// 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 SSE in [2, 4]
$assert not AVX or SSE == 4
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$SIMD_TILE = BATCH_TILE // 8
$SSE_HEADER = {2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <${SSE_HEADER}>

#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>


$ISA = "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE]
void xnn_f16_f32_vcvt_ukernel__${ISA}_int32_x${BATCH_TILE}(
    size_t n,
    const void* input,
    float* output,
    const union xnn_f16_f32_cvt_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 __m128i vsign_mask = _mm_load_si128((const __m128i*) params->sse_int32.sign_mask);
  const __m128i vexp_offset = _mm_load_si128((const __m128i*) params->sse_int32.exp_offset);
  const __m128 vexp_scale = _mm_load_ps(params->sse_int32.exp_scale);
  const __m128i vmagic_bias = _mm_load_si128((const __m128i*) params->sse_int32.magic_bias);
  const __m128i vdenorm_cutoff = _mm_load_si128((const __m128i*) params->sse_int32.denorm_cutoff);

  const uint16_t* i = (const uint16_t*) input;
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) {
      const __m128i vh0 = _mm_loadu_si128((const __m128i*) i);
      $for N in range(1, SIMD_TILE):
        const __m128i vh${N} = _mm_loadu_si128((const __m128i*) (i + ${N * 8}));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m128i vw${2*N} = _mm_unpacklo_epi16(_mm_setzero_si128(), vh${N});
        const __m128i vw${2*N+1} = _mm_unpackhi_epi16(_mm_setzero_si128(), vh${N});

      $for N in range(2*SIMD_TILE):
        const __m128i vsign${N} = _mm_and_si128(vw${N}, vsign_mask);

      $for N in range(2*SIMD_TILE):
        const __m128i vnonsign${N} = _mm_xor_si128(vw${N}, vsign${N});

      $for N in range(2*SIMD_TILE):
        const __m128i vnorm${N} = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign${N}, 3), vexp_offset)), vexp_scale));

      $for N in range(2*SIMD_TILE):
        const __m128i vdenorm${N} = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign${N}, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));

      $for N in range(2*SIMD_TILE):
        const __m128i vmask${N} = _mm_cmpgt_epi32(vnonsign${N}, vdenorm_cutoff);

      $for N in range(2*SIMD_TILE):
        $if SSE == 4:
          const __m128i vf${N} = _mm_or_si128(vsign${N}, _mm_blendv_epi8(vdenorm${N}, vnorm${N}, vmask${N}));
        $else:
          const __m128i vf${N} = _mm_or_si128(vsign${N},
            _mm_or_si128(_mm_and_si128(vmask${N}, vnorm${N}), _mm_andnot_si128(vmask${N}, vdenorm${N})));

      _mm_storeu_ps(output, _mm_castsi128_ps(vf0));
      $for N in range(1, 2*SIMD_TILE):
        _mm_storeu_ps(output + ${N * 4}, _mm_castsi128_ps(vf${N}));
      output += ${BATCH_TILE};
    }
  for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);
    i += 8;

    const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
    const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);

    const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
    const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);

    const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
    const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));

    const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
    $if SSE == 4:
      const __m128i vf_lo = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
    $else:
      const __m128i vf_lo = _mm_or_si128(vsign_lo,
        _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));

    const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
    $if SSE == 4:
      const __m128i vf_hi = _mm_or_si128(vsign_hi, _mm_blendv_epi8(vdenorm_hi, vnorm_hi, vmask_hi));
    $else:
      const __m128i vf_hi = _mm_or_si128(vsign_hi,
        _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));

    _mm_storeu_ps(output, _mm_castsi128_ps(vf_lo));
    _mm_storeu_ps(output + 4, _mm_castsi128_ps(vf_hi));
    output += 8;
  }
  if XNN_UNPREDICTABLE(n != 0) {
    const __m128i vh = _mm_loadu_si128((const __m128i*) i);

    const __m128i vw_lo = _mm_unpacklo_epi16(_mm_setzero_si128(), vh);
    const __m128i vw_hi = _mm_unpackhi_epi16(_mm_setzero_si128(), vh);

    const __m128i vsign_lo = _mm_and_si128(vw_lo, vsign_mask);
    const __m128i vsign_hi = _mm_and_si128(vw_hi, vsign_mask);

    const __m128i vnonsign_lo = _mm_xor_si128(vw_lo, vsign_lo);
    const __m128i vnonsign_hi = _mm_xor_si128(vw_hi, vsign_hi);

    const __m128i vnorm_lo = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_lo, 3), vexp_offset)), vexp_scale));
    const __m128i vnorm_hi = _mm_castps_si128(_mm_mul_ps(_mm_castsi128_ps(_mm_add_epi32(_mm_srli_epi32(vnonsign_hi, 3), vexp_offset)), vexp_scale));

    const __m128i vdenorm_lo = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_lo, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));
    const __m128i vdenorm_hi = _mm_castps_si128(_mm_sub_ps(_mm_castsi128_ps(_mm_or_si128(_mm_srli_epi32(vnonsign_hi, 16), vmagic_bias)), _mm_castsi128_ps(vmagic_bias)));

    const __m128i vmask_lo = _mm_cmpgt_epi32(vnonsign_lo, vdenorm_cutoff);
    $if SSE == 4:
      __m128i vf = _mm_or_si128(vsign_lo, _mm_blendv_epi8(vdenorm_lo, vnorm_lo, vmask_lo));
    $else:
      __m128i vf = _mm_or_si128(vsign_lo,
        _mm_or_si128(_mm_and_si128(vmask_lo, vnorm_lo), _mm_andnot_si128(vmask_lo, vdenorm_lo)));

    if (n & (4 * sizeof(uint16_t))) {
      _mm_storeu_ps(output, _mm_castsi128_ps(vf));
      output += 4;

      const __m128i vmask_hi = _mm_cmpgt_epi32(vnonsign_hi, vdenorm_cutoff);
      $if SSE == 4:
        vf = _mm_or_si128(vsign_hi, _mm_blendv_epi8(vdenorm_hi, vnorm_hi, vmask_hi));
      $else:
        vf = _mm_or_si128(vsign_hi,
          _mm_or_si128(_mm_and_si128(vmask_hi, vnorm_hi), _mm_andnot_si128(vmask_hi, vdenorm_hi)));
    }
    if (n & (2 * sizeof(uint16_t))) {
      _mm_storel_pi((__m64*) output, _mm_castsi128_ps(vf));
      output += 2;

      vf = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(vf), _mm_castsi128_ps(vf)));
    }
    if (n & (1 * sizeof(uint16_t))) {
      _mm_store_ss(output, _mm_castsi128_ps(vf));
    }
  }
}