RAM: drop the renderer's CPU copy of the GPU geometry buffers (~2 GB in heavy regions) #584

New issue

Closed

opened 2026-10-09 02:20:12 +00:00 by sky_line · 1 comment

sky_line commented

2026-10-09 02:20:12 +00:00

Owner

A jemalloc heap profile in a heavy region (2026-10-09; 7.1 GB live heap, unsampled) attributes about 2.1 GB to LLCleanRendererVK::fillFromCleanStores / placeSlot. That is a full CPU copy of the GPU mega-buffers (mCpuVerts, mCpuIdx, mCpuWeights, plus the slot table), including freed ranges and vector growth capacity.

The GPU buffers are authoritative, and the meshstore keeps the quantized per-slot mirror (#573/#582) that the packer reads from, so this third copy only serves placement and repack.

Goal

Remove the CPU shadow of the GPU geometry buffers.

Placement: writes slots straight from the meshstore mirror into the upload staging buffer, at offsets from the existing free-list allocators. Those keep only offset bookkeeping, no bytes.
Compaction / full repack: done on the GPU with vkCmdCopyBuffer between ranges (or into a new buffer) instead of rebuilding a CPU image.
CPU readers of the packed copy (stats, debug dumps, garbage accounting) switch to bookkeeping or the meshstore mirror.

Done when

fillFromCleanStores no longer shows in heap profiles beyond small bookkeeping.
RSS drops by about 2 GB in the heavy region.
Geometry, rigged meshes, RT and shadows are unchanged, and there are no upload hitches.

A jemalloc heap profile in a heavy region (2026-10-09; 7.1 GB live heap, unsampled) attributes about **2.1 GB** to `LLCleanRendererVK::fillFromCleanStores` / `placeSlot`. That is a full CPU copy of the GPU mega-buffers (`mCpuVerts`, `mCpuIdx`, `mCpuWeights`, plus the slot table), including freed ranges and vector growth capacity. The GPU buffers are authoritative, and the meshstore keeps the quantized per-slot mirror (#573/#582) that the packer reads from, so this third copy only serves placement and repack. ## Goal Remove the CPU shadow of the GPU geometry buffers. - **Placement:** writes slots straight from the meshstore mirror into the upload staging buffer, at offsets from the existing free-list allocators. Those keep only offset bookkeeping, no bytes. - **Compaction / full repack:** done on the GPU with `vkCmdCopyBuffer` between ranges (or into a new buffer) instead of rebuilding a CPU image. - **CPU readers** of the packed copy (stats, debug dumps, garbage accounting) switch to bookkeeping or the meshstore mirror. ## Done when - `fillFromCleanStores` no longer shows in heap profiles beyond small bookkeeping. - RSS drops by about 2 GB in the heavy region. - Geometry, rigged meshes, RT and shadows are unchanged, and there are no upload hitches.

sky_line added the

perf

status

in-progress

labels

2026-10-09 02:20:12 +00:00

sky_line added this to the Slipstream project

2026-10-09 02:20:12 +00:00

sky_line added

status

testing

and removed