Skip to content

GitLab

"docs/vscode:/vscode.git/clone" did not exist on "6bd3851289e2bfc3eb57e9582999415f21e7af52"
Unverified Commit ec959387 authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into ck_tile/fmha_receipt_aiter

parents c1e2fef7 0e5e29c4
Hide whitespace changes
Inline Side-by-side
Showing with 2452 additions and 68 deletions
include/ck/utility/debug.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef UTILITY_DEBUG_HPP #ifndef UTILITY_DEBUG_HPP
#define UTILITY_DEBUG_HPP #define UTILITY_DEBUG_HPP
#include "type.hpp"
namespace ck { namespace ck {
namespace debug { namespace debug {
......
include/ck/utility/e8m0.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/type.hpp"
namespace ck {
/**
* @brief Unsigned representation of a conventional biased Float32 exponent.
*
* bias = 127;
*
* E8M0_1 = 0b01111111; => 2^(127-127) = 1
* E8M0_2 = 0b10000000; => 2^(128-127) = 2^1 = 2
* E8M0_3 = 0b10000010; => 2^(130-127) = 2^3 = 8
* E8M0_135 = 0b10000111; => 2^(135-127) = 2^8 = 256
* E8M0_142 = 0b10001110; => 2^(142-127) = 2^15 = 32768
* E8M0_MIN = 0b00000000; => 2^-127
* E8M0_MAX = 0b11111110; => 2^127
* E8M0_NAN = 0b11111111; => NaN
*/
struct e8m0_bexp_t
{
using type = uint8_t;
type data;
constexpr static type bias = 127;
constexpr static type nan_mask = 0xFF;
__host__ __device__ constexpr e8m0_bexp_t() : data{type{}} {}
__host__ __device__ constexpr e8m0_bexp_t(type init) : data{init} {}
__host__ __device__ constexpr e8m0_bexp_t(int init) : data{static_cast<type>(init & nan_mask)}
{
}
__host__ __device__ explicit constexpr e8m0_bexp_t(float scale)
: data{static_cast<type>((bit_cast<uint32_t>(scale) & (nan_mask << 23)) >> 23)}
{
}
__host__ __device__ explicit constexpr operator float() const
{
if(data == nan_mask || data == 0)
{
uint32_t bits = data << 1;
bits |= 1;
bits <<= 22;
return bit_cast<float>(bits);
}
else
{
uint32_t bits = data << 23;
return bit_cast<float>(bits);
}
}
__host__ __device__ constexpr bool operator==(const e8m0_bexp_t& other) const
{
// strict IEEE compliance for NaN
return data == other.data && data != nan_mask;
}
__host__ __device__ constexpr bool is_nan() const { return data == nan_mask; }
};
namespace utils {
template <typename T>
__host__ __device__ inline int get_exponent_value(T x);
template <>
__host__ __device__ inline int get_exponent_value<e8m0_bexp_t>(e8m0_bexp_t x)
{
return x.data;
}
} // namespace utils
} // namespace ck
include/ck/utility/enable_if.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
namespace ck { namespace ck {
#ifndef CK_CODE_GEN_RTC
template <bool B, typename T = void> template <bool B, typename T = void>
using enable_if = std::enable_if<B, T>; using enable_if = std::enable_if<B, T>;
template <bool B, typename T = void> template <bool B, typename T = void>
using enable_if_t = typename std::enable_if<B, T>::type; using enable_if_t = typename std::enable_if<B, T>::type;
#else
template <bool B, class T = void>
struct enable_if
{
};
template <class T>
struct enable_if<true, T>
{
using type = T;
};
template <bool B, class T = void>
using enable_if_t = typename enable_if<B, T>::type;
#endif
} // namespace ck } // namespace ck
include/ck/utility/env.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_CODE_GEN_RTC
#pragma once #pragma once
#include <cstdlib> #include <cstdlib>
...@@ -183,3 +184,4 @@ void UpdateEnvVar(EnvVar, const std::string_view& val) ...@@ -183,3 +184,4 @@ void UpdateEnvVar(EnvVar, const std::string_view& val)
} }
} // namespace ck } // namespace ck
#endif
include/ck/utility/functional.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -120,11 +120,11 @@ constexpr auto conditional_expr(X&& x, Y&& y) ...@@ -120,11 +120,11 @@ constexpr auto conditional_expr(X&& x, Y&& y)
{ {
if constexpr(predicate) if constexpr(predicate)
{ {
return std::forward<X>(x); return ck::forward<X>(x);
} }
else else
{ {
return std::forward<Y>(y); return ck::forward<Y>(y);
} }
} }
......
include/ck/utility/functional4.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_FUNCTIONAL4_HPP #ifndef CK_FUNCTIONAL4_HPP
#define CK_FUNCTIONAL4_HPP #define CK_FUNCTIONAL4_HPP
...@@ -21,7 +21,7 @@ struct unpack_impl<Sequence<Is...>> ...@@ -21,7 +21,7 @@ struct unpack_impl<Sequence<Is...>>
template <typename F, typename X> template <typename F, typename X>
__host__ __device__ constexpr auto operator()(F&& f, X&& x) const __host__ __device__ constexpr auto operator()(F&& f, X&& x) const
{ {
return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...); return ck::forward<F>(f)(ck::forward<X>(x).At(Number<Is>{})...);
} }
}; };
...@@ -35,8 +35,8 @@ struct unpack2_impl<Sequence<Is...>, Sequence<Js...>> ...@@ -35,8 +35,8 @@ struct unpack2_impl<Sequence<Is...>, Sequence<Js...>>
template <typename F, typename X, typename Y> template <typename F, typename X, typename Y>
__host__ __device__ constexpr auto operator()(F&& f, X&& x, Y&& y) const __host__ __device__ constexpr auto operator()(F&& f, X&& x, Y&& y) const
{ {
return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})..., return ck::forward<F>(f)(ck::forward<X>(x).At(Number<Is>{})...,
std::forward<Y>(y).At(Number<Js>{})...); ck::forward<Y>(y).At(Number<Js>{})...);
} }
}; };
...@@ -47,7 +47,7 @@ __host__ __device__ constexpr auto unpack(F&& f, X&& x) ...@@ -47,7 +47,7 @@ __host__ __device__ constexpr auto unpack(F&& f, X&& x)
{ {
using X_ = remove_reference_t<X>; using X_ = remove_reference_t<X>;
return detail::unpack_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type>{}( return detail::unpack_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type>{}(
std::forward<F>(f), std::forward<X>(x)); ck::forward<F>(f), ck::forward<X>(x));
} }
// TODO: properly implement unpack that takes any number of containers // TODO: properly implement unpack that takes any number of containers
...@@ -58,7 +58,7 @@ __host__ __device__ constexpr auto unpack2(F&& f, X&& x, Y&& y) ...@@ -58,7 +58,7 @@ __host__ __device__ constexpr auto unpack2(F&& f, X&& x, Y&& y)
using Y_ = remove_reference_t<Y>; using Y_ = remove_reference_t<Y>;
return detail::unpack2_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type, return detail::unpack2_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type,
typename arithmetic_sequence_gen<0, Y_::Size(), 1>::type>{}( typename arithmetic_sequence_gen<0, Y_::Size(), 1>::type>{}(
std::forward<F>(f), std::forward<X>(x), std::forward<Y>(y)); ck::forward<F>(f), ck::forward<X>(x), ck::forward<Y>(y));
} }
} // namespace ck } // namespace ck
......
include/ck/utility/integral_constant.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -48,4 +48,9 @@ __host__ __device__ constexpr auto operator%(integral_constant<TX, X>, integral_ ...@@ -48,4 +48,9 @@ __host__ __device__ constexpr auto operator%(integral_constant<TX, X>, integral_
return integral_constant<decltype(X % Y), X % Y>{}; return integral_constant<decltype(X % Y), X % Y>{};
} }
template <bool B>
using bool_constant = integral_constant<bool, B>;
using true_type = bool_constant<true>;
using false_type = bool_constant<false>;
} // namespace ck } // namespace ck
include/ck/utility/is_detected.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include "ck/utility/integral_constant.hpp"
namespace ck { namespace ck {
namespace detail { namespace detail {
template <class Default, class AlwaysVoid, template <class...> class Op, class... Args> template <class Default, class AlwaysVoid, template <class...> class Op, class... Args>
struct detector struct detector
{ {
using value_t = std::false_type; using value_t = integral_constant<bool, false>;
using type = Default; using type = Default;
}; };
template <class Default, template <class...> class Op, class... Args> template <class Default, template <class...> class Op, class... Args>
struct detector<Default, std::void_t<Op<Args...>>, Op, Args...> struct detector<Default, ck::void_t<Op<Args...>>, Op, Args...>
{ {
using value_t = std::true_type; using value_t = integral_constant<bool, true>;
using type = Op<Args...>; using type = Op<Args...>;
}; };
} // namespace detail } // namespace detail
...@@ -32,12 +34,12 @@ template <template <class...> class Op, class... Args> ...@@ -32,12 +34,12 @@ template <template <class...> class Op, class... Args>
using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t; using is_detected = typename detail::detector<nonesuch, void, Op, Args...>::value_t;
template <typename T> template <typename T>
using is_pack2_invocable_t = decltype(std::declval<T&>().is_pack2_invocable); using is_pack2_invocable_t = decltype(ck::declval<T&>().is_pack2_invocable);
template <typename T> template <typename T>
using is_pack4_invocable_t = decltype(std::declval<T&>().is_pack4_invocable); using is_pack4_invocable_t = decltype(ck::declval<T&>().is_pack4_invocable);
template <typename T> template <typename T>
using is_pack8_invocable_t = decltype(std::declval<T&>().is_pack8_invocable); using is_pack8_invocable_t = decltype(ck::declval<T&>().is_pack8_invocable);
} // namespace ck } // namespace ck
include/ck/utility/loop_scheduler.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_CODE_GEN_RTC
#include <ostream> #include <ostream>
#endif
#pragma once #pragma once
...@@ -25,6 +28,7 @@ constexpr LoopScheduler make_default_loop_scheduler() ...@@ -25,6 +28,7 @@ constexpr LoopScheduler make_default_loop_scheduler()
} // namespace ck } // namespace ck
#ifndef CK_CODE_GEN_RTC
inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s) inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
{ {
switch(s) switch(s)
...@@ -35,3 +39,4 @@ inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s) ...@@ -35,3 +39,4 @@ inline std::ostream& operator<<(std::ostream& os, const ck::LoopScheduler& s)
} }
return os; return os;
} }
#endif
include/ck/utility/magic_division.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -9,6 +9,10 @@ ...@@ -9,6 +9,10 @@
#include "type.hpp" #include "type.hpp"
#include "tuple.hpp" #include "tuple.hpp"
#ifdef CK_CODE_GEN_RTC
#define INT32_MAX 2147483647
#endif
namespace ck { namespace ck {
// magic number division // magic number division
......
include/ck/utility/math_v2.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -19,7 +19,7 @@ extern "C" __device__ float __ocml_native_recip_f32(float); ...@@ -19,7 +19,7 @@ extern "C" __device__ float __ocml_native_recip_f32(float);
#endif #endif
// math functions for the host, some are implemented by calling C++ std functions // math functions for the host, some are implemented by calling C++ std functions
#ifndef CK_CODE_GEN_RTC
static inline __host__ float abs(float x) { return std::abs(x); }; static inline __host__ float abs(float x) { return std::abs(x); };
static inline __host__ double abs(double x) { return std::abs(x); }; static inline __host__ double abs(double x) { return std::abs(x); };
...@@ -459,7 +459,7 @@ inline __host__ double expm1<double>(double x) ...@@ -459,7 +459,7 @@ inline __host__ double expm1<double>(double x)
{ {
return std::expm1(x); return std::expm1(x);
} }
#endif
// math functions for the HIP kernel, some are implemented by calling hip builtin functions // math functions for the HIP kernel, some are implemented by calling hip builtin functions
static inline __device__ float abs(float x) { return ::abs(x); }; static inline __device__ float abs(float x) { return ::abs(x); };
......
include/ck/utility/mxf4_utils.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
namespace ck::utils {
template <>
__host__ __device__ inline bool is_nan<f4_t>(e8m0_bexp_t const scale,
f4_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale == NumericLimits<e8m0_bexp_t>::QuietNaN();
}
// no infinity representation in ocp_e2m1_mxfp4 will always return false
template <>
__host__ __device__ inline bool is_inf<f4_t>(e8m0_bexp_t const scale [[maybe_unused]],
f4_t const data [[maybe_unused]])
{
// no inf representation for ocp_e2m1_mxfp4
return false;
}
template <>
__host__ __device__ inline bool is_zero<f4_t>(e8m0_bexp_t const scale, f4_t const data)
{
if(is_nan<f4_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
f4_t result = (data & 0b00001111) & NumericUtils<f4_t>::set_sign_mask;
return result == 0b0;
}
template <>
__host__ __device__ inline float to_float<f4_t>(e8m0_bexp_t const scale, f4_t const data)
{
if(is_nan<f4_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<f4_t>(scale, data))
return 0.0f;
f4_t prepared_data = data & 0b00001111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<f4_t>(prepared_data, scale_exp);
}
template <>
__host__ __device__ inline f4_t sat_convert_to_type<f4_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
f4_t res = convert_to_type<f4_t>(value);
if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f4_t>::DataMinSubnorm())
return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
: NumericUtils<f4_t>::positive_zero_mask;
return res;
}
template <>
__host__ __device__ inline f4_t sat_convert_to_type_sr<f4_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<f4_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f4_t>::data_max_negative_normal_mask
: NumericUtils<f4_t>::data_max_positive_normal_mask;
f4_t res = convert_to_type_sr<f4_t>(value, seed);
if(std::abs(to_float<f4_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f4_t>::DataMinSubnorm())
return value < 0 ? NumericUtils<f4_t>::negative_zero_mask
: NumericUtils<f4_t>::positive_zero_mask;
return res;
}
} // namespace ck::utils
include/ck/utility/mxf6_utils.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
namespace ck::utils {
/**
* @brief Checks if an f6_t value is NaN based on the provided scale.
*
* For f6_t data, NaN cannot be represented directly. Instead, this function
* determines NaN by checking if the scale is set to a quiet NaN.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param dataBytes The f6_t value to check (unused in this implementation).
* @return true if the scale indicates a NaN value, false otherwise.
*/
template <>
__host__ __device__ inline bool is_nan<f6_t>(e8m0_bexp_t const scale,
f6_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale.is_nan();
}
/**
* @brief Checks if an bf6_t value is NaN based on the provided scale.
*
* For bf6_t data, NaN cannot be represented directly. Instead, this function
* determines NaN by checking if the scale is set to a quiet NaN.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param dataBytes The bf6_t value to check (unused in this implementation).
* @return true if the scale indicates a NaN value, false otherwise.
*/
template <>
__host__ __device__ inline bool is_nan<bf6_t>(e8m0_bexp_t const scale,
bf6_t const dataBytes [[maybe_unused]])
{
// no need to check for data as it does not have NaN representation
return scale.is_nan();
}
/**
* @brief Checks if an f6_t value is infinite.
*
* Because f6_t does not support infinite values, this function always returns false.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to check.
* @return Always false, as infinity is not represented in f6_t.
*/
template <>
__host__ __device__ inline bool is_inf<f6_t>(e8m0_bexp_t const scale [[maybe_unused]],
f6_t const data [[maybe_unused]])
{
// no inf representation for fp6
return false;
}
/**
* @brief Checks if an bf6_t value is infinite.
*
* Because bf6_t does not support infinite values, this function always returns false.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to check.
* @return Always false, as infinity is not represented in bf6_t.
*/
template <>
__host__ __device__ inline bool is_inf<bf6_t>(e8m0_bexp_t const scale [[maybe_unused]],
bf6_t const data [[maybe_unused]])
{
// no inf representation for bf6
return false;
}
/**
* @brief Checks whether an f6_t value is zero.
*
* If the specified f6_t is NaN, this function returns false.
* Otherwise, it masks out the sign bits and checks if the remaining bits
* are zero.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to check.
* @return true if the value is zero; otherwise false.
*/
template <>
__host__ __device__ inline bool is_zero<f6_t>(e8m0_bexp_t const scale, f6_t const data)
{
if(is_nan<f6_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
f6_t result = (data & 0b00111111) & NumericUtils<f6_t>::set_sign_mask;
return result == 0b0;
}
/**
* @brief Checks whether an bf6_t value is zero.
*
* If the specified bf6_t is NaN, this function returns false.
* Otherwise, it masks out the sign bits and checks if the remaining bits
* are zero.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to check.
* @return true if the value is zero; otherwise false.
*/
template <>
__host__ __device__ inline bool is_zero<bf6_t>(e8m0_bexp_t const scale, bf6_t const data)
{
if(is_nan<bf6_t>(scale, data))
return false;
// no need to check for scale as it does not have a 0 representation
bf6_t result = (data & 0b00111111) & NumericUtils<bf6_t>::set_sign_mask;
return result == 0b0;
}
/**
* @brief Converts an f6_t value to a float based on an e8m0_bexp_t scale factor.
*
* Checks if the f6_t value is NaN or zero before performing the conversion.
* Applies the exponent from the scale to compute the final float result.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param data The f6_t value to convert.
* @return The converted float value.
*/
template <>
__host__ __device__ inline float to_float<f6_t>(e8m0_bexp_t const scale, f6_t const data)
{
if(is_nan<f6_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<f6_t>(scale, data))
return 0.0f;
f6_t prepared_data = data & 0b00111111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<f6_t>(prepared_data, scale_exp);
}
/**
* @brief Converts an bf6_t value to a float based on an e8m0_bexp_t scale factor.
*
* Checks if the bf6_t value is NaN or zero before performing the conversion.
* Applies the exponent from the scale to compute the final float result.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param data The bf6_t value to convert.
* @return The converted float value.
*/
template <>
__host__ __device__ inline float to_float<bf6_t>(e8m0_bexp_t const scale, bf6_t const data)
{
if(is_nan<bf6_t>(scale, data))
return std::numeric_limits<float>::quiet_NaN();
if(is_zero<bf6_t>(scale, data))
return 0.0f;
bf6_t prepared_data = data & 0b00111111;
int scale_exp = get_exponent_value<e8m0_bexp_t>(scale);
return convert_to_float<bf6_t>(prepared_data, scale_exp);
}
/**
* @brief Converts a float to f6_t with saturation.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline f6_t sat_convert_to_type<f6_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<f6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
f6_t res = convert_to_type<f6_t>(value);
if(std::abs(to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f6_t>::DataMinSubnorm())
return sign ? NumericUtils<f6_t>::negative_zero_mask
: NumericUtils<f6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to bf6_t with saturation.
*
* If the input is NaN or exceeds the representable range for bf6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated bf6_t value.
*/
template <>
__host__ __device__ inline bf6_t sat_convert_to_type<bf6_t>(float value)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
{
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
}
if(std::abs(value) > NumericLimits<bf6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
bf6_t res = convert_to_type<bf6_t>(value);
if(std::abs(to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<bf6_t>::DataMinSubnorm())
return sign ? NumericUtils<bf6_t>::negative_zero_mask
: NumericUtils<bf6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to f6_t with saturation and stochastic rounding.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline f6_t sat_convert_to_type_sr<f6_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<f6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<f6_t>::data_max_negative_normal_mask
: NumericUtils<f6_t>::data_max_positive_normal_mask;
f6_t res = convert_to_type_sr<f6_t>(value, seed);
if(std::abs(to_float<f6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<f6_t>::DataMinSubnorm())
return sign ? NumericUtils<f6_t>::negative_zero_mask
: NumericUtils<f6_t>::positive_zero_mask;
return res;
}
/**
* @brief Converts a float to f6_t with saturation and stochastic rounding.
*
* If the input is NaN or exceeds the representable range for f6_t, returns
* the corresponding max normal mask. Handles subnormal cases by returning
* zero with the appropriate sign.
*
* @param value The float value to be converted.
* @return The saturated f6_t value.
*/
template <>
__host__ __device__ inline bf6_t sat_convert_to_type_sr<bf6_t>(float value, uint32_t seed)
{
cvt t;
t.value_float = value;
uint32_t sign = t.value_bitwise >> 31;
if(std::isnan(value))
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
if(std::abs(value) > NumericLimits<bf6_t>::Max()) // covers inf case as well
return sign ? NumericUtils<bf6_t>::data_max_negative_normal_mask
: NumericUtils<bf6_t>::data_max_positive_normal_mask;
bf6_t res = convert_to_type_sr<bf6_t>(value, seed);
if(std::abs(to_float<bf6_t>(NumericLimits<e8m0_bexp_t>::Binary_1(), res)) <
NumericLimits<bf6_t>::DataMinSubnorm())
return sign ? NumericUtils<bf6_t>::negative_zero_mask
: NumericUtils<bf6_t>::positive_zero_mask;
return res;
}
} // namespace ck::utils
include/ck/utility/mxf8_utils.hpp 0 → 100644
View file @ ec959387
#include "ck/utility/data_type.hpp"
#include "ck/utility/mxfp_utils.hpp"
#if defined(__gfx950__) && __HIP_DEVICE_COMPILE__
#define CK_MX_FP8_CVT_FAST_PATH 1
#else
#define CK_MX_FP8_CVT_FAST_PATH 0
#endif
namespace ck {
namespace fp8_impl {
#if CK_MX_FP8_CVT_FAST_PATH
template <ck_fp8_interpretation_t interpret>
static __device__ float cast_to_f32_from_f8_scaled(float scale, fp8_storage_t v)
{
union
{
unsigned int i32val;
unsigned char i8val[4];
} val;
val.i8val[0] = v;
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return __builtin_amdgcn_cvt_scalef32_f32_fp8(val.i32val, scale, 0);
}
else
{
return __builtin_amdgcn_cvt_scalef32_f32_bf8(val.i32val, scale, 0);
}
}
template <ck_fp8_interpretation_t interpret>
static __device__ float2_t cast_to_f32x2_from_f8x2_scaled(float scale, fp8x2_storage_t v)
{
const auto i16val = bit_cast<uint16_t>(v);
static_assert(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interpret == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return __builtin_amdgcn_cvt_scalef32_pk_f32_fp8(i16val, scale, 0);
}
else
{
return __builtin_amdgcn_cvt_scalef32_pk_f32_bf8(i16val, scale, 0);
}
}
template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
static __device__ fp8_storage_t cast_to_f8_from_f32_scaled(float v,
unsigned int rng = 0,
float scale = 1.0f)
{
fp8_storage_t i8data;
union
{
float fval;
unsigned int i32val;
} val;
union
{
uint32_t ival;
vector_type<int16_t, 2>::type v2i16;
fp8_storage_t v4i8[4];
} ret{};
// unsigned int ival = 0;
val.fval = v;
if constexpr(stochastic_rounding)
{
ret.ival =
(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
? __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, val.fval, rng, scale, 0)
: __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, val.fval, rng, scale, 0);
i8data = ret.v4i8[0];
}
else
{
// RNE CVT
// llvm.amdgcn.cvt.scalef32.pk.fp8.f32
// v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
// If fval / scale > max fp8, returns Nan
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
val.fval,
val.fval,
scale,
/*dst_lo_hi_sel*/ false);
}
else
{
// If fval / scale > max bf8, returns Inf
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
val.fval,
val.fval,
scale,
/*dst_lo_hi_sel*/ false);
}
i8data = ret.v4i8[0];
}
return i8data;
}
template <ck_fp8_interpretation_t interpret, bool stochastic_rounding = false>
static __device__ fp8x2_storage_t cast_to_f8_from_f32_scaled(float2_t v,
unsigned int rng = 0,
float scale = 1.0f)
{
union
{
uint32_t ival;
vector_type<int16_t, 2>::type v2i16;
StaticallyIndexedArray<fp8x2_storage_t, 2> v2f8x2;
} ret{};
if constexpr(stochastic_rounding)
{
fp8x2_storage_t f8x2;
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[0], rng, scale, 0);
f8x2[0] = ret.v2f8x2(Number<0>{})[0];
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_fp8_f32(ret.ival, v[1], rng, scale, 0);
f8x2[1] = ret.v2f8x2(Number<0>{})[0];
}
else
{
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[0], rng, scale, 0);
f8x2[0] = ret.v2f8x2(Number<0>{})[0];
ret.ival = __builtin_amdgcn_cvt_scalef32_sr_bf8_f32(ret.ival, v[1], rng, scale, 0);
f8x2[1] = ret.v2f8x2(Number<0>{})[0];
}
return f8x2;
}
else
{
// RNE CVT
// llvm.amdgcn.cvt.scalef32.pk.fp8.f32
// v2i16 old_vdst, float srcA, float srcB, float scale, bool dst_lo_hi_sel
if constexpr(interpret == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
// If fval / scale > max fp8, returns Nan
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_fp8_f32(/*old_vdst*/ ret.v2i16,
v[0],
v[1],
scale,
/*dst_lo_hi_sel*/ false);
}
else
{
// If fval / scale > max bf8, returns Inf
ret.v2i16 = __builtin_amdgcn_cvt_scalef32_pk_bf8_f32(/*old_vdst*/ ret.v2i16,
v[0],
v[1],
scale,
/*dst_lo_hi_sel*/ false);
}
return ret.v2f8x2(Number<0>{});
}
}
#endif // CK_MX_FP8_CVT_FAST_PATH
#if CK_MX_FP8_CVT_FAST_PATH
/**
* \brief convert float to @p fp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is available
*
* \tparam interp interpretation of fp8
* \param f float number
* \param scale scaling factor
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
{
__is_interpret_supported(interp);
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
}
/**
* \brief convert 2xfloat to @p 2xfp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is available
*
* \tparam interp interpretation of fp8
* \param f 2xfloat
* \param scale scaling factor
* \return 2xfp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
float scale)
{
__is_interpret_supported(interp);
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f[0]);
}
return cast_to_f8_from_f32_scaled<interp, stochastic_rounding>(f, rng, scale);
}
#else
/**
* \brief convert float to @p fp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is not available
*
* \tparam interp interpretation of fp8
* \param f float number
* \param scale scaling factor
* \return fp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8_storage_t cvt_float_to_fp8_scaled(const float f, float scale)
{
static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f);
}
if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f / scale, rng);
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
{
return cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f / scale, rng);
}
else
{
__hip_assert(false && "FP8 type is not supported by current target device");
return 0;
}
}
/**
* \brief convert two float to @p 2xfp8_storage_t with scaling
*
* This version is used when the fast path (MX FP8 hardware) is not available
*
* \tparam interp interpretation of fp8
* \param f 2xfloat
* \param scale scaling factor
* \return 2xfp8_storage_t
*/
template <ck_fp8_interpretation_t interp, bool stochastic_rounding = false>
__host__ __device__ static inline fp8x2_storage_t cvt_float_to_fp8_scaled(const float2_t f,
float scale)
{
static_assert(interp == ck_fp8_interpretation_t::CK_E4M3_OCP ||
interp == ck_fp8_interpretation_t::CK_E5M2_OCP,
"Only OCP interpretations are supported");
uint32_t rng = 0;
if constexpr(stochastic_rounding)
{
constexpr int seed = 1254739;
rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&f), f[0]);
}
if constexpr(interp == ck_fp8_interpretation_t::CK_E4M3_OCP)
{
return {cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[0] / scale, rng),
cast_to_f8<float, 3, 4, false, true, stochastic_rounding>(f[1] / scale, rng)};
}
else if constexpr(interp == ck_fp8_interpretation_t::CK_E5M2_OCP)
{
return {cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[0] / scale, rng),
cast_to_f8<float, 2, 5, false, true, stochastic_rounding>(f[1] / scale, rng)};
}
else
{
__hip_assert(false && "FP8 type is not supported by current target device");
return 0;
}
}
#endif // CK_MX_FP8_CVT_FAST_PATH
} // namespace fp8_impl
// Declare a template function for fp8 conversion using SR
template <typename Y, typename X>
__host__ __device__ constexpr Y mxf8_convert_sr(X x, float scale);
// Declare a template function for fp8 conversion using RNE
template <typename Y, typename X>
__host__ __device__ constexpr Y mxf8_convert_rne(X x, float scale);
// convert fp32 to fp8 with rounding to nearest even
template <>
inline __host__ __device__ f8_ocp_t mxf8_convert_rne<f8_ocp_t, float>(float x, float scale)
{
return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32 to bf8 with rounding to nearest even
template <>
inline __host__ __device__ bf8_ocp_t mxf8_convert_rne<bf8_ocp_t, float>(float x, float scale)
{
return bf8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x2 to fp8x2 with rounding to nearest even
template <>
inline __host__ __device__ f8x2_ocp_t mxf8_convert_rne<f8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return f8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x2 to bf8x2 with rounding to nearest even
template <>
inline __host__ __device__ bf8x2_ocp_t mxf8_convert_rne<bf8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return bf8x2_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret>(x, scale)};
}
// convert fp32x16 to fp8x16 with rounding to nearest even
template <>
inline __host__ __device__ f8x16_ocp_t mxf8_convert_rne<f8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
f8x16_ocp_t fp8_1x16;
f8x2_ocp_t fp8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.fp8_2x8[i] = mxf8_convert_rne<f8x2_ocp_t>(in.float_2x8[i], scale); });
return out.fp8_1x16;
}
// convert fp32x16 to bf8x16 with rounding to nearest even
template <>
inline __host__ __device__ bf8x16_ocp_t mxf8_convert_rne<bf8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.bf8_2x8[i] = mxf8_convert_rne<bf8x2_ocp_t>(in.float_2x8[i], scale); });
return out.bf8_1x16;
}
// convert fp32x32 to fp8x32 with rounding to nearest even
template <>
inline __host__ __device__ f8x32_ocp_t mxf8_convert_rne<f8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
f8x32_ocp_t fp8_1x32;
f8x16_ocp_t fp8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.fp8_16x2[i] = mxf8_convert_rne<f8x16_ocp_t>(in.float_16x2[i], scale); });
return out.fp8_1x32;
}
// convert fp32x32 to bf8x32 with rounding to nearest even
template <>
inline __host__ __device__ bf8x32_ocp_t mxf8_convert_rne<bf8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.bf8_16x2[i] = mxf8_convert_rne<bf8x16_ocp_t>(in.float_16x2[i], scale); });
return out.bf8_1x32;
}
// convert fp32 to fp8 with stochastic rounding
template <>
inline __host__ __device__ f8_ocp_t mxf8_convert_sr<f8_ocp_t, float>(float x, float scale)
{
return f8_ocp_t{fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32 to bf8 with stochastic rounding
template <>
inline __host__ __device__ bf8_ocp_t mxf8_convert_sr<bf8_ocp_t, float>(float x, float scale)
{
return bf8_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x2 to fp8x2 with stochastic rounding
template <>
inline __host__ __device__ f8x2_ocp_t mxf8_convert_sr<f8x2_ocp_t, float2_t>(float2_t x, float scale)
{
return f8x2_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<f8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x2 to bf8x2 with stochastic rounding
template <>
inline __host__ __device__ bf8x2_ocp_t mxf8_convert_sr<bf8x2_ocp_t, float2_t>(float2_t x,
float scale)
{
return bf8x2_ocp_t{
fp8_impl::cvt_float_to_fp8_scaled<bf8_ocp_t::default_interpret, true>(x, scale)};
}
// convert fp32x16 to fp8x16 with stochastic rounding
template <>
inline __host__ __device__ f8x16_ocp_t mxf8_convert_sr<f8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
f8x16_ocp_t fp8_1x16;
f8x2_ocp_t fp8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.fp8_2x8[i] = mxf8_convert_sr<f8x2_ocp_t>(in.float_2x8[i], scale); });
return out.fp8_1x16;
}
// convert fp32x16 to bf8x16 with stochastic rounding
template <>
inline __host__ __device__ bf8x16_ocp_t mxf8_convert_sr<bf8x16_ocp_t, float16_t>(float16_t x,
float scale)
{
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} in{x};
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}(
[&](auto i) { out.bf8_2x8[i] = mxf8_convert_sr<bf8x2_ocp_t>(in.float_2x8[i], scale); });
return out.bf8_1x16;
}
// convert fp32x32 to fp8x32 with stochastic rounding
template <>
inline __host__ __device__ f8x32_ocp_t mxf8_convert_sr<f8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
f8x32_ocp_t fp8_1x32;
f8x16_ocp_t fp8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.fp8_16x2[i] = mxf8_convert_sr<f8x16_ocp_t>(in.float_16x2[i], scale); });
return out.fp8_1x32;
}
// convert fp32x32 to bf8x32 with stochastic rounding
template <>
inline __host__ __device__ bf8x32_ocp_t mxf8_convert_sr<bf8x32_ocp_t, float32_t>(float32_t x,
float scale)
{
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} in{x};
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}(
[&](auto i) { out.bf8_16x2[i] = mxf8_convert_sr<bf8x16_ocp_t>(in.float_16x2[i], scale); });
return out.bf8_1x32;
}
} // namespace ck
include/ck/utility/mxfp_utils.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck::utils {
union cvt
{
float value_float;
uint32_t value_bitwise;
};
template <typename DTYPE>
inline bool getDataHasInf()
{
return DTYPE::dataInfo.hasInf;
}
template <typename T>
__host__ __device__ inline bool is_zero(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline bool is_nan(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline bool is_inf(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ inline int get_exponent_value(T x)
{
x >>= NumericUtils<T>::mant;
x &= ((1 << NumericUtils<T>::exp) - 1);
return static_cast<int>(x);
}
template <typename T>
__host__ __device__ inline bool is_subnormal(T x)
{
return get_exponent_value<T>(x) == 0;
}
template <typename T>
__host__ __device__ inline double get_mantissa_value(T x)
{
double mantissa = is_subnormal<T>(x) ? 0.0f : 1.0f;
for(uint i = 0; i < NumericUtils<T>::mant; i++)
{
mantissa += std::pow(2, -int32_t((NumericUtils<T>::mant - i))) * (x & 0b1);
x >>= 1;
}
return mantissa;
}
template <typename T>
__host__ __device__ inline bool get_data_has_inf()
{
return NumericUtils<T>::has_inf;
}
template <typename T>
__host__ __device__ float convert_to_float(T data, int scale_exp)
{
float d_sign =
std::pow(-1, static_cast<float>(data >> (NumericUtils<T>::exp + NumericUtils<T>::mant)));
float d_exp;
if(is_subnormal<T>(data))
d_exp = std::pow(2, 1 - static_cast<int>(NumericUtils<T>::bias));
else
d_exp = std::pow(2, get_exponent_value<T>(data) - static_cast<int>(NumericUtils<T>::bias));
float d_mant = get_mantissa_value<T>(data);
float data_value = d_sign * d_exp * d_mant;
float scale_value = std::pow(
2, static_cast<float>((scale_exp - static_cast<int>(NumericUtils<e8m0_bexp_t>::bias))));
return data_value * scale_value;
}
template <typename T>
__host__ __device__ inline float to_float(e8m0_bexp_t const scale, T const data);
template <typename T>
__host__ __device__ T sat_convert_to_type(float value);
template <typename T>
__host__ __device__ T sat_convert_to_type_sr(float value, uint32_t seed);
template <typename T>
inline T convert_to_type(float value)
{
using bitwise_type = typename NumericUtils<T>::bitwise_type;
if(std::abs(value) > NumericLimits<T>::Max())
{
float max_value = NumericLimits<T>::Max();
cvt t;
// cppcheck-suppress redundantAssignment
t.value_float = max_value;
uint32_t max_bitwise = t.value_bitwise;
// cppcheck-suppress redundantAssignment
t.value_float = value;
bitwise_type sign =
t.value_bitwise >> (NumericUtils<float>::exp + NumericUtils<float>::mant);
bitwise_type exp =
((max_bitwise >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask) -
(NumericUtils<float>::bias - NumericUtils<T>::bias);
bitwise_type mantissa = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t mant_prev = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant_prev &= ((1 << NumericUtils<T>::mant) - 1);
mant_prev--;
mant_prev <<= (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t prev_bit =
((max_bitwise >> NumericUtils<float>::mant) << NumericUtils<float>::mant) | mant_prev;
t.value_bitwise = prev_bit;
float prev_val = t.value_float;
float diff = max_value - prev_val;
float actual_max = max_value + (diff / 2);
if(std::abs(value) < actual_max)
{
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(exp << NumericUtils<T>::mant) | mantissa;
}
else
{
if(!get_data_has_inf<T>())
{
return (1 << (NumericUtils<T>::mant + NumericUtils<T>::exp)) - 1;
}
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(exp << NumericUtils<T>::mant);
}
}
}
const int mfmt = NumericUtils<float>::mant;
uint32_t x;
x = bit_cast<uint32_t>(value);
uint32_t head, mantissa;
int32_t exponent, bias;
uint32_t sign;
head = x & NumericUtils<float>::head_mask;
mantissa = x & NumericUtils<float>::mant_mask;
exponent = (head >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask;
sign = head >> (NumericUtils<float>::mant + NumericUtils<float>::exp);
bias = NumericUtils<float>::bias;
if(x == 0)
{
return 0b0;
}
const int mini_bias = NumericUtils<T>::bias;
const int mini_denormal_act_exponent = 1 - mini_bias;
int act_exponent, out_exponent, exponent_diff;
bool is_subnorm = false;
if(exponent == 0)
{
act_exponent = exponent - bias + 1;
exponent_diff = mini_denormal_act_exponent - act_exponent;
is_subnorm = true;
}
else
{
act_exponent = exponent - bias;
if(act_exponent <= mini_denormal_act_exponent)
{
exponent_diff = mini_denormal_act_exponent - act_exponent;
is_subnorm = true;
}
else
{
exponent_diff = 0;
}
mantissa += (1UL << mfmt);
}
auto shift_amount = (mfmt - NumericUtils<T>::mant + exponent_diff);
shift_amount = (shift_amount >= 64) ? 63 : shift_amount;
bool midpoint = (mantissa & ((1UL << shift_amount) - 1)) == (1UL << (shift_amount - 1));
float min_subnorm = NumericLimits<T>::DataMinSubnorm() * (sign ? -1 : 1);
if(is_subnorm && std::abs(value) < std::abs(min_subnorm))
{
// closer to 0
if(std::abs(value) <= std::abs(min_subnorm - value))
return 0;
else
return 1 | (sign << (NumericUtils<T>::exp + NumericUtils<T>::mant));
}
if(exponent_diff > 0)
mantissa >>= exponent_diff;
else if(exponent_diff == -1)
mantissa <<= -exponent_diff;
bool implicit_one = mantissa & (1 << mfmt);
out_exponent = (act_exponent + exponent_diff) + mini_bias - (implicit_one ? 0 : 1);
uint32_t drop_mask = (1UL << (mfmt - NumericUtils<T>::mant)) - 1;
bool odd = mantissa & (1UL << (mfmt - NumericUtils<T>::mant));
mantissa += (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa) & drop_mask;
if(out_exponent == 0)
{
if((1UL << mfmt) & mantissa)
{
out_exponent = 1;
}
}
else
{
if((1UL << (mfmt + 1)) & mantissa)
{
mantissa >>= 1;
out_exponent++;
}
}
mantissa >>= (mfmt - NumericUtils<T>::mant);
if(out_exponent == 0 && mantissa == 0)
{
return 0;
}
mantissa &= (1UL << NumericUtils<T>::mant) - 1;
return (sign << (NumericUtils<T>::exp + NumericUtils<T>::mant)) |
(out_exponent << NumericUtils<T>::mant) | mantissa;
}
template <typename T>
inline T convert_to_type_sr(float value, uint32_t seed)
{
if(std::abs(value) > NumericLimits<T>::Max())
{
float max_value = NumericLimits<T>::Max();
cvt t;
// cppcheck-suppress redundantAssignment
t.value_float = max_value;
uint max_bitwise = t.value_bitwise;
// cppcheck-suppress redundantAssignment
t.value_float = value;
T sign = t.value_bitwise >> (NumericUtils<float>::exp + NumericUtils<float>::mant);
T exp = ((max_bitwise >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask) -
(NumericUtils<float>::bias - NumericUtils<T>::bias);
uint32_t mant_prev = max_bitwise >> (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant_prev &= ((1UL << NumericUtils<T>::mant) - 1);
mant_prev--;
mant_prev <<= (NumericUtils<float>::mant - NumericUtils<T>::mant);
uint32_t prev_bit =
((max_bitwise >> NumericUtils<float>::mant) << NumericUtils<float>::mant) | mant_prev;
t.value_bitwise = prev_bit;
float prev_val = t.value_float;
float diff = max_value - prev_val;
float actual_max = max_value + (diff / 2);
if(std::abs(value) < actual_max)
{
double d_max_value = static_cast<double>(max_value);
double d_actual_max = static_cast<double>(actual_max);
double d_value = static_cast<double>(value);
double d_is = std::abs(d_max_value - d_actual_max);
double d_seed = static_cast<double>(seed);
double d_prob = 1.0f - (std::abs(d_value - d_max_value) / d_is); // prob to round down
double thresh = UINT_MAX * d_prob;
if(!get_data_has_inf<T>() || d_seed <= thresh)
// return static_cast<T>(satConvertToType(getDataMax<DTYPE>())); //round down time
return sign == 0 ? NumericUtils<f4_t>::data_max_positive_normal_mask
: NumericUtils<f4_t>::data_max_negative_normal_mask;
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) // inf
| (exp << NumericUtils<T>::mant);
}
}
else
{
if(!get_data_has_inf<T>())
return (1 << (NumericUtils<T>::mant + NumericUtils<T>::exp)) - 1;
else
{
exp++;
return sign << ((NumericUtils<T>::exp + NumericUtils<T>::mant)) // inf
| (exp << NumericUtils<T>::mant);
}
}
}
uint32_t f32 = bit_cast<uint32_t>(value);
auto f32_mant = f32 & NumericUtils<float>::mant_mask;
auto head = f32 & NumericUtils<float>::head_mask;
auto f32_exp = (head >> NumericUtils<float>::mant) & NumericUtils<float>::exp_mask;
auto sign_bit = head >> (NumericUtils<float>::mant + NumericUtils<float>::exp);
auto sign = sign_bit << (NumericUtils<T>::exp + NumericUtils<T>::mant);
f32_exp = static_cast<int32_t>(f32_exp) - NumericUtils<float>::bias;
int32_t exp = f32_exp;
auto mant = f32_mant;
bool subnorm = false;
if(f32 == 0)
return 0b0;
if(exp >= NumericUtils<T>::unbiased_exp_min)
{
mant = f32_mant;
}
// if the exponent bit is 8, then the subnormal is exactly the same as f32
else if(exp < NumericUtils<T>::unbiased_exp_min &&
NumericUtils<T>::exp < NumericUtils<float>::exp)
{
subnorm = true;
auto diff = static_cast<uint32_t>(NumericUtils<T>::unbiased_exp_min - exp);
if(diff >= 32)
{
mant = 0;
f32_mant = 0;
}
else
{
f32_mant |= static_cast<uint32_t>(1) << NumericUtils<float>::mant;
f32_mant >>= diff;
}
exp = 0;
mant = f32_mant;
}
uint32_t sr_shift = NumericUtils<T>::sr_shift;
// For stochastic-rounding we add the aligned random value to the
// mantissa and then truncate (RTZ).
mant += seed >> sr_shift;
// Increment exponent when mantissa overflows due to rounding
if(mant >= static_cast<uint32_t>(1) << NumericUtils<float>::mant)
++exp;
mant >>= (NumericUtils<float>::mant - NumericUtils<T>::mant);
mant &= ((1 << NumericUtils<T>::mant) - 1);
auto biased_exp = static_cast<uint32_t>(exp);
if(!subnorm)
biased_exp = static_cast<uint32_t>(exp + NumericUtils<T>::bias);
biased_exp &= ((1 << NumericUtils<T>::exp) - 1);
auto val = sign | biased_exp << NumericUtils<T>::mant | mant;
return val;
}
} // namespace ck::utils
include/ck/utility/random_gen.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include <ck/utility/ignore.hpp>
#include "ck/ck.hpp" #include "ck/ck.hpp"
#ifdef CK_CODE_GEN_RTC
using uint8_t = unsigned char;
using uint16_t = unsigned short;
using uint32_t = unsigned int;
#endif
namespace ck { namespace ck {
// Pseudo random number generator // Pseudo random number generator
// version for fp32 // version for fp32
template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<float, T>{}, bool> = false> template <typename T, uint32_t seed_t, ck::enable_if_t<std::is_same<float, T>{}, bool> = false>
__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t) __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
{ {
uint32_t x = *(reinterpret_cast<uint32_t*>(&val)); uint32_t x = *(reinterpret_cast<uint32_t*>(&val));
...@@ -25,7 +30,7 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = ...@@ -25,7 +30,7 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
} }
// version for fp16 // version for fp16
template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<_Float16, T>{}, bool> = false> template <typename T, uint32_t seed_t, ck::enable_if_t<std::is_same<_Float16, T>{}, bool> = false>
__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t) __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
{ {
uint16_t x = *(reinterpret_cast<uint16_t*>(&val)); uint16_t x = *(reinterpret_cast<uint16_t*>(&val));
...@@ -40,15 +45,14 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = ...@@ -40,15 +45,14 @@ __host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed =
} }
// return 0 if data is not fp16 or fp32 // return 0 if data is not fp16 or fp32
template < template <typename T,
typename T, uint32_t seed_t,
uint32_t seed_t, ck::enable_if_t<!(std::is_same<float, T>{} || std::is_same<_Float16, T>{}), bool> = false>
std::enable_if_t<!(std::is_same<float, T>{} || std::is_same<_Float16, T>{}), bool> = false>
__host__ __device__ uint32_t prand_generator(int id, T val, uint32_t seed = seed_t) __host__ __device__ uint32_t prand_generator(int id, T val, uint32_t seed = seed_t)
{ {
std::ignore = id; ck::ignore = id;
std::ignore = val; ck::ignore = val;
std::ignore = seed; ck::ignore = seed;
return 0; return 0;
} }
......
include/ck/utility/scaled_type_convert.hpp 0 → 100644
View file @ ec959387
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/type_convert.hpp"
#include "ck/utility/mxf8_utils.hpp"
#ifdef CK_USE_NATIVE_MX_SUPPORT
#define CK_USE_NATIVE_MX_SUPPORT 1
#else
#define CK_USE_NATIVE_MX_SUPPORT 0
#endif
namespace ck {
// Declare a template function for scaled conversion
template <typename Y, typename X>
#if CK_USE_OCP_FP8
__host__ __device__ constexpr Y scaled_type_convert(e8m0_bexp_t scale, X x);
#else
__host__ constexpr Y scaled_type_convert(e8m0_bexp_t scale, X x);
#endif
// convert f8_ocp_t to fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float scaled_type_convert<float, f8_ocp_t>(e8m0_bexp_t scale, f8_ocp_t x)
#else
inline __host__ float scaled_type_convert<float, f8_ocp_t>(e8m0_bexp_t scale, f8_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32_from_f8_scaled<f8_ocp_t::default_interpret>(
type_convert<float>(scale), x.data);
#else
return type_convert<float>(scale) * type_convert<float>(x);
#endif
}
// convert bf8_ocp_t to fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float scaled_type_convert<float, bf8_ocp_t>(e8m0_bexp_t scale,
bf8_ocp_t x)
#else
inline __host__ float scaled_type_convert<float, bf8_ocp_t>(e8m0_bexp_t scale, bf8_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32_from_f8_scaled<bf8_ocp_t::default_interpret>(
type_convert<float>(scale), x.data);
#else
return type_convert<float>(scale) * type_convert<float>(x);
#endif
}
// convert 2 x f8_ocp_t to 2 x fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float2_t scaled_type_convert<float2_t, f8x2_ocp_t>(e8m0_bexp_t scale,
f8x2_ocp_t x)
#else
inline __host__ float2_t scaled_type_convert<float2_t, f8x2_ocp_t>(e8m0_bexp_t scale, f8x2_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32x2_from_f8x2_scaled<f8_ocp_t::default_interpret>(
type_convert<float>(scale), x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
#else
return float2_t{scaled_type_convert<float>(scale, x.AsType<f8_ocp_t>()[Number<0>{}]),
scaled_type_convert<float>(scale, x.AsType<f8_ocp_t>()[Number<1>{}])};
#endif
}
// convert 2 x bf8_ocp_t to 2 x fp32
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float2_t scaled_type_convert<float2_t, bf8x2_ocp_t>(e8m0_bexp_t scale,
bf8x2_ocp_t x)
#else
inline __host__ float2_t scaled_type_convert<float2_t, bf8x2_ocp_t>(e8m0_bexp_t scale,
bf8x2_ocp_t x)
#endif
{
#if CK_MX_FP8_CVT_FAST_PATH
return fp8_impl::cast_to_f32x2_from_f8x2_scaled<bf8_ocp_t::default_interpret>(
type_convert<float>(scale), x.AsType<fp8_impl::fp8x2_storage_t>()[Number<0>{}]);
#else
return float2_t{scaled_type_convert<float>(scale, x.AsType<bf8_ocp_t>()[Number<0>{}]),
scaled_type_convert<float>(scale, x.AsType<bf8_ocp_t>()[Number<1>{}])};
#endif
}
// convert 16 x f8_ocp_t to 16 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float16_t scaled_type_convert<float16_t, f8x16_ocp_t>(e8m0_bexp_t scale,
f8x16_ocp_t x)
#else
inline __host__ float16_t scaled_type_convert<float16_t, f8x16_ocp_t>(e8m0_bexp_t scale,
f8x16_ocp_t x)
#endif
{
union
{
f8x16_ocp_t f8_1x16;
f8x2_ocp_t f8_2x8[8];
} in{x};
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}([&](auto i) {
out.float_2x8[i] = scaled_type_convert<float2_t, f8x2_ocp_t>(scale, in.f8_2x8[i]);
});
return out.float_1x16;
}
// convert 16 x bf8_ocp_t to 16 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float16_t scaled_type_convert<float16_t, bf8x16_ocp_t>(e8m0_bexp_t scale,
bf8x16_ocp_t x)
#else
inline __host__ float16_t scaled_type_convert<float16_t, bf8x16_ocp_t>(e8m0_bexp_t scale,
bf8x16_ocp_t x)
#endif
{
union
{
bf8x16_ocp_t bf8_1x16;
bf8x2_ocp_t bf8_2x8[8];
} in{x};
union
{
float16_t float_1x16;
float2_t float_2x8[8];
} out{};
ck::static_for<0, 8, 1>{}([&](auto i) {
out.float_2x8[i] = scaled_type_convert<float2_t, bf8x2_ocp_t>(scale, in.bf8_2x8[i]);
});
return out.float_1x16;
}
// convert 32 x f8_ocp_t to 32 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f8x32_ocp_t>(e8m0_bexp_t scale,
f8x32_ocp_t x)
#else
inline __host__ float32_t scaled_type_convert<float32_t, f8x32_ocp_t>(e8m0_bexp_t scale,
f8x32_ocp_t x)
#endif
{
union
{
f8x32_ocp_t f8_1x32;
f8x16_ocp_t f8_16x2[2];
} in{x};
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}([&](auto i) {
out.float_16x2[i] = scaled_type_convert<float16_t, f8x16_ocp_t>(scale, in.f8_16x2[i]);
});
return out.float_1x32;
}
// convert 32 x bf8_ocp_t to 32 x fp32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ float32_t scaled_type_convert<float32_t, bf8x32_ocp_t>(e8m0_bexp_t scale,
bf8x32_ocp_t x)
#else
inline __host__ float32_t scaled_type_convert<float32_t, bf8x32_ocp_t>(e8m0_bexp_t scale,
bf8x32_ocp_t x)
#endif
{
union
{
bf8x32_ocp_t bf8_1x32;
bf8x16_ocp_t bf8_16x2[2];
} in{x};
union
{
float32_t float_1x32;
float16_t float_16x2[2];
} out{};
ck::static_for<0, 2, 1>{}([&](auto i) {
out.float_16x2[i] = scaled_type_convert<float16_t, bf8x16_ocp_t>(scale, in.bf8_16x2[i]);
});
return out.float_1x32;
}
// convert fp32 to fp8
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8_ocp_t scaled_type_convert<f8_ocp_t, float>(e8m0_bexp_t scale, float x)
#else
inline __host__ f8_ocp_t scaled_type_convert<f8_ocp_t, float>(e8m0_bexp_t scale, float x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32 to bf8
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8_ocp_t scaled_type_convert<bf8_ocp_t, float>(e8m0_bexp_t scale,
float x)
#else
inline __host__ bf8_ocp_t scaled_type_convert<bf8_ocp_t, float>(e8m0_bexp_t scale, float x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x2 to fp8x2
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x2_ocp_t scaled_type_convert<f8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#else
inline __host__ f8x2_ocp_t scaled_type_convert<f8x2_ocp_t, float2_t>(e8m0_bexp_t scale, float2_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x2_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x2_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x2 to bf8x2
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x2_ocp_t scaled_type_convert<bf8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#else
inline __host__ bf8x2_ocp_t scaled_type_convert<bf8x2_ocp_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x2_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x2_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x16 to fp8x16
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x16_ocp_t
scaled_type_convert<f8x16_ocp_t, float16_t>(e8m0_bexp_t scale, float16_t x)
#else
inline __host__ f8x16_ocp_t scaled_type_convert<f8x16_ocp_t, float16_t>(e8m0_bexp_t scale,
float16_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x16_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x16_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x16 to bf8x16
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x16_ocp_t
scaled_type_convert<bf8x16_ocp_t, float16_t>(e8m0_bexp_t scale, float16_t x)
#else
inline __host__ bf8x16_ocp_t scaled_type_convert<bf8x16_ocp_t, float16_t>(e8m0_bexp_t scale,
float16_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x16_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x16_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x32 to fp8x32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ f8x32_ocp_t
scaled_type_convert<f8x32_ocp_t, float32_t>(e8m0_bexp_t scale, float32_t x)
#else
inline __host__ f8x32_ocp_t scaled_type_convert<f8x32_ocp_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<f8x32_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<f8x32_ocp_t>(x, type_convert<float>(scale));
#endif
}
// convert fp32x32 to bf8x32
// @note Host version gives compilation error. Requires extra compiler options.
template <>
#if CK_USE_OCP_FP8
inline __host__ __device__ bf8x32_ocp_t
scaled_type_convert<bf8x32_ocp_t, float32_t>(e8m0_bexp_t scale, float32_t x)
#else
inline __host__ bf8x32_ocp_t scaled_type_convert<bf8x32_ocp_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
#endif
{
#if CK_USE_SR_F8_CONVERSION
return mxf8_convert_sr<bf8x32_ocp_t>(x, type_convert<float>(scale));
#else
return mxf8_convert_rne<bf8x32_ocp_t>(x, type_convert<float>(scale));
#endif
}
// activate for architectures with native MX support
#if CK_USE_NATIVE_MX_SUPPORT
// convert fp4 to fp32
template <>
inline __host__ __device__ float scaled_type_convert<float, f4_t>(e8m0_bexp_t scale, f4_t x)
{
#if defined(__gfx950__)
union
{
float float_array[2];
float2_t float2_array;
} float_values{};
float_values.float2_array =
__builtin_amdgcn_cvt_scalef32_pk_f32_fp4(x, type_convert<float>(scale), 0);
return float_values.float_array[0];
#else
return utils::to_float<f4_t>(scale, x);
#endif
}
// convert vector of 2 fp4 to vector of 2 fp32
template <>
inline __host__ __device__ float2_t scaled_type_convert<float2_t, f4x2_t>(e8m0_bexp_t scale,
f4x2_t x)
{
#if defined(__gfx950__)
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} value{};
value.f4x2_array[0] = x;
return __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(value.bitwise, type_convert<float>(scale), 0);
#else
float2_t ret{utils::to_float<f4_t>(
scale, x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{})),
utils::to_float<f4_t>(
scale, x.template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}))};
return ret;
#endif
}
// convert vector of 32 fp4 to vector of 32 fp32
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f4x32_t>(e8m0_bexp_t scale,
f4x32_t x)
{
#if defined(__gfx950__)
union
{
f4x32_t f4x32_array;
f4x2_t fp4x2[16];
} value{x};
union
{
uint32_t bitwise;
f4x2_t f4x2_array[4];
} bitwise_value{};
float2_t op;
float32_t ret;
// TODO: pack in a loop
bitwise_value.f4x2_array[0] = value.fp4x2[0];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[0] = op[0];
ret[1] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[1];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[2] = op[0];
ret[3] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[2];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[4] = op[0];
ret[5] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[3];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[6] = op[0];
ret[7] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[4];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[8] = op[0];
ret[9] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[5];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[10] = op[0];
ret[11] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[6];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[12] = op[0];
ret[13] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[7];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[14] = op[0];
ret[15] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[8];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[16] = op[0];
ret[17] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[9];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[18] = op[0];
ret[19] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[10];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[20] = op[0];
ret[21] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[11];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[22] = op[0];
ret[23] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[12];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[24] = op[0];
ret[25] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[13];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[26] = op[0];
ret[27] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[14];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[28] = op[0];
ret[29] = op[1];
bitwise_value.f4x2_array[0] = value.fp4x2[15];
op = __builtin_amdgcn_cvt_scalef32_pk_f32_fp4(
bitwise_value.bitwise, type_convert<float>(scale), 0);
ret[30] = op[0];
ret[31] = op[1];
return ret;
#else
union
{
float32_t float32_array;
float float_array[32];
} float_values{};
union
{
__uint128_t bitwise;
f4x2_t f4x2_array[16];
f4x32_t f4x32_array;
} f4_values{bit_cast<__uint128_t>(x)};
// TODO: pack in a loop
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[0].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[1].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[2].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[3].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[4].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[5].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[6].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[7].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[8].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[9].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[10].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[11].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[0] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[1] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[12].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[2] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[3] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[13].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[4] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[5] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[14].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
float_values.float_array[6] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<0>{}));
float_values.float_array[7] = utils::to_float<f4_t>(
scale,
f4_values.f4x2_array[15].template AsType<f4x2_pk_t>()[Number<0>{}].unpack<>(Number<1>{}));
return float_values.float32_array;
#endif
}
// convert fp32 to fp4
template <>
inline __host__ __device__ f4_t scaled_type_convert<f4_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
// convert vector of 2 fp32 to vector of 2 fp4
template <>
inline __host__ __device__ f4x2_t scaled_type_convert<f4x2_t, float2_t>(e8m0_bexp_t scale,
float2_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
// convert vector of 32 fp32 to vector of 32 fp4
template <>
inline __host__ __device__ f4x32_t scaled_type_convert<f4x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F4_CONVERSION
return f4_convert_sr(x, type_convert<float>(scale));
#else
return f4_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a 6-bit floating-point value (f6_t) to a 32-bit float,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param x The f6_t value to be converted.
* @return The converted 32-bit float representation of the input.
*/
template <>
inline __host__ __device__ float scaled_type_convert<float, f6_t>(e8m0_bexp_t scale, f6_t x)
{
#if defined(__gfx950__)
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector =
__builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(in.f6_vector, type_convert<float>(scale));
return out.float_array[0];
#else
return utils::to_float<f6_t>(scale, x);
#endif
}
/**
* @brief Converts a vector of 32 6-bit floating-point values (f6x32_t) to a vector of 32 floats,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The f6x32_t vector to be converted.
* @return The converted float vector representation of the input.
*/
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, f6x32_t>(e8m0_bexp_t scale,
f6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(x, type_convert<float>(scale));
#else
union
{
f6x32_t f6_vector;
f6_t f6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}(
[&](auto i) { out.float_array[i] = utils::to_float<f6_t>(scale, in.f6_array[i]); });
return out.float_vector;
#endif
}
/**
* @brief Converts a 6-bit floating-point value (bf6_t) to a 32-bit float,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param x The bf6_t value to be converted.
* @return The converted 32-bit float representation of the input.
*/
template <>
inline __host__ __device__ float scaled_type_convert<float, bf6_t>(e8m0_bexp_t scale, bf6_t x)
{
#if defined(__gfx950__)
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
out.float_vector =
__builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(in.bf6_vector, type_convert<float>(scale));
return out.float_array[0];
#else
return utils::to_float<bf6_t>(scale, x);
#endif
}
/**
* @brief Converts a vector of 6-bit floating-point values (bf6x32_t) to a vector of 32 floats,
* applying the specified scaling factor.
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The bf6x32_t vector to be converted.
* @return The converted vector of 32 float representation of the input.
*/
template <>
inline __host__ __device__ float32_t scaled_type_convert<float32_t, bf6x32_t>(e8m0_bexp_t scale,
bf6x32_t x)
{
#if defined(__gfx950__)
return __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(x, type_convert<float>(scale));
#else
union
{
bf6x32_t bf6_vector;
bf6_t bf6_array[32];
} in{x};
union
{
float32_t float_vector;
float float_array[32];
} out{};
ck::static_for<0, 32, 1>{}(
[&](auto i) { out.float_array[i] = utils::to_float<bf6_t>(scale, in.bf6_array[i]); });
return out.float_vector;
#endif
}
/**
* @brief Converts a 32-bit float to a 6-bit floating-point value (f6_t), applying the specified
* scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (f6_convert_sr) or round-to-nearest-even (f6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for f6_t.
* @param x The float value to convert.
* @return The converted 6-bit floating-point value (f6_t).
*/
template <>
inline __host__ __device__ f6_t scaled_type_convert<f6_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x, type_convert<float>(scale));
#else
return f6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a vector of 32 floats to a vector of 32 6-bit floating-point values (f6x32_t),
* applying the specified scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (f6_convert_sr) or round-to-nearest-even (f6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The float vector to convert.
* @return The converted vector of 6-bit floating-point values (f6x32_t).
*/
template <>
inline __host__ __device__ f6x32_t scaled_type_convert<f6x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return f6_convert_sr(x, type_convert<float>(scale));
#else
return f6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a 32-bit float to a 6-bit floating-point value (bf6_t), applying the specified
* scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (bf6_convert_sr) or round-to-nearest-even (bf6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t) used for bf6_t.
* @param x The float value to convert.
* @return The converted 6-bit floating-point value (bf6_t).
*/
template <>
inline __host__ __device__ bf6_t scaled_type_convert<bf6_t, float>(e8m0_bexp_t scale, float x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x, type_convert<float>(scale));
#else
return bf6_convert_rne(x, type_convert<float>(scale));
#endif
}
/**
* @brief Converts a vector of 32 floats to a vector of 32 6-bit floating-point values (bf6x32_t),
* applying the specified scale.
*
* Depending on whether CK_USE_SR_F6_CONVERSION is defined, it uses either stochastic rounding
* (bf6_convert_sr) or round-to-nearest-even (bf6_convert_rne).
*
* @param scale The exponent scale factor (e8m0_bexp_t).
* @param x The float vector to convert.
* @return The converted 6-bit floating-point vector (bf6x32_t).
*/
template <>
inline __host__ __device__ bf6x32_t scaled_type_convert<bf6x32_t, float32_t>(e8m0_bexp_t scale,
float32_t x)
{
#if CK_USE_SR_F6_CONVERSION
return bf6_convert_sr(x, type_convert<float>(scale));
#else
return bf6_convert_rne(x, type_convert<float>(scale));
#endif
}
#endif // #if CK_USE_NATIVE_MX_SUPPORT
} // namespace ck
include/ck/utility/sequence.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#ifndef CK_CODE_GEN_RTC
#include <ostream> #include <ostream>
#endif
#include "ck/utility/integral_constant.hpp" #include "ck/utility/integral_constant.hpp"
#include "ck/utility/type.hpp" #include "ck/utility/type.hpp"
...@@ -900,6 +902,7 @@ using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type; ...@@ -900,6 +902,7 @@ using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
} // namespace ck } // namespace ck
#ifndef CK_CODE_GEN_RTC
template <ck::index_t... Is> template <ck::index_t... Is>
std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>) std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
{ {
...@@ -910,3 +913,4 @@ std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>) ...@@ -910,3 +913,4 @@ std::ostream& operator<<(std::ostream& os, const ck::Sequence<Is...>)
os << S::At(S::Size() - ck::Number<1>{}).value << "}"; os << S::At(S::Size() - ck::Number<1>{}).value << "}";
return os; return os;
} }
#endif
include/ck/utility/statically_indexed_array_multi_index.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP #ifndef CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
#define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP #define CK_STATICALLY_INDEXED_ARRAY_MULTI_INDEX_HPP
...@@ -35,10 +35,9 @@ __host__ __device__ constexpr auto to_multi_index(const T& x) ...@@ -35,10 +35,9 @@ __host__ __device__ constexpr auto to_multi_index(const T& x)
// is the alias of the latter. This is because compiler cannot infer the NSize if // is the alias of the latter. This is because compiler cannot infer the NSize if
// using MultiIndex<NSize> // using MultiIndex<NSize>
// TODO: how to fix this? // TODO: how to fix this?
template < template <typename... Ys,
typename... Ys, typename X,
typename X, enable_if_t<!ck::is_integral<X>::value && !ck::is_floating_point<X>::value, bool> = false>
enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
__host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x) __host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x)
{ {
static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same"); static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
...@@ -47,10 +46,9 @@ __host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x) ...@@ -47,10 +46,9 @@ __host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x)
return y; return y;
} }
template < template <typename... Ys,
typename... Ys, typename X,
typename X, enable_if_t<!ck::is_integral<X>::value && !ck::is_floating_point<X>::value, bool> = false>
enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
__host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x) __host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x)
{ {
static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same"); static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
...@@ -59,10 +57,9 @@ __host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x) ...@@ -59,10 +57,9 @@ __host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x)
return y; return y;
} }
template < template <typename... Xs,
typename... Xs, typename Y,
typename Y, enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y) __host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y)
{ {
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same"); static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
...@@ -73,10 +70,9 @@ __host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y) ...@@ -73,10 +70,9 @@ __host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y)
return r; return r;
} }
template < template <typename... Xs,
typename... Xs, typename Y,
typename Y, enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y) __host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y)
{ {
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same"); static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
...@@ -87,10 +83,9 @@ __host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y) ...@@ -87,10 +83,9 @@ __host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y)
return r; return r;
} }
template < template <typename... Xs,
typename... Xs, typename Y,
typename Y, enable_if_t<!ck::is_integral<Y>::value && !ck::is_floating_point<Y>::value, bool> = false>
enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y) __host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y)
{ {
static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same"); static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
...@@ -104,7 +99,7 @@ __host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y) ...@@ -104,7 +99,7 @@ __host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y)
// MultiIndex = scalar * MultiIndex // MultiIndex = scalar * MultiIndex
template <typename... Xs, template <typename... Xs,
typename Y, typename Y,
enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false> enable_if_t<ck::is_integral<Y>::value || ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x) __host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x)
{ {
constexpr index_t NSize = sizeof...(Xs); constexpr index_t NSize = sizeof...(Xs);
...@@ -117,7 +112,7 @@ __host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x) ...@@ -117,7 +112,7 @@ __host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x)
// MultiIndex = MultiIndex * scalar // MultiIndex = MultiIndex * scalar
template <typename... Xs, template <typename... Xs,
typename Y, typename Y,
enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false> enable_if_t<ck::is_integral<Y>::value || ck::is_floating_point<Y>::value, bool> = false>
__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, Y a) __host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, Y a)
{ {
return a * x; return a * x;
......
include/ck/utility/tuple.hpp
View file @ ec959387
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
...@@ -32,7 +32,7 @@ struct TupleElementKeyData ...@@ -32,7 +32,7 @@ struct TupleElementKeyData
template <typename T, template <typename T,
typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value, typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(std::forward<T>(v)) __host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(ck::forward<T>(v))
{ {
} }
...@@ -67,7 +67,7 @@ get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x) ...@@ -67,7 +67,7 @@ get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x)
template <typename Key, typename Data> template <typename Key, typename Data>
__host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x) __host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
{ {
return std::forward(x.mData); return ck::forward(x.mData);
} }
template <typename Indices, typename... Xs> template <typename Indices, typename... Xs>
...@@ -83,13 +83,13 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<I ...@@ -83,13 +83,13 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<I
!is_same<remove_cvref_t<Y>, TupleImpl>::value, !is_same<remove_cvref_t<Y>, TupleImpl>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr TupleImpl(Y&& y) __host__ __device__ constexpr TupleImpl(Y&& y)
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Y>(y))... : TupleElementKeyData<TupleElementKey<Is>, Xs>(ck::forward<Y>(y))...
{ {
} }
template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false> template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
__host__ __device__ constexpr TupleImpl(Ys&&... ys) __host__ __device__ constexpr TupleImpl(Ys&&... ys)
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))... : TupleElementKeyData<TupleElementKey<Is>, Xs>(ck::forward<Ys>(ys))...
{ {
static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys), static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys),
"wrong! inconsistent size"); "wrong! inconsistent size");
...@@ -123,14 +123,14 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X ...@@ -123,14 +123,14 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
template <typename Y, template <typename Y,
typename enable_if<sizeof...(Xs) == 1 && !is_same<remove_cvref_t<Y>, Tuple>::value, typename enable_if<sizeof...(Xs) == 1 && !is_same<remove_cvref_t<Y>, Tuple>::value,
bool>::type = false> bool>::type = false>
__host__ __device__ constexpr Tuple(Y&& y) : base(std::forward<Y>(y)) __host__ __device__ constexpr Tuple(Y&& y) : base(ck::forward<Y>(y))
{ {
} }
template <typename... Ys, template <typename... Ys,
typename enable_if<sizeof...(Ys) == sizeof...(Xs) && sizeof...(Ys) >= 2, bool>::type = typename enable_if<sizeof...(Ys) == sizeof...(Xs) && sizeof...(Ys) >= 2, bool>::type =
false> false>
__host__ __device__ constexpr Tuple(Ys&&... ys) : base(std::forward<Ys>(ys)...) __host__ __device__ constexpr Tuple(Ys&&... ys) : base(ck::forward<Ys>(ys)...)
{ {
} }
...@@ -210,7 +210,7 @@ using tuple_element_t = typename tuple_element<I, TTuple>::type; ...@@ -210,7 +210,7 @@ using tuple_element_t = typename tuple_element<I, TTuple>::type;
template <typename... Xs> template <typename... Xs>
__host__ __device__ constexpr auto make_tuple(Xs&&... xs) __host__ __device__ constexpr auto make_tuple(Xs&&... xs)
{ {
return Tuple<remove_cvref_t<Xs>...>(std::forward<Xs>(xs)...); return Tuple<remove_cvref_t<Xs>...>(ck::forward<Xs>(xs)...);
} }
// https://en.cppreference.com/w/cpp/utility/tuple/tie // https://en.cppreference.com/w/cpp/utility/tuple/tie
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment