// 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 CHANNEL_TILE % 8 == 0
$assert CHANNEL_TILE >= 8
$assert ROW_TILE >= 1
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/math.h>
#include <xnnpack/prelu.h>


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

void xnn_f32_prelu_ukernel__avx_${ROW_TILE}x${CHANNEL_TILE}(
    size_t rows,
    size_t channels,
    const float*restrict input,
    size_t input_stride,
    const float*restrict weights,
    float*restrict output,
    size_t output_stride)
{
  assert(rows != 0);
  assert(channels != 0);
  assert(channels % sizeof(float) == 0);

  const float* i0 = input;
  float* o0 = output;
  $for M in range(1, ROW_TILE):
    const float* i${M} = (const float*) ((uintptr_t) i${M-1} + input_stride);
    float* o${M} = (float*) ((uintptr_t) o${M-1} + output_stride);
    $if M % 2 == 0:
      if XNN_UNPREDICTABLE(rows <= ${M}) {
        i${M} = i${M-1};
        o${M} = o${M-1};
      }
    $else:
      if XNN_UNPREDICTABLE(rows < ${M+1}) {
        i${M} = i${M-1};
        o${M} = o${M-1};
      }

  const size_t input_increment = input_stride * ${ROW_TILE} - channels;
  const size_t output_increment = output_stride * ${ROW_TILE} - channels;

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

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

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vprod${M}x${ABC[C:C+8]} = _mm256_mul_ps(vi${M}x${ABC[C:C+8]}, vw${ABC[C:C+8]});

      $for M in range(ROW_TILE):
        $for C in range(0, CHANNEL_TILE, 8):
          const __m256 vacc${M}x${ABC[C:C+8]} = _mm256_blendv_ps(vi${M}x${ABC[C:C+8]}, vprod${M}x${ABC[C:C+8]}, vi${M}x${ABC[C:C+8]});

      $for M in range(ROW_TILE):
        _mm256_storeu_ps(o${M}, vacc${M}x${ABC[0:8]});
        $for C in range(8, CHANNEL_TILE, 8):
          _mm256_storeu_ps(o${M} + ${C}, vacc${M}x${ABC[C:C+8]});
        o${M} += ${CHANNEL_TILE};
    }
    $if CHANNEL_TILE > 8:
      for (; c >= 8 * sizeof(float); c -= 8 * sizeof(float)) {
        const __m256 vw = _mm256_load_ps(w);
        w += 8;

        $for M in range(ROW_TILE):
          const __m256 vi${M} = _mm256_loadu_ps(i${M});
          i${M} += 8;

        $for M in range(ROW_TILE):
          const __m256 vprod${M} = _mm256_mul_ps(vi${M}, vw);

        $for M in range(ROW_TILE):
          const __m256 vacc${M} = _mm256_blendv_ps(vi${M}, vprod${M}, vi${M});

        $for M in range(ROW_TILE):
          _mm256_storeu_ps(o${M}, vacc${M});
          o${M} += 8;
      }
    if XNN_UNLIKELY(c != 0) {
      assert(c >= 1 * sizeof(float));
      assert(c <= 7 * sizeof(float));
      __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - c));

      const __m256 vw = _mm256_maskload_ps(w, vmask);

      $for M in range(ROW_TILE):
        const __m256 vi${M} = _mm256_maskload_ps(i${M}, vmask);
        i${M} = (const float*) ((uintptr_t) i${M} + c);

      $for M in range(ROW_TILE):
        const __m256 vprod${M} = _mm256_mul_ps(vi${M}, vw);

      $for M in range(ROW_TILE):
        __m256 vacc${M} = _mm256_blendv_ps(vi${M}, vprod${M}, vi${M});

      // _mm256_maskstore_ps(o${M}, vmask, vacc${M}) could be used here, but triggers msan failures (probably an msan bug).
      $for M in range(ROW_TILE):
        __m128 vacc${M}_lo = _mm256_castps256_ps128(vacc${M});
      if (c & (4 * sizeof(float))) {
        $for M in range(ROW_TILE):
          _mm_storeu_ps(o${M}, vacc${M}_lo);

        $for M in range(ROW_TILE):
          vacc${M}_lo = _mm256_extractf128_ps(vacc${M}, 1);

        $for M in range(ROW_TILE):
          o${M} += 4;
      }
      if (c & (2 * sizeof(float))) {
        $for M in range(ROW_TILE):
          _mm_storel_pi((__m64*) o${M}, vacc${M}_lo);

        $for M in range(ROW_TILE):
          vacc${M}_lo = _mm_movehl_ps(vacc${M}_lo, vacc${M}_lo);

        $for M in range(ROW_TILE):
          o${M} += 2;
      }
      if (c & (1 * sizeof(float))) {
        $for M in range(ROW_TILE):
          _mm_store_ss(o${M}, vacc${M}_lo);

        $for M in range(ROW_TILE):
          o${M} += 1;
      }
    }
    $for M in range(ROW_TILE):
      i${M} = (const float*) ((uintptr_t) i${M} + input_increment);
      o${M} = (float*) ((uintptr_t) o${M} + output_increment);
      $if M % 2 == 1:
        if XNN_UNPREDICTABLE(rows < ${ROW_TILE + M+1}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }
      $elif M != 0:
        if XNN_UNPREDICTABLE(rows <= ${ROW_TILE + M}) {
          i${M} = i${M-1};
          o${M} = o${M-1};
        }
    rows = doz(rows, ${ROW_TILE});
  } while (rows != 0);
}