// Auto-generated file. Do not edit!
//   Template: src/qs8-gemm/MRx4c8-sse.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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 <smmintrin.h>

#include <xnnpack/gemm.h>
#include <xnnpack/math.h>


void xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse41(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(mr != 0);
  assert(mr <= 3);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(int8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8);
  const int8_t* a0 = a;
  int8_t* c0 = c;
  const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
  int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    a1 = a0;
    c1 = c0;
  }
  const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
  int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    a2 = a1;
    c2 = c1;
  }

  do {
    __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
    __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
    __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
    __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
    __m128i vacc1x0 = vacc0x0;
    __m128i vacc1x1 = vacc0x1;
    __m128i vacc1x2 = vacc0x2;
    __m128i vacc1x3 = vacc0x3;
    __m128i vacc2x0 = vacc0x0;
    __m128i vacc2x1 = vacc0x1;
    __m128i vacc2x2 = vacc0x2;
    __m128i vacc2x3 = vacc0x3;
    w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));

    size_t k = 0;
    while (k < kc) {
      const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
      const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
      a0 += 8;
      const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
      const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
      a1 += 8;
      const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
      const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
      a2 += 8;

      const __m128i vxb0 = _mm_load_si128((const __m128i*) w);

      vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
      vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
      vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
      const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));

      vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
      vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
      vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
      const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));

      vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
      vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
      vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
      const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));

      vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
      vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
      vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));

      w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
      k += 8 * sizeof(int8_t);
    }

    const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
    const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
    const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
    const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
    const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
    const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);

    __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
    __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
    __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);

    const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
    const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);

    const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
    const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
    const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));

    const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
    const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
    const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);

    const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
    const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
    const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);

    const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
    const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
    const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
    const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
    const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
    const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);

    const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
    const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
    const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);

    const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
    const __m128i vrem0x0123 =
      _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
    const __m128i vrem1x0123 =
      _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
    const __m128i vrem2x0123 =
      _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));

    const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
    const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
    vacc0x0123 =
      _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
    vacc1x0123 =
      _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
    vacc2x0123 =
      _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));

    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
    __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
    __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);

    const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
    const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
    vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
    vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);

    __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);

    if (nc >= 4) {
      *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
      *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
      *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);

      c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
      c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
      c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);

      a0 = (const int8_t*) ((uintptr_t) a0 - kc);
      a1 = (const int8_t*) ((uintptr_t) a1 - kc);
      a2 = (const int8_t*) ((uintptr_t) a2 - kc);

      nc -= 4;
    } else {
      if (nc & 2) {
        *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
        c0 += 2;
        *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
        c1 += 2;
        *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
        c2 += 2;
        vout = _mm_srli_epi32(vout, 16);
      }
      if (nc & 1) {
        *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
        *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
        *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
      }

      nc = 0;
    }
  } while (nc != 0);
}