__lasx_xvsrarni_h_w

LASX intrinsic · group: Shift

Synopsis

__m256i __lasx_xvsrarni_h_w (__m256i a, __m256i b, imm0_31 imm)
#include <lasxintrin.h>
Instruction: xvsrarni.h.w xr, xr, imm
CPU Flags: LASX

Description

Arithmetic right shift with rounding the signed 32-bit elements in a and b by imm, write the lower 16 bits of the shifted result into the lower half of each 128-bit lane of dst, and zero the upper half.

Operation

for (int i = 0; i < 8; i++) {
  if (i < 4) {
    if (imm == 0) {
      dst.half[i] = (s16)(s32)b.word[i];
    } else {
      dst.half[i] = (s16)(((s32)b.word[i] >> imm) +
                          (((s32)b.word[i] >> (imm - 1)) & 0x1));
    }
  } else {
    if (imm == 0) {
      dst.half[i] = (s16)(s32)a.word[i - 4];
    } else {
      dst.half[i] = (s16)(((s32)a.word[i - 4] >> imm) +
                          (((s32)a.word[i - 4] >> (imm - 1)) & 0x1));
    }
  }
}
for (int i = 8; i < 16; i++) {
  if (i < 12) {
    if (imm == 0) {
      dst.half[i] = (s16)(s32)b.word[i - 4];
    } else {
      dst.half[i] = (s16)(((s32)b.word[i - 4] >> imm) +
                          (((s32)b.word[i - 4] >> (imm - 1)) & 0x1));
    }
  } else {
    if (imm == 0) {
      dst.half[i] = (s16)(s32)a.word[i - 8];
    } else {
      dst.half[i] = (s16)(((s32)a.word[i - 8] >> imm) +
                          (((s32)a.word[i - 8] >> (imm - 1)) & 0x1));
    }
  }
}

Tested on real machine.

Latency and Throughput

CPU	µarch	Latency	Throughput (IPC)
3C5000	LA464	4	1
3A6000	LA664	4	2
3C6000	LA664	4	2

---

All intrinsics in alphabetical order