Skip to content

Releases: CeeJayDK/ReShade-Testing-Initiative

6.8.0.7

Choose a tag to compare

@CeeJayDK CeeJayDK released this 29 Sep 22:40
9334331

Full Changelog: v6.8.0.6...v6.8.0.7

v6.8.0.7

Built against ReShade v6.8.0.

Every effect in our test corpus now compiles to DXBC on Linux

133 of 133 (v6.8.0.5: 108, v6.8.0.6: 129). Every effect that compiled before
gives byte-identical output.

  • isnan: vkd3d-shader has no isnan, so the HLSL handed to it gets
    (x != x) instead. Under IEEE rules NaN is the only value not equal to itself.
    Fixes Fubax PerfectPerspective. --hlsl output and the Windows build are
    unchanged.
  • Division by zero: legacy DOF.fx (GP65CJ042DOF) guards its modulo with
    if (sampleCycle != 0), but vkd3d folded n % 0 inside that dead branch
    anyway and refused to compile. A small patch to vkd3d
    (patches/vkd3d/fold-division-by-zero.patch) makes it a warning, as in
    D3DCompiler.

fxstat gets the same fixes

fxstat --dxbc on Linux now handles [fastopt] and isnan too. One effect is
very slow there in performance mode (fxstat's default): prod80
PD80_01B_RT_Correct_Color takes over 5 minutes, in vkd3d's loop unroller.
That was always the case; fxstat just failed on it earlier before. Noted in
docs/HANDOFF.md.

Also in the repo, not in these archives

ShaderLab's own vkd3d build now applies the same patch, so the DOF effect
above renders there too.

6.8.0.6

Choose a tag to compare

@CeeJayDK CeeJayDK released this 29 Sep 20:37
c7b267d

Full Changelog: v6.8.0.5...v6.8.0.6

v6.8.0.6

Built against ReShade v6.8.0.

reshadefx_cli_fixed: faster, and more effects compile to DXBC on Linux

  • --all-entry-points, and -E can be repeated. The effect is parsed once
    and the entry points are compiled in parallel (-j <count>, default all cores).
    -Fo out.cso writes out.<entry point>.cso. SPIR-V for 133 effects: 3.9 s → 0.8 s.
  • Faster front end in every tool, Windows included: tokens no longer copy
    their file name into a new heap string, which cuts the front end's instruction
    count by 14%.
  • [fastopt] on Linux: ReShade marks [loop] loops [fastopt] at shader
    model 4+, and vkd3d-shader (behind --dxbc on Linux) rejected it. It is now
    passed to vkd3d as [loop]. On a 133-effect corpus this fixes 21 of 25 DXBC
    failures. --hlsl output and the Windows build still use [fastopt], as
    ReShade does.

None of this changes compiled output: all 133 effects are byte-identical to
v6.8.0.5 in --hlsl, --glsl, --spirv and --dxbc, except the 21 that
now compile. Details: docs/COMPILE-SPEED.md.

Also in the repo, not in these archives

shaderlab/imgcompare.py: compares two images (absolute difference, PSNR,
SSIMULACRA 2). fxrender.py --reference runs the same comparison on a render,
to check whether an optimized shader is close enough to the original. The
SSIMULACRA 2 scores match libjxl's reference tool to within 0.0001.

6.8.0.5

Choose a tag to compare

@CeeJayDK CeeJayDK released this 21 Sep 23:45

Full Changelog: 6.8.0.4...v6.8.0.5

v6.8.0.5

Built against ReShade v6.8.0.

The repo is now ReShade Testing Initiative (was ReshadeFX-tools). The
release assets are renamed to match — reshade-testing-initiative-linux-x64.zip
and reshade-testing-initiative-windows-x64.zip — as are the build scripts
(build_reshade_testing_initiative.sh / .bat).

New: fxstat

SPIR-V and DXBC instruction statistics, built as a library with a thin CLI on
top rather than a one-off tool.

  • Lane-weighted ALU and transcendental counts, classified from the actual
    opcode stream.
  • --rga <path> compiles each entry point with AMD's Radeon GPU Analyzer and
    reports real GPU ISA: VALU, transcendentals, scalar ALU, texture fetches,
    scratch memory and register usage. RGA is not bundled — it runs AMD's Vulkan
    driver compiler offline, no GPU needed.
  • --dxbc via D3DCompiler on Windows, vkd3d-shader on Linux. Counts from the
    two are not comparable; every report records which compiler produced it.
  • --baseline / --fail-on-regression for CI, and --json.
  • --preset takes uniform values from a ReShade preset.

New: reshadefx_coverage

Walks every effect, entry point and shader model in a corpus and reports what
fails to compile to DXBC and why. Written because "the back end builds" is not
the same as "it handles real shaders".

Removed

reshadefx_stats — superseded by fxstat.

Fixes

  • Performance mode now matches what the runtime does.
  • Windows CI builds and tests all five tools again. fxstat.exe picked up
    SPIRV-Tools import libraries instead of the static ones and failed to launch
    without MSYS2 on PATH; it is now genuinely standalone.

Also in the repo, not in these archives

shaderlab/ — a Linux harness that renders an effect onto a PNG headless
(ReShade + DXVK + ShaderLab under Wine, software Vulkan). It answers what a
shader looks like, where the rest of the repo answers whether it compiles and
what it costs. Source only; see shaderlab/README.md.

6.8.0.4

Choose a tag to compare

@CeeJayDK CeeJayDK released this 09 Sep 15:55
84b8a9e

RESHADE should now be set after the actual Reshade version
And you can override the default preprocessor definitions. This is mainly useful so you can compile in performance mode.

Full Changelog: v6.8.0.3...6.8.0.4

6.8.0.3

Choose a tag to compare

@CeeJayDK CeeJayDK released this 08 Sep 02:07
0210bc2

Full Changelog: v6.8.0.2...v6.8.0.3

6.8.0.2

Choose a tag to compare

@CeeJayDK CeeJayDK released this 07 Sep 11:54
28f7284

Now with --json support
And the ability to choose to only build some of the tools (still defaults to building all 3)

v6.8.0.1

Choose a tag to compare

@CeeJayDK CeeJayDK released this 06 Sep 22:55
94014ac

So far the version is a .number after the Reshade version it's based on and this is based on v6.8.0