// 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 DATATYPE in ["QS8", "QU8"]
$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>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$_MM256_CVTEPX8_EPI32 = {"QS8": "_mm256_cvtepi8_epi32", "QU8": "_mm256_cvtepu8_epi32"}[DATATYPE]
$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE]
$_MM_MIN_EPX8 = {"QS8": "_mm_min_epi8", "QU8": "_mm_min_epu8"}[DATATYPE]
$_MM_MAX_EPX8 = {"QS8": "_mm_max_epi8", "QU8": "_mm_max_epu8"}[DATATYPE]
void xnn_${DATATYPE.lower()}_vadd_minmax_ukernel__avx2_mul32_ld64_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* input_a,
    const ${XINT8_T}* input_b,
    ${XINT8_T}* output,
    const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  const __m256i vbias = _mm256_load_si256((const __m256i*) params->avx2.bias);
  const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier);
  const __m256i vb_multiplier = _mm256_load_si256((const __m256i*) params->avx2.b_multiplier);
  const __m128i vshift = _mm_load_si128((const __m128i*) params->avx2.shift);
  $if BATCH_TILE > 8:
    const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
  $else:
    const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point);
  const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min);
  const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max);

  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a));
    const __m256i vb${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_b));
    $for N in range(8, BATCH_TILE, 8):
      const __m256i va${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input_a + ${N})));
      const __m256i vb${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input_b + ${N})));
    input_a += ${BATCH_TILE};
    input_b += ${BATCH_TILE};

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

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

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

    $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):
      $if N + 8 < BATCH_TILE:
        __m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vout${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]} = ${_MM_PACKXS_EPI16}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+16]}, voutput_min);
      $else:
        vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        vout${ABC[N:N+16]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+16]}, voutput_max);
      $else:
        vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max);

    $if BATCH_TILE >= 16:
      _mm_storeu_si128((__m128i*) output, vout${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) {
    ${"do " if BATCH_TILE > 8 else ""}{
      const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a));
      const __m256i vb${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_b));
      $if BATCH_TILE > 8:
        input_a += 8;
        input_b += 8;

      __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[0:8]}, va_multiplier));

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

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

      $if BATCH_TILE > 8:
        __m128i vout${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));
      $else:
        __m128i vout${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);
      __m128i vout${ABC[0:8]}${ABC[0:8]} = ${_MM_PACKXS_EPI16}(vout${ABC[0:8]}, vout${ABC[0:8]});
      vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MAX_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_min);
      vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MIN_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max);

      $if BATCH_TILE > 8:
        if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) {
          _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
          output += 8;
          n -= 8 * sizeof(${XINT8_T});
        } else {
          if (n & (4 * sizeof(${XINT8_T}))) {
            _mm_storeu_si32(output, 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(${XINT8_T}))) {
            _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]});
            vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
            output += 2;
          }
          if (n & (1 * sizeof(${XINT8_T}))) {
            *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
          }
          n = 0;
        }
      $else:
        if (n & (4 * sizeof(${XINT8_T}))) {
          _mm_storeu_si32(output, 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(${XINT8_T}))) {
          _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]});
          vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
          output += 2;
        }
        if (n & (1 * sizeof(${XINT8_T}))) {
          *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}