Files

T

marfrit a4e7d8ab90 initial seed: retrofit campaign lineage from local working trees

panvk-bifrost campaigns (r1..r4 Vulkan compositor + r5.video1 Vulkan
video decode) shipped before this repo existed; the deliverable
patches live in marfrit-packages, but the reasoning chain, phase docs,
and source-state evidence lived only in local working trees on the
development host.

This retrofit imports:
- mesa-panvk-bifrost/   — r1..r4 era phase docs (iter1..iter18)
                          (libmali stub blobs at iter18/blob/ excluded
                          — 109MB of RE artifacts replaced with a README
                          pointer)
- mesa-panvk-bifrost-video/ — sibling campaign phase docs + probe
- evidence/             — frozen .tgz source snapshots at each milestone
                          (basis for the 0005 patch diff generation)

Future iterations should branch off here from day one, so each iter is
a commit rather than a snapshot. See [[feedback-session-local-process-pins]]
for the process drift this retrofit closes.

Total: 1.9 MB across 124 files.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-23 05:25:37 +02:00

4.6 KiB

Raw Permalink Blame History

Phase 0 — substrate for iter2

Opened 2026-05-19 by mfritsche + claude-noether, immediately after iter1 closed GREEN (phase8_iteration1_close.md).

Locked research question — iter2

Get a minimal Vulkan image-side workload to execute end-to-end on PanVk-Bifrost (ohm / Mali-G52 r1 MC1 / PAN_I_WANT_A_BROKEN_VULKAN_DRIVER=1): create a 4×4 VK_FORMAT_R8G8B8A8_UNORM image with TRANSFER_DST|TRANSFER_SRC usage on device-local memory, transition UNDEFINED → TRANSFER_DST_OPTIMAL, vkCmdClearColorImage to color 0x11223344 (R=0x11 G=0x22 B=0x33 A=0x44), transition TRANSFER_DST_OPTIMAL → TRANSFER_SRC_OPTIMAL, vkCmdCopyImageToBuffer to host-visible staging, fence-wait, verify all 16 pixels read back as 0x44332211 (little-endian uint32). No GPU faults in dmesg, no validation errors.

If GREEN: lock iter3 against a first-triangle graphics pipeline (vertex + fragment shader, fullscreen triangle via gl_VertexIndex, render pass or dynamic rendering, single draw, readback).

If RED: characterize the first failure point and fix or work around in iter2.

Why this shape

iter1 collapsed three of four phase0 hypotheses on the compute side (device init, cmd-buffer recording, shader compilation). iter2 bridges from compute to graphics by adding only image-handling machinery, keeping the same submit/sync skeleton:

VkImage create + bind (new)
Image layout transitions via VkImageMemoryBarrier (new)
vkCmdClearColorImage (new — but it's a transfer op, not a real graphics pipeline)
vkCmdCopyImageToBuffer (new — but again a transfer op)
Optimal tiling (new — Bifrost arranges tiles differently from Valhall in some cases)

Notably not in iter2:

Render pass / dynamic rendering
Vertex + fragment shaders
Graphics pipeline state (rasterizer, viewport, blend, depth)
Vertex buffers / index buffers
Framebuffer

So if iter2 fails, the failure points to image/layout/transfer machinery, not "graphics pipeline" in general. That's a usefully narrow target.

Hypothesis space (where iter2 may fail)

Image creation + memory binding. Bifrost has specific tiling layouts (e.g. block-based tiling). vkGetImageMemoryRequirements may report a size and alignment Mesa's PanVk-Bifrost path can't satisfy, or the allocator may pick a memory type that's not actually usable for an optimal-tiled image.
Layout transitions via image barriers. The Bifrost cache / tiler invalidation hooks may not be wired into the JM submit path consistently. Specifically: UNDEFINED → TRANSFER_DST and TRANSFER_DST → TRANSFER_SRC transitions need to flush L2 / invalidate tile caches, and that's per-arch code that may have rotted in the v6/v7 paths.
vkCmdClearColorImage lowering. PanVk may lower vkCmdClearColorImage to either a real hardware clear (tile-level) or a compute shader (meta clear). Bifrost-specific paths exist (the lone bifrost/panvk_vX_meta_desc_copy.c is descriptor-copy meta — a similar clear-meta path may or may not work).
vkCmdCopyImageToBuffer + tiling decode. Bifrost optimal tiling is non-linear — a copy from optimal-tiled image to a linear buffer needs the tiler to detile correctly. If detile is wrong, the readback will show pixel-shuffled output (recognizable from the pattern of 0x11/0x22/0x33/0x44 bytes).

The clear color 0x11223344 was chosen specifically: each pixel byte is distinct, so a pixel-shuffle bug will show up as wrong-byte-order rather than all-zeros (which would mean clear didn't fire at all).

Phase 0 deliverables

This document. iter2's substrate is lighter than iter1's because iter1 already proved out the broader environment.

In-scope (LOCKED 2026-05-19 for iter2)

Hardware: ohm only.
Format: R8G8B8A8_UNORM, optimal tiling, 4×4, 1 mip, 1 layer.
Operations: image create + bind + 2 layout transitions + clear + image-to-buffer copy.
Verification: all 16 pixels = 0x44332211.

Out-of-scope (LOCKED 2026-05-19 for iter2)

Render pass / dynamic rendering (iter3).
Vertex / fragment shaders (iter3).
Graphics pipeline state (iter3).
WSI / swapchain (iter4+).
Larger / multi-mip / multi-layer / multi-sample images.
Other formats (R5G6B5, R32G32B32A32_SFLOAT, depth/stencil, ASTC). Add later if it makes sense to exercise per-format codepaths.
Sub-region clears, scissored copies.
Compute path (proven in iter1; not revisited).
Upstreaming.

Reference

phase0_findings.md — campaign substrate.
phase8_iteration1_close.md — iter1 close.
iter1/ — compute probe (reusable skeleton for iter2).

4.6 KiB Raw Permalink Blame History Unescape Escape