// 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>

$if SSE == 5:
  #ifdef __GNUC__
    #include <x86intrin.h>
  #else
    #include <immintrin.h>
    #include <ammintrin.h>
  #endif
$else:
  #include <immintrin.h>

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


$ISA = {4: "sse41", 5: "xop"}[SSE]
void xnn_qs8_vadd_minmax_ukernel__${ISA}_mul32_ld32_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 __m128i vzero_point_product = _mm_load_si128((const __m128i*) params->sse2.zero_point_product);
  const __m128i vx_multiplier = _mm_load_si128((const __m128i*) params->sse2.x_multiplier);
  const __m128i vy_multiplier = _mm_load_si128((const __m128i*) params->sse2.y_multiplier);
  const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
  const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
  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);

  for (; n >= ${BATCH_TILE} * sizeof(int8_t); n -= ${BATCH_TILE} * sizeof(int8_t)) {
    const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
    const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
    $for N in range(4, BATCH_TILE, 4):
      const __m128i vx${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + ${N}));
      const __m128i vy${ABC[N:N+4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + ${N}));
    input_x += ${BATCH_TILE};
    input_y += ${BATCH_TILE};

    $if SSE == 5:
      $for N in range(0, BATCH_TILE, 4):
        __m128i vacc${ABC[N:N+4]} = _mm_macc_epi32(vx${ABC[N:N+4]}, vx_multiplier, vzero_point_product);

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = _mm_macc_epi32(vy${ABC[N:N+4]}, vy_multiplier, vacc${ABC[N:N+4]});
    $else:
      $for N in range(0, BATCH_TILE, 4):
        __m128i vacc${ABC[N:N+4]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[N:N+4]}, vx_multiplier));

      $for N in range(0, BATCH_TILE, 4):
        vacc${ABC[N:N+4]} = _mm_add_epi32(vacc${ABC[N:N+4]}, _mm_mullo_epi32(vy${ABC[N:N+4]}, vy_multiplier));

    $for N in range(0, BATCH_TILE, 4):
      const __m128i vrem${ABC[N:N+4]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[N:N+4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[N:N+4]}));

    $for N in range(0, BATCH_TILE, 4):
      vacc${ABC[N:N+4]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[N:N+4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[N:N+4]}, vremainder_threshold));

    $for N in range(0, BATCH_TILE, 8):
      __m128i vout${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) {
    ${"do " if BATCH_TILE > 8 else ""}{
      const __m128i vx${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x));
      const __m128i vy${ABC[0:4]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y));
      const __m128i vx${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_x + 4));
      const __m128i vy${ABC[4:8]} = _mm_cvtepi8_epi32(_mm_loadu_si32(input_y + 4));
      $if BATCH_TILE > 8:
        input_x += 8;
        input_y += 8;

      $if SSE == 5:
        __m128i vacc${ABC[0:4]} = _mm_macc_epi32(vx${ABC[0:4]}, vx_multiplier, vzero_point_product);
        __m128i vacc${ABC[4:8]} = _mm_macc_epi32(vx${ABC[4:8]}, vx_multiplier, vzero_point_product);

        vacc${ABC[0:4]} = _mm_macc_epi32(vy${ABC[0:4]}, vy_multiplier, vacc${ABC[0:4]});
        vacc${ABC[4:8]} = _mm_macc_epi32(vy${ABC[4:8]}, vy_multiplier, vacc${ABC[4:8]});
      $else:
        __m128i vacc${ABC[0:4]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[0:4]}, vx_multiplier));
        __m128i vacc${ABC[4:8]} = _mm_add_epi32(vzero_point_product, _mm_mullo_epi32(vx${ABC[4:8]}, vx_multiplier));

        vacc${ABC[0:4]} = _mm_add_epi32(vacc${ABC[0:4]}, _mm_mullo_epi32(vy${ABC[0:4]}, vy_multiplier));
        vacc${ABC[4:8]} = _mm_add_epi32(vacc${ABC[4:8]}, _mm_mullo_epi32(vy${ABC[4:8]}, vy_multiplier));

      const __m128i vrem${ABC[0:4]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[0:4]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[0:4]}));
      const __m128i vrem${ABC[4:8]} = _mm_add_epi32(_mm_and_si128(vacc${ABC[4:8]}, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vacc${ABC[4:8]}));

      vacc${ABC[0:4]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[0:4]}, vshift), _mm_cmpgt_epi32(vrem${ABC[0:4]}, vremainder_threshold));
      vacc${ABC[4:8]} = _mm_sub_epi32(_mm_sra_epi32(vacc${ABC[4:8]}, vshift), _mm_cmpgt_epi32(vrem${ABC[4:8]}, vremainder_threshold));

      __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 ""}
  }
}