fbarchard

I've got a reproducible, but its a bit confusing ``` git clone https://github.com/google/XNNPACK.git cd XNNPACK bazel build -c opt bench:all ``` has many strict-aliasing errors... another others `bazel build bench:all`...

Agreed that it would be better to fix the warning, but 1. I cant reproduce the issue with clang - it doesnt have the same warning 2. on linux x86...

At the moment the strict aliasing is disabled, and for bazel there are other build errors that prevent the build, so this is lower priority until the build works. But...

Thanks for the offer to upload. Certainly worth consideration. Is it reasonably simple? Would it benefit from support in the main stream xnnpack github? If its not already, each dependency...

I ran into this as well but for immintrin.h theres a work around #define __MM_MALLOC_H 1 https://godbolt.org/z/cr6qGsrfq Theres a bug in the clang-cl 18 that _mm256_fmadd_ps doesnt compile with /arch:AVX...

armv8 (a series) already works with mca... e.g. https://godbolt.org/z/PEzW7Tvra armv7 I expect -m32 already works, but M-series arm uses -mthumb and M55 is thumb2 (a mix of 16/32 bit) for...

Hi, thanks for the question. As you referred to Windows, thats something I'm learning myself. From linux, I've used clangcl via lexan, which is --config=lexan_x86_64 and I've used wine or...

There is no Exynos or Tensor support. I think the field we want is soc. eq Samsung S23: [ro.soc.model]: [SM8550]

I have a Samsung S23 handy, and can confirm SoC is 'unknown' SoC name: Unknown Microarchitectures: 1x Cortex-X3 4x Cortex-A715 3x Cortex-A510 Cores: 0: 1 processor (0), ARM Cortex-X3 1:...

I see this same behavior (cast to int) in sse2, 0armv7, wasmsimd, and hexagon. Also hexagon qfloats do not have a formal NaN or inf. You can find implementations for...