panvk-bifrost/mesa-panvk-bifrost/phase8_iteration2_close.md

Iteration 2 close — GREEN

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

Locked question

(From phase0_findings_iter2.md)

Get a minimal Vulkan image-side workload to execute end-to-end on PanVk-Bifrost: create a 4×4 VK_FORMAT_R8G8B8A8_UNORM image, transition UNDEFINED → TRANSFER_DST, vkCmdClearColorImage to 0x11223344, transition TRANSFER_DST → TRANSFER_SRC, vkCmdCopyImageToBuffer to host-visible staging, fence-wait, verify all 16 pixels read back as 0x44332211.

Result: GREEN

7/7 runs PASS (1 baseline + 1 with VK_LAYER_KHRONOS_validation + 5 stability). All 16 pixels match exactly. No GPU faults, no MMU faults, no kernel-side panfrost messages, no validation-layer warnings or errors.

Evidence: phase0_evidence/iter2_image_clear_run.txt.

What the close tells us

Four image-side hypotheses from phase0_findings_iter2.md were tested. All four work:

Hypothesis	Status at iter2
H1: image creation + memory binding	✗ no — vkCreateImage + vkGetImageMemoryRequirements + bind work for 4×4 RGBA8 optimal-tiled (4096-byte aligned allocation)
H2: layout transitions	✗ no — UNDEFINED→TRANSFER_DST and TRANSFER_DST→TRANSFER_SRC both clean
H3: vkCmdClearColorImage lowering	✗ no — clear lands in image correctly
H4: vkCmdCopyImageToBuffer + Bifrost tile decode	✗ no — all 16 pixels round-trip with no shuffling, no rounding error

The image-side transfer path on PanVk-Bifrost is functional for this minimal case. Combined with iter1, we now know the following work end-to-end:

• Vulkan instance + physical device + logical device + queue
• Buffer create + alloc + bind + map (host-visible)
• Image create + alloc + bind (device-local)
• Image layout transitions via vkCmdPipelineBarrier
• vkCmdClearColorImage (transfer-op level, not via shader)
• vkCmdCopyImageToBuffer with Bifrost tile-layout decode
• Compute pipeline: shader module + pipeline layout + compute pipeline + dispatch
• Command buffer recording + submit + fence wait + memory barriers (memory + image + buffer)

What we still don't know works: graphics pipeline (vertex + fragment + rasterizer + render pass / dynamic rendering).

iter2 in-tree artifacts

• iter2/probe_image_clear.c — ~340 LoC, pure Vulkan 1.0 core
• iter2/Makefile — make builds, make run / make run-validation

Deferred to iter3+ (not in iter2 scope)

• Vertex + fragment shaders
• Render pass and/or dynamic rendering
• Graphics pipeline state (rasterizer, viewport, blend, depth)
• Larger images, mipmaps, layered images, MSAA
• Other formats (R32G32B32A32_SFLOAT, BC/ETC2/ASTC compressed, depth/stencil)
• WSI / swapchain (iter4+)
• TuxRacer / Zink-on-PanVk

Next iter — iter3 lock proposal

Smallest viable graphics workload that exercises the rasterizer + shaders:

Render a single full-screen triangle into a 64×64 R8G8B8A8_UNORM color attachment via dynamic rendering (VK_KHR_dynamic_rendering), using a trivial vertex shader (no vertex buffer — emit positions from gl_VertexIndex) and a trivial fragment shader (output constant color gl_FragCoord-encoded so we can detect rasterizer correctness). Copy attachment to host-visible buffer. Verify: (a) some pixels are written (not all sentinel), (b) at least one pixel has the encoded gl_FragCoord value matching its position.

Justifications:

• 64×64 (not 4×4) so multiple tiles get exercised — Bifrost is a tile-based rasterizer, so single-tile workloads might side-step real tile binning.
• Dynamic rendering instead of render pass — simpler API surface, no framebuffer object, no subpass dependencies. Render pass / framebuffer can be iter3.5 if needed.
• Fullscreen triangle from gl_VertexIndex so no vertex buffer needed — exercises pipeline-state but not vertex-input-state.
• Trivial fragment shader (no textures, no UBO, no SSBO) — exercises rasterization + frag shader output but not descriptor lookups (proven in iter1 anyway).
• gl_FragCoord-encoded color so a wrong-rasterization bug (e.g. swapped-Y framebuffer convention, off-by-pixel) is detectable from pixel data.

If iter3 turns up the first real failure, that's the campaign's first interesting bug. If iter3 also passes, iter4 adds vertex buffer + UBO + a texture sample, and we're well into "actually exercising PanVk-Bifrost" territory.

Pacing: same 8-phase cadence. iter3 phase 0 substrate lock when the operator opens.