Repository navigation
vulkan: use double-precision rcp_d on AMD wave32 devices - #37
Merged
Vasyl-Trefilov merged 1 commit intoNov 15, 2026
Merged
Conversation
Member
|
ty again! |
Vasyl-Trefilov
added a commit
that referenced
this pull request
Nov 15, 2026
… drafters, Windows NVRTC lookup - web UI: terminal-style redesign, message actions, sampler presets, export/import, load options, accent color by model family - serve: batched decode and batched prefill of queued requests on dense CUDA models; cap of cached CUDA graphs - serve suggests calibration; no queue limit by default (--max-waiting) - EAGLE3 speculative decoding, n-gram drafters, --prompt-cache, RDNA3.5 Vulkan work (#36) - cuda: NVRTC found by our own name list, so Windows finds a CUDA 12.x DLL (#35) - vulkan: double-precision rcp_d on AMD wave32 devices (#37) - TQ1_0 and TQ2_0 GGUF types
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
On
main(b99db7d),quant_xqruns in wave32 on AMD RDNA (W64 == false), sorcp_d(a, b)falls back tofdiv(a, b)(a * (1.0 / b)). On AMD Radeon 890M (gfx1150, RADV Mesa 26.2.6), that rounding difference failstests/kernels.rs(vulkan_type_matrixandgemv_partial_superblocks_and_q2_0) withgemm_dense/gemmrelative errors around1.2e-4 .. 1.5e-4.This adds specialization constant
13(RCP_D, set whenvendor_id == 0x1002) sorcp_dusesdouble(a) / double(b)on AMD even in wave32, while keeping NVIDIA wave32fdivbehavior completely untouched.Tested on AMD Radeon 890M (
gfx1150) with./tools/gen_spirv.sh && cargo test --release --lib --bins --tests: all 19 test crates pass (includingtests/kernels.rs: 31 passed, 0 failed, 4 ignored).