Repository navigation
feat(shader): implement the missing 32-bit LDS atomics and fix ds_min/max_f32 - #320
Merged
Merged
Conversation
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 <noreply@anthropic.com>
…/max_f32 ds_rsub(_rtn)_u32, ds_inc_u32, ds_dec_u32, ds_mskor(_rtn)_b32, ds_cmpst(_rtn)_b32/f32, ds_add(_rtn)_f32, ds_min/max_rtn_f32, ds_wrap_rtn_b32 and ds_nop threw in the decoder. The new ones run as a compare-exchange loop over the old value, with the semantics measured on an RDNA2 GPU. ds_min_f32 and ds_max_f32 take one data operand on gfx10, but compared the memory with a second register (the unused data1 field, usually v0) and stored data0 when that compare held. They now keep the minimum/maximum of memory and data like the hardware: a NaN loses to a number, -0 < +0. Both IR ops return the old value, so the _rtn forms share them. Co-Authored-By: Claude <noreply@anthropic.com>
Progress report (709d7fe → c8af889)📈 GPU shader instructions: 54.03% (+1.37%, +16 implemented instructions) ✅ 16 implemented
|
This was referenced Oct 2, 2026
Merged
Merged
9 tasks done
9 tasks done
This was referenced Oct 2, 2026
9 tasks done
oneandonlydean
added a commit
to oneandonlydean/AnyPS5
that referenced
this pull request
Oct 3, 2026
…oykopovar#304 boykopovar#305 boykopovar#306 boykopovar#322 boykopovar#409 boykopovar#416, boykopovar#276 boykopovar#277 boykopovar#275 boykopovar#307, shader ops boykopovar#226 boykopovar#255 boykopovar#260 boykopovar#280 boykopovar#308 boykopovar#309 boykopovar#311 boykopovar#318 boykopovar#320 boykopovar#324 boykopovar#326 boykopovar#379 boykopovar#384 boykopovar#399, boykopovar#261, kernel boykopovar#335 boykopovar#341 boykopovar#349 boykopovar#355, libs boykopovar#227 boykopovar#244 boykopovar#245 boykopovar#273 boykopovar#274 boykopovar#279 boykopovar#286 boykopovar#287 boykopovar#289 boykopovar#296 boykopovar#300 boykopovar#328 boykopovar#358 boykopovar#402) into astrobot Where our PR merged in a revised form, main's version is taken; where astrobot carries the same work plus later local commits, astrobot's code stays and main's additions appear once. Conflicts: - AvPlayer (Export.cpp both libs, include/AvPlayer.hpp, src/Player.cpp, src/Source.cpp, AvPlayerLibrary.cmake, CMakeLists.txt, SceTypes.hpp, tests/GuestAvPlayer.cpp): main's boykopovar#307 (our reviewed squash, which carries 9c8e7fd e9ea5c8 adcede3 fb4eeb9 953057c ac456ee). It links the shared ffmpeg-core target, so astrobot's second FFmpeg build goes: 3rdparty/FFmpeg submodule, core/libs/FFmpeg.cmake and src/Exports.cpp. - libScePngEnc: main's boykopovar#227 (supersedes OverkillLabs' encoder). - libScePad Export.cpp: main's boykopovar#287 tilt correction (supersedes 49d45f5). - libSceAudioPropagation Export.cpp: astrobot's object model, which defines every function boykopovar#289 declares as a stub. - libSceSaveData.native Export.cpp: both (SaveDataHoldWrites and boykopovar#289's unknown exports). - VideoOutDriver.cpp: astrobot's (2546bd1 counters around boykopovar#416's lastFlipVblank store, same value). tests/Flip.cpp: boykopovar#416's pacing test and boykopovar#417's after-GPU test both kept. - boykopovar#278 files (DisplayFormat, VulkanDevice, GpuColorTransfer, ResidentPresent, agc CMakeLists): main's comment-free boykopovar#278 plus astrobot's P5 present-direct (b29728d), which removed DetileImage, ResidentPresent.hpp and the recorder-test hook; the resident present test stays its own executable. - VulkanDevice.cpp/.hpp: astrobot's GDS, local memory probe, sample dumps and P5 code; boykopovar#303/boykopovar#409 comments as main has them; main's shader clock (boykopovar#324) and DepthSurface-free includes. - RequestSerializer.cpp: version stays 8 (6 mesh passthrough, 7 nonConstantImageOffsets, 8 astrobot's later field). boykopovar#303's auto-merged second write and read of nonConstantImageOffsets removed. - SpirvEmitter.hpp, SpirvEmitterState.hpp, Recompiler.hpp/.cpp, StageIO.hpp, SpirvOptimizer.cpp, InstructionTranslator.cpp, ShaderReplay.cpp: local fields kept, boykopovar#303/boykopovar#305 comments as main has them. - SpirvAluEmitterMath.cpp: v_readlane keeps the split-wave path and takes boykopovar#322's wrap at the wave size for the others. - SpirvEmitter.cpp: astrobot's single-lane layout check. - SpirvFlowEmitter.cpp: boykopovar#304 as merged (the second lane emits a shared value itself when the first lane defined nothing). - SpirvImageEmitter.cpp: main's Offset operand placed before MinLod. - boykopovar#304: main's WaveUniformValues (SpirvAnalysis.cpp) replaces astrobot's copy in SpirvWaveExchange.cpp. - Recorder.hpp/.cpp, Draw.hpp, Graphics Draw.cpp, ShaderResources.cpp: astrobot's boykopovar#276/boykopovar#277 with its later snapshot work (per-use budgets, stale-snapshot refresh, profile counters, boykopovar#410, boykopovar#424, fan GS); boykopovar#409/boykopovar#306 comments as main has them. - PacketMemory.cpp: boykopovar#409 as merged plus astrobot's two opt-outs. - tests: ShaderMemory.cpp takes boykopovar#303's and boykopovar#304's new tests; Recorder.cpp takes boykopovar#409's afterRecordedWorkTests and drops main's copies of the boykopovar#276/boykopovar#277 tests astrobot already has; Structurizer.cpp keeps boykopovar#413's case. boykopovar#143 (NGG passthrough) stays as on astrobot: upstream did not touch its files since the boykopovar#258 revert.
9 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
New (these threw "DS opcode is not supported"):
ds_rsub_u32,ds_rsub_rtn_u32,ds_inc_u32,ds_dec_u32,ds_mskor_b32,ds_mskor_rtn_b32,ds_cmpst_b32,ds_cmpst_rtn_b32,ds_cmpst_f32,ds_cmpst_rtn_f32,ds_add_f32,ds_add_rtn_f32,ds_min_rtn_f32,ds_max_rtn_f32,ds_wrap_rtn_b32,ds_nop.Fix:
ds_min_f32/ds_max_f32take a single data operand on gfx10, but the translation read a second one (source2, the unused data1 field, which is usually v0) and stored data0 when that compared below/above memory. They now keep min/max(memory, data).The reference emulators don't implement these, so the semantics were measured on an RDNA2 GPU (Radeon 660M, gfx1035) with raw kernels through ROCm HSA. A Python model of all eight forms matched 512/512 hardware results (
old= memory before,d0/d1= data0/data1):rsub:d0 - oldcmpst_b32:old == d0 ? d1 : old;cmpst_f32does the same with an ordered float comparemskor:(old & ~d0) | d1wrap:old >= d0 ? old - d0 : old + d1add_f32: IEEE addmin/max_f32: IEEE minNum/maxNum, so a NaN in memory or data loses to the number, and −0 < +0inc/decreuse the existing ops.Implementation:
SharedAtomicFMin32/FMax32now return the old value (U32) like the other shared atomics, so the_rtnforms share them; they no longer need the scratch-dword workaround. New IR opsSharedAtomicRsub32/FAdd32/Cmpst32/CmpstF32/Mskor32/Wrap32are emitted through the existingAtomicUpdatecompare-exchange loop, with the same workgroup acquire/release barrierAtomic32uses for LDS.Test
agc_driver_lds_atomics: one wave32 shader, each of the 17 ops on its own LDS dword per lane. It checks the memory afterwards and every_rtnresult. Inputs are random words, small integers (inc/dec/wrap/rsub edges), float specials (±0, ±1, NaN, inf), and equal/unequal pairs for cmpst. The expected values come from a C++ model of the hardware semantics above. Float ops are not checked on lanes with f32 denormals, since those follow the driver's float controls.Tested
Linux, full ctest (99) passes.
Checklist
main; no other open PR implements the same functionsNotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt.mdfiles or images added to the repository (attach them to this PR)