bcolz/c-blosc/blosc/shuffle-sse2.c

/*********************************************************************
  Blosc - Blocked Shuffling and Compression Library

  Author: Francesc Alted <francesc@blosc.org>

  See LICENSES/BLOSC.txt for details about copyright and rights to use.
**********************************************************************/

#include "shuffle-generic.h"
#include "shuffle-sse2.h"

/* Make sure SSE2 is available for the compilation target and compiler. */
#if !defined(__SSE2__)
  #error SSE2 is not supported by the target architecture/platform and/or this compiler.
#endif

#include <emmintrin.h>


/* The next is useful for debugging purposes */
#if 0
#include <stdio.h>
#include <string.h>

static void printxmm(__m128i xmm0)
{
  uint8_t buf[16];

  ((__m128i *)buf)[0] = xmm0;
  printf("%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x,%x\n",
          buf[0], buf[1], buf[2], buf[3],
          buf[4], buf[5], buf[6], buf[7],
          buf[8], buf[9], buf[10], buf[11],
          buf[12], buf[13], buf[14], buf[15]);
}
#endif


/* Routine optimized for shuffling a buffer for a type size of 2 bytes. */
static void
shuffle2_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 2;
  size_t j;
  int k;
  uint8_t* dest_for_jth_element;
  __m128i xmm0[2], xmm1[2];

  for (j = 0; j < vectorizable_elements; j += sizeof(__m128i)) {
    /* Fetch 16 elements (32 bytes) then transpose bytes, words and double words. */
    for (k = 0; k < 2; k++) {
      xmm0[k] = _mm_loadu_si128((__m128i*)(src + (j * bytesoftype) + (k * sizeof(__m128i))));
      xmm0[k] = _mm_shufflelo_epi16(xmm0[k], 0xd8);
      xmm0[k] = _mm_shufflehi_epi16(xmm0[k], 0xd8);
      xmm0[k] = _mm_shuffle_epi32(xmm0[k], 0xd8);
      xmm1[k] = _mm_shuffle_epi32(xmm0[k], 0x4e);
      xmm0[k] = _mm_unpacklo_epi8(xmm0[k], xmm1[k]);
      xmm0[k] = _mm_shuffle_epi32(xmm0[k], 0xd8);
      xmm1[k] = _mm_shuffle_epi32(xmm0[k], 0x4e);
      xmm0[k] = _mm_unpacklo_epi16(xmm0[k], xmm1[k]);
      xmm0[k] = _mm_shuffle_epi32(xmm0[k], 0xd8);
    }
    /* Transpose quad words */
    for (k = 0; k < 1; k++) {
      xmm1[k*2] = _mm_unpacklo_epi64(xmm0[k], xmm0[k+1]);
      xmm1[k*2+1] = _mm_unpackhi_epi64(xmm0[k], xmm0[k+1]);
    }
    /* Store the result vectors */
    dest_for_jth_element = dest + j;
    for (k = 0; k < 2; k++) {
      _mm_storeu_si128((__m128i*)(dest_for_jth_element + (k * total_elements)), xmm1[k]);
    }
  }
}

/* Routine optimized for shuffling a buffer for a type size of 4 bytes. */
static void
shuffle4_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 4;
  size_t i;
  int j;
  uint8_t* dest_for_ith_element;
  __m128i xmm0[4], xmm1[4];

  for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
    /* Fetch 16 elements (64 bytes) then transpose bytes and words. */
    for (j = 0; j < 4; j++) {
      xmm0[j] = _mm_loadu_si128((__m128i*)(src + (i * bytesoftype) + (j * sizeof(__m128i))));
      xmm1[j] = _mm_shuffle_epi32(xmm0[j], 0xd8);
      xmm0[j] = _mm_shuffle_epi32(xmm0[j], 0x8d);
      xmm0[j] = _mm_unpacklo_epi8(xmm1[j], xmm0[j]);
      xmm1[j] = _mm_shuffle_epi32(xmm0[j], 0x04e);
      xmm0[j] = _mm_unpacklo_epi16(xmm0[j], xmm1[j]);
    }
    /* Transpose double words */
    for (j = 0; j < 2; j++) {
      xmm1[j*2] = _mm_unpacklo_epi32(xmm0[j*2], xmm0[j*2+1]);
      xmm1[j*2+1] = _mm_unpackhi_epi32(xmm0[j*2], xmm0[j*2+1]);
    }
    /* Transpose quad words */
    for (j = 0; j < 2; j++) {
      xmm0[j*2] = _mm_unpacklo_epi64(xmm1[j], xmm1[j+2]);
      xmm0[j*2+1] = _mm_unpackhi_epi64(xmm1[j], xmm1[j+2]);
    }
    /* Store the result vectors */
    dest_for_ith_element = dest + i;
    for (j = 0; j < 4; j++) {
      _mm_storeu_si128((__m128i*)(dest_for_ith_element + (j * total_elements)), xmm0[j]);
    }
  }
}

/* Routine optimized for shuffling a buffer for a type size of 8 bytes. */
static void
shuffle8_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 8;
  size_t j;
  int k, l;
  uint8_t* dest_for_jth_element;
  __m128i xmm0[8], xmm1[8];

  for (j = 0; j < vectorizable_elements; j += sizeof(__m128i)) {
    /* Fetch 16 elements (128 bytes) then transpose bytes. */
    for (k = 0; k < 8; k++) {
      xmm0[k] = _mm_loadu_si128((__m128i*)(src + (j * bytesoftype) + (k * sizeof(__m128i))));
      xmm1[k] = _mm_shuffle_epi32(xmm0[k], 0x4e);
      xmm1[k] = _mm_unpacklo_epi8(xmm0[k], xmm1[k]);
    }
    /* Transpose words */
    for (k = 0, l = 0; k < 4; k++, l +=2) {
      xmm0[k*2] = _mm_unpacklo_epi16(xmm1[l], xmm1[l+1]);
      xmm0[k*2+1] = _mm_unpackhi_epi16(xmm1[l], xmm1[l+1]);
    }
    /* Transpose double words */
    for (k = 0, l = 0; k < 4; k++, l++) {
      if (k == 2) l += 2;
      xmm1[k*2] = _mm_unpacklo_epi32(xmm0[l], xmm0[l+2]);
      xmm1[k*2+1] = _mm_unpackhi_epi32(xmm0[l], xmm0[l+2]);
    }
    /* Transpose quad words */
    for (k = 0; k < 4; k++) {
      xmm0[k*2] = _mm_unpacklo_epi64(xmm1[k], xmm1[k+4]);
      xmm0[k*2+1] = _mm_unpackhi_epi64(xmm1[k], xmm1[k+4]);
    }
    /* Store the result vectors */
    dest_for_jth_element = dest + j;
    for (k = 0; k < 8; k++) {
      _mm_storeu_si128((__m128i*)(dest_for_jth_element + (k * total_elements)), xmm0[k]);
    }
  }
}

/* Routine optimized for shuffling a buffer for a type size of 16 bytes. */
static void
shuffle16_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 16;
  size_t j;
  int k, l;
  uint8_t* dest_for_jth_element;
  __m128i xmm0[16], xmm1[16];

  for (j = 0; j < vectorizable_elements; j += sizeof(__m128i)) {
    /* Fetch 16 elements (256 bytes). */
    for (k = 0; k < 16; k++) {
      xmm0[k] = _mm_loadu_si128((__m128i*)(src + (j * bytesoftype) + (k * sizeof(__m128i))));
    }
    /* Transpose bytes */
    for (k = 0, l = 0; k < 8; k++, l +=2) {
      xmm1[k*2] = _mm_unpacklo_epi8(xmm0[l], xmm0[l+1]);
      xmm1[k*2+1] = _mm_unpackhi_epi8(xmm0[l], xmm0[l+1]);
    }
    /* Transpose words */
    for (k = 0, l = -2; k < 8; k++, l++) {
      if ((k%2) == 0) l += 2;
      xmm0[k*2] = _mm_unpacklo_epi16(xmm1[l], xmm1[l+2]);
      xmm0[k*2+1] = _mm_unpackhi_epi16(xmm1[l], xmm1[l+2]);
    }
    /* Transpose double words */
    for (k = 0, l = -4; k < 8; k++, l++) {
      if ((k%4) == 0) l += 4;
      xmm1[k*2] = _mm_unpacklo_epi32(xmm0[l], xmm0[l+4]);
      xmm1[k*2+1] = _mm_unpackhi_epi32(xmm0[l], xmm0[l+4]);
    }
    /* Transpose quad words */
    for (k = 0; k < 8; k++) {
      xmm0[k*2] = _mm_unpacklo_epi64(xmm1[k], xmm1[k+8]);
      xmm0[k*2+1] = _mm_unpackhi_epi64(xmm1[k], xmm1[k+8]);
    }
    /* Store the result vectors */
    dest_for_jth_element = dest + j;
    for (k = 0; k < 16; k++) {
      _mm_storeu_si128((__m128i*)(dest_for_jth_element + (k * total_elements)), xmm0[k]);
    }
  }
}

/* Routine optimized for shuffling a buffer for a type size larger than 16 bytes. */
static void
shuffle16_tiled_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements, const size_t bytesoftype)
{
  size_t j;
  const size_t vecs_per_el_rem = bytesoftype % sizeof(__m128i);
  int k, l;
  uint8_t* dest_for_jth_element;
  __m128i xmm0[16], xmm1[16];

  for (j = 0; j < vectorizable_elements; j += sizeof(__m128i)) {
    /* Advance the offset into the type by the vector size (in bytes), unless this is
    the initial iteration and the type size is not a multiple of the vector size.
    In that case, only advance by the number of bytes necessary so that the number
    of remaining bytes in the type will be a multiple of the vector size. */
    size_t offset_into_type;
    for (offset_into_type = 0; offset_into_type < bytesoftype;
      offset_into_type += (offset_into_type == 0 && vecs_per_el_rem > 0 ? vecs_per_el_rem : sizeof(__m128i))) {

      /* Fetch elements in groups of 256 bytes */
      const uint8_t* const src_with_offset = src + offset_into_type;
      for (k = 0; k < 16; k++) {
        xmm0[k] = _mm_loadu_si128((__m128i*)(src_with_offset + (j + k) * bytesoftype));
      }
      /* Transpose bytes */
      for (k = 0, l = 0; k < 8; k++, l +=2) {
        xmm1[k*2] = _mm_unpacklo_epi8(xmm0[l], xmm0[l+1]);
        xmm1[k*2+1] = _mm_unpackhi_epi8(xmm0[l], xmm0[l+1]);
      }
      /* Transpose words */
      for (k = 0, l = -2; k < 8; k++, l++) {
        if ((k%2) == 0) l += 2;
        xmm0[k*2] = _mm_unpacklo_epi16(xmm1[l], xmm1[l+2]);
        xmm0[k*2+1] = _mm_unpackhi_epi16(xmm1[l], xmm1[l+2]);
      }
      /* Transpose double words */
      for (k = 0, l = -4; k < 8; k++, l++) {
        if ((k%4) == 0) l += 4;
        xmm1[k*2] = _mm_unpacklo_epi32(xmm0[l], xmm0[l+4]);
        xmm1[k*2+1] = _mm_unpackhi_epi32(xmm0[l], xmm0[l+4]);
      }
      /* Transpose quad words */
      for (k = 0; k < 8; k++) {
        xmm0[k*2] = _mm_unpacklo_epi64(xmm1[k], xmm1[k+8]);
        xmm0[k*2+1] = _mm_unpackhi_epi64(xmm1[k], xmm1[k+8]);
      }
      /* Store the result vectors */
      dest_for_jth_element = dest + j;
      for (k = 0; k < 16; k++) {
        _mm_storeu_si128((__m128i*)(dest_for_jth_element + (total_elements * (offset_into_type + k))), xmm0[k]);
      }
    }
  }
}

/* Routine optimized for unshuffling a buffer for a type size of 2 bytes. */
static void
unshuffle2_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 2;
  size_t i;
  int j;
  __m128i xmm0[2], xmm1[2];

  for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
    /* Load 16 elements (32 bytes) into 2 XMM registers. */
    const uint8_t* const src_for_ith_element = src + i;
    for (j = 0; j < 2; j++) {
      xmm0[j] = _mm_loadu_si128((__m128i*)(src_for_ith_element + (j * total_elements)));
    }
    /* Shuffle bytes */
    /* Compute the low 32 bytes */
    xmm1[0] = _mm_unpacklo_epi8(xmm0[0], xmm0[1]);
    /* Compute the hi 32 bytes */
    xmm1[1] = _mm_unpackhi_epi8(xmm0[0], xmm0[1]);
    /* Store the result vectors in proper order */
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (0 * sizeof(__m128i))), xmm1[0]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (1 * sizeof(__m128i))), xmm1[1]);
  }
}

/* Routine optimized for unshuffling a buffer for a type size of 4 bytes. */
static void
unshuffle4_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 4;
  size_t i;
  int j;
  __m128i xmm0[4], xmm1[4];

  for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
    /* Load 16 elements (64 bytes) into 4 XMM registers. */
    const uint8_t* const src_for_ith_element = src + i;
    for (j = 0; j < 4; j++) {
      xmm0[j] = _mm_loadu_si128((__m128i*)(src_for_ith_element + (j * total_elements)));
    }
    /* Shuffle bytes */
    for (j = 0; j < 2; j++) {
      /* Compute the low 32 bytes */
      xmm1[j] = _mm_unpacklo_epi8(xmm0[j*2], xmm0[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm1[2+j] = _mm_unpackhi_epi8(xmm0[j*2], xmm0[j*2+1]);
    }
    /* Shuffle 2-byte words */
    for (j = 0; j < 2; j++) {
      /* Compute the low 32 bytes */
      xmm0[j] = _mm_unpacklo_epi16(xmm1[j*2], xmm1[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm0[2+j] = _mm_unpackhi_epi16(xmm1[j*2], xmm1[j*2+1]);
    }
    /* Store the result vectors in proper order */
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (0 * sizeof(__m128i))), xmm0[0]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (1 * sizeof(__m128i))), xmm0[2]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (2 * sizeof(__m128i))), xmm0[1]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (3 * sizeof(__m128i))), xmm0[3]);
  }
}

/* Routine optimized for unshuffling a buffer for a type size of 8 bytes. */
static void
unshuffle8_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 8;
  size_t i;
  int j;
  __m128i xmm0[8], xmm1[8];

  for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
    /* Load 16 elements (128 bytes) into 8 XMM registers. */
    const uint8_t* const src_for_ith_element = src + i;
    for (j = 0; j < 8; j++) {
      xmm0[j] = _mm_loadu_si128((__m128i*)(src_for_ith_element + (j * total_elements)));
    }
    /* Shuffle bytes */
    for (j = 0; j < 4; j++) {
      /* Compute the low 32 bytes */
      xmm1[j] = _mm_unpacklo_epi8(xmm0[j*2], xmm0[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm1[4+j] = _mm_unpackhi_epi8(xmm0[j*2], xmm0[j*2+1]);
    }
    /* Shuffle 2-byte words */
    for (j = 0; j < 4; j++) {
      /* Compute the low 32 bytes */
      xmm0[j] = _mm_unpacklo_epi16(xmm1[j*2], xmm1[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm0[4+j] = _mm_unpackhi_epi16(xmm1[j*2], xmm1[j*2+1]);
    }
    /* Shuffle 4-byte dwords */
    for (j = 0; j < 4; j++) {
      /* Compute the low 32 bytes */
      xmm1[j] = _mm_unpacklo_epi32(xmm0[j*2], xmm0[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm1[4+j] = _mm_unpackhi_epi32(xmm0[j*2], xmm0[j*2+1]);
    }
    /* Store the result vectors in proper order */
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (0 * sizeof(__m128i))), xmm1[0]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (1 * sizeof(__m128i))), xmm1[4]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (2 * sizeof(__m128i))), xmm1[2]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (3 * sizeof(__m128i))), xmm1[6]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (4 * sizeof(__m128i))), xmm1[1]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (5 * sizeof(__m128i))), xmm1[5]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (6 * sizeof(__m128i))), xmm1[3]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (7 * sizeof(__m128i))), xmm1[7]);
  }
}

/* Routine optimized for unshuffling a buffer for a type size of 16 bytes. */
static void
unshuffle16_sse2(uint8_t* const dest, const uint8_t* const src,
  const size_t vectorizable_elements, const size_t total_elements)
{
  static const size_t bytesoftype = 16;
  size_t i;
  int j;
  __m128i xmm1[16], xmm2[16];

  for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
    /* Load 16 elements (256 bytes) into 16 XMM registers. */
    const uint8_t* const src_for_ith_element = src + i;
    for (j = 0; j < 16; j++) {
      xmm1[j] = _mm_loadu_si128((__m128i*)(src_for_ith_element + (j * total_elements)));
    }
    /* Shuffle bytes */
    for (j = 0; j < 8; j++) {
      /* Compute the low 32 bytes */
      xmm2[j] = _mm_unpacklo_epi8(xmm1[j*2], xmm1[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm2[8+j] = _mm_unpackhi_epi8(xmm1[j*2], xmm1[j*2+1]);
    }
    /* Shuffle 2-byte words */
    for (j = 0; j < 8; j++) {
      /* Compute the low 32 bytes */
      xmm1[j] = _mm_unpacklo_epi16(xmm2[j*2], xmm2[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm1[8+j] = _mm_unpackhi_epi16(xmm2[j*2], xmm2[j*2+1]);
    }
    /* Shuffle 4-byte dwords */
    for (j = 0; j < 8; j++) {
      /* Compute the low 32 bytes */
      xmm2[j] = _mm_unpacklo_epi32(xmm1[j*2], xmm1[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm2[8+j] = _mm_unpackhi_epi32(xmm1[j*2], xmm1[j*2+1]);
    }
    /* Shuffle 8-byte qwords */
    for (j = 0; j < 8; j++) {
      /* Compute the low 32 bytes */
      xmm1[j] = _mm_unpacklo_epi64(xmm2[j*2], xmm2[j*2+1]);
      /* Compute the hi 32 bytes */
      xmm1[8+j] = _mm_unpackhi_epi64(xmm2[j*2], xmm2[j*2+1]);
    }

    /* Store the result vectors in proper order */
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (0 * sizeof(__m128i))), xmm1[0]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (1 * sizeof(__m128i))), xmm1[8]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (2 * sizeof(__m128i))), xmm1[4]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (3 * sizeof(__m128i))), xmm1[12]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (4 * sizeof(__m128i))), xmm1[2]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (5 * sizeof(__m128i))), xmm1[10]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (6 * sizeof(__m128i))), xmm1[6]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (7 * sizeof(__m128i))), xmm1[14]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (8 * sizeof(__m128i))), xmm1[1]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (9 * sizeof(__m128i))), xmm1[9]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (10 * sizeof(__m128i))), xmm1[5]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (11 * sizeof(__m128i))), xmm1[13]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (12 * sizeof(__m128i))), xmm1[3]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (13 * sizeof(__m128i))), xmm1[11]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (14 * sizeof(__m128i))), xmm1[7]);
    _mm_storeu_si128((__m128i*)(dest + (i * bytesoftype) + (15 * sizeof(__m128i))), xmm1[15]);
  }
}

/* Routine optimized for unshuffling a buffer for a type size larger than 16 bytes. */
static void
unshuffle16_tiled_sse2(uint8_t* const dest, const uint8_t* const orig,
  const size_t vectorizable_elements, const size_t total_elements, const size_t bytesoftype)
{
  size_t i;
  const size_t vecs_per_el_rem = bytesoftype % sizeof(__m128i);

  int j;
  uint8_t* dest_with_offset;
  __m128i xmm1[16], xmm2[16];

  /* The unshuffle loops are inverted (compared to shuffle_tiled16_sse2)
     to optimize cache utilization. */
  size_t offset_into_type;
  for (offset_into_type = 0; offset_into_type < bytesoftype;
    offset_into_type += (offset_into_type == 0 && vecs_per_el_rem > 0 ? vecs_per_el_rem : sizeof(__m128i))) {
    for (i = 0; i < vectorizable_elements; i += sizeof(__m128i)) {
      /* Load the first 128 bytes in 16 XMM registers */
      const uint8_t* const src_for_ith_element = orig + i;
      for (j = 0; j < 16; j++) {
        xmm1[j] = _mm_loadu_si128((__m128i*)(src_for_ith_element + (total_elements * (offset_into_type + j))));
      }
      /* Shuffle bytes */
      for (j = 0; j < 8; j++) {
        /* Compute the low 32 bytes */
        xmm2[j] = _mm_unpacklo_epi8(xmm1[j*2], xmm1[j*2+1]);
        /* Compute the hi 32 bytes */
        xmm2[8+j] = _mm_unpackhi_epi8(xmm1[j*2], xmm1[j*2+1]);
      }
      /* Shuffle 2-byte words */
      for (j = 0; j < 8; j++) {
        /* Compute the low 32 bytes */
        xmm1[j] = _mm_unpacklo_epi16(xmm2[j*2], xmm2[j*2+1]);
        /* Compute the hi 32 bytes */
        xmm1[8+j] = _mm_unpackhi_epi16(xmm2[j*2], xmm2[j*2+1]);
      }
      /* Shuffle 4-byte dwords */
      for (j = 0; j < 8; j++) {
        /* Compute the low 32 bytes */
        xmm2[j] = _mm_unpacklo_epi32(xmm1[j*2], xmm1[j*2+1]);
        /* Compute the hi 32 bytes */
        xmm2[8+j] = _mm_unpackhi_epi32(xmm1[j*2], xmm1[j*2+1]);
      }
      /* Shuffle 8-byte qwords */
      for (j = 0; j < 8; j++) {
        /* Compute the low 32 bytes */
        xmm1[j] = _mm_unpacklo_epi64(xmm2[j*2], xmm2[j*2+1]);
        /* Compute the hi 32 bytes */
        xmm1[8+j] = _mm_unpackhi_epi64(xmm2[j*2], xmm2[j*2+1]);
      }

      /* Store the result vectors in proper order */
      dest_with_offset = dest + offset_into_type;
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 0) * bytesoftype), xmm1[0]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 1) * bytesoftype), xmm1[8]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 2) * bytesoftype), xmm1[4]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 3) * bytesoftype), xmm1[12]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 4) * bytesoftype), xmm1[2]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 5) * bytesoftype), xmm1[10]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 6) * bytesoftype), xmm1[6]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 7) * bytesoftype), xmm1[14]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 8) * bytesoftype), xmm1[1]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 9) * bytesoftype), xmm1[9]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 10) * bytesoftype), xmm1[5]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 11) * bytesoftype), xmm1[13]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 12) * bytesoftype), xmm1[3]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 13) * bytesoftype), xmm1[11]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 14) * bytesoftype), xmm1[7]);
      _mm_storeu_si128((__m128i*)(dest_with_offset + (i + 15) * bytesoftype), xmm1[15]);
    }
  }
}

/* Shuffle a block.  This can never fail. */
void
shuffle_sse2(const size_t bytesoftype, const size_t blocksize,
             const uint8_t* const _src, uint8_t* const _dest) {
  const size_t vectorized_chunk_size = bytesoftype * sizeof(__m128i);
  /* If the blocksize is not a multiple of both the typesize and
     the vector size, round the blocksize down to the next value
     which is a multiple of both. The vectorized shuffle can be
     used for that portion of the data, and the naive implementation
     can be used for the remaining portion. */
  const size_t vectorizable_bytes = blocksize - (blocksize % vectorized_chunk_size);
  const size_t vectorizable_elements = vectorizable_bytes / bytesoftype;
  const size_t total_elements = blocksize / bytesoftype;

  /* If the block size is too small to be vectorized,
     use the generic implementation. */
  if (blocksize < vectorized_chunk_size) {
    shuffle_generic(bytesoftype, blocksize, _src, _dest);
    return;
  }

  /* Optimized shuffle implementations */
  switch (bytesoftype)
  {
  case 2:
    shuffle2_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 4:
    shuffle4_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 8:
    shuffle8_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 16:
    shuffle16_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  default:
    if (bytesoftype > sizeof(__m128i)) {
      shuffle16_tiled_sse2(_dest, _src, vectorizable_elements, total_elements, bytesoftype);
    }
    else {
      /* Non-optimized shuffle */
      shuffle_generic(bytesoftype, blocksize, _src, _dest);
      /* The non-optimized function covers the whole buffer,
         so we're done processing here. */
      return;
    }
  }

  /* If the buffer had any bytes at the end which couldn't be handled
     by the vectorized implementations, use the non-optimized version
     to finish them up. */
  if (vectorizable_bytes < blocksize) {
    shuffle_generic_inline(bytesoftype, vectorizable_bytes, blocksize, _src, _dest);
  }
}

/* Unshuffle a block.  This can never fail. */
void
unshuffle_sse2(const size_t bytesoftype, const size_t blocksize,
               const uint8_t* const _src, uint8_t* const _dest) {
  const size_t vectorized_chunk_size = bytesoftype * sizeof(__m128i);
  /* If the blocksize is not a multiple of both the typesize and
     the vector size, round the blocksize down to the next value
     which is a multiple of both. The vectorized unshuffle can be
     used for that portion of the data, and the naive implementation
     can be used for the remaining portion. */
  const size_t vectorizable_bytes = blocksize - (blocksize % vectorized_chunk_size);
  const size_t vectorizable_elements = vectorizable_bytes / bytesoftype;
  const size_t total_elements = blocksize / bytesoftype;


  /* If the block size is too small to be vectorized,
     use the generic implementation. */
  if (blocksize < vectorized_chunk_size) {
    unshuffle_generic(bytesoftype, blocksize, _src, _dest);
    return;
  }

  /* Optimized unshuffle implementations */
  switch (bytesoftype)
  {
  case 2:
    unshuffle2_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 4:
    unshuffle4_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 8:
    unshuffle8_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  case 16:
    unshuffle16_sse2(_dest, _src, vectorizable_elements, total_elements);
    break;
  default:
    if (bytesoftype > sizeof(__m128i)) {
      unshuffle16_tiled_sse2(_dest, _src, vectorizable_elements, total_elements, bytesoftype);
    }
    else {
      /* Non-optimized unshuffle */
      unshuffle_generic(bytesoftype, blocksize, _src, _dest);
      /* The non-optimized function covers the whole buffer,
         so we're done processing here. */
      return;
    }
  }

  /* If the buffer had any bytes at the end which couldn't be handled
     by the vectorized implementations, use the non-optimized version
     to finish them up. */
  if (vectorizable_bytes < blocksize) {
    unshuffle_generic_inline(bytesoftype, vectorizable_bytes, blocksize, _src, _dest);
  }
}