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

#include <wasm_simd128.h>

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


$WASM_V32X4_LANESELECT = "wasm_i32x4_relaxed_laneselect" if RELAXED else "wasm_v128_bitselect"
$WASM_F32X4_OP = {
$  "ADD": "wasm_f32x4_add",
$  "DIV": "wasm_f32x4_div",
$  "MAX": "wasm_f32x4_pmax" if X86 else "wasm_f32x4_max",
$  "MIN": "wasm_f32x4_pmin" if X86 else "wasm_f32x4_min",
$  "MUL": "wasm_f32x4_mul",
$  "SUB": "wasm_f32x4_sub",
$  "SQRDIFF": "wasm_f32x4_sub",
$  "PRELU": "wasm_f32x4_mul",
$}[OP]
$ARCH_SUFFIX = "" if OP not in ["MIN", "MAX"] else "_x86" if X86 else "_arm"
$RELAXED_SUFFIX = "relaxed" if RELAXED else ""
void xnn_f32_v${OP.lower()}_ukernel__wasm${RELAXED_SUFFIX}simd${ARCH_SUFFIX}_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);

  $if BATCH_TILE > 4:
    for (; batch >= ${BATCH_TILE} * sizeof(float); batch -= ${BATCH_TILE} * sizeof(float)) {
      const v128_t va${ABC[0]} = wasm_v128_load(input_a);
      $for N in range(1, SIMD_TILE):
        const v128_t va${ABC[N]} = wasm_v128_load(input_a + ${N * 4});
      input_a += ${BATCH_TILE};

      const v128_t vb${ABC[0]} = wasm_v128_load(input_b);
      $for N in range(1, SIMD_TILE):
        const v128_t vb${ABC[N]} = wasm_v128_load(input_b + ${N * 4});
      input_b += ${BATCH_TILE};

      $for N in range(SIMD_TILE):
        v128_t vacc${ABC[N]} = ${WASM_F32X4_OP}(va${ABC[N]}, vb${ABC[N]});

      $if OP == "SQRDIFF":
        $for N in range(SIMD_TILE):
          vacc${ABC[N]} = wasm_f32x4_mul(vacc${ABC[N]}, vacc${ABC[N]});
      $elif OP == "PRELU":
        $for N in range(SIMD_TILE):
          const v128_t vmask${ABC[N]} = wasm_i32x4_shr(va${ABC[N]}, 31);

        $for N in range(SIMD_TILE):
          vacc${ABC[N]} = ${WASM_V32X4_LANESELECT}(vacc${ABC[N]}, va${ABC[N]}, vmask${ABC[N]});

      wasm_v128_store(output, vacc${ABC[0]});
      $for N in range(1, SIMD_TILE):
        wasm_v128_store(output + ${N * 4}, vacc${ABC[N]});
      output += ${BATCH_TILE};
    }
  for (; batch >= 4 * sizeof(float); batch -= 4 * sizeof(float)) {
    const v128_t va = wasm_v128_load(input_a);
    input_a += 4;

    const v128_t vb = wasm_v128_load(input_b);
    input_b += 4;

    v128_t vacc = ${WASM_F32X4_OP}(va, vb);
    $if OP == "SQRDIFF":
      vacc = wasm_f32x4_mul(vacc, vacc);
    $elif OP == "PRELU":
      const v128_t vmask = wasm_i32x4_shr(va, 31);

      vacc = ${WASM_V32X4_LANESELECT}(vacc, va, vmask);

    wasm_v128_store(output, vacc);
    output += 4;
  }
  if XNN_UNLIKELY(batch != 0) {
    const v128_t va = wasm_v128_load(input_a);
    const v128_t vb = wasm_v128_load(input_b);

    v128_t vacc = ${WASM_F32X4_OP}(va, vb);
    $if OP == "SQRDIFF":
      vacc = wasm_f32x4_mul(vacc, vacc);
    $elif OP == "PRELU":
      const v128_t vmask = wasm_i32x4_shr(va, 31);

      vacc = ${WASM_V32X4_LANESELECT}(vacc, va, vmask);

    if (batch & (2 * sizeof(float))) {
      wasm_v128_store64_lane(output, vacc, 0);
      vacc = wasm_v64x2_shuffle(vacc, vacc, 1, 1);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      wasm_v128_store32_lane(output, vacc, 0);
    }
  }
}