// 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 in ["ABS", "NEG"]
#include <assert.h>

#include <emmintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/unaligned.h>
#include <xnnpack/vunary.h>


$_MM_OP_SI128 = {
$  "ABS": lambda x: "_mm_and_si128(%s, vnonsign_mask)" % x,
$  "NEG": lambda x: "_mm_xor_si128(%s, vsign_mask)" % x,
$}[OP]
$PARAMS = {
$  "ABS": "xnn_f16_abs_params",
$  "NEG": "xnn_f16_neg_params",
$}[OP]
void xnn_f16_v${OP.lower()}_ukernel__sse2_x${BATCH_TILE}(
    size_t batch,
    const void* input,
    void* output,
    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % 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 OP == "ABS":
    const __m128i vnonsign_mask = _mm_load_si128((const __m128i*) params->sse.nonsign_mask);
  $elif OP == "NEG":
    const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse.sign_mask);
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      __m128i vacc${ABC[0]} = _mm_loadu_si128((const __m128i*) i);
      $for N in range(1, SIMD_TILE):
        __m128i vacc${ABC[N]} = _mm_loadu_si128((const __m128i*) (i + ${N*8}));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        vacc${ABC[N]} = ${_MM_OP_SI128("vacc" + ABC[N])};

      _mm_storeu_si128((__m128i*) o, vacc${ABC[0]});
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (o + ${N*8}), vacc${ABC[N]});
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    __m128i vacc = _mm_loadu_si128((const __m128i*) i);
    i += 8;
    vacc = ${_MM_OP_SI128("vacc")};
    _mm_storeu_si128((__m128i*) o, vacc);
    o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    __m128i vacc = _mm_loadu_si128((const __m128i*) i);
    vacc = ${_MM_OP_SI128("vacc")};
    if (batch & (4 * sizeof(uint16_t))) {
      _mm_storel_epi64((__m128i*) o, vacc);
      o += 4;
      vacc = _mm_unpackhi_epi64(vacc, vacc);
    }
    if (batch & (2 * sizeof(uint16_t))) {
      unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vacc));
      o += 2;
      vacc = _mm_srli_epi64(vacc, 32);
    }
    if (batch & (1 * sizeof(uint16_t))) {
      *o = (uint16_t) _mm_extract_epi16(vacc, 0);
    }
  }
}