xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/3x4c8-minmax-fp32-xop-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-igemm/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>

#if defined(__GNUC__) || defined(__clang__)
  #include <x86intrin.h>
#else
  #include <immintrin.h>
  #include <ammintrin.h>
#endif

#include <xnnpack/igemm.h>
#include <xnnpack/math.h>
#include <xnnpack/unaligned.h>


void xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__xop_ld64(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const uint8_t** restrict a,
    const void* restrict w,
    uint8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const uint8_t* zero,
    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(mr != 0);
  assert(mr <= 3);
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (3 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(uint8_t) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  kc = round_up_po2(kc, 8);
  uint8_t* c0 = c;
  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    c1 = c0;
  }
  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    c2 = c1;
  }

  do {
    __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
    __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
    __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
    __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) 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 int32_t*) w + 4;

    size_t p = ks;
    do {
      const uint8_t* restrict a0 = a[0];
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
      }
      const uint8_t* restrict a1 = a[1];
      if XNN_UNPREDICTABLE(a1 != zero) {
        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
      }
      const uint8_t* restrict a2 = a[2];
      if XNN_UNPREDICTABLE(a2 != zero) {
        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
      }
      a += 3;

      size_t k = 0;
      const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
      while (k < kc) {
        const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
        const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
        a0 += 8;
        const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
        const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
        a1 += 8;
        const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
        const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
        a2 += 8;

        const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
        const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);

        vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
        vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
        vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
        const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
        const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);

        vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
        vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
        vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
        const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
        const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);

        vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
        vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
        vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
        const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
        const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);

        vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
        vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
        vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);

        w = (const void*) ((const uint8_t*) w + 32);
        k += 8 * sizeof(uint8_t);
      }
      p -= 3 * sizeof(void*);
    } while (p != 0);

    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);

    __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
    __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
    __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);

    const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
    vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
    vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
    vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);

    const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
    vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
    vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
    vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);

    vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
    vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
    vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);

    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_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);

    __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);

    vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));

    if (nc >= 4) {
      unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
      c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
      unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
      c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
      unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
      c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);

      a = (const uint8_t**restrict) ((uintptr_t) a - ks);

      nc -= 4;
    } else {
      if (nc & 2) {
        unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
        c2 += 2;
        unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
        c1 += 2;
        unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
        c0 += 2;
        vout = _mm_srli_epi32(vout, 16);
      }
      if (nc & 1) {
        *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
        *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
        *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
      }

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