// 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 % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <emmintrin.h>

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


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$_MM_CVTEPX8_EPI32 = {"QS8": "_mm_cvtepi8_epi32", "QU8": "_mm_cvtepu8_epi32"}[DATATYPE]
void xnn_${DATATYPE.lower()}_f32_vcvt_ukernel__sse2_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    float* y,
    const union xnn_${DATATYPE.lower()}_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(${XINT8_T}) == 0);
  assert(x != NULL);
  assert(y != NULL);

  $if DATATYPE == "QS8":
    const __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse2.sign_mask);
  const __m128i vmagic_exp = _mm_load_si128((const __m128i*) params->sse2.magic_exp);
  const __m128 vmagic_bias = _mm_load_ps(params->sse2.magic_bias);
  const __m128 vscale = _mm_load_ps(params->sse2.scale);
  const __m128i vzero = _mm_setzero_si128();
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
      __m128i vx${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) x);
      $for N in range(8, BATCH_TILE, 8):
        __m128i vx${ABC[N:N+8]} = _mm_loadl_epi64((const __m128i*) (x + ${N}));
      x += ${BATCH_TILE};

      $if DATATYPE == "QS8":
        $for N in range(0, BATCH_TILE, 8):
          vx${ABC[N:N+8]} = _mm_xor_si128(vx${ABC[N:N+8]}, vsign_mask);

      $for N in range(0, BATCH_TILE, 8):
        vx${ABC[N:N+8]} = _mm_unpacklo_epi8(vx${ABC[N:N+8]}, vzero);

      $for N in range(0, BATCH_TILE, 8):
        __m128 vy${ABC[N:N+4]} = _mm_castsi128_ps(_mm_unpacklo_epi16(vx${ABC[N:N+8]}, vmagic_exp));
        __m128 vy${ABC[N+4:N+8]} = _mm_castsi128_ps(_mm_unpackhi_epi16(vx${ABC[N:N+8]}, vmagic_exp));

      $for N in range(0, BATCH_TILE, 4):
        vy${ABC[N:N+4]} = _mm_sub_ps(vy${ABC[N:N+4]}, vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        vy${ABC[N:N+4]} = _mm_mul_ps(vy${ABC[N:N+4]}, vscale);

      _mm_storeu_ps(y, vy${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]});
      y += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) {
    __m128i vx = _mm_loadl_epi64((const __m128i*) x);
    $if DATATYPE == "QS8":
      vx = _mm_xor_si128(vx, vsign_mask);
    vx = _mm_unpacklo_epi8(vx, vzero);
    x += 8;

    __m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
    __m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));

    vy_lo = _mm_sub_ps(vy_lo, vmagic_bias);
    vy_hi = _mm_sub_ps(vy_hi, vmagic_bias);

    vy_lo = _mm_mul_ps(vy_lo, vscale);
    vy_hi = _mm_mul_ps(vy_hi, vscale);

    _mm_storeu_ps(y, vy_lo);
    _mm_storeu_ps(y + 4, vy_hi);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(${XINT8_T}));
    assert(n <= 7 * sizeof(${XINT8_T}));

    __m128i vx = _mm_loadl_epi64((const __m128i*) x);
    $if DATATYPE == "QS8":
      vx = _mm_xor_si128(vx, vsign_mask);
    vx = _mm_unpacklo_epi8(vx, vzero);

    __m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp));
    vy = _mm_sub_ps(vy, vmagic_bias);
    vy = _mm_mul_ps(vy, vscale);

    if (n & (4 * sizeof(${XINT8_T}))) {
      _mm_storeu_ps(y, vy);
      vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp));
      vy = _mm_sub_ps(vy, vmagic_bias);
      vy = _mm_mul_ps(vy, vscale);
      y += 4;
    }
    if (n & (2 * sizeof(${XINT8_T}))) {
      _mm_storel_pi((__m64*) y, vy);
      vy = _mm_movehl_ps(vy, vy);
      y += 2;
    }
    if (n & (1 * sizeof(${XINT8_T}))) {
      _mm_store_ss(y, vy);
    }
  }
}