// 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 % 4 == 0
$assert BATCH_TILE >= 4
$SIMD_TILE = BATCH_TILE // 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$assert OP in ["ADD", "DIV", "RDIV", "MAX", "MIN", "MUL", "SUB", "RSUB", "SQRDIFF", "PRELU", "RPRELU"]
#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/common.h"
#include "xnnpack/vbinary.h"


$VOPQ_F32 = {
$  "ADD": lambda x: "vaddq_f32(%s, vb)" % x,
$  "DIV": lambda x: "vdivq_f32(%s, vb)" % x,
$  "RDIV": lambda x: "vdivq_f32(vb, %s)" % x,
$  "MAX": lambda x: "vmaxq_f32(%s, vb)" % x,
$  "MIN": lambda x: "vminq_f32(%s, vb)" % x,
$  "MUL": lambda x: "vmulq_f32(%s, vb)" % x,
$  "SUB": lambda x: "vsubq_f32(%s, vb)" % x,
$  "RSUB": lambda x: "vsubq_f32(vb, %s)" % x,
$  "SQRDIFF": lambda x: "vsubq_f32(%s, vb)" % x,
$  "PRELU": lambda x: "vmulq_f32(%s, vb)" % x,
$  "RPRELU": lambda x: "vmulq_f32(%s, vb)" % x,
$}[OP]
$ISA = "aarch64_neon" if OP in ["DIV", "RDIV"] else "neon"
void xnn_f32_v${OP.lower()}c_ukernel__${ISA}_u${BATCH_TILE}(
    size_t batch,
    const float* input_a,
    const float* input_b,
    float* output,
    const struct xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input_a != NULL);
  assert(input_b != NULL);
  assert(output != NULL);

  const float32x4_t vb = vld1q_dup_f32(input_b);

  $if OP == "RPRELU":
    const uint32x4_t vm = vcltq_s32(vreinterpretq_s32_f32(vb), vmovq_n_s32(0));
  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      $for N in range(SIMD_TILE):
        const float32x4_t va${ABC[N]} = vld1q_f32(input_a); input_a += 4;

      $for N in range(SIMD_TILE):
        float32x4_t vacc${ABC[N]} = ${VOPQ_F32("va" + ABC[N])};

      $if OP == "SQRDIFF":
        $for N in range(SIMD_TILE):
          vacc${ABC[N]} = vmulq_f32(vacc${ABC[N]}, vacc${ABC[N]});
      $elif OP == "PRELU":
        $for N in range(SIMD_TILE):
          const uint32x4_t vm${ABC[N]} = vcltq_s32(vreinterpretq_s32_f32(va${ABC[N]}), vmovq_n_s32(0));

        $for N in range(SIMD_TILE):
          vacc${ABC[N]} = vbslq_f32(vm${ABC[N]}, vacc${ABC[N]}, va${ABC[N]});
      $elif OP == "RPRELU":
        $for N in range(SIMD_TILE):
          vacc${ABC[N]} = vbslq_f32(vm, vacc${ABC[N]}, vb);

      $for N in range(SIMD_TILE):
        vst1q_f32(output, vacc${ABC[N]}); output += 4;
    }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const float32x4_t va = vld1q_f32(input_a); input_a += 4;

    float32x4_t vacc = ${VOPQ_F32("va")};
    $if OP == "SQRDIFF":
      vacc = vmulq_f32(vacc, vacc);
    $elif OP == "PRELU":
      const uint32x4_t vm = vcltq_s32(vreinterpretq_s32_f32(va), vmovq_n_s32(0));
      vacc = vbslq_f32(vm, vacc, va);
    $elif OP == "RPRELU":
      vacc = vbslq_f32(vm, vacc, vb);

    vst1q_f32(output, vacc); output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    const float32x4_t va = vld1q_f32(input_a);

    float32x4_t vacc = ${VOPQ_F32("va")};
    $if OP == "SQRDIFF":
      vacc = vmulq_f32(vacc, vacc);
    $elif OP == "PRELU":
      const uint32x4_t vm = vcltq_s32(vreinterpretq_s32_f32(va), vmovq_n_s32(0));
      vacc = vbslq_f32(vm, vacc, va);
    $elif OP == "RPRELU":
      vacc = vbslq_f32(vm, vacc, vb);

    float32x2_t vacc_lo = vget_low_f32(vacc);
    if (batch & (2 * sizeof(float))) {
      vst1_f32(output, vacc_lo); output += 2;
      vacc_lo = vget_high_f32(vacc);
    }
    if (batch & (1 * sizeof(float))) {
      vst1_lane_f32(output, vacc_lo, 0);
    }
  }
}