// 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 BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vadd.h>


void xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x${BATCH_TILE}(
    size_t n,
    const int8_t* input_x,
    const int8_t* input_y,
    int8_t* output,
    const union xnn_qs8_add_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  const __m256i vzero_point_product = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.zero_point_product));
  const __m256i vx_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.x_multiplier));
  const __m256i vy_multiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.y_multiplier));
  const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
  const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
  $if BATCH_TILE > 8:
    const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
    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));
  $else:
    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_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);

  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
    const __m256i vx${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
    const __m256i vy${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_y));
    $for N in range(8, BATCH_TILE, 8):
      const __m256i vx${ABC[N:N+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_x + ${N})));
      const __m256i vy${ABC[N:N+8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_y + ${N})));
    input_x += ${BATCH_TILE};
    input_y += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 8):
      __m256i vacc${ABC[N:N+8]} = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx${ABC[N:N+8]}, vx_multiplier));

    $for N in range(0, BATCH_TILE, 8):
      vacc${ABC[N:N+8]} = _mm256_add_epi32(vacc${ABC[N:N+8]}, _mm256_mullo_epi32(vy${ABC[N:N+8]}, vy_multiplier));

    $for N in range(0, BATCH_TILE, 8):
      const __m256i vrem${ABC[N:N+8]} = _mm256_add_epi32(_mm256_and_si256(vacc${ABC[N:N+8]}, vremainder_mask), _mm256_srai_epi32(vacc${ABC[N:N+8]}, 31));

    $for N in range(0, BATCH_TILE, 8):
      vacc${ABC[N:N+8]} = _mm256_sub_epi32(_mm256_sra_epi32(vacc${ABC[N:N+8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[N:N+8]}, vremainder_threshold));

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        __m256i vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[N:N+8]}, vacc${ABC[N+8:N+16]}), voutput_zero_point);
      $elif BATCH_TILE > 8:
        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
      $else:
        __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), voutput_zero_point);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_min_epi16(_mm256_max_epi16(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, voutput_min), voutput_max);
      $elif BATCH_TILE > 8:
        vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
      $else:
        vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, voutput_min), voutput_max);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        __m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}), _mm256_extracti128_si256(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0));
      $else:
        __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});

    $if BATCH_TILE >= 16:
      _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
    $else:
      _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
    $for N in range(16, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]});
      $else:
        _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]});
    output += ${BATCH_TILE};
  }
  if XNN_UNLIKELY(n != 0) {
    ${"do " if BATCH_TILE > 8 else ""}{
      const __m256i vx${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_x));
      const __m256i vy${ABC[0:8]} = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_y));
      $if BATCH_TILE > 8:
        input_x += 8;
        input_y += 8;

      __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vzero_point_product, _mm256_mullo_epi32(vx${ABC[0:8]}, vx_multiplier));

      vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_mullo_epi32(vy${ABC[0:8]}, vy_multiplier));

      const __m256i vrem${ABC[0:8]} = _mm256_add_epi32(_mm256_and_si256(vacc${ABC[0:8]}, vremainder_mask), _mm256_srai_epi32(vacc${ABC[0:8]}, 31));

      vacc${ABC[0:8]} = _mm256_sub_epi32(_mm256_sra_epi32(vacc${ABC[0:8]}, vshift), _mm256_cmpgt_epi32(vrem${ABC[0:8]}, vremainder_threshold));

      $if BATCH_TILE > 8:
        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), _mm256_castsi256_si128(voutput_zero_point));
        vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, _mm256_castsi256_si128(voutput_min)), _mm256_castsi256_si128(voutput_max));
      $else:
        __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point);
        vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max);
      __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});

      $if BATCH_TILE > 8:
        if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
          output += 8;
          n -= 8 * sizeof(int8_t);
        } else {
          if (n & (4 * sizeof(int8_t))) {
            *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
            output += 4;
          }
          if (n & (2 * sizeof(int8_t))) {
            *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (n & (1 * sizeof(int8_t))) {
            *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
          }
          n = 0;
        }
      $else:
        if (n & (4 * sizeof(int8_t))) {
          *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
          output += 4;
        }
        if (n & (2 * sizeof(int8_t))) {
          *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (n & (1 * sizeof(int8_t))) {
          *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}