[brkosebay]

Depends on #2; please review it after #2. It is stacked on #2's branch (g5p-pr1), and its 5 commits come after
#2's 8.

What this is. The first feature PR after #2, as agreed: G5P graphics pipelines compiled by our backend, and
non-indexed draws. Together with #2's render job, an A9 renders a cleared target with a triangle on it,
pixel-exact on hardware. It is built only with -Dimagination-g5p-experimental=true, and it edits no apple_g9p
path.

What it builds, and what it refuses by name (G5P refuses: ... on stderr and VK_ERROR_FEATURE_NOT_PRESENT), so
nothing unbuilt runs:

• Built: a vertex and a fragment shader through our backend (SPIR-V → NIR → G5P USC), 32-bit float vertex
  attributes from one vertex buffer, triangle lists/strips/fans and line lists/strips, one instance, a dynamic or
  static viewport and scissor, culling, and varyings (the coefficient program).
• Refused: more than one colour target, blending and write masks, depth/stencil, descriptors and push constants,
  indexed draws, instancing, a scissor smaller than the attachment, rasterizer discard, multiview, MSAA. These are
  the next PRs (the order is at the end).

This is AI-written. Every commit carries Generated-by: LLM. It is for your fork only, not for upstream Mesa.

Commits

1. 1bb4e34b7dc pvr: apple_g5p: SPIR-V to NIR for the G5P backend
   • pvr_g5p_nir.c: a stage's SPIR-V becomes our backend's program the way our A9 driver has always prepared it.
     That means its NIR options and pass sequence (the backend's recognised forms depend on them), the point-size,
     unattached-output and sample-mask lowerings, and the vertex system values.
   • Then pco_g5p_compile_general_link(). The general backend only: no captured-shape matcher, no out-of-process
     fallback.
2. 3b99268eda8 pvr: apple_g5p: graphics pipelines and draws: pvr_arch_g5p_gfx.c, in the shape of your
   pvr_arch_g9p_gfx.c (graphics_pipeline_compile, cmd_draw, emit_draw).
   • A pipeline compiles the fragment program first and links the vertex program to the varyings it reads. It
     uploads both USC programs and three pdsc programs in Apple's DOUT form: the vertex fetch, the fragment launch
     and the coefficient program. It also uploads its two static PPP blocks.
   • A draw records the vertex-fetch data, six per-draw PPP blocks and its VDM words (csbgen). The stream-out
     program is G5P (Apple A9) on the shared PowerVR driver, behind an experimental option #2's device HALT, as your stream_out is.
   • Also: a g5p pointer in pvr_graphics_pipeline, and pvr_pipeline.h includes pvr_g5p.h at its end (both
     #ifdef).
3. aaebfeda7f0 pvr/winsys: apple_g5p: log the control stream and its PPP terminate: with PVR_G5P_WS_LOG, the
   submit also dumps the VDM stream, so a job's words can be compared offline.
4. 176a6c781e9 pvr: apple_g5p: hook graphics pipelines and draws: five one-line hooks (below).
5. 1a183f6a6b5 imagination/g5p: triangle, a hardware test of the draw path: the test program below, in
   src/imagination/g5p/tests/triangle/ (GLSL compiled by its build.sh, not built by meson).

Every edit outside the new files

New files: vulkan/pvr_g5p_nir.c, vulkan/pvr_arch_g5p_gfx.c. G5P files changed: pvr_g5p.h, pvr_arch_g5p.c (a
check that the HALT's code sits at data + 16, the layout the stream-out block points at), the winsys, and
pvr_g5p_hooks.h (four new hooks; load_op is #2's).

File	Lines	What and why
src/imagination/vulkan/meson.build	+2	the two new files, in the G5P-only block
vulkan/pvr_pipeline.h	+10	struct pvr_g5p_pipeline *g5p at the end of pvr_graphics_pipeline, and #include "pvr_g5p.h" at the end of the header, both #ifdef PVR_APPLE_G5P. The header's only __LINE__ users after them are two assert()s, which b_ndebug compiles out
vulkan/pvr_arch_pipeline.c	+2 −2	hooks: pipeline_destroy (a G5P pipeline frees its programs and returns), pipeline_compile (a G5P pipeline is compiled by pvr_arch_g5p_graphics_pipeline_compile())
vulkan/pvr_arch_cmd_buffer.c	+3 −3	hooks: draw (CmdDraw → pvr_arch_g5p_cmd_draw(), then CmdDraw's own end label), draw_indexed (refused by name), load_op in CmdNextSubpass2 (#2's hook: no soft load op; clears and loads are background objects, as on G9P)

Each of the five hooks takes the place of a blank line right before your G9P if, since vk_error() and friends
follow in both files. So the hook design is #2's:

• PVR_G5P_HOOK(name, ...) is one line ≤ 80 columns and expands to nothing without the option. Every line keeps its
  number, and every byte of the option-off driver stays the same.
• With the option, it expands to PVR_G5P_HOOK_<name> in pvr_g5p_hooks.h, which documents each one.
• The draw hook reads CmdDraw's parameters (vertexCount, instanceCount, firstVertex, firstInstance) by
  name and jumps to the label it is given. Passing the four parameters would not fit one 80-column line.
• No #line directives. The hook lines and the new macros are clang-format clean with src/imagination/.clang-format.

git diff aad20d6e7e7 176a6c781e9 -- src/imagination/g9p 'src/imagination/vulkan/*g9p*' is empty.

Gates

Same container and options as #2 (Fedora, GCC 16, aarch64, git archive into fixed paths; release, stripped,
b_ndebug). The gates ran on 3ea65ec6bcd, whose tree is identical to this branch's 176a6c781e9, commit by
commit.

• Option off, byte for byte: PASS.
  • 3ea65ec6bcd (tree = 176a6c781e9) installs the same 165 files as your main (73a49abe247), with identical
    bytes. Commit 5 adds only the test program's files, which meson doesn't build.
  • cmp of the two libvulkan_powervr_mesa.so finds no difference:
    76e29856fe5011d2d63fd45756a6ed7fd9ad62de79a1a107df4d40c431cc7584 in both, the same hash as G5P (Apple A9) on the shared PowerVR driver, behind an experimental option #2's.
• Option on, the offline smoke over our fake node (it runs our kernel's checks on every job; it does not render):
  • smoke, the clear-only render and the triangle are submitted and retired: G5P-PVR-SMOKE PASS,
    G5P-PVR-RENDER PASS, G5P-PVR-TRIANGLE PASS;
  • G5P (Apple A9) on the shared PowerVR driver, behind an experimental option #2's 15 render-job fields still equal our reference job: step6-compare: PASS (15 of 15 fields equal);
  • the draw's words equal our A9 driver's recorded job for the same scene: step7-compare: PASS (28 of 28 fields equal), against both recorded placements (general-triangle, p1-triangle):
    • the VDM stream: 30 words, addresses masked;
    • every program word: the vertex fetch, the fragment launch, the HALTs;
    • the eight PPP blocks;
    • the fetch data;
    • the size words;
    • 17 pointer relations, each address landing on its object;
    • the USC bytes: the vertex shader's 70 and the fragment shader's 32 are byte-identical.

No Apple binaries

• No program is held as bytes. The USC programs come from our backend, from SPIR-V. The PDS programs come from pdsc:
  the vertex fetch (DDMAD, DOUTU, HALT), the fragment launch (DOUTU) and the coefficient program (DOUTI). The
  stream-out program is G5P (Apple A9) on the shared PowerVR driver, behind an experimental option #2's pdsc HALT.

• Every constant initializer in the new and changed files was grepped (static const, array initializers, hex
  literals). The only hits:

  • three named register values (below);
  • zero-initialised work arrays;
  • a const float f[6] of the viewport's computed scale and offset;
  • a launch_data[4] holding the computed DOUTU word;
  • the dump labels.

• What stays Apple-derived: four register values, named #defines in pvr_arch_g5p_gfx.c:

  Name	Value	What it is
  G5P_VS_STATE5	0x1082	the VDM_STATE5 base: PDS temps 2 (the data size and unified store are set per draw)
  G5P_VDM_CAM_SIZE	95	the vertex cache size for vertices of up to 32 dwords
  G5P_WCLAMP	0x1e3ce508	the W clamp of every captured draw (~1e-20f)
  G5P_STREAM_OUT1	0x4022	STREAM_OUT1, your G9P's 0x4022 too

  Everything else is packed by csbgen from field values: the ISP words, region clip, output select, PPP control and
  the index list.

Hardware result (iPad 5, A9, our kernel)

• Flown on g5p-1b cb42bf093fe: this PR's code with the older, padded hook placement. Kernel
  7.3.0-rc1-release8-ga10a8556b98e, with Apple's unmodified GPU firmware, beside a live Hyprland session.
  • G5P-1B-DEVICE RESULT PASS vulkaninfo, smoke, 20 device cycles back to zero, clear-only render exact, triangle exact, ... gpu events 0.
  • The triangle: 1012 inside and 2796 outside exact, edge band 288 (140 red, 148 clear, 0 other). Every pixel
    at least 1 px from an edge is exact. The edge band holds only the two colours.
  • The words the device's driver used (its own dump): step7-compare 28 of 28. VDM 30 words; VS 70 B and FS 32 B
    byte-identical to the reference.
• The hardware run was on cb42bf093fe, not on this branch. This PR's code differs from what flew only in where
  the hooks sit; the option-on smoke and step7-compare above (28 of 28, the same words) ran on this PR's tree itself.
  • git diff cb42bf093fe 176a6c781e9 over pvr_arch_g5p_gfx.c, pvr_g5p_nir.c, pvr_arch_g5p.c and
    src/imagination/g5p is empty.
  • In pvr_g5p.h, one prototype moved under its own heading (the commit split).
  • The winsys diff is G5P (Apple A9) on the shared PowerVR driver, behind an experimental option #2's own (it no longer checks the geometry stream).
  • The rest is hook placement: padded #ifdef blocks became PVR_G5P_HOOK lines, and their bodies moved into
    pvr_g5p_hooks.h unchanged.

Its hardware test program

triangle, in this PR at src/imagination/g5p/tests/triangle/ (commit 5). It is plain Vulkan with no driver
internals, so the same binary runs on your A10 against your G9P path:

src/imagination/g5p/tests/triangle/build.sh /tmp/triangle
/tmp/triangle/triangle --triangle --verify

• It needs glslangValidator, a C compiler, and the Vulkan headers and loader. On the device it needs a
  host-visible memory type, a linear B8G8R8A8_UNORM colour attachment of 64×64, and a render pass object. The
  shaders are GLSL; no SPIR-V binary is committed.

• What it does:

  1. It clears the image, which it pre-fills with 0x5a.
  2. It draws one red triangle from a vertex buffer, with a negative-height viewport.
  3. It checks every pixel at least 1 px from an edge: red inside, the clear colour outside. The edge band is
     reported apart and may hold only those two colours.

  The verdict is G5P-PVR-TRIANGLE PASS|FAIL.

• Reference run on lavapipe (LLVM 22.1.8): G5P-PVR-TRIANGLE PASS, 1012 inside and 2796 outside exact, edge
  band 288 (140 red, 148 clear, 0 other). These are the same counts as the A9's.

• --render --verify checks a clear-only pass. --load --verify (two passes, the second loading the first) is for
  the follow-up LOAD PR.

• --bos PATH reads our kernel's BO accounting (debugfs). It's optional; leave it out on the A10.

After this PR

Each comes as its own PR on this one, with the same gates:

• textures, samplers and blend (on hardware: 13 of 13 exact);
• depth/stencil (on hardware: 5 of 5 rungs exact);
• WSI (on hardware: 300 frames presented into Hyprland);
• LOAD/STORE as a compiled background program (in progress), and uniforms;
• 4× MSAA, MRT, transform feedback;
• full vertex input (formats, instancing, topologies, indexed draws);
• queries and copies.

🤖 Generated with Claude Code