[VMA] Phase 2 — switch suballocator to VMA Virtual Blocks, add budget-aware reclaim & safer commit #345

wildcard · 2025-08-29T00:43:33+02:00

wildcard commented

2025-08-29 00:43:33 +02:00

[V2]
Replace the custom suballocator (std::vector + gap scan) with VMA Virtual Blocks
Suballocator architecture
Old code used, commits vector of ranges + FindFreeRegion() linear scan whereas the new uses per-block VmaVirtualBlock with vmaVirtualAllocate/Free.
Similarly,
the MemoryAllocation::Commit(size, alignment) now uses VB.
MemoryAllocation::Free(begin) → vmaVirtualFree() by stored handle.
TLDR,
Introduce VmaVirtualBlock in MemoryAllocation; RAII destruction in dtor and all the corresponding handling changes to VB.
And all the previous goodies from V1.
[V1]
This patch makes VMA suballocation smarter under pressure and fixes a masking bug when RenderDoc is attached.
Adds a budget-aware check (via vmaGetHeapBudgets) and attempts to reclaim empty internal blocks before allocating new device memory blocks.
Implements a bounded OOM recovery ladder (retry, trim empty blocks, relax DEVICE_LOCAL, fall back to a smaller chunk, then fail).
This is phase one of the VMA changes and after testing, phase 2 shall roll out.

[V2] Replace the custom suballocator (std::vector<Range> + gap scan) with VMA Virtual Blocks Suballocator architecture Old code used, commits vector of ranges + FindFreeRegion() linear scan whereas the new uses per-block VmaVirtualBlock with vmaVirtualAllocate/Free. Similarly, the MemoryAllocation::Commit(size, alignment) now uses VB. MemoryAllocation::Free(begin) → vmaVirtualFree() by stored handle. TLDR, Introduce VmaVirtualBlock in MemoryAllocation; RAII destruction in dtor and all the corresponding handling changes to VB. And all the previous goodies from V1. [V1] This patch makes VMA suballocation smarter under pressure and fixes a masking bug when RenderDoc is attached. Adds a budget-aware check (via vmaGetHeapBudgets) and attempts to reclaim empty internal blocks before allocating new device memory blocks. Implements a bounded OOM recovery ladder (retry, trim empty blocks, relax DEVICE_LOCAL, fall back to a smaller chunk, then fail). This is phase one of the VMA changes and after testing, phase 2 shall roll out.

Lizzie reviewed

2025-08-29 00:46:51 +02:00

src/video_core/vulkan_common/vulkan_memory_allocator.cpp Outdated

					
				@ -303,0 +369,4 @@

				            }

				        }

				    throw vk::Exception(VK_ERROR_OUT_OF_DEVICE_MEMORY);

Sadpath codegen for stack unwinding is not optimal for exceptions
Maybe exploring not having to use them?

Sadpath codegen for stack unwinding is not optimal for exceptions Maybe exploring not having to use them?

we can leave it as it is since if we are oom, we have bigger issues than making code optimal for exception.

Lizzie reviewed

2025-08-29 00:59:37 +02:00

src/video_core/vulkan_common/vulkan_memory_allocator.cpp Outdated

					
				@ -286,0 +309,4 @@

				}

				void MemoryAllocator::FreeEmptyAllocations() {

				        std::vector<MemoryAllocation*> to_release;

boost::container::small_vector

solved it by not using it

oh wait i am blind

wildcard changed title from ~~[VMA] phase 1: Budget aware sub allocation, "Actual" Robust OOM handling.~~ to [VMA] Phase 2 — switch suballocator to VMA Virtual Blocks, add budget-aware reclaim & safer commit

2025-08-29 19:19:07 +02:00

Shinmegumi force-pushed VMA_Changes_v1 from 0942ea1cc9

eden-license / license-header (pull_request) Successful in 27s

Details

to 69de851e84

eden-license / license-header (pull_request) Successful in 27s

Details

2025-08-29 22:53:17 +02:00

Compare

wildcard closed this pull request

2025-09-01 05:06:30 +02:00

wildcard commented