TsundereStation/data/shaders
LibretroAdmin 9e4c33d53c core/d3d12: pre-bake the textured Nearest batch FS slice
Second batch FS template to be pre-baked, after batch_untextured.
Adds the HLSL template + regen recipe + 72 DXBC blobs for the
(texture_mode != GPUTextureMode::Disabled AND m_texture_filter
== GPUTextureFilter::Nearest) slice. No C++ wiring yet - the
activation commit lands next, mirroring the
9e6f933 -> b6d1903 untextured arc.

Variant matrix (post-c532a34 TRANSPARENCY routing):

  * Texture mode (6 combos via 3 `-D` macros):
      (PALETTE_4_BIT=0, PALETTE_8_BIT=0, RAW_TEXTURE=0): Direct16Bit
      (PALETTE_4_BIT=0, PALETTE_8_BIT=0, RAW_TEXTURE=1): RawDirect16Bit
      (PALETTE_4_BIT=1, PALETTE_8_BIT=0, RAW_TEXTURE=0): Palette4Bit
      (PALETTE_4_BIT=1, PALETTE_8_BIT=0, RAW_TEXTURE=1): RawPalette4Bit
      (PALETTE_4_BIT=0, PALETTE_8_BIT=1, RAW_TEXTURE=0): Palette8Bit
      (PALETTE_4_BIT=0, PALETTE_8_BIT=1, RAW_TEXTURE=1): RawPalette8Bit
    PALETTE_4_BIT and PALETTE_8_BIT are mutually exclusive in the
    shadergen #elif chain - no variant has both set. The
    Reserved_Direct16Bit / Reserved_RawDirect16Bit enum values fold
    to their non-Reserved counterparts at the C++ picker level
    (their shader source is identical).

  * USE_DUAL_SOURCE (0/1): drives o_col1 declaration + write.
    Same axis as untextured (b6d1903).
  * INTERP_CENTROID / INTERP_SAMPLE: interp tri-state qualifier.
    Same axis as untextured.
  * NOPERSP (0/1): `noperspective` on v_col0. Same axis as
    untextured.

6 (tex_mode) x 2 (dual) x 3 (interp) x 2 (persp) = 72 blobs. The
MSAA axis (m_multisamples 1/2/4/8/16/32) does NOT multiply this
template - the batch FS samples from the single-sample shadow
VRAM regardless of m_multisamples, so there's no LOAD_TEXTURE_MS
unroll. MSAA only affects the interp qualifier (centroid /
sample), already captured by INTERP_*.

Without c532a34 (TRANSPARENCY -> u_render_mode routing) this
template would have been 6 x 4 x 2 x 3 x 2 = 288 variants on the
[render_mode] axis - 4x bigger. Worth the routing commit.

Files in this commit (74 total, no C++ touched):

  data/shaders/d3d12/batch_textured_nearest.ps.hlsl
    Hand-translated mirror of shadergen's textured-Nearest output:
    BatchUBOData cbuffer, Texture2D + SamplerState bindings at
    register t0/s0 (matches m_batch_root_signature - see
    CreateRootSignatures in gpu_hw_d3d12.cpp), full
    ApplyDithering / ApplyTextureWindow / ApplyUpscaledTextureWindow
    / FloatToIntegerCoords / RGBA8ToRGBA5551 / SampleFromVRAM
    helpers, and the main() body with palette / direct / raw /
    semitransparent / discard logic. The post-c532a34 runtime
    branches on u_render_mode (premultiply, output arms,
    OnlyOpaque/OnlyTransparent discards) and u_uv_limits (UV
    clamping), u_dithering, u_interlacing, u_pgxp_depth, u_uv_limits
    cbuffer scalars are all wired the same way as the Vulkan
    template at batch_textured_nearest.frag.glsl.

  tools/regen_d3d12_dxbc.py
    TEMPLATE_VARIANTS gains a batch_textured_nearest.ps.hlsl entry
    with all 72 variant tuples. Variant suffix encoding:
    pXrY_d{0,1}_{none,centroid,sample}_n{0,1} where X is the
    palette mode digit (4 / 8 / 0), Y is RawTextureBit, the d/n
    digits encode the dual_source / nopersp bits, and the interp
    word names the qualifier.

  src/common/d3d12/embedded_dxbc/batch_textured_nearest_ps_*.inc
    72 new .inc files, ~2592 KiB total on disk, ~400 KiB embedded
    DXBC in the .so. Per-variant DXBC sizes range from 4132 bytes
    (RawDirect16Bit + no dual + no qual) to 6656 bytes
    (Palette8Bit + dual + centroid + nopersp); average ~5500
    bytes. The raw variants are smaller because RAW_TEXTURE=1
    skips the colour-modulation path; the palette variants are
    larger because they unroll the index extraction + palette
    sample.

What does NOT change in this commit:

  - GPU_HW_D3D12::GetBatchPipeline still calls
    GetBatchFragmentShader for textured slices, which runs
    shadergen + D3DCompile. The 72 pre-baked Nearest blobs are
    not yet consulted - they sit on disk waiting for the
    activation commit (parallel to b6d1903 for untextured).

  - src/common/d3d12/embedded_shaders.{h,cpp} untouched here -
    they need 72 new extern declarations / includes to actually
    surface the .inc blob symbols to the C++ linker, but adding
    those without C++ consumers would yield -Wunused warnings
    that the build flags treat as errors. Both land together in
    the activation commit.

  - m_texture_filter == Nearest sessions still go through
    D3DCompile until the activation commit lands. The picker
    function PickBatchUntexturedFSBytecode (gpu_hw_d3d12.cpp) is
    untouched - the textured analogue lands with the activation.

Test:
  - python3 tools/regen_d3d12_dxbc.py --fxc /tmp/fxc_lite/fxc_lite.exe
    --wine /usr/bin/wine: 88 .inc files generated, 72 of which are
    the new batch_textured_nearest_ps_* set. Build clean across all
    72 variants - no fxc warnings or errors on any combination.
  - Hand-compile of representative variants
    (PALETTE_4_BIT=1 / PALETTE_8_BIT=1 / RAW_TEXTURE=1 / dual /
     INTERP_SAMPLE / NOPERSP combos): all produce valid DXBC
    blobs at expected sizes.

Cumulative pre-bake variant count progress for D3D12 batch FS:

  Pre-arc:                                    75 264 (every axis a -D)
  Post-cbuffer-routing of:
    DITHERING/INTERLACING/UV_LIMITS/PGXP_DEPTH 4 704
  Post-c532a34 (TRANSPARENCY -> u_render_mode) 1 176
  Post-c01f8ae (untextured re-bake, slice live)  12 baked
  Post-this-commit (Nearest .inc on disk):
    Untextured slice (live):    12
    Nearest slice (on disk):    72
    Remaining filters (Bilinear/JINC2/xBR): 3 templates x 72 = 216

Next:
  - Activation: 144 new extern decls + 72 new #includes in
    embedded_shaders.{h,cpp}, plus a PickBatchNearestFSBytecode-
    style helper in gpu_hw_d3d12.cpp and the GetBatchPipeline
    branch that routes m_texture_filter == Nearest +
    texture_mode != Disabled to the new picker. Parallel to
    b6d1903 for untextured.
  - Then the remaining 3 textured filter templates (Bilinear /
    JINC2 / xBR) each as a foundation + activation pair.
2026-05-19 09:12:23 +02:00
..
d3d12 core/d3d12: pre-bake the textured Nearest batch FS slice 2026-05-19 09:12:23 +02:00
vulkan core: route TRANSPARENCY through the batch UBO 2026-05-19 08:13:03 +02:00