// 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.

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

#include <arm_neon.h>

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


void xnn_qs8_requantize_rndnu__neon_mull(
    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 int32_t multiplier = ((int32_t) scale_bits & INT32_C(0x007FFFFF)) | INT32_C(0x00800000);
  const int32_t shift = 127 + 23 - (scale_bits >> 23);
  assert(shift >= 24);
  assert(shift < 56);

#if defined(__aarch64__)
  const int32x4_t vmultiplier = vdupq_n_s32(multiplier);
#else
  const int32x2_t vmultiplier = vdup_n_s32(multiplier);
#endif
  const int16x8_t vzero_point = vdupq_n_s16((int16_t) zero_point);
  const int64x2_t vshift = vdupq_n_s64(-shift);
  const int8x16_t vqmin = vdupq_n_s8(qmin);
  const int8x16_t vqmax = vdupq_n_s8(qmax);
  for (; n != 0; n -= 16) {
    const int32x4_t x = vld1q_s32(input);
    const int32x4_t y = vld1q_s32(input + 4);
    const int32x4_t z = vld1q_s32(input + 8);
    const int32x4_t w = vld1q_s32(input + 12);
    input += 16;

#if defined(__aarch64__)
    const int64x2_t x01_product = vmull_s32(vget_low_s32(x), vget_low_s32(vmultiplier));
    const int64x2_t x23_product = vmull_high_s32(x, vmultiplier);
    const int64x2_t y01_product = vmull_s32(vget_low_s32(y), vget_low_s32(vmultiplier));
    const int64x2_t y23_product = vmull_high_s32(y, vmultiplier);
    const int64x2_t z01_product = vmull_s32(vget_low_s32(z), vget_low_s32(vmultiplier));
    const int64x2_t z23_product = vmull_high_s32(z, vmultiplier);
    const int64x2_t w01_product = vmull_s32(vget_low_s32(w), vget_low_s32(vmultiplier));
    const int64x2_t w23_product = vmull_high_s32(w, vmultiplier);
#else
    const int64x2_t x01_product = vmull_s32(vget_low_s32(x), vmultiplier);
    const int64x2_t x23_product = vmull_s32(vget_high_s32(x), vmultiplier);
    const int64x2_t y01_product = vmull_s32(vget_low_s32(y), vmultiplier);
    const int64x2_t y23_product = vmull_s32(vget_high_s32(y), vmultiplier);
    const int64x2_t z01_product = vmull_s32(vget_low_s32(z), vmultiplier);
    const int64x2_t z23_product = vmull_s32(vget_high_s32(z), vmultiplier);
    const int64x2_t w01_product = vmull_s32(vget_low_s32(w), vmultiplier);
    const int64x2_t w23_product = vmull_s32(vget_high_s32(w), vmultiplier);
#endif

    const int64x2_t x01_scaled = vrshlq_s64(x01_product, vshift);
    const int64x2_t x23_scaled = vrshlq_s64(x23_product, vshift);
    const int64x2_t y01_scaled = vrshlq_s64(y01_product, vshift);
    const int64x2_t y23_scaled = vrshlq_s64(y23_product, vshift);
    const int64x2_t z01_scaled = vrshlq_s64(z01_product, vshift);
    const int64x2_t z23_scaled = vrshlq_s64(z23_product, vshift);
    const int64x2_t w01_scaled = vrshlq_s64(w01_product, vshift);
    const int64x2_t w23_scaled = vrshlq_s64(w23_product, vshift);

#ifdef __aarch64__
    const int32x4_t x_scaled = vuzp1q_s32(vreinterpretq_s32_s64(x01_scaled), vreinterpretq_s32_s64(x23_scaled));
    const int32x4_t y_scaled = vuzp1q_s32(vreinterpretq_s32_s64(y01_scaled), vreinterpretq_s32_s64(y23_scaled));
    const int32x4_t z_scaled = vuzp1q_s32(vreinterpretq_s32_s64(z01_scaled), vreinterpretq_s32_s64(z23_scaled));
    const int32x4_t w_scaled = vuzp1q_s32(vreinterpretq_s32_s64(w01_scaled), vreinterpretq_s32_s64(w23_scaled));

    const int16x8_t xy_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(x_scaled), y_scaled), vzero_point);
    const int16x8_t zw_packed = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(z_scaled), w_scaled), vzero_point);
    const int8x16_t xyzw_packed = vqmovn_high_s16(vqmovn_s16(xy_packed), zw_packed);
#else
    const int32x4_t x_scaled = vcombine_s32(vmovn_s64(x01_scaled), vmovn_s64(x23_scaled));
    const int32x4_t y_scaled = vcombine_s32(vmovn_s64(y01_scaled), vmovn_s64(y23_scaled));
    const int32x4_t z_scaled = vcombine_s32(vmovn_s64(z01_scaled), vmovn_s64(z23_scaled));
    const int32x4_t w_scaled = vcombine_s32(vmovn_s64(w01_scaled), vmovn_s64(w23_scaled));

    const int16x8_t xy_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(x_scaled), vqmovn_s32(y_scaled)), vzero_point);
    const int16x8_t zw_packed = vqaddq_s16(vcombine_s16(vqmovn_s32(z_scaled), vqmovn_s32(w_scaled)), vzero_point);
    const int8x16_t xyzw_packed = vcombine_s8(vqmovn_s16(xy_packed), vqmovn_s16(zw_packed));
#endif

    const int8x16_t xyzw_clamped = vmaxq_s8(vminq_s8(xyzw_packed, vqmax), vqmin);

    // AArch32 version:
    //   8x VMULL.S32 Qd, Dm, Dn
    //   8x VRSHL.S32 Qd, Qm, Qn
    //   8x VMOVN.S64 Dd, Qm
    //   4x VQMOVN.S32 Dd, Qm
    //   2x VQADD.S16 Qd, Qm, Qn
    //   2x VQMOVUN.S16 Dd, Qm
    //   1x VMAX.U8 Qd, Qm, Qn
    //   1x VMIN.U8 Qd, Qm, Qn
    // ---------------------
    // 34 instructions total
    //
    // AArch64 version:
    //   4x SMULL Vd.2D, Vn.2S, Vm.2S
    //   4x SMULL2 Vd.2D, Vn.4S, Vm.4S
    //   8x SRSHL Vd.2D, Vn.2D, Vm.2D
    //   4x UZP1 Vd.4S, Vn.4S, Vm.4S
    //   2x SQXTN Vd.4H, Vn.4S
    //   2x SQXTN2 Vd.8H, Vn.4S
    //   2x SQADD Vd.8H, Vn.8H, Vm.8H
    //   1x SQXTN Vd.8B, Vn.8H
    //   1x SQXTN2 Vd.16B, Vn.8H
    //   1x SMIN Vd.16B, Vn.16B, Vm.16B
    //   1x SMAX Vd.16B, Vn.16B, Vm.16B
    // ---------------------
    // 30 instructions total

    vst1q_s8(output, xyzw_clamped);
    output += 16;
  }
}