xref: /aosp_15_r20/external/XNNPACK/src/qu8-vcvt/gen/vcvt-ssse3-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-vcvt/ssse3.c.in
//   Generator: tools/xngen
//
// 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.

#include <assert.h>

#include <tmmintrin.h>

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


void xnn_qu8_vcvt_ukernel__ssse3_x16(
    size_t n,
    const uint8_t* x,
    uint8_t* y,
    const union xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(uint8_t) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point);
  const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier);
  const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point);
  const __m128i vzero = _mm_setzero_si128();
  for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
    const __m128i vx0 = _mm_loadu_si128((const __m128i*) x);
    x += 16;

    __m128i vacc0 = _mm_unpacklo_epi8(vx0, vzero);
    __m128i vacc1 = _mm_unpackhi_epi8(vx0, vzero);

    vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0);
    vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1);

    vacc0 = _mm_slli_epi16(vacc0, 7);
    vacc1 = _mm_slli_epi16(vacc1, 7);

    vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier);
    vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier);

    vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point);
    vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point);

    const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1);

    _mm_storeu_si128((__m128i*) y, vy0);
    y += 16;
  }
  for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
    const __m128i vx = _mm_loadu_si128((const __m128i*) x);
    x += 16;

    __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
    __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
    vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
    vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
    vacc_lo = _mm_slli_epi16(vacc_lo, 7);
    vacc_hi = _mm_slli_epi16(vacc_hi, 7);
    vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
    vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
    vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
    vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);

    const __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
    _mm_storeu_si128((__m128i*) y, vy);
    y += 16;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(uint8_t));
    assert(n <= 15 * sizeof(uint8_t));

    const __m128i vx = _mm_loadu_si128((const __m128i*) x);

    __m128i vacc_lo = _mm_unpacklo_epi8(vx, vzero);
    __m128i vacc_hi = _mm_unpackhi_epi8(vx, vzero);
    vacc_lo = _mm_sub_epi16(vinput_zero_point, vacc_lo);
    vacc_hi = _mm_sub_epi16(vinput_zero_point, vacc_hi);
    vacc_lo = _mm_slli_epi16(vacc_lo, 7);
    vacc_hi = _mm_slli_epi16(vacc_hi, 7);
    vacc_lo = _mm_mulhrs_epi16(vacc_lo, vmultiplier);
    vacc_hi = _mm_mulhrs_epi16(vacc_hi, vmultiplier);
    vacc_lo = _mm_adds_epi16(vacc_lo, voutput_zero_point);
    vacc_hi = _mm_adds_epi16(vacc_hi, voutput_zero_point);

    __m128i vy = _mm_packus_epi16(vacc_lo, vacc_hi);
    if (n & (8 * sizeof(uint8_t))) {
      _mm_storel_epi64((__m128i*) y, vy);
      vy = _mm_unpackhi_epi64(vy, vy);
      y += 8;
    }
    if (n & (4 * sizeof(uint8_t))) {
      unaligned_store_u32(y, (uint32_t) _mm_cvtsi128_si32(vy));
      vy = _mm_srli_epi64(vy, 32);
      y += 4;
    }
    uint32_t vy_lo = (uint32_t) _mm_cvtsi128_si32(vy);
    if (n & (2 * sizeof(uint8_t))) {
      unaligned_store_u16(y, (uint16_t) vy_lo);
      vy_lo >>= 16;
      y += 2;
    }
    if (n & (1 * sizeof(uint8_t))) {
      *y = (uint8_t) vy_lo;
    }
  }
}