src/x8-lut/ssse3.c.in

// Copyright 2021 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$assert BATCH_TILE >= 16
$assert BATCH_TILE % 16 == 0
$SIMD_TILE = BATCH_TILE // 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

$if AVX:
  #include <immintrin.h>
$else:
  #include <tmmintrin.h>

#include <xnnpack/lut.h>
#include <xnnpack/common.h>


void xnn_x8_lut_ukernel__${"avx" if AVX else "ssse3"}_x${BATCH_TILE}(
    size_t n,
    const uint8_t* x,
    uint8_t* y,
    const uint8_t t[restrict XNN_MIN_ELEMENTS(256)])
{
  assert(n != 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m128i vt0 = _mm_load_si128((const __m128i*) t);
  $for T in range(1, 16):
    const __m128i vt${ABC[T]} = _mm_load_si128((const __m128i*) (t + ${T * 16}));

  const __m128i vtable0 = vt0;
  $for T in range(1, 8):
    const __m128i vtable${ABC[T]} = _mm_xor_si128(vt${ABC[T-1]}, vt${ABC[T]});
  $for T in range(8, 16):
    const __m128i vtable${ABC[T]} = _mm_xor_si128(_mm_xor_si128(vt${ABC[T-1]}, vt${ABC[T]}), vtable${ABC[T-8]});

  const __m128i voffset = _mm_set1_epi8(16);
  $if BATCH_TILE > 16:
    for (; n >= ${BATCH_TILE} * sizeof(uint8_t); n -= ${BATCH_TILE} * sizeof(uint8_t)) {
      __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
      $for N in range(1, SIMD_TILE):
        __m128i vx${N} = _mm_loadu_si128((const __m128i*) (x + ${N * 16}));
      x += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        __m128i vy${N} = _mm_shuffle_epi8(vtable0, vx${N});

      $for T in range(1, 9):
        $for N in range(SIMD_TILE):
          vx${N} = _mm_sub_epi8(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = _mm_xor_si128(vy${N}, _mm_shuffle_epi8(vtable${ABC[T]}, vx${N}));

      $for T in range(9, 16):
        $for N in range(SIMD_TILE):
          vx${N} = _mm_subs_epi8(vx${N}, voffset);
        $for N in range(SIMD_TILE):
          vy${N} = _mm_xor_si128(vy${N}, _mm_shuffle_epi8(vtable${ABC[T]}, vx${N}));

      _mm_storeu_si128((__m128i*) y, vy0);
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (y + ${N * 16}), vy${N});
      y += ${BATCH_TILE};
    }
  for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
    __m128i vx = _mm_loadu_si128((const __m128i*) x);
    x += 16;

    __m128i vy = _mm_shuffle_epi8(vtable0, vx);

    $for T in range(1, 9):
      vx = _mm_sub_epi8(vx, voffset);
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = _mm_subs_epi8(vx, voffset);
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable${ABC[T]}, vx));

    _mm_storeu_si128((__m128i*) y, vy);
    y += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    __m128i vx = _mm_loadu_si128((const __m128i*) x);

    __m128i vy = _mm_shuffle_epi8(vtable0, vx);

    $for T in range(1, 9):
      vx = _mm_sub_epi8(vx, voffset);
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable${ABC[T]}, vx));

    $for T in range(9, 16):
      vx = _mm_subs_epi8(vx, voffset);
      vy = _mm_xor_si128(vy, _mm_shuffle_epi8(vtable${ABC[T]}, vx));

    if (n & (8 * sizeof(uint8_t))) {
      _mm_storel_epi64((__m128i*) y, vy);
      vy = _mm_unpackhi_epi64(vy, vy);
      y += 8;
    }
    if (n & (4 * sizeof(uint8_t))) {
      $if AVX:
        _mm_storeu_si32(y, vy);
      $else:
        *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vy);
      vy = _mm_srli_epi64(vy, 32);
      y += 4;
    }
    $if AVX:
      if (n & (2 * sizeof(uint8_t))) {
        *((uint16_t*) y) = (uint16_t) _mm_extract_epi16(vy, 0);
        vy = _mm_srli_epi32(vy, 16);
        y += 2;
      }
      if (n & (1 * sizeof(uint8_t))) {
        *y = (uint8_t) _mm_extract_epi8(vy, 0);
      }
    $else:
      uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
      if (n & (2 * sizeof(uint8_t))) {
        *((uint16_t*) y) = (uint16_t) vy_lo;
        vy_lo >>= 16;
        y += 2;
      }
      if (n & (1 * sizeof(uint8_t))) {
        *y = (uint8_t) vy_lo;
      }
  }
}