src/x8-lut/scalar.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 >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

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


void xnn_x8_lut_ukernel__scalar_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);

  $if BATCH_TILE > 1:
    for (; n >= ${BATCH_TILE} * sizeof(uint8_t); n -= ${BATCH_TILE} * sizeof(uint8_t)) {
      $for N in range(BATCH_TILE):
        const size_t vx${N} = (size_t) x[${N}];
      x += ${BATCH_TILE};

      $for N in range(BATCH_TILE):
        const uint32_t vt${N} = (uint32_t) t[vx${N}];

      $for N in range(BATCH_TILE):
        y[${N}] = (uint8_t) vt${N};
      y += ${BATCH_TILE};
    }
    if XNN_UNLIKELY(n != 0) {
      $if BATCH_TILE > 2:
        do {
          const size_t vx = (size_t) *x++;
          const uint32_t vt = (uint32_t) t[vx];
          *y++ = (uint8_t) vt;
          n -= sizeof(uint8_t);
        } while (n != 0);
      $else:
        const size_t vx = (size_t) *x;
        const uint32_t vt = (uint32_t) t[vx];
        *y = (uint8_t) vt;
    }
  $else:
    do {
      const size_t vx = (size_t) *x++;
      const uint32_t vt = (uint32_t) t[vx];
      *y++ = (uint8_t) vt;
      n -= sizeof(uint8_t);
    } while (n != 0);
}