// 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>


static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

$ISA = {0: "avx", 3: "fma3"}[FMA]
void xnn_f32_dwconv_minmax_ukernel_up${CHANNEL_TILE}x${KERNEL_TILE}__${ISA}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
    size_t channels,
    size_t output_width,
    const float** input,
    const float* weights,
    float* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const float* zero,
    const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(channels != 0);
  assert(output_width != 0);

  const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
  const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
  do {
    $for K in range(KERNEL_TILE):
      const float* i${K} = input[${K}];
      assert(i${K} != NULL);
      if XNN_UNPREDICTABLE(i${K} != zero) {
        i${K} = (const float*) ((uintptr_t) i${K} + input_offset);
      }
    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const float* w = weights;
    for (; c >= ${CHANNEL_TILE}; c -= ${CHANNEL_TILE}) {
      __m256 vacc${ABC[0:8]}p0 = _mm256_load_ps(w);
      $for C in range(8, CHANNEL_TILE, 8):
        __m256 vacc${ABC[C:C+8]}p0 = _mm256_load_ps(w + ${C});

      $for K in range(KERNEL_TILE):

        const __m256 vi${K}x${ABC[0:8]} = _mm256_loadu_ps(i${K});
        $for C in range(8, CHANNEL_TILE, 8):
          const __m256 vi${K}x${ABC[C:C+8]} = _mm256_loadu_ps(i${K} + ${C});
        i${K} += ${CHANNEL_TILE};

        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vk${K}x${ABC[C:C+8]} = _mm256_load_ps(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${K} = _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]});
          $elif FMA == 3:
            vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}, vacc${ABC[C:C+8]}p${K % ACCUMULATORS});
          $else:
            vacc${ABC[C:C+8]}p${K % ACCUMULATORS} = _mm256_add_ps(vacc${ABC[C:C+8]}p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x${ABC[C:C+8]}, vk${K}x${ABC[C:C+8]}));

      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${A} = _mm256_add_ps(vacc${ABC[C:C+8]}p${A}, vacc${ABC[C:C+8]}p${A + ACC_SLICE});
          $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);

      _mm256_storeu_ps(output, vacc${ABC[0:8]});
      $for C in range(8, CHANNEL_TILE, 8):
        _mm256_storeu_ps(output + ${C}, vacc${ABC[C:C+8]});
      output += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 8:
      for (; c >= 8; c -= 8) {
        __m256 vacc01234567p0 = _mm256_load_ps(w);
        $for K in range(KERNEL_TILE):

          const __m256 vi${K}x01234567 = _mm256_loadu_ps(i${K});
          i${K} += 8;

          const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE});
          $if 1 <= K < ACCUMULATORS:
            __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
          $elif FMA == 3:
            vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
          $else:
            vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

        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_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
            $ACC_SLICE *= 2

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

        _mm256_storeu_ps(output, vacc01234567);
        output += 8;
      }
    if XNN_UNLIKELY(c != 0) {
      assert(c >= 1);
      assert(c <= 7);
      __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[7 - c]);

      __m256 vacc01234567p0 = _mm256_load_ps(w);
      $for K in range(KERNEL_TILE):

        const __m256 vi${K}x01234567 = _mm256_maskload_ps(i${K}, vmask);
        const __m256 vk${K}x01234567 = _mm256_load_ps(w + ${(K + 1) * CHANNEL_TILE});
        $if 1 <= K < ACCUMULATORS:
          __m256 vacc01234567p${K} = _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567);
        $elif FMA == 3:
          vacc01234567p${K % ACCUMULATORS} = _mm256_fmadd_ps(vi${K}x01234567, vk${K}x01234567, vacc01234567p${K % ACCUMULATORS});
        $else:
          vacc01234567p${K % ACCUMULATORS} = _mm256_add_ps(vacc01234567p${K % ACCUMULATORS}, _mm256_mul_ps(vi${K}x01234567, vk${K}x01234567));

      $if ACCUMULATORS > 1:
        // Add up all accumulators to vacc${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_add_ps(vacc01234567p${A}, vacc01234567p${A + ACC_SLICE});
          $ACC_SLICE *= 2

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

      // _mm256_maskstore_ps(output, vmask, vacc01234567); output += c; could be used here, but triggers msan failures (probably an msan bug).
      __m128 vacc0123 = _mm256_castps256_ps128(vacc01234567);
      if (c & 4) {
        _mm_storeu_ps(output, vacc0123);
        vacc0123 = _mm256_extractf128_ps(vacc01234567, 1);
        output += 4;
      }
      if (c & 2) {
        _mm_storel_pi((__m64*) output, vacc0123);
        vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
        output += 2;
      }
      if (c & 1) {
        _mm_store_ss(output, vacc0123);
        output += 1;
      }
    }

    output = (float*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}