/usr/local/lib64/python3.6/site-packages/torch/include/ATen/cpu/vec/vec256
Edit: /usr/local/lib64/python3.6/site-packages/torch/include/ATen/cpu/vec/vec256/vec256_float.h (13867B)
#pragma once
// DO NOT DEFINE STATIC DATA IN THIS HEADER!
// See Note [Do not compile initializers with AVX]
#include
#include
#if defined(CPU_CAPABILITY_AVX2) && !defined(_MSC_VER)
#include
#endif
namespace at {
namespace vec {
// See Note [Acceptable use of anonymous namespace in header]
namespace {
#if defined(CPU_CAPABILITY_AVX2) && !defined(_MSC_VER)
template <> class Vectorized {
private:
__m256 values;
public:
using value_type = float;
using size_type = int;
static constexpr size_type size() {
return 8;
}
Vectorized() {}
Vectorized(__m256 v) : values(v) {}
Vectorized(float val) {
values = _mm256_set1_ps(val);
}
Vectorized(float val1, float val2, float val3, float val4,
float val5, float val6, float val7, float val8) {
values = _mm256_setr_ps(val1, val2, val3, val4, val5, val6, val7, val8);
}
operator __m256() const {
return values;
}
template
static Vectorized blend(const Vectorized& a, const Vectorized& b) {
return _mm256_blend_ps(a.values, b.values, mask);
}
static Vectorized blendv(const Vectorized& a, const Vectorized& b,
const Vectorized& mask) {
return _mm256_blendv_ps(a.values, b.values, mask.values);
}
template
static Vectorized arange(float base = 0.f, step_t step = static_cast(1)) {
return Vectorized(
base, base + step, base + 2 * step, base + 3 * step,
base + 4 * step, base + 5 * step, base + 6 * step, base + 7 * step);
}
static Vectorized set(const Vectorized& a, const Vectorized& b,
int64_t count = size()) {
switch (count) {
case 0:
return a;
case 1:
return blend<1>(a, b);
case 2:
return blend<3>(a, b);
case 3:
return blend<7>(a, b);
case 4:
return blend<15>(a, b);
case 5:
return blend<31>(a, b);
case 6:
return blend<63>(a, b);
case 7:
return blend<127>(a, b);
}
return b;
}
static Vectorized loadu(const void* ptr, int64_t count = size()) {
if (count == size())
return _mm256_loadu_ps(reinterpret_cast(ptr));
__at_align__ float tmp_values[size()];
// Ensure uninitialized memory does not change the output value See https://github.com/pytorch/pytorch/issues/32502
// for more details. We do not initialize arrays to zero using "={0}" because gcc would compile it to two
// instructions while a loop would be compiled to one instruction.
for (auto i = 0; i < size(); ++i) {
tmp_values[i] = 0.0;
}
std::memcpy(
tmp_values, reinterpret_cast(ptr), count * sizeof(float));
return _mm256_loadu_ps(tmp_values);
}
void store(void* ptr, int64_t count = size()) const {
if (count == size()) {
_mm256_storeu_ps(reinterpret_cast(ptr), values);
} else if (count > 0) {
float tmp_values[size()];
_mm256_storeu_ps(reinterpret_cast(tmp_values), values);
std::memcpy(ptr, tmp_values, count * sizeof(float));
}
}
const float& operator[](int idx) const = delete;
float& operator[](int idx) = delete;
int zero_mask() const {
// returns an integer mask where all zero elements are translated to 1-bit and others are translated to 0-bit
__m256 cmp = _mm256_cmp_ps(values, _mm256_set1_ps(0.0f), _CMP_EQ_OQ);
return _mm256_movemask_ps(cmp);
}
Vectorized isnan() const {
return _mm256_cmp_ps(values, _mm256_set1_ps(0.0f), _CMP_UNORD_Q);
}
Vectorized map(float (*const 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 abs() const {
auto mask = _mm256_set1_ps(-0.f);
return _mm256_andnot_ps(mask, values);
}
Vectorized angle() const {
const auto zero_vec = _mm256_set1_ps(0.f);
const auto nan_vec = _mm256_set1_ps(NAN);
const auto not_nan_mask = _mm256_cmp_ps(values, values, _CMP_EQ_OQ);
const auto nan_mask = _mm256_cmp_ps(not_nan_mask, zero_vec, _CMP_EQ_OQ);
const auto pi = _mm256_set1_ps(c10::pi);
const auto neg_mask = _mm256_cmp_ps(values, zero_vec, _CMP_LT_OQ);
auto angle = _mm256_blendv_ps(zero_vec, pi, neg_mask);
angle = _mm256_blendv_ps(angle, nan_vec, nan_mask);
return angle;
}
Vectorized real() const {
return *this;
}
Vectorized imag() const {
return _mm256_set1_ps(0);
}
Vectorized conj() const {
return *this;
}
Vectorized acos() const {
return Vectorized(Sleef_acosf8_u10(values));
}
Vectorized asin() const {
return Vectorized(Sleef_asinf8_u10(values));
}
Vectorized atan() const {
return Vectorized(Sleef_atanf8_u10(values));
}
Vectorized atan2(const Vectorized &b) const {
return Vectorized(Sleef_atan2f8_u10(values, b));
}
Vectorized copysign(const Vectorized &sign) const {
return Vectorized(Sleef_copysignf8(values, sign));
}
Vectorized erf() const {
return Vectorized(Sleef_erff8_u10(values));
}
Vectorized erfc() const {
return Vectorized(Sleef_erfcf8_u15(values));
}
Vectorized erfinv() const {
return map(calc_erfinv);
}
Vectorized exp() const {
return Vectorized(Sleef_expf8_u10(values));
}
Vectorized expm1() const {
return Vectorized(Sleef_expm1f8_u10(values));
}
Vectorized fmod(const Vectorized& q) const {
return Vectorized(Sleef_fmodf8(values, q));
}
Vectorized log() const {
return Vectorized(Sleef_logf8_u10(values));
}
Vectorized log2() const {
return Vectorized(Sleef_log2f8_u10(values));
}
Vectorized log10() const {
return Vectorized(Sleef_log10f8_u10(values));
}
Vectorized log1p() const {
return Vectorized(Sleef_log1pf8_u10(values));
}
Vectorized frac() const;
Vectorized sin() const {
return Vectorized(Sleef_sinf8_u10(values));
}
Vectorized sinh() const {
return Vectorized(Sleef_sinhf8_u10(values));
}
Vectorized cos() const {
return Vectorized(Sleef_cosf8_u10(values));
}
Vectorized cosh() const {
return Vectorized(Sleef_coshf8_u10(values));
}
Vectorized ceil() const {
return _mm256_ceil_ps(values);
}
Vectorized floor() const {
return _mm256_floor_ps(values);
}
Vectorized hypot(const Vectorized &b) const {
return Vectorized(Sleef_hypotf8_u05(values, b));
}
Vectorized i0() const {
return map(calc_i0);
}
Vectorized i0e() const {
return map(calc_i0e);
}
Vectorized igamma(const Vectorized &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 igammac(const Vectorized &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 neg() const {
return _mm256_xor_ps(_mm256_set1_ps(-0.f), values);
}
Vectorized nextafter(const Vectorized &b) const {
return Vectorized(Sleef_nextafterf8(values, b));
}
Vectorized round() const {
return _mm256_round_ps(values, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
}
Vectorized tan() const {
return Vectorized(Sleef_tanf8_u10(values));
}
Vectorized tanh() const {
return Vectorized(Sleef_tanhf8_u10(values));
}
Vectorized trunc() const {
return _mm256_round_ps(values, (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC));
}
Vectorized lgamma() const {
return Vectorized(Sleef_lgammaf8_u10(values));
}
Vectorized sqrt() const {
return _mm256_sqrt_ps(values);
}
Vectorized reciprocal() const {
return _mm256_div_ps(_mm256_set1_ps(1), values);
}
Vectorized rsqrt() const {
return _mm256_div_ps(_mm256_set1_ps(1), _mm256_sqrt_ps(values));
}
Vectorized pow(const Vectorized &b) const {
return Vectorized(Sleef_powf8_u10(values, b));
}
// Comparison using the _CMP_**_OQ predicate.
// `O`: get false if an operand is NaN
// `Q`: do not raise if an operand is NaN
Vectorized operator==(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_EQ_OQ);
}
Vectorized operator!=(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_NEQ_UQ);
}
Vectorized operator<(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_LT_OQ);
}
Vectorized operator<=(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_LE_OQ);
}
Vectorized operator>(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_GT_OQ);
}
Vectorized operator>=(const Vectorized& other) const {
return _mm256_cmp_ps(values, other.values, _CMP_GE_OQ);
}
Vectorized eq(const Vectorized& other) const;
Vectorized ne(const Vectorized& other) const;
Vectorized gt(const Vectorized& other) const;
Vectorized ge(const Vectorized& other) const;
Vectorized lt(const Vectorized& other) const;
Vectorized le(const Vectorized& other) const;
};
template <>
Vectorized inline operator+(const Vectorized& a, const Vectorized& b) {
return _mm256_add_ps(a, b);
}
template <>
Vectorized inline operator-(const Vectorized& a, const Vectorized& b) {
return _mm256_sub_ps(a, b);
}
template <>
Vectorized inline operator*(const Vectorized& a, const Vectorized& b) {
return _mm256_mul_ps(a, b);
}
template <>
Vectorized inline operator/(const Vectorized& a, const Vectorized& b) {
return _mm256_div_ps(a, b);
}
// frac. Implement this here so we can use subtraction
Vectorized Vectorized::frac() const {
return *this - this->trunc();
}
// Implements the IEEE 754 201X `maximum` operation, which propagates NaN if
// either input is a NaN.
template <>
Vectorized inline maximum(const Vectorized& a, const Vectorized& b) {
Vectorized max = _mm256_max_ps(a, b);
Vectorized isnan = _mm256_cmp_ps(a, b, _CMP_UNORD_Q);
// Exploit the fact that all-ones is a NaN.
return _mm256_or_ps(max, isnan);
}
// Implements the IEEE 754 201X `minimum` operation, which propagates NaN if
// either input is a NaN.
template <>
Vectorized inline minimum(const Vectorized& a, const Vectorized& b) {
Vectorized min = _mm256_min_ps(a, b);
Vectorized isnan = _mm256_cmp_ps(a, b, _CMP_UNORD_Q);
// Exploit the fact that all-ones is a NaN.
return _mm256_or_ps(min, isnan);
}
template <>
Vectorized inline clamp(const Vectorized& a, const Vectorized& min, const Vectorized& max) {
return _mm256_min_ps(max, _mm256_max_ps(min, a));
}
template <>
Vectorized inline clamp_max(const Vectorized& a, const Vectorized& max) {
return _mm256_min_ps(max, a);
}
template <>
Vectorized inline clamp_min(const Vectorized& a, const Vectorized& min) {
return _mm256_max_ps(min, a);
}
template <>
Vectorized inline operator&(const Vectorized& a, const Vectorized& b) {
return _mm256_and_ps(a, b);
}
template <>
Vectorized inline operator|(const Vectorized& a, const Vectorized& b) {
return _mm256_or_ps(a, b);
}
template <>
Vectorized inline operator^(const Vectorized& a, const Vectorized& b) {
return _mm256_xor_ps(a, b);
}
Vectorized Vectorized::eq(const Vectorized& other) const {
return (*this == other) & Vectorized(1.0f);
}
Vectorized Vectorized::ne(const Vectorized& other) const {
return (*this != other) & Vectorized(1.0f);
}
Vectorized Vectorized::gt(const Vectorized& other) const {
return (*this > other) & Vectorized(1.0f);
}
Vectorized Vectorized::ge(const Vectorized& other) const {
return (*this >= other) & Vectorized(1.0f);
}
Vectorized Vectorized::lt(const Vectorized& other) const {
return (*this < other) & Vectorized(1.0f);
}
Vectorized Vectorized::le(const Vectorized& other) const {
return (*this <= other) & Vectorized(1.0f);
}
template <>
inline void convert(const float* src, float* dst, int64_t n) {
int64_t i;
#pragma unroll
for (i = 0; i <= (n - Vectorized::size()); i += Vectorized::size()) {
_mm256_storeu_ps(dst + i, _mm256_loadu_ps(src + i));
}
#pragma unroll
for (; i < n; i++) {
dst[i] = src[i];
}
}
template <>
Vectorized inline fmadd(const Vectorized& a, const Vectorized& b, const Vectorized& c) {
return _mm256_fmadd_ps(a, b, c);
}
#endif
}}}