// Auto-generated file. Do not edit!
//   Template: src/f32-qs8-vcvt/wasmsimd-magic.c.in
//   Generator: tools/xngen
//
// Copyright 2021 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.

#include <assert.h>

#include <arm_neon.h>

#include "xnnpack/common.h"
#include "xnnpack/intrinsics-polyfill.h"
#include "xnnpack/math.h"
#include "xnnpack/vcvt.h"


void xnn_f32_qs8_vcvt_ukernel__wasmsimd_magic_u8(
    size_t batch,
    const float* input,
    int8_t* output,
    const struct xnn_f32_qs8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(batch != 0);
  assert(batch % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const float output_min_less_zero_point = (float) ((int32_t) -128 - (int32_t) params->scalar.output_zero_point);
  const v128_t vscale = wasm_v128_load32_splat(&params->scalar.scale);
  const v128_t vmagic_bias = wasm_f32x4_const_splat(12582912.0f);
  const v128_t vmagic_min = wasm_u32x4_splat(float_as_uint32(12582912.0f + output_min_less_zero_point));
  const v128_t vmagic_bias_less_zero_point = wasm_i32x4_splat(INT32_C(0x4B400000) - (int32_t) params->scalar.output_zero_point);
  const v128_t voutput_max = wasm_u8x16_const_splat(127);
  XNN_FORCE_REALIZATION(vscale);
  XNN_FORCE_REALIZATION(vmagic_bias);
  XNN_FORCE_REALIZATION(vmagic_min);
  XNN_FORCE_REALIZATION(vmagic_bias_less_zero_point);
  XNN_FORCE_REALIZATION(voutput_max);
  for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) {
    v128_t vx_lo = wasm_v128_load(input);
    v128_t vx_hi = wasm_v128_load(input + 4);
    input += 8;

    vx_lo = wasm_f32x4_mul(vx_lo, vscale);
    vx_hi = wasm_f32x4_mul(vx_hi, vscale);

    vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
    vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);

    v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
    v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);

    vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
    vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);

    const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);

    v128_t vy = wasm_i8x16_narrow_i16x8(vacc, vacc);
    vy = wasm_i8x16_min(vy, voutput_max);
    wasm_v128_store64_lane(output, vy, 0);
    output += 8;
  }
  if XNN_UNLIKELY(batch != 0) {
    assert(batch >= 1 * sizeof(float));
    assert(batch <= 7 * sizeof(float));
    v128_t vx_lo = wasm_v128_load(input);
    const float* x_hi = (const float*) ((uintptr_t) input + (batch & (4 * sizeof(float))));
    v128_t vx_hi = wasm_v128_load(x_hi);

    vx_lo = wasm_f32x4_mul(vx_lo, vscale);
    vx_hi = wasm_f32x4_mul(vx_hi, vscale);

    vx_lo = wasm_f32x4_add(vx_lo, vmagic_bias);
    vx_hi = wasm_f32x4_add(vx_hi, vmagic_bias);

    v128_t vacc_lo = wasm_i32x4_max(vx_lo, vmagic_min);
    v128_t vacc_hi = wasm_i32x4_max(vx_hi, vmagic_min);

    vacc_lo = wasm_i32x4_sub(vacc_lo, vmagic_bias_less_zero_point);
    vacc_hi = wasm_i32x4_sub(vacc_hi, vmagic_bias_less_zero_point);

    const v128_t vacc = wasm_i16x8_narrow_i32x4(vacc_lo, vacc_hi);

    v128_t vy = wasm_i8x16_narrow_i16x8(vacc, vacc);
    vy = wasm_i8x16_min(vy, voutput_max);

    if (batch & (4 * sizeof(float))) {
      wasm_v128_store32_lane(output, vy, 0);
      vy = wasm_u64x2_shr(vy, 32);
      output += 4;
    }
    if (batch & (2 * sizeof(float))) {
      wasm_v128_store16_lane(output, vy, 0);
      vy = wasm_u32x4_shr(vy, 16);
      output += 2;
    }
    if (batch & (1 * sizeof(float))) {
      wasm_v128_store8_lane(output, vy, 0);
    }
  }
}