src/math/cvt-f32-f16-wasmsimd.c

// 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 <stddef.h>
#include <stdint.h>

#include <wasm_simd128.h>

#include <xnnpack/math-stubs.h>


void xnn_math_f32_f16_cvt__wasmsimd(
    size_t n,
    const float* input,
    void* output)
{
  assert(n % (8 * sizeof(uint16_t)) == 0);

  const v128_t vscale_to_inf = wasm_f32x4_const_splat(0x1.0p+112f);
  const v128_t vscale_to_zero = wasm_f32x4_const_splat(0x1.0p-110f);
  const v128_t vexp_bias = wasm_i32x4_const_splat(0x07800000);
  const v128_t vexpw_max = wasm_i32x4_const_splat(0x7F800000);
  const v128_t vbias_min = wasm_i32x4_const_splat(0x40008000);
  const v128_t vexph_mask = wasm_i32x4_const_splat(0x7C00);
  const v128_t vmanth_mask = wasm_i32x4_const_splat(0x0FFF);
  const v128_t vnanh = wasm_i16x8_const_splat(0x7E00);

  uint16_t* o = (uint16_t*) output;
  for (; n != 0; n -= 8 * sizeof(uint16_t)) {
    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;
  }
}