// 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 SSE in [2, 4]
$assert not AVX or SSE == 4
$SSE_HEADER = {2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
$assert BATCH_TILE % 8 == 0
$assert BATCH_TILE >= 8
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <${SSE_HEADER}>

#include <xnnpack/vadd.h>


$ISA = "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE]
void xnn_qs8_vaddc_minmax_ukernel__${ISA}_mul16_ld64_x${BATCH_TILE}(
    size_t n,
    const int8_t* input_a,
    const int8_t* input_b,
    int8_t* output,
    const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  const __m128i va_multiplier_lo = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_lo);
  const __m128i va_multiplier_hi = _mm_load_si128((const __m128i*) params->sse2.a_multiplier_hi);
  const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
  const __m128i vshift = _mm_cvtsi32_si128((int) params->sse2.shift);
  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);

  __m128i vbias = _mm_cvtsi32_si128(params->sse2.b_multiplier[0] * (int32_t) *input_b);
  vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
  vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse2.bias));
  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
    $if SSE == 4:
      const __m128i va${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
      $for N in range(8, BATCH_TILE, 8):
        const __m128i va${ABC[N:N+8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) (input_a + ${N})));
    $else:
      __m128i va${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) input_a);
      $for N in range(8, BATCH_TILE, 8):
        __m128i va${ABC[N:N+8]} = _mm_loadl_epi64((const __m128i*) (input_a + ${N}));
    input_a += ${BATCH_TILE};

    $if SSE < 4:
      $for N in range(0, BATCH_TILE, 8):
        va${ABC[N:N+8]} = _mm_srai_epi16(_mm_unpacklo_epi8(va${ABC[N:N+8]}, va${ABC[N:N+8]}), 8);

    $for N in range(0, BATCH_TILE, 8):
      __m128i vaprod${ABC[N:N+8]}hi = _mm_mulhi_epu16(va${ABC[N:N+8]}, va_multiplier_lo);
      const __m128i vaprod${ABC[N:N+8]}lo = _mm_mullo_epi16(va${ABC[N:N+8]}, va_multiplier_lo);

    $for N in range(0, BATCH_TILE, 8):
      vaprod${ABC[N:N+8]}hi = _mm_add_epi16(vaprod${ABC[N:N+8]}hi, _mm_mullo_epi16(va${ABC[N:N+8]}, va_multiplier_hi));

    $for N in range(0, BATCH_TILE, 8):
      vaprod${ABC[N:N+8]}hi = _mm_sub_epi16(vaprod${ABC[N:N+8]}hi, _mm_and_si128(_mm_srai_epi16(va${ABC[N:N+8]}, 15), va_multiplier_lo));

    $for N in range(0, BATCH_TILE, 8):
      __m128i vacc${ABC[N:N+4]} = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod${ABC[N:N+8]}lo, vaprod${ABC[N:N+8]}hi));
      __m128i vacc${ABC[N+4:N+8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod${ABC[N:N+8]}lo, vaprod${ABC[N:N+8]}hi));

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[N:N+4]}, vrounding), vshift);

    $for N in range(0, BATCH_TILE, 8):
      __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[N:N+4]}, vacc${ABC[N+4:N+8]}), voutput_zero_point);

    $for N in range(0, BATCH_TILE, 8):
      vout${ABC[N:N+8]} = _mm_max_epi16(vout${ABC[N:N+8]}, voutput_min);

    $for N in range(0, BATCH_TILE, 8):
      vout${ABC[N:N+8]} = _mm_min_epi16(vout${ABC[N:N+8]}, voutput_max);

    $for N in range(0, BATCH_TILE, 16):
      $if N + 8 < BATCH_TILE:
        const __m128i vout${ABC[N:N+16]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N+8:N+16]});
      $else:
        const __m128i vout${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 ""}{
      $if SSE == 4:
        const __m128i va${ABC[0:8]} = _mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) input_a));
      $else:
        __m128i va${ABC[0:8]} = _mm_loadl_epi64((const __m128i*) input_a);
      $if BATCH_TILE > 8:
        input_a += 8;

      $if SSE < 4:
        va${ABC[0:8]} = _mm_srai_epi16(_mm_unpacklo_epi8(va${ABC[0:8]}, va${ABC[0:8]}), 8);

      __m128i vaprod${ABC[0:8]}hi = _mm_mulhi_epu16(va${ABC[0:8]}, va_multiplier_lo);
      const __m128i vaprod${ABC[0:8]}lo = _mm_mullo_epi16(va${ABC[0:8]}, va_multiplier_lo);

      vaprod${ABC[0:8]}hi = _mm_add_epi16(vaprod${ABC[0:8]}hi, _mm_mullo_epi16(va${ABC[0:8]}, va_multiplier_hi));

      vaprod${ABC[0:8]}hi = _mm_sub_epi16(vaprod${ABC[0:8]}hi, _mm_and_si128(_mm_srai_epi16(va${ABC[0:8]}, 15), va_multiplier_lo));

      __m128i vacc${ABC[0:4]} = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vaprod${ABC[0:8]}lo, vaprod${ABC[0:8]}hi));
      __m128i vacc${ABC[4:8]} = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vaprod${ABC[0:8]}lo, vaprod${ABC[0:8]}hi));

      vacc${ABC[0:4]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[0:4]}, vrounding), vshift);
      vacc${ABC[4:8]} = _mm_sra_epi32(_mm_add_epi32(vacc${ABC[4:8]}, vrounding), vshift);

      __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(vacc${ABC[0:4]}, vacc${ABC[4:8]}), voutput_zero_point);
      vout${ABC[0:8]} = _mm_max_epi16(vout${ABC[0:8]}, voutput_min);
      vout${ABC[0:8]} = _mm_min_epi16(vout${ABC[0:8]}, 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))) {
            $if SSE == 4:
              *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
            $else:
              *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
          }
          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))) {
          $if SSE == 4:
            *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
          $else:
            *output = (int32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
        }
    }${" while (n != 0);" if BATCH_TILE > 8 else ""}
  }
}