ZstdGpuMemsetMemcpy: Increase threadgroup size and binary-search scalarization optimizations.#100
Open
Jonathan-Weinstein-AMD wants to merge 7 commits intomicrosoft:developmentfrom
Conversation
…o 64. For one thing, enables wave64 on RDNA3/RDNA4. Also, all waves in a threadgroup are launched on the same compute-unit, which can have L0/K$ benefits, so in general having a larger threadgroup size can help, and performance for the insects archive does seem to improve more when trying 128 threads. Performance of [Memcpy RAW blocks, Memset RLE blocks]: new_duration/old_duration for insects archive, clocks were fixed at some non-specific frequency, but timings were fairly consistent: 7900 XTX: ~= 0.60 RTX 4080: ~= 0.67
The insects archive has 233 raw blocks that sum to 18,860,773 bytes. It is quite likely then that all threads in a wave write to the same block. Detect this and scalarize the second zstdgpu_BinarySearch to lookup the frame index. Performance of [Memcpy RAW blocks, Memset RLE blocks]: new_duration/old_duration for insects archive, clocks were fixed at some non-specific frequency, but timings were fairly consistent: 7900 XTX: - Against last commit (tgsize already 64): ~= 0.62 - Against upstream: ~= 0.38
… too. This is a bit fiddley and I don't super like the added `if (blockSize == 0) continue;` in `ParseFrame`. Performance of [Memcpy RAW blocks, Memset RLE blocks]: new_duration/old_duration for insects archive, clocks were fixed at some non-specific frequency, but timings were fairly consistent: 7900 XTX: - Against last commit (tgsize already 64 and had WaveActiveAllEqual from frame lookup): ~= 0.77 - Against upstream: ~= 0.29
…ressedBlocks" This reverts commit 273d1af.
…Optimization: remove `i >= lds[r]` check.
…zero-decompressed-size Raw/RLE blocks. No per-thread LDS used (still some LDS, a word per group so don't have to care about wave size or things of that nature). K$ vs LDS latency is perhaps comparable. I think I decided LDS at first since I thought to do a binary search within the a tail of `min(Constants.blockCount - groupLeaderBlockIdx, numActiveThreads)` when assuming there are no empty Raw/RLE blocks at that point.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Increasing
kzstdgpu_TgSizeX_MemsetMemcpyfrom 32 to 64 yields nice performance improvements on both an AMD 7900 XTX and an NV RTX 4080. Reasons for this on AMD may include going from wave32 to wave64, which can help in ways like thes_clauseinstruction can, and all waves (though still just one in this case, but it has more lanes) of a threadgroup are launched on the same compute-unit; each compute-unit has its own L0/K$, and it is likely adjacent threadgroups touch adjacent memory.In the insects archive, the threadgroup size change alone nets between a 30-40% duration reduction on both GPUs; clocks were fixed at some non-specific frequency in both cases, but timings were fairly consistent.
Increasing the threadgroup size to 128 or 256 could further improve performance, but I decide to stop at 64 now, and that's what xbox seems to do.
The insects archive has 233 raw blocks that sum to 18,860,773 bytes. The second optimization (f570ccb) in this PR tries to detect the (likely) case all lanes in a wave store within the same block (via
WaveActiveAllEqual(globalBlockIdx)), and if so, does the frame lookup in a uniform context. This seems pretty neutral on RTX 4080, but on 7900 XTX it further reduces duration by about 38% on top of the threadgroup size change.The last optimization then also tries to scalarize the first binary search too. This absorbs the previous optimization. This also seems pretty neutral on RTX 4080, but on 7900 XTX it seems about a 23% duration reduction on top of the previous change.
In total, all three changes net about a 70% duration reduction for the insects archive on 7900 XTX (at some low-ish fixed-clock frequency).
There's no rush for changes like these.