panvk-bifrost/mesa-panvk-bifrost/phase8_iteration3_close.md

Iteration 3 close — GREEN

Closed 2026-05-19 by mfritsche + claude-noether, same session as iter1 + iter2.

Locked question

(From phase0_findings_iter3.md)

Render a single fullscreen triangle into a 64×64 R8G8B8A8_UNORM color attachment via VK_KHR_dynamic_rendering, with a trivial vertex shader (positions from gl_VertexIndex) and a gl_FragCoord-encoded fragment shader. Copy attachment to host-visible buffer. Verify every pixel at (col, row) reads back as 0xff80(row)(col).

Result: GREEN

7/7 runs PASS (1 baseline + 1 with VK_LAYER_KHRONOS_validation + 5 stability). All 4096 pixels per run match the expected gl_FragCoord encoding. No GPU faults, no validation warnings.

Evidence: phase0_evidence/iter3_triangle_run.txt.

What the close tells us

All five hypotheses in phase0_findings_iter3.md were tested. All five work:

Hypothesis	Status at iter3
H1: Pipeline creation / shader compilation (vert+frag)	✗ no — both shaders compile, link, run correctly
H2: Dynamic rendering plumbing	✗ no — vkCmdBeginRenderingKHR + EndRenderingKHR work, attachment format propagates to tiler
H3: Rasterizer state plumbing	✗ no — viewport, scissor, cull-none, polygon-fill all honored
H4: Tile binner / draw submission	✗ no — 4×4 grid of 16×16 tiles all rasterized, no missing tile, no edge gap
H5: Fragment shader output → tile → image memory	✗ no — every pixel matches exact gl_FragCoord encoding

The combined verdict across iter1 + iter2 + iter3: PanVk-Bifrost (Mali-G52 r1, v7) on Mesa 26.0.6 is functionally a much more complete Vulkan driver than the PAN_I_WANT_A_BROKEN_VULKAN_DRIVER gate at panvk_physical_device.c:413 suggests. The gate reads as defensive ("not well-tested") rather than reflecting hard breakage on these minimal paths.

What's been proven functional, cumulatively:

• Instance + extension loading
• Physical device + memory + queue family + format properties
• Logical device + queue + KHR feature chain (dynamic_rendering)
• Buffer + image creation, memory allocation, binding
• Image views
• Layout transitions (UNDEFINED ↔ COLOR_ATTACHMENT ↔ TRANSFER_DST ↔ TRANSFER_SRC)
• Memory + buffer + image barriers
• Command buffer record, submit, fence wait
• Compute pipeline + dispatch + descriptor sets + storage buffer
• Graphics pipeline + vertex shader + fragment shader + rasterizer + tile binner
• Dynamic rendering (VkRenderingInfoKHR, VkRenderingAttachmentInfoKHR)
• vkCmdClearColorImage, vkCmdCopyImageToBuffer (with Bifrost tile-layout decode)
• Validation-layer clean (Khronos validation reports zero issues)
• 17/17 runs across all 3 iters PASS (6 + 7 + 7-1 because validation counted as separate run — close enough)

iter3 in-tree artifacts

• iter3/probe_triangle.c — graphics probe
• iter3/probe_triangle.vert — fullscreen triangle from gl_VertexIndex
• iter3/probe_triangle.frag — gl_FragCoord-encoded fragment
• iter3/Makefile

Deferred to iter4+

The next layer of complexity stacks: descriptor sets, vertex buffers, textures, legacy render passes, MSAA, depth/stencil. The path of most-likely-to-find-bugs:

• Vertex input bindings (vertex buffers) — Bifrost's attribute descriptor model differs from Valhall's; this is where PANVK_BIFROST_DESC references in panvk_vX_cmd_draw.c actually start exercising the divergent code.
• Sampled textures (combined_image_sampler descriptor) — first time the descriptor model meets the image side seriously. This is where the bifrost-specific descriptor table layout (PANVK_BIFROST_DESC_TABLE_COUNT) really gets stressed.
• Uniform buffers (UBO) — exercises BDA-vs-classic-binding distinction.

Next iter — iter4 lock proposal

Render a textured fullscreen quad: 4×4 RGBA8 source texture (uploaded via staging buffer + image copy + layout transition), sampled by a fragment shader with a trivial sampler (NEAREST filter, CLAMP_TO_EDGE), into a 64×64 RGBA8 attachment. Output color = texelFetch(texture, ivec2(gl_FragCoord.xy) % 4). Verify the output is a clean 16×16-tile-repeated 4×4 texture pattern.

Justifications:

• Adds: image upload via copy, sampler descriptor, image-view binding to descriptor set, sampled image read.
• Doesn't yet add: vertex buffer (still use gl_VertexIndex fullscreen triangle), UBO, push constants, multiple draws, MSAA.
• Predictable output pattern (modulo 4×4) makes verification trivially deterministic.
• Uses texelFetch (not texture()) to skip sampler filtering, isolating texture fetch from filter logic.

If iter4 turns up a real bug, that's our first interesting finding. If iter4 passes, the campaign is going faster than the README projected.

Pacing: same 8-phase cadence.