Ben Niu

Add /alternatename:main for ARM64EC

1

According to https://docs.microsoft.com/en-us/cpp/build/arm64ec-windows-abi-conventions?view=msvc-170#arm64ec-function-name-decoration, the ARM64EC name decoration is different from the x64 and ARM64, so we need to update the /alternatename:main for ARM64EC.

Reduce FP16 kernel size by replacing ldr pairs with ldp

1

Summary: Replace ldr pairs with ldp for smaller kernel code. Without the change, the kernel assemblies have a total of 1954 instructions. With the change, the number of kernel instructions...

Enable arm64 convolution for fbgemm through the reference convolution APIs

1

Summary: This diff adds convolution support to arm64 fbgemm by reusing existing reference implementations. 1. Introduced conv_requant_ref that invokes the reference conv_ref and requantize_u8acc32_ref and added it in places where...

Slightly improve requantize_ AVX2 code performance

1

Summary: 1. Tweaked permute_mask_v so that we can remove pair-wise duplication by _mm256_moveldup_ps 2. Removed max_v since _mm256_min_epu8(xyzw_packed_v, max_v) is just xyzw_packed_v. Differential Revision: D86371930