diff options
Diffstat (limited to 'results/classifier/deepseek-r1:14b/output/other/844')
| -rw-r--r-- | results/classifier/deepseek-r1:14b/output/other/844 | 45 |
1 files changed, 45 insertions, 0 deletions
diff --git a/results/classifier/deepseek-r1:14b/output/other/844 b/results/classifier/deepseek-r1:14b/output/other/844 new file mode 100644 index 000000000..5235eebcf --- /dev/null +++ b/results/classifier/deepseek-r1:14b/output/other/844 @@ -0,0 +1,45 @@ + +Close gap for x86_64-v3 ABI in TCG - CPU support for fma, f16c, avx, avx2 features required +Description of problem: +There are 3 additional ABIs defined by a collaboration of vendors for the `x86_64` architecture, over the original baseline: + +* https://gitlab.com/x86-psABIs/x86-64-ABI/-/blob/master/x86-64-ABI/low-level-sys-info.tex + +This is no problem for KVM assuming suitable host hardware, but TCG is currently unable to support more than the original baseline and the `x86_64-v2` step. + +For `x86_64-v3` there are some gaps in its emulation coverage. This can be seen by taking `Nehalem` which is a good fit for `x86_64-v2`, and requesting the extra v3 features: + +``` +$ qemu-system-x86_64 -accel tcg -cpu Nehalem,+avx,+avx2,+bmi1,+bmi2,+f16c,+fma,+abm,+movbe +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.fma [bit 12] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.avx [bit 28] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.f16c [bit 29] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.07H:EBX.avx2 [bit 5] +``` + +IOW, the strict bare minimum TCG needs in order to satisfy `x86_64-v3` is `fma`, `f16c`, `avx` and `avx2` support + +If we want to fully support a named CPU model satisfying v3, then `Haswell` is the closest and that has a few additional gaps + +``` +$ qemu-system-x86_64 -accel tcg -cpu Haswell-noTSX +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.fma [bit 12] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.pcid [bit 17] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.x2apic [bit 21] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.tsc-deadline [bit 24] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.avx [bit 28] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.01H:ECX.f16c [bit 29] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.07H:EBX.avx2 [bit 5] +qemu-system-x86_64: warning: TCG doesn't support requested feature: CPUID.07H:EBX.invpcid [bit 10] + +``` + +Those additional gaps wouldn't impact ability to execute binaries build for the `x86_64-v3` ABI though, so not as important. + +The reason `x86_64-v3` compatibility in TCG matters is because sooner or later some Linux OS are going to set this as the baseline for their compiler toolchain. There is a proposal to set this in `Fedora ELN`, which is what feeds in to a possible future `RHEL-10`. + +I imagine adding these extra features would be non-negligible work in TCG / take some time to complete. + +Thus I file this bug for the purpose of suggesting these 4 specific missing features be considered a priority to address, compared to other missing CPU features in TCG that might be considered more of a 'nice to have'. + +eg looking further the `x86_64-v4` baseline brings in a requirement for `avx512f`, `avx512bw`, `avx512cd`, `avx512dq`, `avx512vl` which TCG also lacks, but I don't think they really need to be considered important at this point in time. |