// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 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.

#include <assert.h>
#include <stdint.h>
#include <stddef.h>

#include <tmmintrin.h>

#include <xnnpack/math.h>
#include <xnnpack/requantization-stubs.h>


void xnn_qs8_requantize_rndna__ssse3(
    size_t n,
    const int32_t* input,
    float scale,
    int8_t zero_point,
    int8_t qmin,
    int8_t qmax,
    int8_t* output)
{
  assert(n % 16 == 0);
  assert(scale < 1.0f);
  assert(scale >= 0x1.0p-32f);

  const uint32_t scale_bits = float_as_uint32(scale);
  const uint32_t multiplier = (scale_bits & UINT32_C(0x007FFFFF)) | UINT32_C(0x00800000);
  const uint32_t shift = 127 + 23 - (scale_bits >> 23);
  assert(shift >= 24);
  assert(shift < 56);
  const uint64_t rounding = UINT64_C(1) << (shift - 1);

  const __m128i vmultiplier = _mm_set1_epi32(multiplier);
  const __m128i vzero_point = _mm_set1_epi16((short) zero_point);
  const __m128i vqmin = _mm_set1_epi8((char) qmin);
  const __m128i vqmax = _mm_set1_epi8((char) qmax);
  const __m128i vshift = _mm_cvtsi32_si128((int) shift);
  const __m128i vrounding = _mm_set1_epi64x(rounding);
  for (; n != 0; n -= 16) {
    const __m128i x = _mm_loadu_si128((const __m128i*) input);
    const __m128i y = _mm_loadu_si128((const __m128i*) (input + 4));
    const __m128i z = _mm_loadu_si128((const __m128i*) (input + 8));
    const __m128i w = _mm_loadu_si128((const __m128i*) (input + 12));
    input += 16;

    const __m128i x_abs0123 = _mm_abs_epi32(x);
    const __m128i y_abs0123 = _mm_abs_epi32(y);
    const __m128i z_abs0123 = _mm_abs_epi32(z);
    const __m128i w_abs0123 = _mm_abs_epi32(w);

    const __m128i x_abs1032 = _mm_shuffle_epi32(x_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i y_abs1032 = _mm_shuffle_epi32(y_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i z_abs1032 = _mm_shuffle_epi32(z_abs0123, _MM_SHUFFLE(2, 3, 0, 1));
    const __m128i w_abs1032 = _mm_shuffle_epi32(w_abs0123, _MM_SHUFFLE(2, 3, 0, 1));

    const __m128i x_absmul02 = _mm_mul_epu32(x_abs0123, vmultiplier);
    const __m128i y_absmul02 = _mm_mul_epu32(y_abs0123, vmultiplier);
    const __m128i z_absmul02 = _mm_mul_epu32(z_abs0123, vmultiplier);
    const __m128i w_absmul02 = _mm_mul_epu32(w_abs0123, vmultiplier);

    const __m128i x_absmul13 = _mm_mul_epu32(x_abs1032, vmultiplier);
    const __m128i y_absmul13 = _mm_mul_epu32(y_abs1032, vmultiplier);
    const __m128i z_absmul13 = _mm_mul_epu32(z_abs1032, vmultiplier);
    const __m128i w_absmul13 = _mm_mul_epu32(w_abs1032, vmultiplier);

    const __m128i x_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(x_absmul02, vrounding), vshift);
    const __m128i x_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(x_absmul13, vrounding), vshift);
    const __m128i y_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(y_absmul02, vrounding), vshift);
    const __m128i y_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(y_absmul13, vrounding), vshift);
    const __m128i z_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(z_absmul02, vrounding), vshift);
    const __m128i z_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(z_absmul13, vrounding), vshift);
    const __m128i w_abs_scaled02 = _mm_srl_epi64(_mm_add_epi64(w_absmul02, vrounding), vshift);
    const __m128i w_abs_scaled13 = _mm_srl_epi64(_mm_add_epi64(w_absmul13, vrounding), vshift);

    const __m128i x_abs_scaled0213 = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
    const __m128i y_abs_scaled0213 = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
    const __m128i z_abs_scaled0213 = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));
    const __m128i w_abs_scaled0213 = _mm_castps_si128(
        _mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, 0, 2, 0)));

    const __m128i x_abs_scaled = _mm_shuffle_epi32(x_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
    const __m128i y_abs_scaled = _mm_shuffle_epi32(y_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
    const __m128i z_abs_scaled = _mm_shuffle_epi32(z_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));
    const __m128i w_abs_scaled = _mm_shuffle_epi32(w_abs_scaled0213, _MM_SHUFFLE(3, 1, 2, 0));

    const __m128i x_scaled = _mm_sign_epi32(x_abs_scaled, x);
    const __m128i y_scaled = _mm_sign_epi32(y_abs_scaled, y);
    const __m128i z_scaled = _mm_sign_epi32(z_abs_scaled, z);
    const __m128i w_scaled = _mm_sign_epi32(w_abs_scaled, w);

    const __m128i xy_packed = _mm_adds_epi16(_mm_packs_epi32(x_scaled, y_scaled), vzero_point);
    const __m128i zw_packed = _mm_adds_epi16(_mm_packs_epi32(z_scaled, w_scaled), vzero_point);
    const __m128i xy_clamped = _mm_max_epi16(_mm_min_epi16(xy_packed, vqmax), vqmin);
    const __m128i zw_clamped = _mm_max_epi16(_mm_min_epi16(zw_packed, vqmax), vqmin);
    const __m128i xyzw_clamped = _mm_packs_epi16(xy_clamped, zw_clamped);

    // 4x PABSD
    // 8x PSHUFD
    // 8x PMULUDQ
    // 8x PSRLQ
    // 8x PADDQ
    // 4x SHUFPS
    // 4x PSIGND
    // 2x PACKSSDW
    // 2x PADDSW
    // 2x PMAXSW
    // 2x PMINSW
    // 1x PACKSSWB
    // ---------------------
    // 53 instructions total

    _mm_storeu_si128((__m128i*) output, xyzw_clamped);
    output += 16;
  }
}