shader_recompiler: Reach in-game status in Control with fragment input fixes

[w1naenator]

Summary

Complete the fragment interpolation path for Liverpool shaders and correctly
translate fragment inputs across AMD and KHR Vulkan implementations.

This fixes incorrect lighting and dark rendering observed in Control while
preserving the alternate front-face encoding required by other games.

Changelog

Fragment interpolation

• Implement specification-driven GCN V_INTERP_P1_F32,
  V_INTERP_P2_F32, and V_INTERP_MOV_F32 behavior.
• Preserve the native GCN P0, P10, and P20 coefficient semantics.
• Handle constant and flat shader inputs without applying artificial
  zero-slope interpolation.
• Add support for perspective and no-perspective interpolation at:
  • Center
  • Centroid
  • Sample
• Implement BaryCoordPullModel:
  • Use BaryCoordPullModelAMD when
    VK_AMD_shader_explicit_vertex_parameter is available.
  • Reconstruct (I/W, J/W, 1/W) using KHR barycentrics and
    FragCoord.w otherwise.
• Support scalar per-vertex fragment parameters.
• Generate properly qualified KHR barycentric inputs instead of using
  interpolation-function approximations.

Primitive and provoking-vertex handling

• Carry fragment primitive topology and provoking-vertex state into shader
  runtime specialization.
• Map per-vertex parameters according to the host API's primitive numbering.
• Handle point, line, triangle, fan, strip, and adjacency topologies.
• Account for odd/even triangle-strip vertex ordering where required.
• Query VK_EXT_provoking_vertex feature support instead of relying only on
  extension presence.
• Query
  triStripVertexOrderIndependentOfProvokingVertex from the KHR barycentric
  properties.
• Add the PrimitiveId input only when topology remapping requires primitive
  parity.

Front-face encoding

• Read FRONT_FACE_ALL_BITS from SPI_BARYC_CNTL.
• Carry the selected encoding through fragment runtime specialization.
• Represent Vulkan FrontFacing as a boolean IR attribute.
• Encode the guest FRONT_FACE VGPR according to Liverpool state:
  • Raw +1.0f / -1.0f bits when FRONT_FACE_ALL_BITS is disabled.
  • Integer 1 / 0 when FRONT_FACE_ALL_BITS is enabled.
• Avoid the unconditional sign encoding that caused regressions in geometry
  using the alternate mode.

This part is based on review suggestions from Roamic, Raphael, and
IndecisiveTurtle.

Sample coverage

• Load SampleCoverage when enabled by the pixel-shader input registers.
• Use a fast single-sample path without declaring SampleMask.
• Mask multisample coverage to the active sample count.
• Return zero coverage for helper invocations.
• Carry the framebuffer sample count through fragment specialization.

Shader resource-table synchronization

• Record the guest-memory ranges traversed by the flattened SRT walker.
• Synchronize GPU-modified SRT ranges before flattening and shader
  specialization.
• Support nested SRT pointers while retaining parent/child relationships.
• Refresh flattened user data through the pipeline cache after required
  buffer-cache downloads.
• Serialize SRT metadata field-by-field, including memory reservations,
  instead of serializing the raw structure.

Renderer and descriptor robustness

• Expand compute image-copy detection to support variants with an optional
  small bounds buffer.
• Validate unique source and destination buffers before replacing a compute
  dispatch with an image copy.
• Clamp shader buffers to actual mapped memory intervals.
• Ignore fully unmapped buffer descriptors safely.
• Invalidate aliased cached images after both formatted and raw
  storage-buffer writes.
• Preserve the correct descriptor type for null or invalid storage images.
• Reject reserved sampler descriptor encodings before creating Vulkan
  samplers.
• Bind a null sampler for invalid descriptors exposed by untaken dynamic
  shader branches.

Vulkan and SPIR-V

• Prefer VK_AMD_shader_explicit_vertex_parameter when available.
• Use VK_KHR_fragment_shader_barycentric as the portable fallback,
  including Vulkan 1.4 core availability.
• Request interpolation, geometry, barycentric, and sample-rate capabilities
  only when required by the generated shader.
• Add missing no-perspective centroid and pull-model IR attributes.
• Add a boolean GetAttributeU1 IR path for boolean Vulkan built-ins.

Shader cache

• Include topology, sample count, provoking-vertex state, and front-face
  encoding in fragment runtime specialization.
• Update shader binary and metadata cache versions to invalidate
  incompatible cached shaders.
• Serialize the new SRT metadata explicitly.

[w1naenator]

Control completes level loading and reaches in-game. The framerate is ~30 FPS, and the game looks very close to the real PS4.

[raphaelthegreat]

The barycentric changes need their separate PR. Also when making multi-patch PRs like this avoid having any "merge main" commits it makes it harder to review

[w1naenator]

The barycentric changes need their separate PR. Also when making multi-patch PRs like this avoid having any "merge main" commits it makes it harder to review

I'll do it when this PR became acceptable.

[w1naenator]

Ok, needed regression tests again.

[bigol83]

No more device lost in games, Control still looks good

Yakuza 6 still looks like this though

[w1naenator]

hmm

No more device lost in games, Control still looks good

Yakuza 6 still looks like this though

But it is beautiful, no?
Ok, I try to understand where it goes wrong.

[bigol83]

hmm

No more device lost in games, Control still looks good
Yakuza 6 still looks like this though

But it is beautiful, no? Ok, I try to understand where it goes wrong.

Zombie Kiryu is great 😄

Anyway there is also a bit of performance loss in Kingdom Hearts 3, like 3-4 fps slower compared to main

[w1naenator]

hmm

No more device lost in games, Control still looks good
Yakuza 6 still looks like this though

But it is beautiful, no? Ok, I try to understand where it goes wrong.

Zombie Kiryu is great 😄

Anyway there is also a bit of performance loss in Kingdom Hearts 3, like 3-4 fps slower compared to main

I haven't Yakuza 6, is it still zombie? Maybe you can provide log?

[bigol83]

Still an issue

[w1naenator]

ok, i found that I am bro

Still an issue

I found last night that I had broken V_INTERP_MOV_F32 in commit d8993a5f. I’ve fixed it, done some cleanup, and a new commit will follow soon.

[w1naenator]

@bigol83 Can you, please test again?

[w1naenator]

Seems like it gives something good to my favorite game too.

[w1naenator]

With it, FIST shows intro now. ^^

It crashes when skipping it... shad_log.zip

With this #4561 menu and loading screen is reachable. But intro skipped.

[tochnonechelovek]

No notable changes or regressions in my titles.