// Copyright 2019 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 CHANNEL_TILE % 8 == 0
$assert KERNEL_TILE >= 2
$assert ACCUMULATORS >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

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


void xnn_f16_dwconv_minmax_ukernel_${KERNEL_TILE}p${CHANNEL_TILE}c__fma3${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const void** input,
    const void* weights,
    void* output,
    intptr_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const void* zero,
    const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m256 vmax = _mm256_load_ps(params->avx.max);
  const __m256 vmin = _mm256_load_ps(params->avx.min);

  uint16_t* o = (uint16_t*) output;
  do {
    $for K in range(KERNEL_TILE):
      const uint16_t* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const uint16_t*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const void**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const uint16_t* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m256 vacc${ABC[0:8]}p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
      $for C in range(8, CHANNEL_TILE, 8):
        __m256 vacc${ABC[C:C+8]}p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + ${C})));

      $for K in range(KERNEL_TILE):

        const __m256 vi${K}x${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${K}));
        $for C in range(8, CHANNEL_TILE, 8):
          const __m256 vi${K}x${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i${K} + ${C})));
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vk${K}x${ABC[C:C+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + ${(K + 1) * CHANNEL_TILE + C})));
        $for C in range(0, CHANNEL_TILE, 8):
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc${ABC[C:C+8]}p${K} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}), _MM_FROUND_TO_NEAREST_INT));
          $else:
            vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS}), _MM_FROUND_TO_NEAREST_INT));

      w += ${(KERNEL_TILE + 1) * CHANNEL_TILE};

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:CHANNEL_TILE]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              $for C in range(0, CHANNEL_TILE, 8):
                vacc${ABC[C:C+8]}p${A} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE}), _MM_FROUND_TO_NEAREST_INT));
          $ACC_SLICE *= 2

      $for C in range(0, CHANNEL_TILE, 8):
        __m256 vacc${ABC[C:C+8]} = _mm256_max_ps(vacc${ABC[C:C+8]}p0, vmin);
      $for C in range(0, CHANNEL_TILE, 8):
        vacc${ABC[C:C+8]} = _mm256_min_ps(vacc${ABC[C:C+8]}, vmax);

      _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc${ABC[0:8]}, _MM_FROUND_TO_NEAREST_INT));
      $for C in range(8, CHANNEL_TILE, 8):
        _mm_storeu_si128((__m128i*) (o + ${C}), _mm256_cvtps_ph(vacc${ABC[C:C+8]}, _MM_FROUND_TO_NEAREST_INT));
      o += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 8:
      for (; c >= 8; c -= 8) {
        __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
        $for K in range(KERNEL_TILE):

          const __m256 vi${K}x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${K}));
          i${K} += 8;

          const __m256 vk${K}x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + ${(K + 1) * CHANNEL_TILE})));
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi${K}x01234567, vk${K}x01234567), _MM_FROUND_TO_NEAREST_INT));
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS}), _MM_FROUND_TO_NEAREST_INT));

        w += 8;

        $if ACCUMULATORS > 1:
          // Add up all accumulators to vacc${ABC[0:8]}p0
          $ACC_SLICE = 1
          $while ACC_SLICE < ACCUMULATORS:
            $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
              $if A + ACC_SLICE < ACCUMULATORS:
                vacc01234567p${A} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE}), _MM_FROUND_TO_NEAREST_INT));
            $ACC_SLICE *= 2

        __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
        vacc01234567 = _mm256_min_ps(vacc01234567, vmax);

        _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc01234567, _MM_FROUND_TO_NEAREST_INT));
        o += 8;
      }
    if XNN_UNLIKELY(c != 0) {
      assert(c >= 1);
      assert(c <= 7);

      __m256 vacc01234567p0 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) w));
      $for K in range(KERNEL_TILE):

        const __m256 vi${K}x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i${K}));

        const __m256 vk${K}x01234567 = _mm256_cvtph_ps(_mm_load_si128((const __m128i*) (w + ${(K + 1) * CHANNEL_TILE})));
        $if 1 <= K < ACCUMULATORS:
          __m256 vacc01234567p${K} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vi${K}x01234567, vk${K}x01234567), _MM_FROUND_TO_NEAREST_INT));
        $else:
          vacc01234567p${K % ACCUMULATORS} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS}), _MM_FROUND_TO_NEAREST_INT));

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${ABC[0:8]}p0
        $ACC_SLICE = 1
        $while ACC_SLICE < ACCUMULATORS:
          $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
            $if A + ACC_SLICE < ACCUMULATORS:
              vacc01234567p${A} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE}), _MM_FROUND_TO_NEAREST_INT));
          $ACC_SLICE *= 2

      __m256 vacc01234567 = _mm256_max_ps(vacc01234567p0, vmin);
      vacc01234567 = _mm256_min_ps(vacc01234567, vmax);

      __m128i vh01234567 = _mm256_cvtps_ph(vacc01234567, _MM_FROUND_TO_NEAREST_INT);
      if (c & 4) {
        _mm_storel_epi64((__m128i*) o, vh01234567);
        vh01234567 = _mm_unpackhi_epi64(vh01234567, vh01234567);
        o += 4;
      }
      if (c & 2) {
        _mm_storeu_si32(o, vh01234567);
        vh01234567 = _mm_srli_epi64(vh01234567, 32);
        o += 2;
      }
      if (c & 1) {
        *o = (uint16_t) _mm_extract_epi16(vh01234567, 0);
        o += 1;
      }
    }

    o = (uint16_t*) ((uintptr_t) o + output_increment);
  } while (--output_width != 0);
}