I am trying to do the following: Take render pass 1, render to attachment set S.
Take render pass 2, render to attachment set S again.
There is a one to one explanation on how to do this in the official Khronos github:
Under Graphics to Graphics dependency. However, trying to implement the same code as in the example isn't working.
This is what I see:
This is what I should see:
What I think is currently happening is that the depth buffer is not being synchronized properly and so depth testing isn't working correctly. I create both of my render passes like this:
// Create a render pass from the information of a render target
vk::UniqueRenderPass CreateRenderPass(HardwareInterface& hi,
const std::vector<vk::AttachmentDescription>& attachment_descriptions)
{
auto instance = hi.GetInstance();
auto phys_device = hi.GetPhysicalDevice();
auto device = hi.GetDevice();
// Create a reference for each image target the renderpass will render to
std::vector<vk::AttachmentReference> attachment_references(
attachment_descriptions.size());
for(uint i = 0; i < attachment_descriptions.size() - 1; i++)
{
vk::AttachmentReference color_attachment_ref(
i, vk::ImageLayout::eColorAttachmentOptimal);
attachment_references[i] = color_attachment_ref;
}
// Last attachment is depth
vk::AttachmentReference depth_attachment_ref(
attachment_descriptions.size() - 1,
vk::ImageLayout::eDepthStencilAttachmentOptimal);
attachment_references[attachment_references.size() - 1] = depth_attachment_ref;
const uint subpass_num = 1;
vector<vk::SubpassDescription> subpasses(subpass_num);
for(auto& sub_pass: subpasses)
{
// No input attachment (3rd and 4th parameters)
sub_pass = vk::SubpassDescription();
sub_pass.pipelineBindPoint = vk::PipelineBindPoint::eGraphics;
sub_pass.inputAttachmentCount = 0;
sub_pass.pInputAttachments = nullptr;
sub_pass.colorAttachmentCount = attachment_references.size() - 1;
sub_pass.pColorAttachments = attachment_references.data();
sub_pass.pResolveAttachments = nullptr;
sub_pass.pDepthStencilAttachment = &attachment_references.back();
}
vector<vk::SubpassDependency> dependencies(subpass_num);
uint current_dependency = 0;
for(auto& dependency: dependencies)
{
// Set the dependency of this subpass
uint32_t prev = current_dependency == 0 ? (uint32_t)VK_SUBPASS_EXTERNAL
: current_dependency - 1;
uint32_t next = current_dependency;
current_dependency++;
// Execution and memory dependencies between subpasses
dependency = vk::SubpassDependency();
dependency.srcSubpass = prev;
dependency.dstSubpass = next;
dependency.srcStageMask =
vk::PipelineStageFlagBits::eColorAttachmentOutput
| vk::PipelineStageFlagBits::eEarlyFragmentTests
| vk::PipelineStageFlagBits::eLateFragmentTests;
dependency.dstStageMask =
vk::PipelineStageFlagBits::eColorAttachmentOutput
| vk::PipelineStageFlagBits::eEarlyFragmentTests
| vk::PipelineStageFlagBits::eLateFragmentTests;
dependency.srcAccessMask =
vk::AccessFlagBits::eDepthStencilAttachmentWrite |
vk::AccessFlagBits::eDepthStencilAttachmentRead;
dependency.dstAccessMask =
vk::AccessFlagBits::eColorAttachmentRead |
vk::AccessFlagBits::eColorAttachmentWrite |
vk::AccessFlagBits::eDepthStencilAttachmentRead |
vk::AccessFlagBits::eDepthStencilAttachmentWrite;
}
vk::RenderPassCreateInfo render_pass_info(
{},
attachment_descriptions.size(),
attachment_descriptions.data(),
subpasses.size(),
subpasses.data(),
dependencies.size(),
dependencies.data());
auto [result, render_pass] = device.createRenderPassUnique(render_pass_info);
if(result != vk::Result::eSuccess)
Log::RecordLogError("Failed to create render pass!");
return move(render_pass);
}
I have read and reread the example and I have tried to make my code as identical to it as possible but alas, the issue still happens. The way I created the working image is by having both renders occur under the same render pass, but I don't want to do it that way.
The problem was being caused because when i created the attachment descriptors, I set the storeOp
and loadOp
enums of the renderpass to eDontCare
so the depth buffer was UB for me.
As suggested in the comments by @solidpixel. The solution is to make the first eStore
and the second 'eLoad'.