// 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
$ABC = "01234567456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ADD", "DIV", "RDIV", "MAX", "MIN", "MUL", "SUB", "RSUB", "SQRDIFF"]
$assert ACTIVATION in ["LINEAR", "MINMAX"]
#include <assert.h>

#include <immintrin.h>

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


$_MM256_OP_PS = {
$  "ADD": lambda x: "_mm256_add_ps(%s, vb)" % x,
$  "DIV": lambda x: "_mm256_div_ps(%s, vb)" % x,
$  "RDIV": lambda x: "_mm256_div_ps(vb, %s)" % x,
$  "MAX": lambda x: "_mm256_max_ps(%s, vb)" % x,
$  "MIN": lambda x: "_mm256_min_ps(%s, vb)" % x,
$  "MUL": lambda x: "_mm256_mul_ps(%s, vb)" % x,
$  "SUB": lambda x: "_mm256_sub_ps(%s, vb)" % x,
$  "RSUB": lambda x: "_mm256_sub_ps(vb, %s)" % x,
$  "SQRDIFF": lambda x: "_mm256_sub_ps(%s, vb)" % x,
$}[OP]
$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
$PARAMS = {"LINEAR": "xnn_f16_default_params", "MINMAX": "xnn_f16_minmax_params"}[ACTIVATION]
void xnn_f16_v${OP.lower()}c${SUFFIX}_ukernel__f16c_x${BATCH_TILE}(
    size_t n,
    const void* restrict a_ptr,
    const void* restrict b_ptr,
    void* restrict y_ptr,
    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint16_t) == 0);
  assert(a_ptr != NULL);
  assert(b_ptr != NULL);
  assert(y_ptr != NULL);

  const uint16_t* a = (const uint16_t*) a_ptr;
  const uint16_t* b = (const uint16_t*) b_ptr;
  uint16_t* y = (uint16_t*) y_ptr;

  $if ACTIVATION == "MINMAX":
    const __m256 vy_min = _mm256_load_ps(params->avx.min);
    const __m256 vy_max = _mm256_load_ps(params->avx.max);

  const __m256 vb = _mm256_cvtph_ps(_mm_set1_epi16((short) *b));
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m256 va${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
      $for N in range(8, BATCH_TILE, 8):
        const __m256 va${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (a + ${N})));
      a += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 8):
        __m256 vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va" + ABC[N:N+8])}, _MM_FROUND_NO_EXC));

      $if OP == "SQRDIFF":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy${ABC[N:N+8]}, vy${ABC[N:N+8]}), _MM_FROUND_NO_EXC));

      $if ACTIVATION == "MINMAX":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_max_ps(vy${ABC[N:N+8]}, vy_min);

        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_min_ps(vy${ABC[N:N+8]}, vy_max);

      _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy${ABC[0:8]}, _MM_FROUND_NO_EXC));
      $for N in range(8, BATCH_TILE, 8):
        _mm_storeu_si128((__m128i*) (y + ${N}), _mm256_cvtps_ph(vy${ABC[N:N+8]}, _MM_FROUND_NO_EXC));
      y += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));
    a += 8;

    __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va")}, _MM_FROUND_NO_EXC));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_NO_EXC));

    $if ACTIVATION == "MINMAX":
      vy = _mm256_max_ps(vy, vy_min);
      vy = _mm256_min_ps(vy, vy_max);

    _mm_storeu_si128((__m128i*) y, _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC));
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    const __m256 va = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) a));

    __m256 vy = _mm256_cvtph_ps(_mm256_cvtps_ph(${_MM256_OP_PS("va")}, _MM_FROUND_NO_EXC));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_NO_EXC));

    $if ACTIVATION == "MINMAX":
      vy = _mm256_max_ps(vy, vy_min);
      vy = _mm256_min_ps(vy, vy_max);

    __m128i vh = _mm256_cvtps_ph(vy, _MM_FROUND_NO_EXC);
    if (n & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) y, vh);
      vh = _mm_unpackhi_epi64(vh, vh);
      y += 4;
    }
    if (n & (2 * sizeof(uint16_t))) {
      _mm_storeu_si32(y, vh);
      vh = _mm_srli_epi64(vh, 32);
      y += 2;
    }
    if (n & (1 * sizeof(uint16_t))) {
      *y = (uint16_t) _mm_extract_epi16(vh, 0);
    }
  }
}