// 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
#include <assert.h>

#include <immintrin.h>

#include "xnnpack/common.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/vunary.h"


void xnn_f16_velu_ukernel__avx2_rr1_p3_u${BATCH_TILE}(
    size_t batch,
    const xnn_float16* input,
    xnn_float16* output,
    const struct xnn_f16_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(batch != 0);
  assert(batch % sizeof(uint16_t) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.0A4000p+3f);
  const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
  const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E430p-1f);
  const __m256 vc3 = _mm256_set1_ps(0x1.5554DCp-3f);
  const __m256 vc2 = _mm256_set1_ps(0x1.01EBB2p-1f);
  const __m256 vc1 = _mm256_set1_ps(0x1.0002F2p+0f);

  XNN_FORCE_REALIZATION(vsat_cutoff);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vlog2e);
  XNN_FORCE_REALIZATION(vminus_ln2);
  XNN_FORCE_REALIZATION(vc3);
  XNN_FORCE_REALIZATION(vc2);
  XNN_FORCE_REALIZATION(vc1);

  const __m256 vprescale = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.prescale));
  const __m256 valpha = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.alpha));
  const __m256 vbeta = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) &params->scalar.beta));

  const uint16_t* i = (const uint16_t*) input;
  uint16_t* o = (uint16_t*) output;
  $if BATCH_TILE > 8:
    for (; batch >= ${BATCH_TILE} * sizeof(uint16_t); batch -= ${BATCH_TILE} * sizeof(uint16_t)) {
      __m256 vx0 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
      $for N in range(1, SIMD_TILE):
        __m256 vx${N} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8})));
      i += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        const __m256 vz${N} = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx${N}, vprescale));

      $for N in range(SIMD_TILE):
        __m256 vn${N} = _mm256_fmadd_ps(vz${N}, vlog2e, vmagic_bias);

      $for N in range(SIMD_TILE):
        __m256 vs${N} = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn${N}), 23));
        vn${N} = _mm256_sub_ps(vn${N}, vmagic_bias);

      $for N in range(SIMD_TILE):
        __m256 vt${N} = _mm256_fmadd_ps(vn${N}, vminus_ln2, vz${N});

      $for N in range(SIMD_TILE):
        __m256 vp${N} = _mm256_fmadd_ps(vc3, vt${N}, vc2);

      $for N in range(SIMD_TILE):
        vp${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vc1);
        vt${N} = _mm256_mul_ps(vt${N}, valpha);

      $for N in range(SIMD_TILE):
        vt${N} = _mm256_mul_ps(vt${N}, vs${N});
        vs${N} = _mm256_fmsub_ps(vs${N}, valpha, valpha);

      $for N in range(SIMD_TILE):
        const __m256 ve${N} = _mm256_fmadd_ps(vp${N}, vt${N}, vs${N});
        vx${N} = _mm256_mul_ps(vx${N}, vbeta);

      $for N in range(SIMD_TILE):
        const __m256 vy${N} = _mm256_blendv_ps(vx${N}, ve${N}, vx${N});

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy0, _MM_FROUND_TO_NEAREST_INT));
      $for N in range(1, SIMD_TILE):
        _mm_storeu_si128((__m128i*) (o + ${N * 8}), _mm256_cvtps_ph(vy${N}, _MM_FROUND_TO_NEAREST_INT));
      o += ${BATCH_TILE};
    }
  for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
    __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));
    i += 8;

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);
    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);
    vt = _mm256_mul_ps(vt, valpha);
    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_fmsub_ps(vs, valpha, valpha);
    const __m256 ve = _mm256_fmadd_ps(vp, vt, vs);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

    _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vy, _MM_FROUND_TO_NEAREST_INT));
    o += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(uint16_t));
    assert(batch <= 7 * sizeof(uint16_t));
    __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i));

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
    __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);
    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = _mm256_fmadd_ps(vc3, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);
    vt = _mm256_mul_ps(vt, valpha);
    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_fmsub_ps(vs, valpha, valpha);
    const __m256 ve = _mm256_fmadd_ps(vp, vt, vs);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

    __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);
    }
  }
}