across separate render encoders within a single Metal command buffer (Vulkan submit).
Add MVKCommandEncodingContext to track information across multiple
MVKCommandEncoders, and use it to track temporary visibility buffer and offset.
Add support for more than one temporary visibility buffer per MTLCommandBuffer
when current temporary visibility buffer is exhausted.
offsets across separate render encoders within a single Vulkan command buffer.
Monotonically increase the offset into the temporary visibility buffer
after each query in a MVKCommandEncoder (ie- Vulkan command buffer).
e1ac50c0 marks temporary buffers as volatile, but it only mlocks() them for the duration
of memcpy() of initial data, leaving Metal runtime a window to snatch it before the buffer
is no longer in use. Simply putting the munlock() in a completion handler raises questions
about possible edge cases where multiple, non page-aligned allocations are live.
This patch fixes it by keeping the whole buffer non-volatile while is supports any active
allocations.
Fixes rendering issues in Sekiro: Shadows Die Twice menus.
MVKDeviceMemory track original Vulkan memory coherency request, not just Metal
coherency status.
MVKDeviceMemory map() and unmap() only triggers host flush if coherent Vulkan memory
was requested, not just when Metal memory is coherent, because this can be changed
for dedicated image resources.
Redesign MVKDeviceMemory flushToDevice() and pullFromDevice()
to be consistent operation with each other.
Don't trigger automatic texture flush during vkCmdCopyImage(), as it can
cause an unexpected update from mapped memory, resulting in a race condition.
MVKBeginRenderPass::setContents() use MVKSmallVector::assign()
to add attachments and clear colors.
MVKFramebuffer add getAttachments() and remove getAttachment()
and getAttachmentCount().
Add mvkCmdBeginRenderPass() and call from both
vkCmdBeginRenderPass() and vkCmdBeginRenderPass2KHR().
MVKCmdBeginRenderPassBase and MVKCmdBeginRenderPassBase
constructors pass VkSubpassBeginInfo.
Remove overloaded MVKCmdBeginRenderPassBase::setContent().
Don't allow memoryless storage for VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT.
Memoryless storage for VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT descriptors
requires support for, and use of, shader framebuffer fetch.
Remove compile-time checking for MVK_APPLE_SILICON before disabling memoryless storage
for buffers and images, since the requirement exists for all platforms, and memoryless
storage will already otherwise be disabled on other platforms anyway.
Remove framebuffer reference in MVKCommandEncoder. Instead,
1) non-imageless: copy the attachment vector from
VkFramebufferCreateInfo if it does not contain the
VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT_KHR flag.
or
2) imageless: copy the attachment vector from
VkRenderPassAttachmentBeginInfo when calling
vkBeginRenderPass/vkBeginRenderPass2.
Extend MVKCommandUse enum to identify additional command buffer uses.
MVKQueueCommandBufferSubmission track use type.
MVKQueue pass use type in functions that result in creation of a MTLCommandBuffer.
Add mvkMTLCommandBufferLabel() to map use types to label strings.
On command buffer submission failure, if MVKConfiguration::resumeLostDevice enabled,
do not release waits on VkDevice, and do not return VK_ERROR_DEVICE_LOST, unless
VkPhysicalDevice is also lost.
MVKQueueCommandBufferSubmission retain() and release() fence and signal semaphores
to ensure they live long enough for the submission to finish using them.
Previously we had loosened the behavior so that function proc addresses are returned
for an extension function even if that extension is disabled, if the API was artificially
restricted during testing, to bypass CTS tests that appeared to assume certain
extension functions were available.
This loosening in turn caused the dEQP-VK.api.version_check.entry_points test to fail,
so the restriction is reapplied and only functions from enabled extensions are published.
mvkMTLSamplerAddressModeFromVkSamplerAddressMode() support
MTLSamplerAddressModeClampToBorderColor and
MTLSamplerAddressModeMirrorClampToEdge if available.
Add MVKSampler::getMTLSamplerAddressMode() to test for device feature availability.
Remove special border color handling in MVKSampler::newMTLSamplerDescriptor().
Adjust VkPhysicalDeviceLimits values:
- minMemoryMapAlignment raised to MIN of 64 per Vulkan spec.
- pointSizeRange upper limit to 511 per Metal.
- lineWidthGranularity to 0 per Vulkan spec when wideLines disabled.
- maxSamplerLodBias set to 4 to cover setting in shader.
to each other to find cached shader, only consider current shader stage resources.
Rename more references to ShaderConverterContext to ShaderConversionConfig.
Separate reference counting from MVKVulkanAPIObject into a new
MVKReferenceCountingMixin class, so reference counting can
potentially be used in other areas of the class hierarchy.
Replace MVKConfigurableObject with MVKConfigurableMixin for the same reason.
Remove some inline definitions as unnecessary in the files being revised here.
Rename kMVKShaderStageMax to kMVKShaderStageCount
to clarify it's common use in array sizing and for loops.
Change mvkConfig() to return reference instead of pointer
to improve syntactic sugar to use . instead of ->.
Update MoltenVK version to 1.1.4.
Update What's New document.
I believe SPIRV_CROSS_NAMESPACE_OVERRIDE should be replaced by SPIRV_CROSS_NAMESPACE defined in SPIRV-Cross spirv_common.hpp. This ensures that SPIRV_CROSS_NAMESPACE_OVERRIDE usage in MVK is optional.
