The second of V6's three commits: everything the fragment stage samples. Plan
sections 4.3 (textures and mip generation) and 4.4 (descriptors).
The descriptor table is the piece that retires GL_ARB_bindless_texture. One
update-after-bind, partially-bound, variable-count combined-image-sampler array
of 16384; registration appends exactly one vkUpdateDescriptorSets and nothing is
written at draw time, so steady state is zero descriptor writes per frame. A
slot is a (view, sampler) pair, exactly like a bindless handle, which is why the
CPU data model needs no change at all - GpuTextureSlot already carries the index
and V2 already moved every batch onto it.
Eviction is retirement-gated and the slot is scrubbed on the way out. Returning
a slot the moment a texture is deleted would let the LRU alias a live draw onto
a new texture, so the release is filed through the ledger; and when it runs the
slot is first overwritten with the default 1x1 white. A stale view descriptor
sitting in a partially-bound array is legal right up until something reads it,
at which point it is a use-after-free with no error attached. Writing the dummy
makes that impossible rather than unlikely.
The CPU block-compression codec is the slice's other substantial piece, and it
exists because Vulkan cannot blit into a compressed image. DAT surfaces arrive
as DXT1/3/5 with no mips, so the chain has to be decoded, box filtered and
re-encoded here. That is not merely a substitute for the missing blit: the GL
path calls glGenerateMipmap on compressed array textures, whose result is
explicitly implementation-defined, so this is the first time that part of the
pipeline has had a defined answer.
Two properties matter more than quality, and both are tested. It is
deterministic - integer arithmetic end to end, endpoints from the block's
bounding box, nearest-palette selection, no dithering and no iterative fit -
because the offline pixel gate compares captures from separate processes and a
chain that varied run to run would make every textured surface look like a
regression. And it preserves BC1's one-bit cut-out: a block containing any texel
below the alpha threshold is encoded in three-colour mode, because retail's
foliage and grates ARE that mode and quantising those texels to an opaque colour
would fill in every leaf. Plan 4.3's escape hatch stands if quality ever trips a
gate: store the affected textures as RGBA8 and blit their mips.
Uncompressed images do take the blit chain, added to the upload queue. Each
source level moves to TRANSFER_SRC for its blit and back to TRANSFER_DST
afterwards; leaving the chain in mixed layouts would be one barrier cheaper and
would then force the batch's final shader-read transition to name a different
old layout per level, so ending every level the same way is what keeps that
transition one barrier per image.
The upload queue now records the layout each image is in on ENTRY to a batch
rather than always naming UNDEFINED. UNDEFINED lets the driver discard existing
contents, which is right for a fresh image and wrong for the incremental
array-layer fills that mirror ManagedGLTextureArray - discarding there would
erase every layer uploaded earlier.
Render targets are single-sampled per the contract and carry SAMPLED usage
alongside COLOR_ATTACHMENT, so a paperdoll or appraisal view can be registered
into the table and drawn by the retained UI the moment its pass ends.
VulkanBackbufferAttachments owns the two attachments the swapchain does not: the
multisampled colour scratch that resolves into the swapchain image, and the
transient depth/stencil. Both are TRANSIENT_ATTACHMENT because nothing reads
either after the frame. Stencil is not optional - issue #117's portal punch
needs the aspect, which is why the V5 gate prefers D32_SFLOAT_S8_UINT over a
depth-only format.
Every format stays UNORM, and that is the V3 audit's finding rather than a
default. The plan previously specified an sRGB swapchain "matching the GL
FramebufferSrgb contract"; that contract does not exist, the renderer is plain
UNORM end to end, and shipping _SRGB would have brightened every frame and
passed silently until V7.
VulkanPipelineLayouts is extracted from V5's capability probe rather than
written beside it, and the probe now calls it. The probe's whole value is
proving the layouts the live backend builds can be built on this device; two
similar-looking definitions would have quietly ended that the first time one of
them changed.
Gates: Release build clean, App suite 4037 passed / 3 skipped (4014 at V6a plus
23 new), offline pixel gate PASS against the parent baseline at a differing
fraction of 4.26e-05 - 24 pixels of 563,200, one above the campaign's recorded
15-23 same-commit noise band and about 23x under the 0.001 threshold, on a
commit that changes no GL code path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The first of V6's three commits, and the half of the Vulkan backend that has
nothing to do with drawing: where memory comes from, how per-frame data reaches
the GPU, and what makes it safe to reuse either.
Plan sections: 4.2 (bindings layer and the no-VMA decision), 4.3 (memory:
arena, staging ring, per-frame data), 4.8 (sync and the frame).
The allocator is hand-rolled, roughly as 4.2 sizes it. Silk ships no VMA, and a
third-party binding would be a native binary to carry across win-x64, linux-x64
and CI lavapipe for an allocation profile that is genuinely tame: two mesh arena
buffers, one staging ring, a per-flight ring buffer each, a few render targets
and a texture pool. What a custom allocator buys instead is exact accounting -
every byte is attributable to a memory type and a block - which is what
GpuMemoryTracker will want and what VMA would obscure.
Placement, block policy and heap choice are pure types with no Vulkan handle in
sight: VulkanMemoryBlockFreeList is first-fit with coalescing on release,
VulkanMemoryTypePool decides when a request is large enough to warrant a block
of its own, and VulkanMemoryTypeSelection maps each GpuMemoryResidency onto a
preference order of property masks. VulkanDeviceMemoryAllocator turns their
answers into vkAllocateMemory and one persistent vkMapMemory per host-visible
block. That split is deliberate: an allocator's real failure modes are
arithmetic - a mis-coalesced neighbour, an alignment that eats a block's tail, a
double release that quietly corrupts the used-byte count - and arithmetic does
not need a GPU to be wrong. Twenty-two tests cover exactly those.
The HostWritable row of the selection table is the campaign's CPU win stated as
data. It prefers a memory type that is both DEVICE_LOCAL and HOST_VISIBLE -
resizable BAR, present on the RX 9070 XT - so per-frame data is written once,
straight into memory the GPU reads, and falls back to ordinary host-visible
coherent memory when no such type exists. GpuCapabilityRecord's
SupportsPersistentlyMappedRings is the first capability that is true on this
backend and false on GL.
Mapping is per block, never per allocation, because Vulkan permits a memory
object to be mapped once - mapping per buffer would need one VkDeviceMemory per
buffer, which is precisely the allocation-count explosion the design exists to
avoid.
VulkanRingBufferState is markedly simpler than its GL sibling, and the
difference IS the point. GlRingBufferState has to track a dirty watermark and
prove its upload never overlaps an in-flight read, because a ring allocation
there writes into a managed array that is later copied into a GL buffer. Here
the allocation hands back memory the GPU reads directly: there is no upload step
to track. What is left is a cursor.
VulkanUploadQueue accumulates transfers rather than issuing them, for two
reasons that both come from Vulkan rather than from taste: copies must be
recorded into a command buffer, and they must be recorded outside a
dynamic-rendering block. So requests queue and drain at the one moment both hold
- immediately before a pass begins - which is the direct analogue of the GL
backend's flush-before-every-draw discipline at the granularity Vulkan needs.
The drain emits one batched buffer barrier for the whole batch, one of the four
to six 4.8 budgets per frame.
Staging exhaustion falls back to a temporary dedicated buffer retired through
the ledger. Section 4.3 already specifies that for oversized uploads; extending
it to "the ring is full of unretired frames" is the same shape and is a policy
rather than a workaround - the transfer stays correct and ordered, it just costs
one allocation.
VulkanFrameFlightController is the mechanical port 4.8 promised. GL's array of
fences becomes one timeline semaphore whose value is the frame serial, "has this
slot retired?" becomes "is the counter at least serial minus two?", and the
SortedDictionary retirement ledger keeps its keys because those keys were
already frame serials. One subtlety is worth stating: a release is filed against
the frame currently being RECORDED, not the last one completed, because commands
already recorded into the open frame may still read the resource. A test pins
that, since getting it wrong frees memory a pending command buffer reads and the
symptom would appear somewhere else entirely.
Frame acquire ordering is the other subtlety. TryBeginFrame waits on the flight
slot BEFORE acquiring its swapchain image, so the slot's acquire semaphore is
provably idle - signalling a semaphore a pending submit still waits on is the
classic Vulkan deadlock. When the acquire fails the serial is still signalled
through an empty submit, because a serial that never completes makes every later
frame wait forever.
The device is a partial class split along the V6 commit boundary: everything
here is memory and frames, while textures and the descriptor table (V6b) and
pipelines, passes and readback (V6c) throw with the slice named rather than
returning something that fails later and further away. Nothing constructs this
device yet - VulkanBringUpHost still presents its clear colour - so the GL path
executes not one new statement.
VK_EXT_debug_utils naming arrives with the allocator rather than at V6c, because
every resource wants a name from birth and the campaign has already spent days
on defects only visible from outside the API. It stays optional: absent
extension means every call is a no-op and no call site checks.
Gates: Release build clean, App suite 4014 passed / 3 skipped (3981 baseline
plus 33 new). One Issue181WallPressEquilibriumTests failure in the full run is
the known #250 zero-allocation flake and passes on a single run. Offline pixel
gate against the parent is a tripwire here - the backend is dark and no GL code
path changed - and is reported with the slice.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Erik