// 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 BATCH_TILE % 16 == 0
$assert BATCH_TILE >= 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"]
$assert ACTIVATION in ["LINEAR", "MINMAX"]
#include <assert.h>

#include <immintrin.h>

#include <xnnpack/common.h>
#include <xnnpack/intrinsics-polyfill.h>
#include <xnnpack/vbinary.h>


$_MM512_OP_PS = {
$  "ADD": lambda x, y: "_mm512_add_ps(%s, %s)" % (x, y),
$  "DIV": lambda x, y: "_mm512_div_ps(%s, %s)" % (x, y),
$  "MAX": lambda x, y: "_mm512_max_ps(%s, %s)" % (x, y),
$  "MIN": lambda x, y: "_mm512_min_ps(%s, %s)" % (x, y),
$  "MUL": lambda x, y: "_mm512_mul_ps(%s, %s)" % (x, y),
$  "SUB": lambda x, y: "_mm512_sub_ps(%s, %s)" % (x, y),
$  "SQRDIFF": lambda x, y: "_mm512_sub_ps(%s, %s)" % (x, y),
$}[OP]
$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION]
$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION]
void xnn_f32_v${OP.lower()}${SUFFIX}_ukernel__avx512f_x${BATCH_TILE}(
    size_t n,
    const float* a,
    const float* b,
    float* y,
    const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(a != NULL);
  assert(b != NULL);
  assert(y != NULL);

  $if ACTIVATION == "MINMAX":
    const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
    const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));

  for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
    const __m512 va${ABC[0:16]} = _mm512_loadu_ps(a);
    $for N in range(16, BATCH_TILE, 16):
      const __m512 va${ABC[N:N+16]} = _mm512_loadu_ps(a + ${N});
    a += ${BATCH_TILE};

    const __m512 vb${ABC[0:16]} = _mm512_loadu_ps(b);
    $for N in range(16, BATCH_TILE, 16):
      const __m512 vb${ABC[N:N+16]} = _mm512_loadu_ps(b + ${N});
    b += ${BATCH_TILE};

    $for N in range(0, BATCH_TILE, 16):
      __m512 vy${ABC[N:N+16]} = ${_MM512_OP_PS("va" + ABC[N:N+16], "vb" + ABC[N:N+16])};

    $if OP == "SQRDIFF":
      $for N in range(0, BATCH_TILE, 16):
        vy${ABC[N:N+16]} = _mm512_mul_ps(vy${ABC[N:N+16]}, vy${ABC[N:N+16]});

    $if ACTIVATION == "MINMAX":
      $for N in range(0, BATCH_TILE, 16):
        vy${ABC[N:N+16]} = _mm512_max_ps(vy${ABC[N:N+16]}, vy_min);

      $for N in range(0, BATCH_TILE, 16):
        vy${ABC[N:N+16]} = _mm512_min_ps(vy${ABC[N:N+16]}, vy_max);

    _mm512_storeu_ps(y, vy${ABC[0:16]});
    $for N in range(16, BATCH_TILE, 16):
      _mm512_storeu_ps(y + ${N}, vy${ABC[N:N+16]});
    y += ${BATCH_TILE};
  }
  $if BATCH_TILE > 16:
    for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
      const __m512 va = _mm512_loadu_ps(a);
      a += 16;

      const __m512 vb = _mm512_loadu_ps(b);
      b += 16;

      __m512 vy = ${_MM512_OP_PS("va", "vb")};
      $if OP == "SQRDIFF":
        vy = _mm512_mul_ps(vy, vy);
      $if ACTIVATION == "MINMAX":
        vy = _mm512_max_ps(vy, vy_min);
        vy = _mm512_min_ps(vy, vy_max);
      _mm512_storeu_ps(y, vy);
      y += 16;
    }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 15 * sizeof(float));
    // Prepare mask for valid 32-bit elements (depends on n).
    n >>= 2 /* log2(sizeof(float)) */;
    const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));

    const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
    const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);

    __m512 vy = ${_MM512_OP_PS("va", "vb")};
    $if OP == "SQRDIFF":
      vy = _mm512_mul_ps(vy, vy);
    $if ACTIVATION == "MINMAX":
      vy = _mm512_max_ps(vy, vy_min);
      vy = _mm512_min_ps(vy, vy_max);
    _mm512_mask_storeu_ps(y, vmask, vy);
  }
}