Apple GPUs do not support rendering/writing to an attachment and then reading
from that attachment within a single Metal renderpass. On Apple Silicon, restart
the Metal renderpass if an input attachment is cleared inside renderpass.
Don't clear render area when restarting Metal renderpass, as it should not occur,
and itself causes a recursive loop restarting the renderpass as a result.
Add MVKCommandUse::kMVKCommandUseRestartSubpass.
MVKCommandEncoder::beginMetalRenderPass() pass MVKCommandUse to help determine
Metal renderpass attachment load and clearing options.
A recent change for resolving color formats mistakenly
broke clearing depth and stencil formats.
Remove optionality of all function arguments in
MVKRenderPassAttachment::populateMTLRenderPassAttachmentDescriptor()
to avoid bad calls being masked by optional function arguments.
Add appropriate canResolveFormat values for all calls to that function.
Rename MVKRenderPassAttachment::shouldUseClearAttachment() to
shouldClearAttachment(), pass whether testing for stencil clearing,
and check stencil clearing distinct from depth clearing.
Refactor MVKRenderSubpass::populateClearAttachments() to work with this.
Remove MVKRenderPassAttachment::getAttachmentStencilLoadOp() as obsolete.
Disable MVK_ALLOW_METAL_FENCES by default.
Disable use of MTLEvent on NVIDIA.
By default, use MTLEvent for VkSemaphore everywhere except NVIDIA.
By default, use CPU synchronization on NVIDIA.
These changes fix a large number of CTS synchronization test failures.
Prefer MTLEvent over MTLFence for VkSemaphore, because MTLEvent handles
sync across MTLCommandBuffers and MTLCommandQueues, except on NVIDIA GPUs,
which have demonstrated trouble with MTLEvents, prefer MTLFence.
Add MVKDevice::VkSemaphoreStyle enum.
Per cdavis5e observation:
Since an MVKArrayRef<T> is just a pointer and a size, passing it around by
reference doesn't buy much. It might actually hurt performance, since cache
lines might be evicted to make room for the MVKArrayRef in memory.
For this reason, LLVM passes its ArrayRefs by value.
Unrelatedly, update runcts to log errors by default.
Metal does not support resolving all formats that support MSAA, whereas Vulkan
assumes any MSAA format can be resolved. We fix that by running an optional
post-renderpass compute shader that resolves such textures by simply taking the
first sample as the resolved sample. This works to fix all failing CTS tests,
because such formats are all integer formats, and Vulkan allows an arbitrary
single sample value to be selected.
If we need to resolve, but the Metal format doesn't support it,
cause the Metal renderpass to store the MSAA attachment results.
MVKRenderSubpass don't establish Metal resolve attachment textures if format
is not natively resolvable, and encode Metal renderpass store actions accordingly.
MVKCommandEncodingPool add MTLComputePipelineStates to run simple resolve
compute shaders on attachments that cannot be resolved in Metal renderpass.
Add MVKRenderSubpass::resolveUnresolvableAttachments() and call from
MVKCommandEncoder::endMetalRenderEncoding(), before subpass index is updated.
Rename MVKCommandEncodingPool::getClearStateIndex() to
getRenderpassLoadStoreStateIndex() and remove MVK_MACOS restriction on
clearing shaders to allow compatibility with resolve shader handling.
MVKRenderPassAttachment remove validation of whether a format can be resolved.
MVKPixelFormats::getMTLTextureUsage() add read and write usage as appropriate
to allow compute shader to run to resolve formats not natively resolvable.
MVKPixelFormats remove obsolete unit test code.
MVKImageView clean up access functions and obsolete constructor
use of MVKConfiguration.
On Apple GPUs, MVKCmdPipelineBarrier restarts Metal renderpass when encountered
inside self-dependent Vulkan subpass where the same attachment acts as both
input attachment and render attachment.
Move update of MVKSwapchainImageAvailability::acquisitionID to the acquisition
time instead of the become available again time, so other images will be preferred
if either all images are available or no images are available.
See for reference spec description of VkGraphicsPipelineCreateInfo, and CTS test
dEQP-VK.api.pipeline.pipeline_invalid_pointers_unused_structs.graphics
Add MVKPhysicalDeviceMetalFeatures::counterSamplingPoints
to track platform availability of GPU counters.
MVKPhysicalDevice creates and manages MTLCounterSets and checks for and enables
flags within MVKPhysicalDeviceMetalFeatures::counterSamplingPoints.
Add abstract MVKGPUCounterQueryPool class as parent of MVKTimestampQueryPool
and MVKPipelineStatisticsQueryPool concrete classes and refactor access to host
and command copy tracking data to allow extraction from MTLCounterSampleBuffer.
MVKTimestampQueryPool uses MTLCounterSampleBuffer if supported, otherwise reverts
to using host data for timestamps.
MVKCommandEncoder encodes Vulkan timestamp commands either as Metal staged or
command timestamps, depending on whether the GPU is tile-based or immediate-mode.
For Metal stage counters, we use a light-weight dummy BLIT encoder to mark
timestamp commands executed in the previous Metal encoding pass.
Add MVKDevice::getDummyBlitMTLBuffer() to supply a dummy single-byte buffer that
can be used by a stand-alone MTLBlitCommandEncoder as dummy work to mark timestamps.
If alpha-to-coverage is enabled, we must enable the fragment shader first color output,
even without a color attachment present or in use, so that coverage can be calculated.
VK_FORMAT_E5B9G9R9_UFLOAT_PACK32 on macOS Apple Silicon.
On Apple Silicon (iOS/tvOs/macOS M1), format VK_FORMAT_E5B9G9R9_UFLOAT_PACK32
is fully supported as a color attachment except that format components cannot
be individually write-enabled. All components must either be write-enabled or
write-disabled together. This is causing several hundred Vulkan CTS blending
tests to fail on M1. The least intrusive behavioural change to allow the CTS
tests to report Not Supported instead, is to disable blending for this format.
Depth/stencil attachments that are also sampled might have an artificially forced
swizzled texture view. The original unswizzled texture must be used for attachments.
Add MVKImageViewPlane::getUnswizzledMTLTexture().
Move MVKImageView::validateSwizzledMTLPixelFormat() to
MVKImageViewPlane::initSwizzledMTLPixelFormat(), make it non-static, and use
instance member content instead of passing all the data as function arguments.
The same set of CTS tests either fails or passes on different GPUs
based on whether or not we adjust float clear colors by one ULP.
Add MVKFloatRounding enum.
Add MVKPhysicalDeviceMetalFeatures::clearColorFloatRounding.
Disable ULP adjustment for clear colors on Apple Silicon.
For consistency and to simplify bookkeepping, calculate
clear color ULP adjustment from bit width of format component.
Update MoltenVK version to 1.1.5.
Update VK_MVK_MOLTENVK_SPEC_VERSION to 32.
MVKOcclusionQueryCommandEncoderState::beginOcclusionQuery() check for visibility
buffer exhaustion, and if needed, log an error and disable further visibility
tracking for the remainder of the current MTLCommandBuffer.
Create visibility buffer if needed during MVKCommandEncoder::beginMetalRenderPass.
Simplify MVKCommandEncodingContext to PODS.
