Quelle  avx512.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0-or-later
/* -*- linux-c -*- --------------------------------------------------------
 *
 *   Copyright (C) 2016 Intel Corporation
 *
 *   Author: Gayatri Kammela <gayatri.kammela@intel.com>
 *   Author: Megha Dey <megha.dey@linux.intel.com>
 *
 *   Based on avx2.c: Copyright 2012 Yuanhan Liu All Rights Reserved
 *   Based on sse2.c: Copyright 2002 H. Peter Anvin - All Rights Reserved
 *
 * -----------------------------------------------------------------------
 */

/*
 * AVX512 implementation of RAID-6 syndrome functions
 *
 */

#include <linux/raid/pq.h>
#include "x86.h"

static const struct raid6_avx512_constants {
 u64 x1d[8];
} raid6_avx512_constants __aligned(512/8) = {
 { 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL,
   0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL,
   0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL,
   0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL,},
};

static int raid6_have_avx512(void)
{
 return boot_cpu_has(X86_FEATURE_AVX2) &&
  boot_cpu_has(X86_FEATURE_AVX) &&
  boot_cpu_has(X86_FEATURE_AVX512F) &&
  boot_cpu_has(X86_FEATURE_AVX512BW) &&
  boot_cpu_has(X86_FEATURE_AVX512VL) &&
  boot_cpu_has(X86_FEATURE_AVX512DQ);
}

static void raid6_avx5121_gen_syndrome(int disks, size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = disks - 3;         /* Highest data disk */
 p = dptr[z0+1];         /* XOR parity */
 q = dptr[z0+2];         /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0\n\t"
       "vpxorq %%zmm1,%%zmm1,%%zmm1" /* Zero temp */
       :
       : "m" (raid6_avx512_constants.x1d[0]));

 for (d = 0; d < bytes; d += 64) {
  asm volatile("prefetchnta %0\n\t"
        "vmovdqa64 %0,%%zmm2\n\t"     /* P[0] */
        "prefetchnta %1\n\t"
        "vmovdqa64 %%zmm2,%%zmm4\n\t" /* Q[0] */
        "vmovdqa64 %1,%%zmm6"
        :
        : "m" (dptr[z0][d]), "m" (dptr[z0-1][d]));
  for (z = z0-2; z >= 0; z--) {
   asm volatile("prefetchnta %0\n\t"
         "vpcmpgtb %%zmm4,%%zmm1,%%k1\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm6,%%zmm2,%%zmm2\n\t"
         "vpxorq %%zmm6,%%zmm4,%%zmm4\n\t"
         "vmovdqa64 %0,%%zmm6"
         :
         : "m" (dptr[z][d]));
  }
  asm volatile("vpcmpgtb %%zmm4,%%zmm1,%%k1\n\t"
        "vpmovm2b %%k1,%%zmm5\n\t"
        "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
        "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
        "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
        "vpxorq %%zmm6,%%zmm2,%%zmm2\n\t"
        "vpxorq %%zmm6,%%zmm4,%%zmm4\n\t"
        "vmovntdq %%zmm2,%0\n\t"
        "vpxorq %%zmm2,%%zmm2,%%zmm2\n\t"
        "vmovntdq %%zmm4,%1\n\t"
        "vpxorq %%zmm4,%%zmm4,%%zmm4"
        :
        : "m" (p[d]), "m" (q[d]));
 }

 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}

static void raid6_avx5121_xor_syndrome(int disks, int start, int stop,
           size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = stop;  /* P/Q right side optimization */
 p = dptr[disks-2]; /* XOR parity */
 q = dptr[disks-1]; /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0"
       : : "m" (raid6_avx512_constants.x1d[0]));

 for (d = 0 ; d < bytes ; d += 64) {
  asm volatile("vmovdqa64 %0,%%zmm4\n\t"
        "vmovdqa64 %1,%%zmm2\n\t"
        "vpxorq %%zmm4,%%zmm2,%%zmm2"
        :
        : "m" (dptr[z0][d]),  "m" (p[d]));
  /* P/Q data pages */
  for (z = z0-1 ; z >= start ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vmovdqa64 %0,%%zmm5\n\t"
         "vpxorq %%zmm5,%%zmm2,%%zmm2\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4"
         :
         : "m" (dptr[z][d]));
  }
  /* P/Q left side optimization */
  for (z = start-1 ; z >= 0 ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4"
         :
         : );
  }
  asm volatile("vpxorq %0,%%zmm4,%%zmm4\n\t"
  /* Don't use movntdq for r/w memory area < cache line */
        "vmovdqa64 %%zmm4,%0\n\t"
        "vmovdqa64 %%zmm2,%1"
        :
        : "m" (q[d]), "m" (p[d]));
 }

 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}

const struct raid6_calls raid6_avx512x1 = {
 raid6_avx5121_gen_syndrome,
 raid6_avx5121_xor_syndrome,
 raid6_have_avx512,
 "avx512x1",
 .priority = 2  /* Prefer AVX512 over priority 1 (SSE2 and others) */
};

/*
 * Unrolled-by-2 AVX512 implementation
 */
static void raid6_avx5122_gen_syndrome(int disks, size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = disks - 3;         /* Highest data disk */
 p = dptr[z0+1];         /* XOR parity */
 q = dptr[z0+2];         /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0\n\t"
       "vpxorq %%zmm1,%%zmm1,%%zmm1" /* Zero temp */
       :
       : "m" (raid6_avx512_constants.x1d[0]));

 /* We uniformly assume a single prefetch covers at least 64 bytes */
 for (d = 0; d < bytes; d += 128) {
  asm volatile("prefetchnta %0\n\t"
        "prefetchnta %1\n\t"
        "vmovdqa64 %0,%%zmm2\n\t"      /* P[0] */
        "vmovdqa64 %1,%%zmm3\n\t"      /* P[1] */
        "vmovdqa64 %%zmm2,%%zmm4\n\t"  /* Q[0] */
        "vmovdqa64 %%zmm3,%%zmm6"      /* Q[1] */
        :
        : "m" (dptr[z0][d]), "m" (dptr[z0][d+64]));
  for (z = z0-1; z >= 0; z--) {
   asm volatile("prefetchnta %0\n\t"
         "prefetchnta %1\n\t"
         "vpcmpgtb %%zmm4,%%zmm1,%%k1\n\t"
         "vpcmpgtb %%zmm6,%%zmm1,%%k2\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpmovm2b %%k2,%%zmm7\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
         "vmovdqa64 %0,%%zmm5\n\t"
         "vmovdqa64 %1,%%zmm7\n\t"
         "vpxorq %%zmm5,%%zmm2,%%zmm2\n\t"
         "vpxorq %%zmm7,%%zmm3,%%zmm3\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6"
         :
         : "m" (dptr[z][d]), "m" (dptr[z][d+64]));
  }
  asm volatile("vmovntdq %%zmm2,%0\n\t"
        "vmovntdq %%zmm3,%1\n\t"
        "vmovntdq %%zmm4,%2\n\t"
        "vmovntdq %%zmm6,%3"
        :
        : "m" (p[d]), "m" (p[d+64]), "m" (q[d]),
          "m" (q[d+64]));
 }

 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}

static void raid6_avx5122_xor_syndrome(int disks, int start, int stop,
           size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = stop;  /* P/Q right side optimization */
 p = dptr[disks-2]; /* XOR parity */
 q = dptr[disks-1]; /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0"
       : : "m" (raid6_avx512_constants.x1d[0]));

 for (d = 0 ; d < bytes ; d += 128) {
  asm volatile("vmovdqa64 %0,%%zmm4\n\t"
        "vmovdqa64 %1,%%zmm6\n\t"
        "vmovdqa64 %2,%%zmm2\n\t"
        "vmovdqa64 %3,%%zmm3\n\t"
        "vpxorq %%zmm4,%%zmm2,%%zmm2\n\t"
        "vpxorq %%zmm6,%%zmm3,%%zmm3"
        :
        : "m" (dptr[z0][d]), "m" (dptr[z0][d+64]),
          "m" (p[d]), "m" (p[d+64]));
  /* P/Q data pages */
  for (z = z0-1 ; z >= start ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm7,%%zmm7,%%zmm7\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpcmpgtb %%zmm6,%%zmm7,%%k2\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpmovm2b %%k2,%%zmm7\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
         "vmovdqa64 %0,%%zmm5\n\t"
         "vmovdqa64 %1,%%zmm7\n\t"
         "vpxorq %%zmm5,%%zmm2,%%zmm2\n\t"
         "vpxorq %%zmm7,%%zmm3,%%zmm3\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6"
         :
         : "m" (dptr[z][d]),  "m" (dptr[z][d+64]));
  }
  /* P/Q left side optimization */
  for (z = start-1 ; z >= 0 ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm7,%%zmm7,%%zmm7\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpcmpgtb %%zmm6,%%zmm7,%%k2\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpmovm2b %%k2,%%zmm7\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6"
         :
         : );
  }
  asm volatile("vpxorq %0,%%zmm4,%%zmm4\n\t"
        "vpxorq %1,%%zmm6,%%zmm6\n\t"
        /* Don't use movntdq for r/w
      * memory area < cache line
      */
        "vmovdqa64 %%zmm4,%0\n\t"
        "vmovdqa64 %%zmm6,%1\n\t"
        "vmovdqa64 %%zmm2,%2\n\t"
        "vmovdqa64 %%zmm3,%3"
        :
        : "m" (q[d]), "m" (q[d+64]), "m" (p[d]),
          "m" (p[d+64]));
 }

 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}

const struct raid6_calls raid6_avx512x2 = {
 raid6_avx5122_gen_syndrome,
 raid6_avx5122_xor_syndrome,
 raid6_have_avx512,
 "avx512x2",
 .priority = 2  /* Prefer AVX512 over priority 1 (SSE2 and others) */
};

#ifdef CONFIG_X86_64

/*
 * Unrolled-by-4 AVX2 implementation
 */
static void raid6_avx5124_gen_syndrome(int disks, size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = disks - 3;         /* Highest data disk */
 p = dptr[z0+1];         /* XOR parity */
 q = dptr[z0+2];         /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0\n\t"
       "vpxorq %%zmm1,%%zmm1,%%zmm1\n\t"       /* Zero temp */
       "vpxorq %%zmm2,%%zmm2,%%zmm2\n\t"       /* P[0] */
       "vpxorq %%zmm3,%%zmm3,%%zmm3\n\t"       /* P[1] */
       "vpxorq %%zmm4,%%zmm4,%%zmm4\n\t"       /* Q[0] */
       "vpxorq %%zmm6,%%zmm6,%%zmm6\n\t"       /* Q[1] */
       "vpxorq %%zmm10,%%zmm10,%%zmm10\n\t"    /* P[2] */
       "vpxorq %%zmm11,%%zmm11,%%zmm11\n\t"    /* P[3] */
       "vpxorq %%zmm12,%%zmm12,%%zmm12\n\t"    /* Q[2] */
       "vpxorq %%zmm14,%%zmm14,%%zmm14"        /* Q[3] */
       :
       : "m" (raid6_avx512_constants.x1d[0]));

 for (d = 0; d < bytes; d += 256) {
  for (z = z0; z >= 0; z--) {
  asm volatile("prefetchnta %0\n\t"
        "prefetchnta %1\n\t"
        "prefetchnta %2\n\t"
        "prefetchnta %3\n\t"
        "vpcmpgtb %%zmm4,%%zmm1,%%k1\n\t"
        "vpcmpgtb %%zmm6,%%zmm1,%%k2\n\t"
        "vpcmpgtb %%zmm12,%%zmm1,%%k3\n\t"
        "vpcmpgtb %%zmm14,%%zmm1,%%k4\n\t"
        "vpmovm2b %%k1,%%zmm5\n\t"
        "vpmovm2b %%k2,%%zmm7\n\t"
        "vpmovm2b %%k3,%%zmm13\n\t"
        "vpmovm2b %%k4,%%zmm15\n\t"
        "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
        "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
        "vpaddb %%zmm12,%%zmm12,%%zmm12\n\t"
        "vpaddb %%zmm14,%%zmm14,%%zmm14\n\t"
        "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
        "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
        "vpandq %%zmm0,%%zmm13,%%zmm13\n\t"
        "vpandq %%zmm0,%%zmm15,%%zmm15\n\t"
        "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
        "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
        "vpxorq %%zmm13,%%zmm12,%%zmm12\n\t"
        "vpxorq %%zmm15,%%zmm14,%%zmm14\n\t"
        "vmovdqa64 %0,%%zmm5\n\t"
        "vmovdqa64 %1,%%zmm7\n\t"
        "vmovdqa64 %2,%%zmm13\n\t"
        "vmovdqa64 %3,%%zmm15\n\t"
        "vpxorq %%zmm5,%%zmm2,%%zmm2\n\t"
        "vpxorq %%zmm7,%%zmm3,%%zmm3\n\t"
        "vpxorq %%zmm13,%%zmm10,%%zmm10\n\t"
        "vpxorq %%zmm15,%%zmm11,%%zmm11\n"
        "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
        "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
        "vpxorq %%zmm13,%%zmm12,%%zmm12\n\t"
        "vpxorq %%zmm15,%%zmm14,%%zmm14"
        :
        : "m" (dptr[z][d]), "m" (dptr[z][d+64]),
          "m" (dptr[z][d+128]), "m" (dptr[z][d+192]));
  }
  asm volatile("vmovntdq %%zmm2,%0\n\t"
        "vpxorq %%zmm2,%%zmm2,%%zmm2\n\t"
        "vmovntdq %%zmm3,%1\n\t"
        "vpxorq %%zmm3,%%zmm3,%%zmm3\n\t"
        "vmovntdq %%zmm10,%2\n\t"
        "vpxorq %%zmm10,%%zmm10,%%zmm10\n\t"
        "vmovntdq %%zmm11,%3\n\t"
        "vpxorq %%zmm11,%%zmm11,%%zmm11\n\t"
        "vmovntdq %%zmm4,%4\n\t"
        "vpxorq %%zmm4,%%zmm4,%%zmm4\n\t"
        "vmovntdq %%zmm6,%5\n\t"
        "vpxorq %%zmm6,%%zmm6,%%zmm6\n\t"
        "vmovntdq %%zmm12,%6\n\t"
        "vpxorq %%zmm12,%%zmm12,%%zmm12\n\t"
        "vmovntdq %%zmm14,%7\n\t"
        "vpxorq %%zmm14,%%zmm14,%%zmm14"
        :
        : "m" (p[d]), "m" (p[d+64]), "m" (p[d+128]),
          "m" (p[d+192]), "m" (q[d]), "m" (q[d+64]),
          "m" (q[d+128]), "m" (q[d+192]));
 }

 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}

static void raid6_avx5124_xor_syndrome(int disks, int start, int stop,
           size_t bytes, void **ptrs)
{
 u8 **dptr = (u8 **)ptrs;
 u8 *p, *q;
 int d, z, z0;

 z0 = stop;  /* P/Q right side optimization */
 p = dptr[disks-2]; /* XOR parity */
 q = dptr[disks-1]; /* RS syndrome */

 kernel_fpu_begin();

 asm volatile("vmovdqa64 %0,%%zmm0"
       :: "m" (raid6_avx512_constants.x1d[0]));

 for (d = 0 ; d < bytes ; d += 256) {
  asm volatile("vmovdqa64 %0,%%zmm4\n\t"
        "vmovdqa64 %1,%%zmm6\n\t"
        "vmovdqa64 %2,%%zmm12\n\t"
        "vmovdqa64 %3,%%zmm14\n\t"
        "vmovdqa64 %4,%%zmm2\n\t"
        "vmovdqa64 %5,%%zmm3\n\t"
        "vmovdqa64 %6,%%zmm10\n\t"
        "vmovdqa64 %7,%%zmm11\n\t"
        "vpxorq %%zmm4,%%zmm2,%%zmm2\n\t"
        "vpxorq %%zmm6,%%zmm3,%%zmm3\n\t"
        "vpxorq %%zmm12,%%zmm10,%%zmm10\n\t"
        "vpxorq %%zmm14,%%zmm11,%%zmm11"
        :
        : "m" (dptr[z0][d]), "m" (dptr[z0][d+64]),
          "m" (dptr[z0][d+128]), "m" (dptr[z0][d+192]),
          "m" (p[d]), "m" (p[d+64]), "m" (p[d+128]),
          "m" (p[d+192]));
  /* P/Q data pages */
  for (z = z0-1 ; z >= start ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm7,%%zmm7,%%zmm7\n\t"
         "vpxorq %%zmm13,%%zmm13,%%zmm13\n\t"
         "vpxorq %%zmm15,%%zmm15,%%zmm15\n\t"
         "prefetchnta %0\n\t"
         "prefetchnta %2\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpcmpgtb %%zmm6,%%zmm7,%%k2\n\t"
         "vpcmpgtb %%zmm12,%%zmm13,%%k3\n\t"
         "vpcmpgtb %%zmm14,%%zmm15,%%k4\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpmovm2b %%k2,%%zmm7\n\t"
         "vpmovm2b %%k3,%%zmm13\n\t"
         "vpmovm2b %%k4,%%zmm15\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
         "vpaddb %%zmm12,%%zmm12,%%zmm12\n\t"
         "vpaddb %%Zmm14,%%zmm14,%%zmm14\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
         "vpandq %%zmm0,%%zmm13,%%zmm13\n\t"
         "vpandq %%zmm0,%%zmm15,%%zmm15\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
         "vpxorq %%zmm13,%%zmm12,%%zmm12\n\t"
         "vpxorq %%zmm15,%%zmm14,%%zmm14\n\t"
         "vmovdqa64 %0,%%zmm5\n\t"
         "vmovdqa64 %1,%%zmm7\n\t"
         "vmovdqa64 %2,%%zmm13\n\t"
         "vmovdqa64 %3,%%zmm15\n\t"
         "vpxorq %%zmm5,%%zmm2,%%zmm2\n\t"
         "vpxorq %%zmm7,%%zmm3,%%zmm3\n\t"
         "vpxorq %%zmm13,%%zmm10,%%zmm10\n\t"
         "vpxorq %%zmm15,%%zmm11,%%zmm11\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
         "vpxorq %%zmm13,%%zmm12,%%zmm12\n\t"
         "vpxorq %%zmm15,%%zmm14,%%zmm14"
         :
         : "m" (dptr[z][d]), "m" (dptr[z][d+64]),
           "m" (dptr[z][d+128]),
           "m" (dptr[z][d+192]));
  }
  asm volatile("prefetchnta %0\n\t"
        "prefetchnta %1\n\t"
        :
        : "m" (q[d]), "m" (q[d+128]));
  /* P/Q left side optimization */
  for (z = start-1 ; z >= 0 ; z--) {
   asm volatile("vpxorq %%zmm5,%%zmm5,%%zmm5\n\t"
         "vpxorq %%zmm7,%%zmm7,%%zmm7\n\t"
         "vpxorq %%zmm13,%%zmm13,%%zmm13\n\t"
         "vpxorq %%zmm15,%%zmm15,%%zmm15\n\t"
         "vpcmpgtb %%zmm4,%%zmm5,%%k1\n\t"
         "vpcmpgtb %%zmm6,%%zmm7,%%k2\n\t"
         "vpcmpgtb %%zmm12,%%zmm13,%%k3\n\t"
         "vpcmpgtb %%zmm14,%%zmm15,%%k4\n\t"
         "vpmovm2b %%k1,%%zmm5\n\t"
         "vpmovm2b %%k2,%%zmm7\n\t"
         "vpmovm2b %%k3,%%zmm13\n\t"
         "vpmovm2b %%k4,%%zmm15\n\t"
         "vpaddb %%zmm4,%%zmm4,%%zmm4\n\t"
         "vpaddb %%zmm6,%%zmm6,%%zmm6\n\t"
         "vpaddb %%zmm12,%%zmm12,%%zmm12\n\t"
         "vpaddb %%zmm14,%%zmm14,%%zmm14\n\t"
         "vpandq %%zmm0,%%zmm5,%%zmm5\n\t"
         "vpandq %%zmm0,%%zmm7,%%zmm7\n\t"
         "vpandq %%zmm0,%%zmm13,%%zmm13\n\t"
         "vpandq %%zmm0,%%zmm15,%%zmm15\n\t"
         "vpxorq %%zmm5,%%zmm4,%%zmm4\n\t"
         "vpxorq %%zmm7,%%zmm6,%%zmm6\n\t"
         "vpxorq %%zmm13,%%zmm12,%%zmm12\n\t"
         "vpxorq %%zmm15,%%zmm14,%%zmm14"
         :
         : );
  }
  asm volatile("vmovntdq %%zmm2,%0\n\t"
        "vmovntdq %%zmm3,%1\n\t"
        "vmovntdq %%zmm10,%2\n\t"
        "vmovntdq %%zmm11,%3\n\t"
        "vpxorq %4,%%zmm4,%%zmm4\n\t"
        "vpxorq %5,%%zmm6,%%zmm6\n\t"
        "vpxorq %6,%%zmm12,%%zmm12\n\t"
        "vpxorq %7,%%zmm14,%%zmm14\n\t"
        "vmovntdq %%zmm4,%4\n\t"
        "vmovntdq %%zmm6,%5\n\t"
        "vmovntdq %%zmm12,%6\n\t"
        "vmovntdq %%zmm14,%7"
        :
        : "m" (p[d]),  "m" (p[d+64]), "m" (p[d+128]),
          "m" (p[d+192]), "m" (q[d]),  "m" (q[d+64]),
          "m" (q[d+128]), "m" (q[d+192]));
 }
 asm volatile("sfence" : : : "memory");
 kernel_fpu_end();
}
const struct raid6_calls raid6_avx512x4 = {
 raid6_avx5124_gen_syndrome,
 raid6_avx5124_xor_syndrome,
 raid6_have_avx512,
 "avx512x4",
 .priority = 2  /* Prefer AVX512 over priority 1 (SSE2 and others) */
};
#endif