Fix barriers

5 files changed, 38 insertions, 27 deletions

diff --git a/include/vk-api.hpp b/include/vk-api.hpp
index 1127bd7..864632e 100644
--- a/include/vk-api.hpp
+++ b/include/vk-api.hpp
@@ -14,7 +14,7 @@ struct buffer {
     VkDeviceAddress     device_address;
 };
 
-struct Pipeline {
+struct pipeline {
     std::vector<VkShaderModule> shader_modules;
     VkPipeline                  handle;
     VkPipelineBindPoint         bind_point;
@@ -136,9 +136,9 @@ class vkapi {
         void destroy_framebuffers(std::vector<VkFramebuffer>& framebuffers);
 
 
-        Pipeline create_compute_pipeline(const char* shader_name);
-        Pipeline create_graphics_pipeline(const char* shader_name, VkShaderStageFlagBits shader_stages, VkRenderPass render_pass, std::vector<VkDynamicState> dynamic_states);
-        void destroy_pipeline(Pipeline &pipeline);
+        pipeline create_compute_pipeline(const char* shader_name);
+        pipeline create_graphics_pipeline(const char* shader_name, VkShaderStageFlagBits shader_stages, VkRenderPass render_pass, std::vector<VkDynamicState> dynamic_states);
+        void destroy_pipeline(pipeline &pipeline);
 
 
         std::vector<VkDescriptorSet> create_descriptor_sets(VkDescriptorSetLayout descriptor_sets_layout, size_t descriptor_sets_count);
@@ -160,13 +160,13 @@ class vkapi {
 
         void image_barrier(VkCommandBuffer command_buffer, VkImageLayout src_layout, VkImageLayout dst_layout, VkPipelineStageFlagBits src_stage, VkPipelineStageFlagBits dst_stage, VkAccessFlags src_access, VkAccessFlags dst_access, image image);
 
-        void begin_render_pass(VkCommandBuffer command_buffer, VkRenderPass render_pass, VkFramebuffer framebuffer, VkExtent2D size, Pipeline pipeline);
+        void begin_render_pass(VkCommandBuffer command_buffer, VkRenderPass render_pass, VkFramebuffer framebuffer, VkExtent2D size, pipeline pipeline);
         void end_render_pass(VkCommandBuffer command_buffer);
 
-        void run_compute_pipeline(VkCommandBuffer command_buffer, Pipeline pipeline, size_t group_count_x, size_t group_count_y, size_t group_count_z);
+        void run_compute_pipeline(VkCommandBuffer command_buffer, pipeline pipeline, size_t group_count_x, size_t group_count_y, size_t group_count_z);
 
-        void draw(VkCommandBuffer command_buffer, Pipeline pipeline, uint32_t vertex_count, uint32_t vertex_offset);
-        void draw(VkCommandBuffer command_buffer, Pipeline pipeline, buffer index_buffer, uint32_t primitive_count, uint32_t indices_offset, uint32_t vertices_offset);
+        void draw(VkCommandBuffer command_buffer, pipeline pipeline, uint32_t vertex_count, uint32_t vertex_offset);
+        void draw(VkCommandBuffer command_buffer, pipeline pipeline, buffer index_buffer, uint32_t primitive_count, uint32_t indices_offset, uint32_t vertices_offset);
 
         void blit_full(VkCommandBuffer command_buffer, image src_image, image dst_image);
 
diff --git a/include/vk-renderer.hpp b/include/vk-renderer.hpp
index 99521a6..2ec9ef8 100644
--- a/include/vk-renderer.hpp
+++ b/include/vk-renderer.hpp
@@ -54,16 +54,16 @@ class vkrenderer {
 
         std::vector<VkCommandBuffer>    command_buffers;
 
-        Pipeline                        compute_pipeline;
+        pipeline                        compute_pipeline;
 
         buffer                          staging_buffer;
 
 
-        Pipeline                        tonemapping_pipeline;
+        pipeline                        tonemapping_pipeline;
 
         std::vector<buffer>             ui_vertex_buffers;
         std::vector<buffer>             ui_index_buffers;
-        Pipeline                        ui_pipeline;
+        pipeline                        ui_pipeline;
         VkRenderPass                    render_pass;
         std::vector<VkFramebuffer>      framebuffers;
 
@@ -75,7 +75,7 @@ class vkrenderer {
         swapchain                       swapchain;
 
         const uint32_t                  min_swapchain_image_count = 3;
-        const uint32_t                  virtual_frames_count = 2;
+        static constexpr uint32_t       virtual_frames_count = 2;
 };
 
 #endif // !__VK_RENDERER_HPP_
diff --git a/shaders/compute.comp b/shaders/compute.comp
index c8257cf..6ca193e 100644
--- a/shaders/compute.comp
+++ b/shaders/compute.comp
@@ -270,6 +270,9 @@ ray generate_camera_ray(ivec2 coords, uint seed) {
 }
 
 void main() {
+    if (any(greaterThanEqual(gl_GlobalInvocationID.xy, uvec2(bufs.scene.width, bufs.scene.height))))
+        return;
+
     ivec2 coords = ivec2(gl_GlobalInvocationID.xy) * bufs.scene.downscale_factor;
 
     uint seed = uint(coords.x * uint(1973) + coords.y * uint(9277) + bufs.scene.sample_index * uint(26699)) | uint(1);
diff --git a/src/vk-api.cpp b/src/vk-api.cpp
index 4bfae42..020efc6 100644
--- a/src/vk-api.cpp
+++ b/src/vk-api.cpp
@@ -442,7 +442,7 @@ VkRenderPass vkapi::create_render_pass(std::vector<VkFormat>& color_attachments_
         attachment_description.stencilLoadOp                            = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
         attachment_description.stencilStoreOp                           = VK_ATTACHMENT_STORE_OP_DONT_CARE;
         attachment_description.initialLayout                            = initial_layout;
-        attachment_description.finalLayout                              = final_layout;
+        attachment_description.finalLayout                              = initial_layout;
 
         attachments_description.push_back(attachment_description);
 
@@ -527,8 +527,8 @@ void vkapi::destroy_framebuffers(std::vector<VkFramebuffer>& framebuffers) {
 }
 
 
-Pipeline vkapi::create_compute_pipeline(const char* shader_name) {
-    Pipeline pipeline {
+pipeline vkapi::create_compute_pipeline(const char* shader_name) {
+    pipeline pipeline {
         .bind_point = VK_PIPELINE_BIND_POINT_COMPUTE
     };
 
@@ -579,8 +579,8 @@ Pipeline vkapi::create_compute_pipeline(const char* shader_name) {
     return std::move(pipeline);
 }
 
-Pipeline vkapi::create_graphics_pipeline(const char* shader_name, VkShaderStageFlagBits shader_stages, VkRenderPass render_pass, std::vector<VkDynamicState> dynamic_states) {
-    Pipeline pipeline {
+pipeline vkapi::create_graphics_pipeline(const char* shader_name, VkShaderStageFlagBits shader_stages, VkRenderPass render_pass, std::vector<VkDynamicState> dynamic_states) {
+    pipeline pipeline {
         .bind_point = VK_PIPELINE_BIND_POINT_GRAPHICS
     };
 
@@ -707,7 +707,7 @@ Pipeline vkapi::create_graphics_pipeline(const char* shader_name, VkShaderStageF
     return pipeline;
 }
 
-void vkapi::destroy_pipeline(Pipeline &pipeline) {
+void vkapi::destroy_pipeline(pipeline &pipeline) {
     for (auto& shader_module: pipeline.shader_modules) {
         vkDestroyShaderModule(context.device, shader_module, nullptr);
     }
@@ -986,7 +986,7 @@ void vkapi::image_barrier(VkCommandBuffer command_buffer, VkImageLayout src_layo
 }
 
 
-void vkapi::begin_render_pass(VkCommandBuffer command_buffer, VkRenderPass render_pass, VkFramebuffer framebuffer, VkExtent2D size, Pipeline pipeline) {
+void vkapi::begin_render_pass(VkCommandBuffer command_buffer, VkRenderPass render_pass, VkFramebuffer framebuffer, VkExtent2D size, pipeline pipeline) {
     VkRect2D render_area = {
         {
             0,
@@ -1014,7 +1014,7 @@ void vkapi::end_render_pass(VkCommandBuffer command_buffer) {
 }
 
 
-void vkapi::run_compute_pipeline(VkCommandBuffer command_buffer, Pipeline pipeline, size_t group_count_x, size_t group_count_y, size_t group_count_z) {
+void vkapi::run_compute_pipeline(VkCommandBuffer command_buffer, pipeline pipeline, size_t group_count_x, size_t group_count_y, size_t group_count_z) {
     vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_COMPUTE, bindless_descriptor.pipeline_layout, 0, 1, &bindless_descriptor.set, 0, nullptr);
 
     vkCmdBindPipeline(command_buffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline.handle);
@@ -1022,14 +1022,14 @@ void vkapi::run_compute_pipeline(VkCommandBuffer command_buffer, Pipeline pipeli
     vkCmdDispatch(command_buffer, group_count_x, group_count_y, group_count_z);
 }
 
-void vkapi::draw(VkCommandBuffer command_buffer, Pipeline pipeline, uint32_t vertex_count, uint32_t vertex_offset) {
+void vkapi::draw(VkCommandBuffer command_buffer, pipeline pipeline, uint32_t vertex_count, uint32_t vertex_offset) {
     vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, bindless_descriptor.pipeline_layout, 0, 1, &bindless_descriptor.set, 0, nullptr);
     vkCmdBindPipeline(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline.handle);
 
     vkCmdDraw(command_buffer, vertex_count, 1, vertex_offset, 0);
 }
 
-void vkapi::draw(VkCommandBuffer command_buffer, Pipeline pipeline, buffer index_buffer, uint32_t index_count, uint32_t index_offset, uint32_t vertex_offset) {
+void vkapi::draw(VkCommandBuffer command_buffer, pipeline pipeline, buffer index_buffer, uint32_t index_count, uint32_t index_offset, uint32_t vertex_offset) {
     vkCmdBindIndexBuffer(command_buffer, index_buffer.handle, 0, VK_INDEX_TYPE_UINT16);
     vkCmdBindDescriptorSets(command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, bindless_descriptor.pipeline_layout, 0, 1, &bindless_descriptor.set, 0, nullptr);
 
diff --git a/src/vk-renderer.cpp b/src/vk-renderer.cpp
index 76185d0..7dd566f 100644
--- a/src/vk-renderer.cpp
+++ b/src/vk-renderer.cpp
@@ -213,17 +213,17 @@ void vkrenderer::reset_accumulation() {
     auto cmd_buf = command_buffers[virtual_frame_index];
     VkClearColorValue clear_color = { 0.f, 0.f, 0.f, 1.f };
 
-    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, accumulation_images[0]);
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, VK_ACCESS_TRANSFER_WRITE_BIT, accumulation_images[0]);
 
     vkCmdClearColorImage(cmd_buf, accumulation_images[0].handle, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &clear_color, 1, &accumulation_images[0].subresource_range);
 
-    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, accumulation_images[0]);
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, accumulation_images[0]);
 
-    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, accumulation_images[1]);
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, VK_ACCESS_TRANSFER_WRITE_BIT, accumulation_images[1]);
 
     vkCmdClearColorImage(cmd_buf, accumulation_images[1].handle, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, &clear_color, 1, &accumulation_images[1].subresource_range);
 
-    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, accumulation_images[1]);
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, accumulation_images[1]);
 }
 
 void vkrenderer::ui() {
@@ -288,6 +288,10 @@ void vkrenderer::ui() {
 void vkrenderer::compute(uint32_t width, uint32_t height) {
     auto cmd_buf = command_buffers[virtual_frame_index];
 
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, swapchain.images[swapchain_image_index]);
+    
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, accumulation_images[1]);
+
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 0, 8, (void*)&scene_buffer_addr);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 8, 8, (void*)&bvh_buffer_addr);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 16, 8, (void*)&indices_buffer_addr);
@@ -297,11 +301,13 @@ void vkrenderer::compute(uint32_t width, uint32_t height) {
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 48, 8, (void*)&materials_buffer_addr);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 56, 4, (void*)&accumulation_images[0].bindless_storage_index);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 60, 4, (void*)&accumulation_images[1].bindless_storage_index);
-    api.run_compute_pipeline(cmd_buf, compute_pipeline, (width / 8) + 1, (height / 8) + 1, 1);
 
+    api.run_compute_pipeline(cmd_buf, compute_pipeline, (width / 8) + 1, (height / 8) + 1, 1);
 
     api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, swapchain.images[swapchain_image_index]);
 
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, accumulation_images[0]);
+
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 0, 8, (void*)&scene_buffer_addr);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 56, 4, (void*)&swapchain.images[swapchain_image_index].bindless_storage_index);
     vkCmdPushConstants(cmd_buf, api.bindless_descriptor.pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, 60, 4, (void*)&accumulation_images[0].bindless_storage_index);
@@ -312,6 +318,8 @@ void vkrenderer::compute(uint32_t width, uint32_t height) {
 void vkrenderer::finish_frame() {
     auto cmd_buf = command_buffers[virtual_frame_index];
 
+    api.image_barrier(cmd_buf, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_ACCESS_SHADER_WRITE_BIT, 0, swapchain.images[swapchain_image_index]);
+
     api.end_record(cmd_buf);
 
     api.submit(cmd_buf, acquire_semaphores[virtual_frame_index], execution_semaphores[virtual_frame_index], submission_fences[virtual_frame_index]);