VREV for byte swap on Arm Neon?

[Triang3l]

The Arm Neon versions of byte swaps (volk_*_byteswap.h) in VOLK use shifts/OR or lookup tables, somewhat similar to the x86 versions. However, Neon has a dedicated instruction for byte swaps — VREV, usable as vrev16q_u8 for 8-in-16, vrev32q_u8 for 8-in-32, and vrev64q_u8 for 8-in-64. Are there performance/compatibility reasons for not using it, or is that more of not knowing about the instruction when the code was written?

[jdemel]

Could you point to the exact code that you have in mind?

Are you refering to this?

volk/kernels/volk/volk_16u_byteswap.h

Lines 223 to 239 in 237a6fc

	static inline void volk_16u_byteswap_neon(uint16_t* intsToSwap, unsigned int num_points)
	{
	unsigned int number;
	unsigned int eighth_points = num_points / 8;
	uint16x8_t input, output;
	uint16_t* inputPtr = intsToSwap;
	for (number = 0; number < eighth_points; number++) {
	input = vld1q_u16(inputPtr);
	output = vsriq_n_u16(output, input, 8);
	output = vsliq_n_u16(output, input, 8);
	vst1q_u16(inputPtr, output);
	inputPtr += 8;
	}
	volk_16u_byteswap_generic(inputPtr, num_points - eighth_points * 8);
	}

This implementation is essentially 7 years old. Are you interested in contributing an optimized version of this code?

[Triang3l]

Yes, and the vtbl4/vqtbl1q implementations for 32-bit and 64-bit swaps. I can try setting up the environment on my phone and write direct vrev versions, and possibly run some speed comparisons, as well as tests, in the weekend.