Permutation

__m256i __lasx_xvpermi_w (__m256i a, __m256i b, imm0_255 imm)

Synopsis

__m256i __lasx_xvpermi_w (__m256i a, __m256i b, imm0_255 imm)
#include <lasxintrin.h>
Instruction: xvpermi.w xr, xr, imm
CPU Flags: LASX

Description

Permute words from a and b with indices recorded in imm and store into dst.

Operation

dst.word[0] = b.word[imm & 0x3];
dst.word[1] = b.word[(imm >> 2) & 0x3];
dst.word[2] = a.word[(imm >> 4) & 0x3];
dst.word[3] = a.word[(imm >> 6) & 0x3];
dst.word[4] = b.word[4 + (imm & 0x3)];
dst.word[5] = b.word[4 + ((imm >> 2) & 0x3)];
dst.word[6] = a.word[4 + ((imm >> 4) & 0x3)];
dst.word[7] = a.word[4 + ((imm >> 6) & 0x3)];

Tested on real machine.

Latency and Throughput

CPU Latency Throughput (IPC)
3A6000 1 4
3C5000 1 2

__m256i __lasx_xvpermi_d (__m256i a, imm0_255 imm)

Synopsis

__m256i __lasx_xvpermi_d (__m256i a, imm0_255 imm)
#include <lasxintrin.h>
Instruction: xvpermi.d xr, xr, imm
CPU Flags: LASX

Description

Permute double words from a and b with indices recorded in imm and store into dst.

Operation

dst.dword[0] = a.dword[imm & 0x3];
dst.dword[1] = a.dword[(imm >> 2) & 0x3];
dst.dword[2] = a.dword[(imm >> 4) & 0x3];
dst.dword[3] = a.dword[(imm >> 6) & 0x3];

Tested on real machine.

Latency and Throughput

CPU Latency Throughput (IPC)
3A6000 3 4
3C5000 3 2

__m256i __lasx_xvpermi_q (__m256i a, __m256i b, imm0_255 imm)

Synopsis

__m256i __lasx_xvpermi_q (__m256i a, __m256i b, imm0_255 imm)
#include <lasxintrin.h>
Instruction: xvpermi.d xr, xr, imm
CPU Flags: LASX

Description

Permute quad words from a and b with indices recorded in imm and store into dst.

Operation

if ((imm & 0x4) && MACHINE_3C5000) {
  // Caveat: observed in 3C5000
  dst.qword[0] = 0;
} else {
  dst.qword[0] = (imm & 2) ? a.qword[imm & 0x1] : b.qword[imm & 0x1];
}
if ((imm & 0x80) && MACHINE_3C5000) {
  // Caveat: observed in 3C5000
  dst.qword[1] = 0;
} else {
  dst.qword[1] =
      (imm & 0x20) ? a.qword[(imm >> 4) & 0x1] : b.qword[(imm >> 4) & 0x1];
}

Tested on real machine.

Latency and Throughput

CPU Latency Throughput (IPC)
3A6000 3 4
3C5000 3 2

__m256i __lasx_xvperm_w (__m256i a, __m256i b)

Synopsis

__m256i __lasx_xvperm_w (__m256i a, __m256i b)
#include <lasxintrin.h>
Instruction: xvperm.w xr, xr, xr
CPU Flags: LASX

Description

Permute words from a with indices recorded in b and store into dst.

Operation

for (int i = 0; i < 8; i++) {
  dst.word[i] = a.word[b.word[i] % 0x8];
}

Tested on real machine.

Latency and Throughput

CPU Latency Throughput (IPC)
3A6000 3 4
3C5000 3 2