// 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 <arm_neon.h>

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


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$XINT8X8_T = {"QS8": "int8x8_t", "QU8": "uint8x8_t"}[DATATYPE]
$XINT8X16_T = {"QS8": "int8x16_t", "QU8": "uint8x16_t"}[DATATYPE]
$VLD1Q_DUP_X8 = {"QS8": "vld1q_dup_s8", "QU8": "vld1q_dup_u8"}[DATATYPE]
$VLD1_DUP_X8 = {"QS8": "vld1_dup_s8", "QU8": "vld1_dup_u8"}[DATATYPE]
$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE]
$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE]
$VST1_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE]
$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE]
$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[DATATYPE]
$VCOMBINE_X8 = {"QS8": "vcombine_s8", "QU8": "vcombine_u8"}[DATATYPE]
$VGET_LOW_X8 = {"QS8": "vget_low_s8", "QU8": "vget_low_u8"}[DATATYPE]
$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE]
$VREINTERPRET_U32_X8 = {"QS8": "vreinterpret_u32_s8", "QU8": "vreinterpret_u32_u8"}[DATATYPE]
$VMINQ_X8 = {"QS8": "vminq_s8", "QU8": "vminq_u8"}[DATATYPE]
$VMIN_X8 = {"QS8": "vmin_s8", "QU8": "vmin_u8"}[DATATYPE]
$VMAXQ_X8 = {"QS8": "vmaxq_s8", "QU8": "vmaxq_u8"}[DATATYPE]
$VMAX_X8 = {"QS8": "vmax_s8", "QU8": "vmax_u8"}[DATATYPE]
void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__neon_x${BATCH_TILE}(
    size_t n,
    const float* x,
    ${XINT8_T}* y,
    const union xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const float32x4_t vscale = vld1q_dup_f32(&params->neon.scale);
  const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neon.magic_bias);
  const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(&params->neon.magic_bias_less_zero_point);
  $if BATCH_TILE > 8:
    const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(&params->neon.output_min);
    const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(&params->neon.output_max);
  $else:
    const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(&params->neon.output_min);
    const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(&params->neon.output_max);
  $if BATCH_TILE > 8:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(0, BATCH_TILE, 4):
        float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4;

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vscale);

      $for N in range(0, BATCH_TILE, 4):
        vx${ABC[N:N+4]} = vaddq_f32(vx${ABC[N:N+4]}, vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        const int32x4_t vacc${ABC[N:N+4]} = vqsubq_s32(vreinterpretq_s32_f32(vx${ABC[N:N+4]}), vmagic_bias_less_zero_point);

      $for N in range(0, BATCH_TILE, 8):
        const int16x8_t vacc${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[N:N+4]}), vqmovn_s32(vacc${ABC[N+4:N+8]}));

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          ${XINT8X16_T} vy${ABC[N:N+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${ABC[N+8:N+16]}));
        $else:
          ${XINT8X8_T} vy${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${ABC[N:N+8]});

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          vy${ABC[N:N+16]} = ${VMAXQ_X8}(vy${ABC[N:N+16]}, voutput_min);
        $else:
          vy${ABC[N:N+8]} = ${VMAX_X8}(vy${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_min));

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          vy${ABC[N:N+16]} = ${VMINQ_X8}(vy${ABC[N:N+16]}, voutput_max);
        $else:
          vy${ABC[N:N+8]} = ${VMIN_X8}(vy${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_max));

      $for N in range(0, BATCH_TILE, 16):
        $if N + 8 < BATCH_TILE:
          ${VST1Q_X8}(y, vy${ABC[N:N+16]}); y += 16;
        $else:
          ${VST1_X8}(y, vy${ABC[N:N+8]}); y += 8;
    }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    float32x4_t vx_lo = vld1q_f32(x); x += 4;
    float32x4_t vx_hi = vld1q_f32(x); x += 4;

    vx_lo = vmulq_f32(vx_lo, vscale);
    vx_hi = vmulq_f32(vx_hi, vscale);

    vx_lo = vaddq_f32(vx_lo, vmagic_bias);
    vx_hi = vaddq_f32(vx_hi, vmagic_bias);

    const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
    const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);

    const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));

    ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc);
    $if BATCH_TILE > 8:
      vy = ${VMAX_X8}(vy, ${VGET_LOW_X8}(voutput_min));
      vy = ${VMIN_X8}(vy, ${VGET_LOW_X8}(voutput_max));
    $else:
      vy = ${VMAX_X8}(vy, voutput_min);
      vy = ${VMIN_X8}(vy, voutput_max);
    ${VST1_X8}(y, vy); y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 7 * sizeof(float));
    float32x4_t vx_lo = vld1q_f32(x);
    const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float))));
    float32x4_t vx_hi = vld1q_f32(x_hi);

    vx_lo = vmulq_f32(vx_lo, vscale);
    vx_hi = vmulq_f32(vx_hi, vscale);

    vx_lo = vaddq_f32(vx_lo, vmagic_bias);
    vx_hi = vaddq_f32(vx_hi, vmagic_bias);

    const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point);
    const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point);

    const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi));

    ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc);
    $if BATCH_TILE > 8:
      vy = ${VMAX_X8}(vy, ${VGET_LOW_X8}(voutput_min));
      vy = ${VMIN_X8}(vy, ${VGET_LOW_X8}(voutput_max));
    $else:
      vy = ${VMAX_X8}(vy, voutput_min);
      vy = ${VMIN_X8}(vy, voutput_max);

    if (n & (4 * sizeof(float))) {
      vst1_lane_u32((void*) y, ${VREINTERPRET_U32_X8}(vy), 0); y += 4;
      vy = ${VEXT_X8}(vy, vy, 4);
    }
    if (n & (2 * sizeof(float))) {
      vst1_lane_u16((void*) y, ${VREINTERPRET_U16_X8}(vy), 0); y += 2;
      vy = ${VEXT_X8}(vy, vy, 2);
    }
    if (n & (1 * sizeof(float))) {
      ${VST1_LANE_X8}(y, vy, 0);
    }
  }
}