__lsx_vhsubw_h_b

LSX intrinsic · group: Integer Computation

Synopsis

__m128i __lsx_vhsubw_h_b (__m128i a, __m128i b)
#include <lsxintrin.h>
Instruction: vhsubw.h.b vr, vr, vr
CPU Flags: LSX

Description

Subtract even-positioned signed 8-bit elements in b from odd-positioned signed 8-bit elements in a, sign-extending both to 16-bit, to get the 16-bit result.

Operation

for (int i = 0; i < 8; i++) {
  dst.half[i] = (s16)(s8)a.byte[2 * i + 1] - (s16)(s8)b.byte[2 * i];
}

// Expands to:

if (0) {
  dst.half[0] = ((s16)((s8)a.byte[1])) - ((s16)((s8)b.byte[0]));
  dst.half[1] = ((s16)((s8)a.byte[3])) - ((s16)((s8)b.byte[2]));
  dst.half[2] = ((s16)((s8)a.byte[5])) - ((s16)((s8)b.byte[4]));
  dst.half[3] = ((s16)((s8)a.byte[7])) - ((s16)((s8)b.byte[6]));
  dst.half[4] = ((s16)((s8)a.byte[9])) - ((s16)((s8)b.byte[8]));
  dst.half[5] = ((s16)((s8)a.byte[11])) - ((s16)((s8)b.byte[10]));
  dst.half[6] = ((s16)((s8)a.byte[13])) - ((s16)((s8)b.byte[12]));
  dst.half[7] = ((s16)((s8)a.byte[15])) - ((s16)((s8)b.byte[14]));
}

Tested on real machine.

Latency and Throughput

CPU	µarch	Latency	Throughput (IPC)
3C5000	LA464	2	2
3A6000	LA664	2	2
3C6000	LA664	2	2
2K1000LA	LA264	2	1
2K3000	LA364E	2	1

---

All intrinsics in alphabetical order