// Copyright 2022 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
$assert BATCH_TILE % 16 == 0
$SIMD_TILE = BATCH_TILE // 16
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
#include <assert.h>

#include <wasm_simd128.h>

#include <xnnpack/common.h>
#include <xnnpack/vcvt.h>


$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE]
$WASM_X16X8_LOAD8X8 = {"QS8": "wasm_i16x8_load8x8", "QU8": "wasm_u16x8_load8x8"}[DATATYPE]
$WASM_X16X8_EXTEND_LOW_X8X16 = {"QS8": "wasm_i16x8_extend_low_i8x16", "QU8": "wasm_u16x8_extend_low_u8x16"}[DATATYPE]
$WASM_X16X8_EXTEND_HIGH_X8X16 = {"QS8": "wasm_i16x8_extend_high_i8x16", "QU8": "wasm_u16x8_extend_high_u8x16"}[DATATYPE]
$WASM_I16X8_Q15MULR = "__builtin_wasm_relaxed_q15mulr_s_i16x8" if RELAXED else "wasm_i16x8_q15mulr_sat"
$WASM_X8X16_NARROW_I16X8 = {"QS8": "wasm_i8x16_narrow_i16x8", "QU8": "wasm_u8x16_narrow_i16x8"}[DATATYPE]
$ISA = "wasmrelaxedsimd" if RELAXED else "wasmsimd"
void xnn_${DATATYPE.lower()}_vlrelu_ukernel__${ISA}_arm_x${BATCH_TILE}(
    size_t n,
    const ${XINT8_T}* x,
    ${XINT8_T}* y,
    const union xnn_${DATATYPE.lower()}_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(${XINT8_T}) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const v128_t vinput_zero_point = wasm_v128_load64_splat(params->wasmsimd_arm.input_zero_point);
  const v128_t vpositive_multiplier = wasm_v128_load64_splat(params->wasmsimd_arm.positive_multiplier);
  const v128_t vnegative_multiplier = wasm_v128_load64_splat(params->wasmsimd_arm.negative_multiplier);
  const v128_t voutput_zero_point = wasm_v128_load64_splat(params->wasmsimd_arm.output_zero_point);
  for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) {
    v128_t vx${ABC[0]} = wasm_v128_load(x);
    $for N in range(1, SIMD_TILE):
      v128_t vx${ABC[N]} = wasm_v128_load(x + ${N * 16});
    x += ${BATCH_TILE};

    $for N in range(SIMD_TILE):
      v128_t vacc${ABC[2*N]} = wasm_i16x8_sub(vinput_zero_point, ${WASM_X16X8_EXTEND_LOW_X8X16}(vx${ABC[N]}));
      v128_t vacc${ABC[2*N+1]} = wasm_i16x8_sub(vinput_zero_point, ${WASM_X16X8_EXTEND_HIGH_X8X16}(vx${ABC[N]}));
      v128_t vmultiplier${ABC[2*N]} = wasm_i16x8_shr(vacc${ABC[2*N]}, 15);
      v128_t vmultiplier${ABC[2*N+1]} = wasm_i16x8_shr(vacc${ABC[2*N+1]}, 15);

    $for N in range(2*SIMD_TILE):
      vacc${ABC[N]} = wasm_i16x8_shl(vacc${ABC[N]}, 7);
      vmultiplier${ABC[N]} = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier${ABC[N]});

    $for N in range(2*SIMD_TILE):
      vacc${ABC[N]} = ${WASM_I16X8_Q15MULR}(vacc${ABC[N]}, vmultiplier${ABC[N]});

    $for N in range(2*SIMD_TILE):
      vacc${ABC[N]} = wasm_i16x8_add_sat(vacc${ABC[N]}, voutput_zero_point);

    $for N in range(SIMD_TILE):
      const v128_t vy${ABC[N]} = ${WASM_X8X16_NARROW_I16X8}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]});

    wasm_v128_store(y, vy${ABC[0]});
    $for N in range(1, SIMD_TILE):
      wasm_v128_store((y + ${N * 16}), vy${ABC[N]});
    y += ${BATCH_TILE};
  }
  for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) {
    const v128_t vx = ${WASM_X16X8_LOAD8X8}(x);
    v128_t vacc = wasm_i16x8_sub(vinput_zero_point, vx);
    v128_t vmultiplier = wasm_i16x8_shr(vacc, 15);
    vacc = wasm_i16x8_shl(vacc, 7);
    vmultiplier = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier);
    vacc = ${WASM_I16X8_Q15MULR}(vacc, vmultiplier);
    vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);
    x += 8;

    const v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc);
    wasm_v128_store64_lane(y, vy, 0);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(${XINT8_T}));
    assert(n <= 7 * sizeof(${XINT8_T}));

    const v128_t vx = ${WASM_X16X8_LOAD8X8}(x);
    v128_t vacc = wasm_i16x8_sub(vinput_zero_point, vx);
    v128_t vmultiplier = wasm_i16x8_shr(vacc, 15);
    vacc = wasm_i16x8_shl(vacc, 7);
    vmultiplier = wasm_v128_bitselect(vpositive_multiplier, vnegative_multiplier, vmultiplier);
    vacc = ${WASM_I16X8_Q15MULR}(vacc, vmultiplier);
    vacc = wasm_i16x8_add_sat(vacc, voutput_zero_point);

    v128_t vy = ${WASM_X8X16_NARROW_I16X8}(vacc, vacc);
    if (n & (4 * sizeof(${XINT8_T}))) {
      wasm_v128_store32_lane(y, vy, 0);
      vy = wasm_u64x2_shr(vy, 32);
      y += 4;
    }
    if (n & (2 * sizeof(${XINT8_T}))) {
      wasm_v128_store16_lane(y, vy, 0);
      vy = wasm_u32x4_shr(vy, 16);
      y += 2;
    }
    if (n & (1 * sizeof(${XINT8_T}))) {
      wasm_v128_store8_lane(y, vy, 0);
    }
  }
}