]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
Allow all RDNA2 archs to use sdot4 intrinsic (llama/8629)
authorJeroen Mostert <redacted>
Tue, 23 Jul 2024 08:50:40 +0000 (10:50 +0200)
committerGeorgi Gerganov <redacted>
Sat, 27 Jul 2024 15:26:12 +0000 (18:26 +0300)
commit5937ee2959ea5bb906a1b418ee1c05271c5dc387
treed591856f804c463a456dbcb0357349d91c89166b
parentaad0ddec8f29dbdf85eb46815318171072dbac83
Allow all RDNA2 archs to use sdot4 intrinsic (llama/8629)

The check gating the use of `__builtin_amdgc_sdot4` specifically checks for gfx1030. This causes a severe perf regression for anything gfx103? that's not gfx1030 and not using `HSA_OVERRIDE_GFX_VERSION` (if you've built ROCm to support it). We already have a generic RDNA2 define, let's use it.
src/ggml-cuda/common.cuh