// 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.

$assert VARIANT in ["LD128", "EXTENDED"]
$assert MR <= 4
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/gemm.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/math.h>


$GEMM_SUFFIX = "_xw" if VARIANT == "EXTENDED" else ""
void xnn_qs8_gemm${GEMM_SUFFIX}_minmax_ukernel_${MR}x8c8__avx2(
    size_t mr,
    size_t nc,
    size_t kc,
    const int8_t* restrict a,
    size_t a_stride,
    const void* restrict w,
    int8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_qs8_gemm${GEMM_SUFFIX}_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(mr != 0);
  assert(mr <= ${MR});
  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;
  $for M in range(1, MR):
    const int8_t* a${M} = (const int8_t*) ((uintptr_t) a${M-1} + a_stride);
    int8_t* c${M} = (int8_t*) ((uintptr_t) c${M-1} + cm_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(mr <= ${M}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $elif M + 1 == MR:
      if XNN_UNPREDICTABLE(mr != ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(mr < ${M+1}) {
        a${M} = a${M-1};
        c${M} = c${M-1};
      }

  do {
    const __m128i vbias0x0 = _mm_loadu_si32(w);
    const __m128i vbias0x1 = _mm_loadu_si32((const void*) ((uintptr_t) w + sizeof(int32_t)));
    __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
    $for N in range(2, 8, 2):
      const __m128i vbias0x${N} = _mm_loadu_si32((const void*) ((uintptr_t) w + ${N} * sizeof(int32_t)));
      const __m128i vbias0x${N+1} = _mm_loadu_si32((const void*) ((uintptr_t) w + ${N+1} * sizeof(int32_t)));
      __m256i vacc0x${N}${N+1} = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x${N}), vbias0x${N+1}, 1);
    $for M in range(1, MR):
      $for N in range(0, 8, 2):
        __m256i vacc${M}x${N}${N+1} = vacc0x${N}${N+1};
    w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));

    size_t k = 0;
    while (k < kc) {
      $for M in range(MR):
        const __m128i va${M} = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a${M}));
        const __m256i vxa${M} = _mm256_cvtepi8_epi16(va${M});
        a${M} += 8;

      $for N in range(0, 8, 2):
        $if VARIANT == "EXTENDED":
          $if N == 0:
            const __m256i vxb${N}${N+1} = _mm256_load_si256((const __m256i*) w);
          $else:
            const __m256i vxb${N}${N+1} = _mm256_load_si256((const __m256i*) ((uintptr_t) w + ${N * 8} * sizeof(int16_t)));
        $else:
          $if N == 0:
            const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) w);
          $else:
            const __m128i vb${N}${N+1} = _mm_load_si128((const __m128i*) ((uintptr_t) w + ${N * 8} * sizeof(int8_t)));
          const __m256i vxb${N}${N+1} = _mm256_cvtepi8_epi16(vb${N}${N+1});

        $for M in range(MR):
          vacc${M}x${N}${N+1} = _mm256_add_epi32(vacc${M}x${N}${N+1}, _mm256_madd_epi16(vxa${M}, vxb${N}${N+1}));

      $if VARIANT == "EXTENDED":
        w = (const void*) ((uintptr_t) w + 64 * sizeof(int16_t));
      $else:
        w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
      k += 8 * sizeof(int8_t);
    }

    $for M in range(MR):
      const __m256i vacc${M}x0213 = _mm256_hadd_epi32(vacc${M}x01, vacc${M}x23);
      const __m256i vacc${M}x4657 = _mm256_hadd_epi32(vacc${M}x45, vacc${M}x67);

    $for M in range(MR):
      const __m256i vacc${M}x02461357 = _mm256_hadd_epi32(vacc${M}x0213, vacc${M}x4657);

    const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
    $for M in range(MR):
      __m256i vacc${M}x01234567 = _mm256_permutevar8x32_epi32(vacc${M}x02461357, vpermute_mask);

    const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier));
    const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));

    $for M in range(MR):
      const __m256i vacc${M}x11335577 = _mm256_shuffle_epi32(vacc${M}x01234567, _MM_SHUFFLE(3, 3, 1, 1));

    $for M in range(MR):
      const __m256i vprod${M}x0246 = _mm256_add_epi64(_mm256_mul_epi32(vacc${M}x01234567, vmultiplier), vrounding);

    $for M in range(MR):
      const __m256i vprod${M}x1357 = _mm256_add_epi64(_mm256_mul_epi32(vacc${M}x11335577, vmultiplier), vrounding);

    $for M in range(MR):
      const __m256i vq31prod${M}x0246 = _mm256_srli_epi64(vprod${M}x0246, 31);
      const __m256i vq31prod${M}x1357 = _mm256_add_epi64(vprod${M}x1357, vprod${M}x1357);

    $for M in range(MR):
      const __m256i vq31prod${M}x01234567 = _mm256_blend_epi16(vq31prod${M}x0246, vq31prod${M}x1357, 0xCC);

    const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
    $for M in range(MR):
      const __m256i vrem${M}x01234567 =
        _mm256_add_epi32(_mm256_and_si256(vq31prod${M}x01234567, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod${M}x01234567));

    const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
    const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
    $for M in range(MR):
      vacc${M}x01234567 =
        _mm256_sub_epi32(_mm256_sra_epi32(vq31prod${M}x01234567, vshift), _mm256_cmpgt_epi32(vrem${M}x01234567, vremainder_threshold));

    const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
    $for M in range(0, MR, 2):
      __m256i vacc${M}${min(M+1, MR-1)}x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc${M}x01234567, vacc${min(M+1, MR-1)}x01234567), voutput_zero_point);

    $for M in range(0, MR, 2):
      vacc${M}${min(M+1, MR-1)}x01234567 = _mm256_permute4x64_epi64(vacc${M}${min(M+1, MR-1)}x01234567, _MM_SHUFFLE(3, 1, 2, 0));

    const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
    const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
    $for M in range(0, MR, 2):
      vacc${M}${min(M+1, MR-1)}x01234567 = _mm256_min_epi16(_mm256_max_epi16(vacc${M}${min(M+1, MR-1)}x01234567, voutput_min), voutput_max);

    $if MR > 2:
      __m256i vout = _mm256_packs_epi16(vacc0${min(1, MR-1)}x01234567, vacc${min(2, MR-1)}${min(3, MR-1)}x01234567);
    $else:
      __m256i vout = _mm256_packs_epi16(vacc0${min(1, MR-1)}x01234567, vacc0${min(1, MR-1)}x01234567);
    __m128i vout_lo = _mm256_castsi256_si128(vout);
    __m128i vout_hi = _mm256_extracti128_si256(vout, 1);

    if (nc >= 8) {
      _mm_storel_epi64((__m128i*) c0, vout_lo);
      $if MR > 1:
        _mm_storel_epi64((__m128i*) c1, vout_hi);
      $if MR > 2:
        _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
      $if MR > 3:
        _mm_storeh_pi((__m64*) c3, _mm_castsi128_ps(vout_hi));

      $for M in range(MR):
        c${M} = (int8_t*) ((uintptr_t) c${M} + cn_stride);

      $for M in range(MR):
        a${M} = (const int8_t*) ((uintptr_t) a${M} - kc);

      nc -= 8;
    } else {
      if (nc & 4) {
        _mm_storeu_si32(c0, vout_lo);
        $if MR > 1:
          _mm_storeu_si32(c1, vout_hi);
        $if MR > 2:
          *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout_lo, 2);
        $if MR > 3:
          *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout_hi, 2);

        $for M in range(MR):
          c${M} += 4;

        vout_lo = _mm_srli_epi64(vout_lo, 32);
        vout_hi = _mm_srli_epi64(vout_hi, 32);
      }
      if (nc & 2) {
        *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout_lo, 0);
        $if MR > 1:
          *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout_hi, 0);
        $if MR > 2:
          *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout_lo, 4);
        $if MR > 3:
          *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout_hi, 4);

        $for M in range(MR):
          c${M} += 2;

        vout_lo = _mm_srli_epi32(vout_lo, 16);
        vout_hi = _mm_srli_epi32(vout_hi, 16);
      }
      if (nc & 1) {
        *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
        $if MR > 1:
          *c1 = (uint8_t) _mm_extract_epi8(vout_hi, 0);
        $if MR > 2:
          *c2 = (uint8_t) _mm_extract_epi8(vout_lo, 8);
        $if MR > 3:
          *c3 = (uint8_t) _mm_extract_epi8(vout_hi, 8);
      }

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