// 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", "PRELU", "RPRELU"]
#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,
$  "PRELU": lambda x: "_mm256_mul_ps(%s, vb)" % x,
$  "RPRELU": lambda x: "_mm256_mul_ps(%s, vb)" % x,
$}[OP]
void xnn_f16_v${OP.lower()}c_ukernel__f16c_u${BATCH_TILE}(
    size_t batch,
    const xnn_float16* restrict input_a,
    const xnn_float16* restrict input_b,
    xnn_float16* restrict output,
    const struct xnn_f16_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const uint16_t* a = (const uint16_t*) input_a;
  const uint16_t* b = (const uint16_t*) input_b;
  uint16_t* o = (uint16_t*) output;

  const __m256 vb = _mm256_cvtph_ps(_mm_set1_epi16((short) *b));
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m256 va${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_TO_NEAREST_INT));

      $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_TO_NEAREST_INT));
      $elif OP == "PRELU":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(va${ABC[N:N+8]}, vy${ABC[N:N+8]}, va${ABC[N:N+8]}), _MM_FROUND_TO_NEAREST_INT));
      $elif OP == "RPRELU":
        $for N in range(0, BATCH_TILE, 8):
          vy${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(vb, vy${ABC[N:N+8]}, vb), _MM_FROUND_TO_NEAREST_INT));

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
      $for N in range(8, BATCH_TILE, 8):
        _mm_storeu_si128((__m128i*) (o + ${N}), _mm256_cvtps_ph(vy${ABC[N:N+8]}, _MM_FROUND_TO_NEAREST_INT));
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 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_TO_NEAREST_INT));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_TO_NEAREST_INT));
    $elif OP == "PRELU":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(va, vy, va), _MM_FROUND_TO_NEAREST_INT));
    $elif OP == "RPRELU":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(vb, vy, vb), _MM_FROUND_TO_NEAREST_INT));

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT));
    o += 8;
  }
  if XNN_UNLIKELY(batch != 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_TO_NEAREST_INT));
    $if OP == "SQRDIFF":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vy, vy), _MM_FROUND_TO_NEAREST_INT));
    $elif OP == "PRELU":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(va, vy, va), _MM_FROUND_TO_NEAREST_INT));
    $elif OP == "RPRELU":
      vy = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_blendv_ps(vb, vy, vb), _MM_FROUND_TO_NEAREST_INT));

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