NHL 27 (PPSA34063) on Windows / AMD RX 6700 XT: reaches match loading, ~1.3 FPS

[poonasor]

Windows 11, i7-7700K, 64 GB RAM, AMD Radeon RX 6700 XT (Adrenalin 25.10.45.05, amdvlk64 B026363).

State: title screen → START → legal/set-up → main menu → Play Now → team select → sides, teams and jerseys → match loading screen (X-Factor screen, music playing). It aborts while loading the match. Menus render correctly apart from one background; speed is ~1.3–2 FPS. Not playable.

Base: all of this was measured on my fork's integration branch poonasor/AnyPS5@dev (8d541cfdd), which is upstream 379d8b4e plus the commits below. It is well behind current main and has not been rebased yet, so the list below is what was needed on that base, not a claim about current main. Several of the functions have since been implemented upstream by others.

On current main (d146ec1b6) the title stops in the relinker:

FAIL: Short guest TLS instruction is followed by an instruction that cannot move (offset 0x747c68d)

PR for that one: #4273. Nothing past the relinker has been run on current main yet.

Written with AI assistance (Claude); runs and observations are from my machine.

Launch

ANYPS5_OUTPUT_4K=1 ANYPS5_NO_SHADER_CACHE=1 APS5_LOOP_GUARD=200000 APS5_STORE_AT_FLIP=1 PPSA34063-app0.exe

ANYPS5_OUTPUT_4K is a fork-only switch (see 8). The loop guard avoids the GPU timeout (25), and it only takes effect without the shader disk cache because it is not part of the cache key. About 12 minutes to the title screen; the first frames after START take 4 to 85 minutes depending on the run (24).

The engine module Engine.Render.Core2.PlatformPs5.prx sits in the title's root, not in sce_module/; I copied it into sce_module/ so the relinker converts it. The executables in my dump were SELF containers with plaintext segments and were rebuilt into plain ELFs before relinking.

Blockers, in the order they were hit

Commits are on the fork. "Workaround" means the abort was replaced by a logged shortcut that is not correct behaviour; those are not meant for upstream as they are.

#	Stop	Cause	Change
1	Relinker: Short guest TLS instruction is followed by an instruction that cannot move at 0x747c68d	Stack-protector epilogue: fs:xor rax,[rbx] followed by jne __stack_chk_fail	3b3e6c76: relative branches after a short TLS access move into the stub
2	Not-implemented aborts at load	sceKernelGetModuleInfo2, sceKernelGetModuleList2, __stack_chk_fail/__stack_chk_guard/__error/rename, sceConvertKeycodeGetCharacter, sce::Json::Value::setNullAccessCallBack, sceAudioOut2UserGetSupportedAttributes	a7a372d1, bceff867, 6826d490, 40003b49, 5d8d74c8, ef95c786
3	Execute of address 0 on Job0, return address in the engine module	The engine module was mapped twice: once at start-up from sce_module/ (imports bound by the entry stub) and again, unbound, by sceKernelLoadStartModule("/app0/Engine....prx")	d8552d1c: Windows dlopen returns the already loaded relinked module
4–6	sceKernelGetPageTableStats, sceNetResolverAbort, sceHttpSetRequestStatusCallback not implemented		224f2912 (fixed synthetic pool, values not verified on hardware), a61a4747 (PS4 signature assumed), 9a6456a8
7	SubmissionManagerPs5::waitForFrame fails with 0x8002003C every second	First dispatch waited for its shader behind the FIFO background preparation of thousands of registered shaders	d610edd6: an awaited shader is prepared ahead of the queue
8	Read of 0x74 on a Job thread ~45 s in	The game picks a hardware profile from sceVideoOutGetOutputStatus + sceKernelIsTrinityMode; with a 1080p display it picks PS5_BASE and its settings lookup returns null. Why is not understood	Workaround ba61a98e: opt-in ANYPS5_OUTPUT_4K=1 reports a 4K display
9–10	sceNpEntitlementAccessGetPftFlag, sceCoredumpGetStopInfoCpu not implemented	Call sites: (int32_t* flag); (buffer, 0x10) expecting 0x81180003 outside a coredump handler	7446811b, d04f5d60
11	storage image access to depth/stencil surface ... is not implemented	The game's transient-target pool reuses one address for a depth target and later for unrelated storage images (other extent or format); depth surfaces stay registered	f35fa82c, 6d28c597: storage views use the sampled path's rule (extent and depth format must match, otherwise reused memory)
12	Read fault on thread Heartbeat	The game scans the recorded stacks of all its threads; a finished, not yet joined Windows thread's stack was already freed	fedb61cb: the host thread parks until it is joined or detached
13	sceAjmBatchJobInitialize (Opus ... nonzero third parameter word)	Mono 48 kHz Opus decoder initialised with a third word of 255	3b35621b: 255 accepted for mono (meaning assumed)
16	incomplete rect-list primitive	Rect-list draw whose count is not a multiple of 3	7ed0b6af: the incomplete trailing primitive is dropped
17	After START: access violation at amdvlk64.dll+0x9e251f inside vkCreateGraphicsPipelines	One main-menu fragment shader is 662,679 SPIR-V words with 216 OpTerminateInvocation/OpKill exits (one per guest-memory read, from StopBdaInvocationIf). AMD's compiler crashes on it; reproduced offline in seconds with a stand-alone pipeline-creation tool. The module passes the validation layer	eb2e2309: a pixel shader records a failed read in a private flag, returns, and terminates once in a wrapper entry point (primitive-ordered shaders excluded)
18	CMASK fast clears of a mipmapped, 3D or array color target are unsupported	Unimplemented	Workaround c1f56c8a: the fast clear is ignored with one log line
19	color views of several array slices are unsupported	Unimplemented (layered colour view)	Workaround 7b5a0deb: renders into the first slice
20	Validation: VUID-VkShaderModuleCreateInfo-pCode-08737, OpVectorShuffle with a non-vector operand	The SPIR-V specialization folded any composite built from extracts of one value into a shuffle, also when the value was a struct or array	4e93b53f
21	Validation: VUID-VkImageViewCreateInfo-usage-02275	A sampled view over a storage image inherited the storage usage	a15035f2
22	After Play Now: storage image access to depth/stencil surface again	Here it is a real alias: a 512x512 depth target with an R32 storage image of the same extent bound over it in a compute dispatch	Workaround c65cbfac: the storage image is served from guest memory with one log line; it does not share the depth texels
23	Process vanishes; Windows reports 170 GB of commit	All 22,000+ registered shaders were prepared ahead and kept while a driver compile blocked rendering	707630ee: only the first 1024 are prepared ahead, the rest on first use

Open

• 14: ~1.3–2 FPS from the intro video on. The emulated GPU queue is busy for the whole frame; several wait-on-memory packets take 30–80 ms each. Not analysed further.
• 24: the menu shader takes the AMD driver ~18 minutes to compile offline (30+ in the game). Measured on variants of the module: stubbing the eight BDA read helpers brings it to 21 s, so the time goes into those helpers being inlined at 455 call sites (DontInline is ignored). 45f4d270 (byte reads in a loop) and 8d541cfd (no allocation per timed equeue wait; the in-game compile was also starved by heap-lock contention) did not make a visible difference in the game.
• 25: GPU timeout (Windows event 4101) and device loss right after START in two of two runs without APS5_LOOP_GUARD, none in two runs with it. The game's breadcrumbs point at hairResolve / ShaderProgram_StrandHairRasterizerFineCs, Dispatch 144 x 1 x 1. Suspected: a compute shader loop that does not terminate under emulation. The shader has not been isolated.
• 26: access violation in the engine module (Engine.Render.Core2.PlatformPs5.prx+0x90cd6, walking a linked list in GPU-visible memory) ~25 minutes after Play Now. Did not recur in 68 minutes with APS5_STORE_AT_FLIP=1. One run each way.
• 27 (current stop), during match loading:

  compute shader 0x1153367800: AGC graphics: sampling the stencil plane of depth surface 0x11c0b90000 (2348x1320, vk format 130) as a 2348x1320 texture of guest format 1 (vk 9), tile mode 3, dimension 1, levels 0-0, slice 0 ... is not implemented
  

• The team-select background renders as speckle noise; the UI over it is correct.
• sceSysmoduleLoadModule logs unknown id: 267 in every run without visible effect.

Not tested

NVIDIA, Intel, Linux; anything on current main past the relinker; gameplay; audio beyond menu music; controller input beyond menu navigation.

[ShootanoJumper]

Hi, I am also working on NHL 27. I have gotten into a match and I have also got some of the HUD to show. I am still trying to fix an issue where everything just renders as white.

I am on Windows 11, Ryzen 7 5700x, RTX 4060, 16 GB DDR4.

[poonasor]

i've seen the white background during the team selection.. whats your framerate at?

frame_057.bmp

[ShootanoJumper]

I got a max of 3 frames in the menu but I got less than a frame while loading in. It took a while to load the match.

[SandstromPL]

@poonasor #5075 brings your stop-13 change (3b35621b, Opus mapping family 255 for a mono stream) to main, with you as co-author. Since #4273 is merged, main should get past the relinker now. If you rerun NHL 27 on current main with #5075, a note on #5075 saying whether the match loading still stops at the Opus initialization would help it get reviewed. (#5517 builds multichannel family 255 on top of it, for Metal Gear Solid 4.)