Skip to content

Correct the conversion from float/double to half on CUDA #3627

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
christophe-murphy merged 1 commit into from
Jan 16, 2025
Merged

Correct the conversion from float/double to half on CUDA #3627

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 25 additions & 19 deletions src/backend/common/half.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,7 @@ using uint16_t = unsigned short;
#define AF_CONSTEXPR constexpr
#else
#include <af/compilers.h>
#include <algorithm>
#include <cmath>
#include <cstdint>
#include <cstring>
Expand Down Expand Up @@ -245,9 +246,9 @@ AF_CONSTEXPR __DH__ native_half_t int2half_impl(T value) noexcept {
/// \return binary representation of half-precision value
template<std::float_round_style R = std::round_to_nearest>
__DH__ native_half_t float2half_impl(float value) noexcept {
uint32_t bits = 0; // = *reinterpret_cast<uint32*>(&value);
// //violating strict aliasing!
std::memcpy(&bits, &value, sizeof(float));
alignas(std::max(alignof(uint32_t), alignof(float))) float _value = value;
uint32_t bits = *reinterpret_cast<uint32_t*>(&_value);

constexpr uint16_t base_table[512] = {
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000,
Expand Down Expand Up @@ -337,9 +338,10 @@ __DH__ native_half_t float2half_impl(float value) noexcept {
24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24,
24, 24, 24, 24, 24, 24, 24, 13};
uint16_t hbits =
base_table[bits >> 23] +
static_cast<uint16_t>((bits & 0x7FFFFF) >> shift_table[bits >> 23]);
alignas(std::max(alignof(uint16_t), alignof(native_half_t)))
uint16_t hbits =
base_table[bits >> 23] +
static_cast<uint16_t>((bits & 0x7FFFFF) >> shift_table[bits >> 23]);
AF_IF_CONSTEXPR(R == std::round_to_nearest)
hbits +=
(((bits & 0x7FFFFF) >> (shift_table[bits >> 23] - 1)) |
Expand Down Expand Up @@ -367,7 +369,8 @@ __DH__ native_half_t float2half_impl(float value) noexcept {
(((bits >> 23) <= 358) & ((bits >> 23) != 256))) &
(hbits < 0xFC00) & (hbits >> 15)) -
((hbits == 0x7C00) & ((bits >> 23) != 255));
return hbits;

return *reinterpret_cast<native_half_t*>(&hbits);
}

/// Convert IEEE double-precision to half-precision.
Expand All @@ -379,11 +382,11 @@ __DH__ native_half_t float2half_impl(float value) noexcept {
/// \return binary representation of half-precision value
template<std::float_round_style R>
__DH__ native_half_t float2half_impl(double value) {
uint64_t bits{0}; // = *reinterpret_cast<uint64*>(&value); //violating
// strict aliasing!
std::memcpy(&bits, &value, sizeof(double));
alignas(std::max(alignof(uint64_t), alignof(double))) double _value = value;
uint64_t bits = *reinterpret_cast<uint64_t*>(&_value);
uint32_t hi = bits >> 32, lo = bits & 0xFFFFFFFF;
uint16_t hbits = (hi >> 16) & 0x8000;
alignas(std::max(alignof(uint16_t), alignof(native_half_t)))
uint16_t hbits = (hi >> 16) & 0x8000;
hi &= 0x7FFFFFFF;
int exp = hi >> 20;
if (exp == 2047)
Expand Down Expand Up @@ -420,7 +423,8 @@ __DH__ native_half_t float2half_impl(double value) {
~(hbits >> 15) & (s | g);
else AF_IF_CONSTEXPR(R == std::round_toward_neg_infinity) hbits +=
(hbits >> 15) & (g | s);
return hbits;

return *reinterpret_cast<native_half_t*>(&hbits);
}

__DH__ inline float half2float_impl(native_half_t value) noexcept {
Expand Down Expand Up @@ -790,14 +794,14 @@ __DH__ inline float half2float_impl(native_half_t value) noexcept {
1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024,
1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024};

uint16_t value_bits = 0;
std::memcpy(&value_bits, &value, sizeof(uint16_t));
uint32_t bits =
alignas(std::max(alignof(uint16_t), alignof(native_half_t)))
native_half_t _value = value;
uint16_t value_bits = *reinterpret_cast<uint16_t*>(&_value);

alignas(std::max(alignof(uint32_t), alignof(float))) uint32_t bits =
mantissa_table[offset_table[value_bits >> 10] + (value_bits & 0x3FF)] +
exponent_table[value_bits >> 10];
float out = 0.0f;
std::memcpy(&out, &bits, sizeof(float));
return out;
return *reinterpret_cast<float*>(&bits);
}

#endif // __CUDACC_RTC__
Expand Down Expand Up @@ -872,7 +876,9 @@ AF_CONSTEXPR T half2int(native_half_t value) {
else AF_IF_CONSTEXPR(std::is_same<T, int>::value) {
return __half2int_rn(value);
}
else { return __half2uint_rn(value); }
else {
return __half2uint_rn(value);
}
#elif defined(AF_ONEAPI)
return static_cast<T>(value);
#else
Expand Down