Previously, the tests were using AcquireNextImage to trigger the fence
and using it to guard the vkFreeCommandBuffers call. However, the
acquire fence tells you when the PE is done with an image and when you
can start rendering to it. This meant that the test was waiting for the
start of the previous frame's rendering not the end before deleting the
previous frame's command buffer. This meant that the previous frame's
rendering could still be active on the GPU when the command buffer was
freed. By triggering the fence off of the vkQueueSubmit, we wait for
the end of the previous frame's rendering before freeing its command
buffer.
VK-GL-CTS issue: 756
Component: Vulkan
Affects:
dEQP-VK.wsi.*.incremental_present.*
Change-Id: I87b25eb2a25e4ae48c8c527d5d63c45d89a31a30
deUint32 imageIndex;
// Acquire next image
- VK_CHECK(m_vkd.acquireNextImageKHR(*m_device, *m_swapchain, foreverNs, currentAcquireSemaphore, fence, &imageIndex));
+ VK_CHECK(m_vkd.acquireNextImageKHR(*m_device, *m_swapchain, foreverNs, currentAcquireSemaphore, (vk::VkFence)0, &imageIndex));
// Create command buffer
{
¤tRenderSemaphore
};
- VK_CHECK(m_vkd.queueSubmit(m_queue, 1u, &submitInfo, (vk::VkFence)0));
+ VK_CHECK(m_vkd.queueSubmit(m_queue, 1u, &submitInfo, fence));
}
// Present frame