// Copyright 2022 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 % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP == "SQR"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vunary.h>


void xnn_f16_vsqr_ukernel__f16c_x${BATCH_TILE}(
    size_t n,
    const void* input,
    void* output,
    const union xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      __m256 vacc${ABC[0]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
      $for N in range(1, SIMD_TILE):
        __m256 vacc${ABC[N]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N*8})));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        vacc${ABC[N]} = _mm256_mul_ps(vacc${ABC[N]}, vacc${ABC[N]});

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc${ABC[0]}, _MM_FROUND_NO_EXC));
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (o + ${N*8}), _mm256_cvtps_ph(vacc${ABC[N]}, _MM_FROUND_NO_EXC));
      o += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;
    vacc = _mm256_mul_ps(vacc, vacc);
    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC));
    o += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    vacc = _mm256_mul_ps(vacc, vacc);
    __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC);
    if (n & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vh);
      o += 4;
      vh = _mm_unpackhi_epi64(vh, vh);
    }
    if (n & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(o, vh);
      o += 2;
      vh = _mm_srli_epi64(vh, 32);
    }
    if (n & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}