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

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


void xnn_f32_prelu_ukernel__wasm_${ROW_TILE}x${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;

  const float vzero = 0.0f;
  do {
    const float* w = weights;
    size_t c = channels;
    $if CHANNEL_TILE > 1:
      for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) {
        $for C in range(CHANNEL_TILE):
          const float vw${ABC[C]} = w[${C}];

        $for M in range(ROW_TILE):
          $for C in range(CHANNEL_TILE):
            float vi${M}x${ABC[C]} = i${M}[${C}];
          i${M} += ${CHANNEL_TILE};

        $for M in range(ROW_TILE):
          $for C in range(CHANNEL_TILE):
            float vacc${M}x${ABC[C]} = __builtin_wasm_max_f32(vi${M}x${ABC[C]}, vzero);
            vi${M}x${ABC[C]} = __builtin_wasm_min_f32(vi${M}x${ABC[C]}, vzero);

        $for M in range(ROW_TILE):
          $for C in range(CHANNEL_TILE):
            vacc${M}x${ABC[C]} += vi${M}x${ABC[C]} * vw${ABC[C]};

        $for M in range(ROW_TILE):
          $for C in range(CHANNEL_TILE):
            o${M}[${C}] = vacc${M}x${ABC[C]};
          o${M} += ${CHANNEL_TILE};

        w += ${CHANNEL_TILE};
      }
      for (; c != 0; c -= sizeof(float)) {
        const float vw = *w++;

        $for M in range(ROW_TILE):
          float vi${M} = *i${M}++;

        $for M in range(ROW_TILE):
          float vacc${M} = __builtin_wasm_max_f32(vi${M}, vzero);
          vi${M} = __builtin_wasm_min_f32(vi${M}, vzero);

        $for M in range(ROW_TILE):
          vacc${M} += vi${M} * vw;

        $for M in range(ROW_TILE):
          *o${M}++ = vacc${M};
      }
    $else:
      do {
        const float vw = *w++;

        $for M in range(ROW_TILE):
          float vi${M} = *i${M}++;

        $for M in range(ROW_TILE):
          float vacc${M} = __builtin_wasm_max_f32(vi${M}, vzero);
          vi${M} = __builtin_wasm_min_f32(vi${M}, vzero);

        $for M in range(ROW_TILE):
          vacc${M} += vi${M} * vw;

        $for M in range(ROW_TILE):
          *o${M}++ = vacc${M};

        c -= sizeof(float);
      } while (c != 0);
    $for M in range(ROW_TILE):
      i${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);
}