OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert "[PyTorch] Improve aarch64 performance of bfloat16 ops (#166028)"

Browse Source 
This reverts commit 3e77a2b478.

Otherwise it fails ARM build with older compilers with errors that looks
as follows:
```
vec128_bfloat16_neon.h:666:12: error: operation not permitted on type ‘bfloat16_t’
  666 |   return (-x) * y - z;
```

For more self-contained example see https://godbolt.org/z/bbY4xWh45
(that compiles the same code using clang-15 and clang-19)
This commit is contained in:
Nikita Shulga 2025-10-29 13:35:59 -07:00
parent fc540cefd4
commit 2c9f877fa7
1 changed files with 1 additions and 92 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

93
aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h
View File

@ -354,47 +354,9 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
Vectorized frac() const;
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
#ifdef __ARM_FEATURE_BF16
Vectorized<c10::BFloat16> neg() const {
return -values;
}
Vectorized<c10::BFloat16> reciprocal() const {
return 1.0f / values;
}
Vectorized<c10::BFloat16> operator==(
const Vectorized<c10::BFloat16>& other) const {
return values == other.values;
}
Vectorized<c10::BFloat16> operator!=(
const Vectorized<c10::BFloat16>& other) const {
return values != other.values;
}
Vectorized<c10::BFloat16> operator<(
const Vectorized<c10::BFloat16>& other) const {
return values < other.values;
}
Vectorized<c10::BFloat16> operator<=(
const Vectorized<c10::BFloat16>& other) const {
return values <= other.values;
}
Vectorized<c10::BFloat16> operator>(
const Vectorized<c10::BFloat16>& other) const {
return values > other.values;
}
Vectorized<c10::BFloat16> operator>=(
const Vectorized<c10::BFloat16>& other) const {
return values >= other.values;
}
#else
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator==)
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator!=)
@ -402,7 +364,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<=)
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>)
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>=)
#endif
#undef DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
#undef DEFINE_BINARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
@ -451,52 +412,28 @@ template <>
Vectorized<c10::BFloat16> inline operator+(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
return x + y;
#else
return binary_operator_via_float(std::plus<Vectorized<float>>(), a, b);
#endif
}
template <>
Vectorized<c10::BFloat16> inline operator-(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
return x - y;
#else
return binary_operator_via_float(std::minus<Vectorized<float>>(), a, b);
#endif
}
template <>
Vectorized<c10::BFloat16> inline operator*(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
return x * y;
#else
return binary_operator_via_float(std::multiplies<Vectorized<float>>(), a, b);
#endif
}
template <>
Vectorized<c10::BFloat16> inline operator/(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
return x / y;
#else
return binary_operator_via_float(std::divides<Vectorized<float>>(), a, b);
#endif
}
// frac. Implement this here so we can use subtraction
@ -607,19 +544,12 @@ Vectorized<c10::BFloat16> inline fmadd(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b,
const Vectorized<c10::BFloat16>& c) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
bfloat16x8_t z = c;
return x * y + z;
#else
// NOTE [BF16 FMA]: There isn't an FMA that accumulates into BF16! Also,
// vbfmlalbq_f32 and vbfmlaltq_f32 take the even and odd-numbered
// elements, not the bottom and top half, so they don't seem
// particularly useful here. Ideally we would include dot product in
// the Vectorized interface...
return a * b + c;
#endif
}
template <>
@ -627,15 +557,8 @@ Vectorized<c10::BFloat16> inline fnmadd(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b,
const Vectorized<c10::BFloat16>& c) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
bfloat16x8_t z = c;
return (-x) * y + z;
#else
// See NOTE [BF16 FMA] above.
return -a * b + c;
#endif
}
template <>
@ -643,15 +566,8 @@ Vectorized<c10::BFloat16> inline fmsub(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b,
const Vectorized<c10::BFloat16>& c) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
bfloat16x8_t z = c;
return x * y - z;
#else
// See NOTE [BF16 FMA] above.
return a * b - c;
#endif
}
template <>
@ -659,15 +575,8 @@ Vectorized<c10::BFloat16> inline fnmsub(
const Vectorized<c10::BFloat16>& a,
const Vectorized<c10::BFloat16>& b,
const Vectorized<c10::BFloat16>& c) {
#ifdef __ARM_FEATURE_BF16
bfloat16x8_t x = a;
bfloat16x8_t y = b;
bfloat16x8_t z = c;
return (-x) * y - z;
#else
// See NOTE [BF16 FMA] above.
return -a * b - c;
#endif
}
#endif // !defined(C10_MOBILE) && defined(__aarch64__)