyInfo dependency_info = {};
         dependency_info.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO;
         dependency_info.bufferMemoryBarrierCount = barriers.size();

diff --git a/src/PBRDeferred.cpp b/src/PBRDeferred.cpp
@@ -553,6 +553,12 @@ void PBRDeferred::update_animation_transforms(frame_context_t &frame_context)
     }
 
     frame_context.cmd_pre_render.begin(0);
+
+    // barriers
+    VkBuffer buffers[] = {frame_context.bone_buffer->handle(), frame_context.morph_param_buffer->handle()};
+    vierkant::barrier(frame_context.cmd_pre_render.handle(), buffers, 2,VK_PIPELINE_STAGE_2_TRANSFER_BIT,
+                      VK_ACCESS_2_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_2_TRANSFER_BIT,
+                      VK_ACCESS_2_TRANSFER_WRITE_BIT);
     vierkant::staging_copy_context_t staging_context = {};
     staging_context.staging_buffer = frame_context.staging_anim;
     staging_context.command_buffer = frame_context.cmd_pre_render.handle();

diff --git a/src/gpu_culling.cpp b/src/gpu_culling.cpp
@@ -280,12 +280,16 @@ draw_cull_result_t gpu_cull(const vierkant::gpu_cull_context_ptr &context, const
     vierkant::barrier(context->cull_cmd_buffer.handle(), draw_buffers.data(), draw_buffers.size(),
                       VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT, VK_ACCESS_2_SHADER_WRITE_BIT, src_stage, src_access);
 
-    context->draw_cull_result_buffer->barrier(context->cull_cmd_buffer.handle(), VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
-                                              VK_ACCESS_2_SHADER_WRITE_BIT, VK_PIPELINE_STAGE_2_TRANSFER_BIT,
-                                              VK_ACCESS_2_TRANSFER_READ_BIT);
+    VkBuffer count_buffers[] = {params.draws_counts_out_pre->handle(), params.draws_counts_out_post->handle()};
+    vierkant::barrier(context->cull_cmd_buffer.handle(), count_buffers, 2, VK_PIPELINE_STAGE_2_TRANSFER_BIT,
+                      VK_ACCESS_2_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT,
+                      VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT);
+
     context->draw_cull_result_buffer_host->barrier(context->cull_cmd_buffer.handle(), src_stage, src_access, dst_stage,
                                                    dst_access);
     // copy result into host-visible buffer
+    context->draw_cull_result_buffer->barrier(context->cull_cmd_buffer.handle(), src_stage, src_access,
+                                              VK_PIPELINE_STAGE_2_TRANSFER_BIT, VK_ACCESS_2_TRANSFER_READ_BIT);
     context->draw_cull_result_buffer->copy_to(context->draw_cull_result_buffer_host, context->cull_cmd_buffer.handle());
 
     if(params.query_pool)

diff --git a/src/mesh_compute.cpp b/src/mesh_compute.cpp
@@ -51,7 +51,11 @@ mesh_compute_context_handle create_mesh_compute_context(const vierkant::DevicePt
                                                       VK_COMMAND_POOL_CREATE_TRANSIENT_BIT |
                                                               VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT);
 
-    ret->cmd_buffer = vierkant::CommandBuffer(device, ret->command_pool.get());
+    vierkant::CommandBuffer::create_info_t cmd_buf_info = {};
+    cmd_buf_info.device = device;
+    cmd_buf_info.command_pool = ret->command_pool.get();
+    cmd_buf_info.name = "mesh_compute_context_t::cmd_buffer";
+    ret->cmd_buffer = vierkant::CommandBuffer(cmd_buf_info);
 
     vierkant::Buffer::create_info_t internal_buffer_info = {};
     internal_buffer_info.device = device;
@@ -60,8 +64,13 @@ mesh_compute_context_handle create_mesh_compute_context(const vierkant::DevicePt
                                  VK_BUFFER_USAGE_TRANSFER_DST_BIT;
     internal_buffer_info.mem_usage = VMA_MEMORY_USAGE_GPU_ONLY;
 
+    internal_buffer_info.name = "mesh_compute_context_t::skin_param_buffer";
     ret->skin_param_buffer = vierkant::Buffer::create(internal_buffer_info);
+
+    internal_buffer_info.name = "mesh_compute_context_t::bone_buffer";
     ret->bone_buffer = vierkant::Buffer::create(internal_buffer_info);
+
+    internal_buffer_info.name = "mesh_compute_context_t::morph_param_buffer";
     ret->morph_param_buffer = vierkant::Buffer::create(internal_buffer_info);
 
     vierkant::Buffer::create_info_t result_buffer_info = {};
@@ -256,6 +265,12 @@ mesh_compute_result_t mesh_compute(const mesh_compute_context_handle &context, c
                              params.query_index_start);
     }
 
+    // barriers
+    VkBuffer buffers[] = {context->bone_buffer->handle(), context->skin_param_buffer->handle(),
+                          context->morph_param_buffer->handle()};
+    vierkant::barrier(context->cmd_buffer.handle(), buffers, 3, VK_PIPELINE_STAGE_2_TRANSFER_BIT,
+                      VK_ACCESS_2_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_2_TRANSFER_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT);
+
     // staging copies of bones + params
     vierkant::staging_copy_context_t staging_context = {};
     staging_context.command_buffer = context->cmd_buffer.handle();