xref: /aosp_15_r20/external/XNNPACK/src/f32-f16-vcvt/gen/vcvt-wasmsimd-x16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/f32-f16-vcvt/wasmsimd.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 <wasm_simd128.h>

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


void xnn_f32_f16_vcvt_ukernel__wasmsimd_x16(
    size_t n,
    const float* input,
    void* output,
    const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(input != NULL);
  assert(output != NULL);

  const v128_t vexp_bias = wasm_v128_load64_splat(params->wasmsimd.exp_bias);
  const v128_t vscale_to_inf = wasm_v128_load64_splat(params->wasmsimd.scale_to_inf);
  const v128_t vexpw_max = wasm_v128_load64_splat(params->wasmsimd.expw_max);
  const v128_t vscale_to_zero = wasm_v128_load64_splat(params->wasmsimd.scale_to_zero);
  const v128_t vbias_min = wasm_v128_load64_splat(params->wasmsimd.bias_min);
  const v128_t vmanth_mask = wasm_v128_load64_splat(params->wasmsimd.manth_mask);
  const v128_t vexph_mask = wasm_v128_load64_splat(params->wasmsimd.exph_mask);
  const v128_t vnanh = wasm_v128_load64_splat(params->wasmsimd.nanh);

  uint16_t* o = (uint16_t*) output;
  for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
    const v128_t vx0 = wasm_v128_load(input);
    const v128_t vx1 = wasm_v128_load(input + 4);
    const v128_t vx2 = wasm_v128_load(input + 8);
    const v128_t vx3 = wasm_v128_load(input + 12);
    input += 16;

    const v128_t vabsx0 = wasm_f32x4_abs(vx0);
    const v128_t vabsx1 = wasm_f32x4_abs(vx1);
    const v128_t vabsx2 = wasm_f32x4_abs(vx2);
    const v128_t vabsx3 = wasm_f32x4_abs(vx3);

    const v128_t vsignx0 = wasm_v128_xor(vx0, vabsx0);
    const v128_t vsignx1 = wasm_v128_xor(vx1, vabsx1);
    const v128_t vsignx2 = wasm_v128_xor(vx2, vabsx2);
    const v128_t vsignx3 = wasm_v128_xor(vx3, vabsx3);

    v128_t vbias0 = wasm_i32x4_add(vabsx0, vexp_bias);
    v128_t vbias1 = wasm_i32x4_add(vabsx1, vexp_bias);
    v128_t vbias2 = wasm_i32x4_add(vabsx2, vexp_bias);
    v128_t vbias3 = wasm_i32x4_add(vabsx3, vexp_bias);

    v128_t vf0 = wasm_f32x4_mul(vabsx0, vscale_to_inf);
    v128_t vf1 = wasm_f32x4_mul(vabsx1, vscale_to_inf);
    v128_t vf2 = wasm_f32x4_mul(vabsx2, vscale_to_inf);
    v128_t vf3 = wasm_f32x4_mul(vabsx3, vscale_to_inf);

    const v128_t vnanmaskw0 = wasm_i32x4_gt(vabsx0, vexpw_max);
    const v128_t vnanmaskw1 = wasm_i32x4_gt(vabsx1, vexpw_max);
    const v128_t vnanmaskw2 = wasm_i32x4_gt(vabsx2, vexpw_max);
    const v128_t vnanmaskw3 = wasm_i32x4_gt(vabsx3, vexpw_max);

    vbias0 = wasm_v128_and(vbias0, vexpw_max);
    vbias1 = wasm_v128_and(vbias1, vexpw_max);
    vbias2 = wasm_v128_and(vbias2, vexpw_max);
    vbias3 = wasm_v128_and(vbias3, vexpw_max);

    vf0 = wasm_f32x4_mul(vf0, vscale_to_zero);
    vf1 = wasm_f32x4_mul(vf1, vscale_to_zero);
    vf2 = wasm_f32x4_mul(vf2, vscale_to_zero);
    vf3 = wasm_f32x4_mul(vf3, vscale_to_zero);

    const v128_t vnanmaskh0 = wasm_i16x8_narrow_i32x4(vnanmaskw0, vnanmaskw1);
    const v128_t vnanmaskh1 = wasm_i16x8_narrow_i32x4(vnanmaskw2, vnanmaskw3);

    const v128_t vsignh0 = wasm_i16x8_narrow_i32x4(vsignx0, vsignx1);
    const v128_t vsignh1 = wasm_i16x8_narrow_i32x4(vsignx2, vsignx3);

    vbias0 = wasm_i16x8_max(vbias0, vbias_min);
    vbias1 = wasm_i16x8_max(vbias1, vbias_min);
    vbias2 = wasm_i16x8_max(vbias2, vbias_min);
    vbias3 = wasm_i16x8_max(vbias3, vbias_min);

    vf0 = wasm_f32x4_add(vf0, vbias0);
    vf1 = wasm_f32x4_add(vf1, vbias1);
    vf2 = wasm_f32x4_add(vf2, vbias2);
    vf3 = wasm_f32x4_add(vf3, vbias3);

    v128_t vexpw0 = wasm_i32x4_shr(vf0, 13);
    v128_t vexpw1 = wasm_i32x4_shr(vf1, 13);
    v128_t vexpw2 = wasm_i32x4_shr(vf2, 13);
    v128_t vexpw3 = wasm_i32x4_shr(vf3, 13);

    const v128_t vmantw0 = wasm_v128_and(vf0, vmanth_mask);
    const v128_t vmantw1 = wasm_v128_and(vf1, vmanth_mask);
    const v128_t vmantw2 = wasm_v128_and(vf2, vmanth_mask);
    const v128_t vmantw3 = wasm_v128_and(vf3, vmanth_mask);

    vexpw0 = wasm_v128_and(vexpw0, vexph_mask);
    vexpw1 = wasm_v128_and(vexpw1, vexph_mask);
    vexpw2 = wasm_v128_and(vexpw2, vexph_mask);
    vexpw3 = wasm_v128_and(vexpw3, vexph_mask);

    const v128_t vnonsignw0 = wasm_i32x4_add(vmantw0, vexpw0);
    const v128_t vnonsignw1 = wasm_i32x4_add(vmantw1, vexpw1);
    const v128_t vnonsignw2 = wasm_i32x4_add(vmantw2, vexpw2);
    const v128_t vnonsignw3 = wasm_i32x4_add(vmantw3, vexpw3);

    const v128_t vnonsignh0 = wasm_i16x8_narrow_i32x4(vnonsignw0, vnonsignw1);
    const v128_t vnonsignh1 = wasm_i16x8_narrow_i32x4(vnonsignw2, vnonsignw3);

    const v128_t vabsh0 = wasm_v128_bitselect(vnanh, vnonsignh0, vnanmaskh0);
    const v128_t vabsh1 = wasm_v128_bitselect(vnanh, vnonsignh1, vnanmaskh1);

    const v128_t vh0 = wasm_v128_or(vabsh0, vsignh0);
    const v128_t vh1 = wasm_v128_or(vabsh1, vsignh1);

    wasm_v128_store(o, vh0);
    wasm_v128_store(o + 8, vh1);
    o += 16;
  }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    const v128_t vx_lo = wasm_v128_load(input);
    const v128_t vx_hi = wasm_v128_load(input + 4);
    input += 8;

    const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
    const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);

    const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
    const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
    v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
    v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
    v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
    v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
    const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
    const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);

    vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
    vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
    vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
    vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
    const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
    const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);

    vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
    vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);

    vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
    vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);

    v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
    v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
    const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
    const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);

    vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
    vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);

    const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
    const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);

    const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);

    const v128_t vabsh = wasm_v128_bitselect(vnanh, vnonsignh, vnanmaskh);

    const v128_t vh = wasm_v128_or(vabsh, vsignh);

    wasm_v128_store(o, vh);
    o += 8;
  }
  if XNN_UNPREDICTABLE(n != 0) {
    const v128_t vx_lo = wasm_v128_load(input);
    const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float))));
    const v128_t vx_hi = wasm_v128_load(input_hi);

    const v128_t vabsx_lo = wasm_f32x4_abs(vx_lo);
    const v128_t vabsx_hi = wasm_f32x4_abs(vx_hi);

    const v128_t vsignx_lo = wasm_v128_xor(vx_lo, vabsx_lo);
    const v128_t vsignx_hi = wasm_v128_xor(vx_hi, vabsx_hi);
    v128_t vbias_lo = wasm_i32x4_add(vabsx_lo, vexp_bias);
    v128_t vbias_hi = wasm_i32x4_add(vabsx_hi, vexp_bias);
    v128_t vf_lo = wasm_f32x4_mul(vabsx_lo, vscale_to_inf);
    v128_t vf_hi = wasm_f32x4_mul(vabsx_hi, vscale_to_inf);
    const v128_t vnanmaskw_lo = wasm_i32x4_gt(vabsx_lo, vexpw_max);
    const v128_t vnanmaskw_hi = wasm_i32x4_gt(vabsx_hi, vexpw_max);

    vbias_lo = wasm_v128_and(vbias_lo, vexpw_max);
    vbias_hi = wasm_v128_and(vbias_hi, vexpw_max);
    vf_lo = wasm_f32x4_mul(vf_lo, vscale_to_zero);
    vf_hi = wasm_f32x4_mul(vf_hi, vscale_to_zero);
    const v128_t vnanmaskh = wasm_i16x8_narrow_i32x4(vnanmaskw_lo, vnanmaskw_hi);
    const v128_t vsignh = wasm_i16x8_narrow_i32x4(vsignx_lo, vsignx_hi);

    vbias_lo = wasm_i16x8_max(vbias_lo, vbias_min);
    vbias_hi = wasm_i16x8_max(vbias_hi, vbias_min);

    vf_lo = wasm_f32x4_add(vf_lo, vbias_lo);
    vf_hi = wasm_f32x4_add(vf_hi, vbias_hi);

    v128_t vexpw_lo = wasm_i32x4_shr(vf_lo, 13);
    v128_t vexpw_hi = wasm_i32x4_shr(vf_hi, 13);
    const v128_t vmantw_lo = wasm_v128_and(vf_lo, vmanth_mask);
    const v128_t vmantw_hi = wasm_v128_and(vf_hi, vmanth_mask);

    vexpw_lo = wasm_v128_and(vexpw_lo, vexph_mask);
    vexpw_hi = wasm_v128_and(vexpw_hi, vexph_mask);

    const v128_t vnonsignw_lo = wasm_i32x4_add(vmantw_lo, vexpw_lo);
    const v128_t vnonsignw_hi = wasm_i32x4_add(vmantw_hi, vexpw_hi);

    const v128_t vnonsignh = wasm_i16x8_narrow_i32x4(vnonsignw_lo, vnonsignw_hi);

    const v128_t vabsh = wasm_v128_bitselect(vnanh, vnonsignh, vnanmaskh);

    v128_t vh = wasm_v128_or(vabsh, vsignh);

    if (n & (4 * sizeof(float))) {
      *((double*) o) = wasm_f64x2_extract_lane(vh, 0);
      vh = wasm_v64x2_shuffle(vh, vh, 1, 1);
      o += 4;
    }
    if (n & (2 * sizeof(float))) {
      *((float*) o) = (float) wasm_f32x4_extract_lane(vh, 0);
      vh = wasm_i64x2_shr(vh, 32);
      o += 2;
    }
    const uint32_t vh_lo = wasm_i32x4_extract_lane(vh, 0);
    if (n & (1 * sizeof(float))) {
      *o = (uint16_t) vh_lo;
    }
  }
}