On devices whose default subgroup is 64 lanes (RADV, AMD on Windows) a
wave32 compute program was compiled for a 64-wide host subgroup, so the two
guest waves of a 64-thread workgroup shared one subgroup: ballots, EXEC,
v_mbcnt and lane reads of the second wave saw the first one's lanes.
With VK_EXT_subgroup_size_control, when compute can require 32 lanes, the
dispatch recompiles wave32 programs for a 32-wide host subgroup
(VulkanDevice::ComputeTarget) and the pipeline requires that size. The
result records the width it was compiled for (hostSubgroupSize, also in the
disk cache).
Co-Authored-By: Claude <[email protected]>
The WriteLane IR takes {value, lane, previous}, as the translation and the
read-lane eliminator use it, but the SPIR-V emitter compared the lane id with
the previous value and selected between the lane index and the value. Every
v_writelane therefore wrote its value to all lanes (or the lane index to the
lane whose old value happened to equal its id).
Co-Authored-By: Claude <[email protected]>
image_sample_*_o takes its texel offset from an address VGPR. The emitter
accepted only an immediate there and failed with "requires a constant texel
offset for image sampling" for any other, including an offset the program
moves into the VGPR under EXEC.
An offset that resolves to a constant at its packed address position stays a
ConstOffset. Any other becomes the Offset image operand, which Vulkan allows
on sampling instructions only with VK_KHR_maintenance8's maintenance8 feature
and the ImageGatherExtended capability. The device enables maintenance8 when
available and reports it through SpirvTarget::nonConstantImageOffsets, which
the cache key, the request serialization (version 6) and the SPIRV-Tools
validator options follow. Without it the sample still fails.
agc_shader_replay --maintenance8 recompiles a request for such a device.
Replacing the headless device with the windowed one dropped the import
registry without freeing its buffers and memory, and the replaced device
is kept alive, so the imported pinned host memory stayed allocated and
the new device ran out of memory importing the same ranges again.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Tests resolved libstdc++/libgcc/winpthread through PATH, so another MinGW
installation earlier in PATH (e.g. Git for Windows) made them fail with
STATUS_ENTRYPOINT_NOT_FOUND (0xc0000139).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Bn7CrhG7pTPR6DT13tN6Me
Physical device selection took the first Vulkan device with graphics,
compute and presentation support, so on systems with an integrated GPU
(e.g. Ryzen 7000 desktop CPUs) the enumeration order decided whether the
integrated or the discrete GPU was used. Rank candidates by device type.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Bn7CrhG7pTPR6DT13tN6Me
Every draw copied its index and vertex ranges from guest memory into
fresh buffers and scanned the indices, although the draws of a frame read
the same meshes again: about 10 GiB of vertex copies per 10 s at
PPSA21564's title.
CopyDrawInput now keeps the copy in the recorder's draw snapshot cache
and binds it again while the write tracker proves the range unchanged
since the copy was taken, with the same rules as the storage-buffer
snapshots: UnchangedSince on the collect made right before the copy and
an unchanged allocation registry. As GuestMemory::Read does, the flush
hook runs first, so pending GPU results over the range are stored and
stamped before the check. Unwatched memory is copied every time. An index
copy keeps its highest index; the device limit is still checked for every
draw.
The cache key includes the snapshot's use (storage, vertex, 16-bit or
32-bit index), so each use keeps its own buffer with its own usage. A
vertex copy runs from the buffer's start to its highest fetched element,
so draws over one shared vertex pool ask for many sizes: a vertex
snapshot also serves a shorter read at the same address while that
read's bytes are unchanged, and a new one replaces the shorter ones.
Vertex and index snapshots have their own least recently used pool
(1 GiB, 16384 entries), so storage copies, most of which change every
frame, do not push them out before their next use.
Tests: agc_driver_recorder_tests copies index and vertex ranges of
write-watched memory and requires a reuse (with the stored highest index)
while the bytes are unchanged, a new copy with the current bytes after a
CPU store, a MarkWritten store and a registry mutation, separate copies
per use, no reuse without a recorder, prefix reuse for vertex reads only,
and storage snapshots not evicting vertex snapshots.
A recorded draw binds a CPU copy of each read-only guest buffer it reads
in place, made for every draw. PPSA21564's title-stage draws bind the
same large light-list buffers written by compute passes (up to 16 MiB
each) hundreds of times per frame, so about 4 ms of every draw went into
copying about 30 MiB into freshly allocated host buffers.
The recorder now keeps the snapshots, least recently used first, under a
256 MiB budget and 1024 entries. A later draw of the same range binds the
kept, immutable copy while GuestMemory::UnchangedSince holds for the
write generation collected right before the copy and the guest allocation
registry generation is unchanged. A CPU store (stamped by a collect), a
driver or GPU store (MarkWritten) or a registry mutation makes the next
draw copy again. Unwatched memory has no generation and is copied for
every draw, as before.
A recency list (each entry holding its own position) gives the eviction
victim and moves a used entry to the back in O(1).
Tests: agc_driver_recorder_tests binds a write-watched, host-imported
range twice and requires the same snapshot, then a new copy with the new
bytes after a CPU store, a MarkWritten store and a registry mutation; and
checks the least recently used eviction order at the entry cap, the byte
budget and a CPU store dropping a kept snapshot.
A conditional inside a loop with one target in the loop body and the
other outside it was taken for a break and emitted without a merge. When
the outside target is not the loop's merge but a tail of blocks that
reaches it, SPIR-V has no construct for that branch: spirv-val rejects
the module ("Selection must be structured") and the driver is free to
compile it into anything.
Such a branch now merges at its in-body target when the tail is private
to it (every tail block is dominated by the tail's first block, stays in
the loop construct and leaves only for the loop's merge), so the tail is
an "if (c) { tail; break; }" the loop already allows. Merges are given
to these after every other construct's, and one already taken leaves the
branch as before.
All 13 PPSA21564 programs whose V#s are read from GPU memory (the
deferred lighting and ray-traced passes) had two such branches in their
traversal loops. With this change none of the 74 dumped title programs
has one, and the 209 modules of a title-to-cinematic run validate apart
from three unrelated image-offset and entry-point errors.
Tests: agc_structurizer_tests structures a loop that exits through one
tail block and a loop with two exit tails, and requires every merge-less
conditional to target its loop's merge or continue.
An MRT slot below the highest written one that the draw does not write is a VK_ATTACHMENT_UNUSED
reference with a disabled blend; color targets keep their slot for export mappings, pipeline and
pass keys. DB_DEPTH_VIEW SLICE_START moves the depth and stencil planes by whole slices, so each
slice of a depth array is its own depth surface.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Sign-in, web browser and player dialogs finish cancelled; the message dialog finishes at once;
NpAuth, entitlement consumption, reachability, account age and the NpUtility bandwidth test
report signed out; sceKernelMtypeprotect changes the protection (the memory type has no host
equivalent); the APR ForEach resolvers resolve every path and report each failure in its own
result entry. vswprintf's stub is dropped in favor of the existing wide-formatter
implementation, which already covers it on every platform.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Co-Authored-By: Claude Sonnet 5 <[email protected]>
The VOP1, VOP2 and VOPC SDWA and DPP decoders read the modifier word at
wordIndex + 1 before SetRdnaRawWords checked that it exists, so a code
span ending at such an instruction was read one word past its end.
AddressSanitizer reports a heap-buffer-overflow in decodeVop2Sdwa on a
synthetic program. The six decoders now throw "truncated <encoding>
instruction" first, as DecodeRdnaVop3 and DecodeRdnaVop3p already do.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01R5WJS4ZdQynAJesCZKH5hb
VOP3 compares have a scalar destination, so checkNativeVop3Modifiers sent
them through the carry/scalar-destination branch, which rejects neg. abs
already passed and both are applied to the sources by the VOPC path, so
v_cmp_lt_f32_e64 vcc_lo, -v1, v2 threw "VOP3 source modifiers are not
implemented". Float compares now accept neg there; integer ones still don't.
Co-Authored-By: Claude <[email protected]>
Fifty VOPC opcodes threw in the decoder: cmpx f/o/u/tru for f32, cmpx f/t
for i32/u32, the 64-bit integer compares except the seven already present,
and the f16 f/o/u/nge/nlg/ngt/nle/nlt/tru compares and their cmpx forms.
64-bit gt/le/ge are built from SLessThan64/ULessThan64 by swapping and
negating, as the IR has no other 64-bit orderings. emitFloatOrderedCompare
gains the f16 and cmpx forms.
Co-Authored-By: Claude <[email protected]>
These VOP3 opcodes threw in the decoder. v_add_nc_i32 and v_sub_nc_i32 are
what compilers emit for signed saturating add/sub, so their clamp saturates
to the int32 range. The 16-bit mads honour op_sel on their 16-bit sources;
the u32_u16/i32_i16 forms take a 32-bit addend and write all 32 bits, so
they do not get the 16-bit destination select.
Co-Authored-By: Claude <[email protected]>
s_icache_inv, s_incperflevel, s_decperflevel, s_ttracedata_imm, s_wait_idle,
s_gl1_inv, s_dcache_inv, s_dcache_wb, buffer_gl0_inv and buffer_gl1_inv threw
in the decoder. SPIR-V has no cache control, so they translate like s_clause
and s_waitcnt, which already emit nothing. The SMEM and MUBUF forms carry no
operands, so no pass sees a phantom base or destination register.
Co-Authored-By: Claude <[email protected]>
The VOP2/VOP1 forms already translate; the VOP3 encodings were rejected by the
decoder. VOP3 f16 add/sub/subrev/mul/min/max now take abs and neg, which
readF16AsF32 already applies. Clamp and omod stay rejected for them because
writeF16 does not apply result modifiers.
Co-Authored-By: Claude <[email protected]>
MixPrecision and SdwaDestination dispatch hand-assembled RDNA through the
recompiler and compare every lane with a CPU model, but were never in ctest.
Move them to tests/execution/, which is globbed into one test per file with
the same target names, so new execution tests add a file and no CMake lines.
They exit 77 (skipped) when the host has no Vulkan device; the Linux runner
installs lavapipe and sets ANYPS5_REQUIRE_VULKAN=1 so they cannot skip there.
Co-Authored-By: Claude <[email protected]>
Explicit v_interp_p1/p2 interpolation (fragment-shader barycentrics) reads every interpolated
parameter per vertex, and the rect-list shaders refused every per-vertex parameter. Their
tessellation stage emits each corner's parameter value, which is what a per-vertex read of the
generated triangles needs. Only custom parameters (custom interpolation or passthrough), whose
fourth vertex has no value of its own, stay refused; FragmentParameter carries that flag.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
A sampled read through a surface's keys after a uniform key fill with a clear code reads the
fill's clear value, not the image's pending results (cachedTexture's clear-through-keys). The
stage-A fast path returned the record's view of the pending image without that check, so a
read between the fill and the image's next refresh saw the previous results.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
An image's filledKeys stands for "the pending results predate the title's
key fill": sampled reads through those keys bind the clear value, and
(with APS5_KEYS_FILL_WRITEBACK_KEEP=1) a write-back leaves the keys at the
fill's code. A refresh that clears the image under the new keys ends that
hold (upload resets it), but the immediate clear of APS5_KEYS_FILL_CLEAR=1
did not. Results drawn after such a fill were then treated as dead: reads
through the keys saw the clear and a write-back kept the keys clear over
them. The immediate clear now ends the hold as the refresh does.
