VECTOR_WIDTH Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/sve/vec_float.h lines 30–523

template <>
class Vectorized<float> {
 private:
  vls_float32_t values;

 public:
  using value_type = float;
  using size_type = int;
  static constexpr size_type size() {
    return VECTOR_WIDTH / sizeof(float);
  }
  Vectorized() {
    values = svdup_n_f32(0);
  }
  Vectorized(svfloat32_t v) : values(v) {}
  Vectorized(float val) {
    values = svdup_n_f32(val);
  }
  template <
      typename... Args,
      typename = std::enable_if_t<(sizeof...(Args) == size())>>
  Vectorized(Args... vals) {
    __at_align__ float buffer[size()] = {vals...};
    values = svld1_f32(ptrue, buffer);
  }
  operator svfloat32_t() const {
    return values;
  }
  template <uint64_t mask>
  static Vectorized<float> blend(
      const Vectorized<float>& a,
      const Vectorized<float>& b) {
    // Build an array of flags: each element is 1 if the corresponding bit in
    // 'mask' is set, 0 otherwise.
    __at_align__ int32_t flag_arr[size()];
    for (int i = 0; i < size(); i++) {
      flag_arr[i] = (mask & (1ULL << i)) ? 1 : 0;
    }
    // Load the flag array into an SVE int32 vector.
    svint32_t int_mask = svld1_s32(svptrue_b32(), flag_arr);
    // Compare each lane of int_mask to 0; returns an svbool_t predicate where
    // true indicates a nonzero flag.
    svbool_t blend_mask = svcmpne_n_s32(svptrue_b32(), int_mask, 0);
    // Use svsel to select elements from b where the predicate is true, else
    // from a.
    svfloat32_t result = svsel_f32(blend_mask, b.values, a.values);
    return Vectorized<float>(result);
  }
  static Vectorized<float> blendv(
      const Vectorized<float>& a,
      const Vectorized<float>& b,
      const Vectorized<float>& mask_) {
    svbool_t mask =
        svcmpeq_s32(ptrue, svreinterpret_s32_f32(mask_), ALL_S32_TRUE_MASK);
    return svsel_f32(mask, b, a);
  }
  template <typename step_t>
  static Vectorized<float> arange(
      float base = 0.f,
      step_t step = static_cast<step_t>(1)) {
    __at_align__ float buffer[size()];
    for (int64_t i = 0; i < size(); i++) {
      buffer[i] = base + i * step;
    }
    return svld1_f32(ptrue, buffer);
  }
  static Vectorized<float> set(
      const Vectorized<float>& a,
      const Vectorized<float>& b,
      int64_t count = size()) {
    if (count == 0) {
      return a;
    } else if (count < size()) {
      return svsel_f32(svwhilelt_b32(0ull, count), b, a);
    }
    return b;
  }
  static Vectorized<float> loadu(const void* ptr, int64_t count = size()) {
    if (count == size())
      return svld1_f32(ptrue, reinterpret_cast<const float*>(ptr));
    svbool_t pg = svwhilelt_b32(0ull, count);
    return svld1_f32(pg, reinterpret_cast<const float*>(ptr));
  }
  void store(void* ptr, int64_t count = size()) const {
    if (count == size()) {
      svst1_f32(ptrue, reinterpret_cast<float*>(ptr), values);
    } else {
      svbool_t pg = svwhilelt_b32(0ull, count);
      svst1_f32(pg, reinterpret_cast<float*>(ptr), values);
    }
  }
  const float& operator[](int idx) const = delete;
  float& operator[](int idx) = delete;
  int64_t zero_mask() const {
    // returns an integer mask where all zero elements are translated to 1-bit
    // and others are translated to 0-bit
    int64_t mask = 0;
    __at_align__ int32_t mask_array[size()];

    svbool_t svbool_mask = svcmpeq_f32(ptrue, values, ZERO_F32);
    svst1_s32(
        ptrue,
        mask_array,
        svsel_s32(svbool_mask, ALL_S32_TRUE_MASK, ALL_S32_FALSE_MASK));
    for (int64_t i = 0; i < size(); ++i) {
      if (mask_array[i])
        mask |= (1ull << i);
    }
    return mask;
  }
  Vectorized<float> isnan() const {
    // NaN check
    svbool_t mask = svcmpuo_f32(ptrue, values, ZERO_F32);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }
  bool has_inf_nan() const {
    return svptest_any(
        ptrue,
        svcmpuo_f32(ptrue, svsub_f32_x(ptrue, values, values), ZERO_F32));
  }
  Vectorized<float> map(float (*f)(float)) const {
    __at_align__ float tmp[size()];
    store(tmp);
    for (int64_t i = 0; i < size(); ++i) {
      tmp[i] = f(tmp[i]);
    }
    return loadu(tmp);
  }
  Vectorized<float> abs() const {
    return svabs_f32_x(ptrue, values);
  }
  Vectorized<float> angle() const {
    const auto nan_vec = svdup_n_f32(NAN);
    const auto nan_mask = svcmpuo_f32(ptrue, values, ZERO_F32);
    const auto pi = svdup_n_f32(c10::pi<float>);

    const auto neg_mask = svcmplt_f32(ptrue, values, ZERO_F32);
    auto angle = svsel_f32(neg_mask, pi, ZERO_F32);
    angle = svsel_f32(nan_mask, nan_vec, angle);
    return angle;
  }
  Vectorized<float> real() const {
    return values;
  }
  Vectorized<float> imag() const {
    return Vectorized<float>(0.f);
  }
  Vectorized<float> conj() const {
    return values;
  }
  Vectorized<float> acos() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_acosfx_u10sve(values)), map(std::acos));
  }
  Vectorized<float> acosh() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_acoshfx_u10sve(values)), map(std::acosh));
  }
  Vectorized<float> asin() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_asinfx_u10sve(values)), map(std::asin));
  }
  Vectorized<float> asinh() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_asinhfx_u10sve(values)), map(std::asinh));
  }
  Vectorized<float> atan() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_atanfx_u10sve(values)), map(std::atan));
  }
  Vectorized<float> atanh() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_atanhfx_u10sve(values)), map(std::atanh));
  }
  Vectorized<float> atan2(const Vectorized<float>& b) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_atan2fx_u10sve(values, b)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_b[size()];
          store(tmp);
          b.store(tmp_b);
          for (int64_t i = 0; i < size(); i++) {
            tmp[i] = std::atan2(tmp[i], tmp_b[i]);
          }
          return loadu(tmp);
        });
  }
  Vectorized<float> copysign(const Vectorized<float>& sign) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_copysignfx_sve(values, sign)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_sign[size()];
          store(tmp);
          sign.store(tmp_sign);
          for (int64_t i = 0; i < size(); ++i) {
            tmp[i] = std::copysign(tmp[i], tmp_sign[i]);
          }
          return loadu(tmp);
        });
  }
  Vectorized<float> erf() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_erffx_u10sve(values)), map(std::erf));
  }
  Vectorized<float> erfc() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_erfcfx_u15sve(values)), map(std::erfc));
  }
  Vectorized<float> erfinv() const {
    return map(calc_erfinv);
  }
  Vectorized<float> exp() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_expfx_u10sve(values)), map(std::exp));
  }
  Vectorized<float> exp2() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_exp2fx_u10sve(values)), map(std::exp2));
  }
  Vectorized<float> expm1() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_expm1fx_u10sve(values)), map(std::expm1));
  }
  // Implementation copied from Arm Optimized Routines:
  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/sve/expf.c
  Vectorized<float> exp_u20() const {
    // special case to handle special inputs that are too large or too small
    // i.e. where there's at least one element x, s.t. |x| >= 87.3...
    svbool_t is_special_case = svacgt(svptrue_b32(), values, 0x1.5d5e2ap+6f);
    if (svptest_any(svptrue_b32(), is_special_case)) {
      return exp();
    }
    const svfloat32_t ln2_hi = svdup_n_f32(0x1.62e4p-1f);
    const svfloat32_t ln2_lo = svdup_n_f32(0x1.7f7d1cp-20f);
    const svfloat32_t c1 = svdup_n_f32(0.5f);
    const svfloat32_t inv_ln2 = svdup_n_f32(0x1.715476p+0f);

    const float shift = 0x1.803f8p17f;

    /* n = round(x/(ln2/N)).  */
    svfloat32_t z = svmad_x(svptrue_b32(), inv_ln2, values, shift);
    svfloat32_t n = svsub_x(svptrue_b32(), z, shift);

    /* r = x - n*ln2/N.  */
    svfloat32_t r = values;
    r = svmls_x(svptrue_b32(), r, n, ln2_hi);
    r = svmls_x(svptrue_b32(), r, n, ln2_lo);

    /* scale = 2^(n/N).  */
    svfloat32_t scale = svexpa(svreinterpret_u32(z));

    /* poly(r) = exp(r) - 1 ~= r + 0.5 r^2.  */
    svfloat32_t r2 = svmul_x(svptrue_b32(), r, r);
    svfloat32_t poly = svmla_x(svptrue_b32(), r, r2, c1);
    return svmla_x(svptrue_b32(), scale, scale, poly);
  }
  Vectorized<float> fexp_u20() const {
    return exp_u20();
  }
  Vectorized<float> fmod(const Vectorized<float>& q) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_fmodfx_sve(values, q)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_q[size()];
          store(tmp);
          q.store(tmp_q);
          for (int64_t i = 0; i < size(); ++i) {
            tmp[i] = std::fmod(tmp[i], tmp_q[i]);
          }
          return loadu(tmp);
        });
  }
  Vectorized<float> hypot(const Vectorized<float>& b) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_hypotfx_u05sve(values, b)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_b[size()];
          store(tmp);
          b.store(tmp_b);
          for (int64_t i = 0; i < size(); i++) {
            tmp[i] = std::hypot(tmp[i], tmp_b[i]);
          }
          return loadu(tmp);
        });
  }
  Vectorized<float> i0() const {
    return map(calc_i0);
  }
  Vectorized<float> i0e() const {
    return map(calc_i0e);
  }
  Vectorized<float> digamma() const {
    return map(calc_digamma);
  }
  Vectorized<float> igamma(const Vectorized<float>& x) const {
    __at_align__ float tmp[size()];
    __at_align__ float tmp_x[size()];
    store(tmp);
    x.store(tmp_x);
    for (int64_t i = 0; i < size(); i++) {
      tmp[i] = calc_igamma(tmp[i], tmp_x[i]);
    }
    return loadu(tmp);
  }
  Vectorized<float> igammac(const Vectorized<float>& x) const {
    __at_align__ float tmp[size()];
    __at_align__ float tmp_x[size()];
    store(tmp);
    x.store(tmp_x);
    for (int64_t i = 0; i < size(); i++) {
      tmp[i] = calc_igammac(tmp[i], tmp_x[i]);
    }
    return loadu(tmp);
  }
  Vectorized<float> nextafter(const Vectorized<float>& b) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_nextafterfx_sve(values, b)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_b[size()];
          store(tmp);
          b.store(tmp_b);
          for (int64_t i = 0; i < size(); ++i) {
            tmp[i] = std::nextafter(tmp[i], tmp_b[i]);
          }
          return loadu(tmp);
        });
  }
  Vectorized<float> log() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_logfx_u10sve(values)), map(std::log));
  }
  Vectorized<float> log2() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_log2fx_u10sve(values)), map(std::log2));
  }
  Vectorized<float> log10() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_log10fx_u10sve(values)), map(std::log10));
  }
  Vectorized<float> log1p() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_log1pfx_u10sve(values)), map(std::log1p));
  }
  Vectorized<float> frac() const;
  Vectorized<float> sin() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_sinfx_u10sve(values)), map(std::sin));
  }
  Vectorized<float> sinh() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_sinhfx_u10sve(values)), map(std::sinh));
  }
  Vectorized<float> cos() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_cosfx_u10sve(values)), map(std::cos));
  }
  Vectorized<float> cosh() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_coshfx_u10sve(values)), map(std::cosh));
  }
  Vectorized<float> ceil() const {
    return svrintp_f32_x(ptrue, values);
  }
  Vectorized<float> floor() const {
    return svrintm_f32_x(ptrue, values);
  }
  Vectorized<float> neg() const {
    return svneg_f32_x(ptrue, values);
  }
  Vectorized<float> round() const {
    return svrinti_f32_x(ptrue, values);
  }
  Vectorized<float> tan() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_tanfx_u10sve(values)), map(std::tan));
  }
  // Implementation is picked from
  // https://github.com/ARM-software/ComputeLibrary/blob/v25.01/src/core/NEON/SVEMath.inl#L179
  Vectorized<float> tanh() const {
    // Constants used for the tanh calculation.
    const svfloat32_t CONST_1 =
        svdup_n_f32(1.f); // Constant 1.0f for the tanh formula.
    const svfloat32_t CONST_2 = svdup_n_f32(
        2.f); // Constant 2.0f for the tanh formula (used in exp(2x)).
    const svfloat32_t CONST_MIN_TANH = svdup_n_f32(
        -10.f); // Minimum threshold for input values to prevent overflow.
    const svfloat32_t CONST_MAX_TANH = svdup_n_f32(
        10.f); // Maximum threshold for input values to prevent overflow.

    // Step 1: Clamp the values within the range [-10, 10] to prevent overflow
    // during exponentiation. The tanh function approaches ±1 rapidly as the
    // input grows large, so we limit the input range to avoid numerical
    // instability. svmax_f32_z ensures values are greater than -10, and
    // svmin_f32_z ensures they are less than 10.
    svfloat32_t x = svmin_f32_z(
        ptrue, svmax_f32_z(ptrue, values, CONST_MIN_TANH), CONST_MAX_TANH);

    // Step 2: Calculate exp(2 * x), where x is the clamped value.
    // svmul_f32_z computes 2 * x, and exp_u20() computes the exponential of
    // the result (via Vectorized<float>, then auto-converts back to
    // svfloat32_t).
    svfloat32_t exp2x =
        Vectorized<float>(svmul_f32_z(ptrue, CONST_2, x)).exp_u20();

    // Step 3: Calculate the numerator of the tanh function, which is exp(2x)
    // - 1.
    svfloat32_t num = svsub_f32_z(ptrue, exp2x, CONST_1);

    // Step 4: Calculate the denominator of the tanh function, which is exp(2x)
    // + 1.
    svfloat32_t den = svadd_f32_z(ptrue, exp2x, CONST_1);

    // Step 5: Calculate the tanh function as the ratio of the numerator and
    // denominator: num / den.
    svfloat32_t tanh = svdiv_f32_z(ptrue, num, den);

    // Return the calculated tanh values.
    return tanh;
  }
  Vectorized<float> trunc() const {
    return svrintz_f32_x(ptrue, values);
  }
  Vectorized<float> lgamma() const {
    return USE_SLEEF(
        Vectorized<float>(Sleef_lgammafx_u10sve(values)), map(std::lgamma));
  }
  Vectorized<float> sqrt() const {
    return svsqrt_f32_x(ptrue, values);
  }
  Vectorized<float> reciprocal() const {
    return svdivr_f32_x(ptrue, values, ONE_F32);
  }
  Vectorized<float> rsqrt() const {
    return svdivr_f32_x(ptrue, svsqrt_f32_x(ptrue, values), ONE_F32);
  }
  Vectorized<float> pow(const Vectorized<float>& b) const {
    USE_SLEEF(
        { return Vectorized<float>(Sleef_powfx_u10sve(values, b)); },
        {
          __at_align__ float tmp[size()];
          __at_align__ float tmp_b[size()];
          store(tmp);
          b.store(tmp_b);
          for (int64_t i = 0; i < size(); i++) {
            tmp[i] = std::pow(tmp[i], tmp_b[i]);
          }
          return loadu(tmp);
        });
  }
  // Comparison using the _CMP_**_OQ predicate.
  //   `O`: get false if an operand is NaN
  //   `Q`: do not raise if an operand is NaN
  Vectorized<float> operator==(const Vectorized<float>& other) const {
    svbool_t mask = svcmpeq_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> operator!=(const Vectorized<float>& other) const {
    svbool_t mask = svcmpne_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> operator<(const Vectorized<float>& other) const {
    svbool_t mask = svcmplt_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> operator<=(const Vectorized<float>& other) const {
    svbool_t mask = svcmple_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> operator>(const Vectorized<float>& other) const {
    svbool_t mask = svcmpgt_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> operator>=(const Vectorized<float>& other) const {
    svbool_t mask = svcmpge_f32(ptrue, values, other);
    return svsel_f32(mask, ALL_F32_TRUE_MASK, ALL_F32_FALSE_MASK);
  }

  Vectorized<float> eq(const Vectorized<float>& other) const;
  Vectorized<float> ne(const Vectorized<float>& other) const;
  Vectorized<float> gt(const Vectorized<float>& other) const;
  Vectorized<float> ge(const Vectorized<float>& other) const;
  Vectorized<float> lt(const Vectorized<float>& other) const;
  Vectorized<float> le(const Vectorized<float>& other) const;
};