1
0
Fork 0
mirror of https://github.com/juce-framework/JUCE.git synced 2026-01-10 23:44:24 +00:00
JUCE/modules/juce_dsp/containers/juce_SIMDRegister.h

455 lines
22 KiB
C++

/*
==============================================================================
This file is part of the JUCE library.
Copyright (c) 2017 - ROLI Ltd.
JUCE is an open source library subject to commercial or open-source
licensing.
By using JUCE, you agree to the terms of both the JUCE 5 End-User License
Agreement and JUCE 5 Privacy Policy (both updated and effective as of the
27th April 2017).
End User License Agreement: www.juce.com/juce-5-licence
Privacy Policy: www.juce.com/juce-5-privacy-policy
Or: You may also use this code under the terms of the GPL v3 (see
www.gnu.org/licenses).
JUCE IS PROVIDED "AS IS" WITHOUT ANY WARRANTY, AND ALL WARRANTIES, WHETHER
EXPRESSED OR IMPLIED, INCLUDING MERCHANTABILITY AND FITNESS FOR PURPOSE, ARE
DISCLAIMED.
==============================================================================
*/
namespace juce
{
namespace dsp
{
#ifndef DOXYGEN
// This class is needed internally.
template <typename Scalar>
struct CmplxSIMDOps;
#endif
//==============================================================================
/**
A wrapper around the platform's native SIMD register type.
This class is only availabe on SIMD machines. Use JUCE_USE_SIMD to query
if SIMD is avaialble for your system.
SIMDRegister<Type> is a templated class representing the native
vectorized version of FloatingType. SIMDRegister supports all numerical
primitive types and std:complex<float> and std::complex<double> supports
and most operations of the corresponding primitive
type. Additionally, SIMDRegister can be accessed like an array to extract
the individual elements.
If you are using SIMDRegister as a pointer, then you must ensure that the
memory is suffeciently aligned for SIMD vector operations. Failing to do so
will result in crashes or very slow code. Use SIMDRegister::isSIMDAligned
to query if a pointer is suffeciently aligned for SIMD vector operations.
Note that using SIMDRegister without enabling optimizations will result
in code with very poor performance.
*/
template <typename Type>
struct SIMDRegister
{
//==============================================================================
/** The type that represents the individual constituents of the SIMD Register */
typedef Type ElementType;
/** STL compatible value_type definition (same as ElementType). */
typedef ElementType value_type;
/** The corresponding primitive integer type, for example, this will be int32_t
if type is a float. */
typedef typename SIMDInternal::MaskTypeFor<ElementType>::type MaskType;
//==============================================================================
// Here are some types which are needed internally
/** The native primitive type (used internally). */
typedef typename SIMDInternal::PrimitiveType<ElementType>::type PrimitiveType;
/** The native operations for this platform and type combination (used internally) */
typedef SIMDNativeOps<PrimitiveType> NativeOps;
/** The native type (used internally). */
typedef typename NativeOps::vSIMDType vSIMDType;
/** The corresponding integer SIMDRegister type (used internally). */
typedef SIMDRegister<MaskType> vMaskType;
/** The internal native type for the corresponding mask type (used internally). */
typedef typename vMaskType::vSIMDType vMaskSIMDType;
/** Wrapper for operations which need to be handled differently for complex
and scalar types (used internally). */
typedef CmplxSIMDOps<ElementType> CmplxOps;
//==============================================================================
/** The size in bytes of this register. */
static constexpr size_t SIMDRegisterSize = sizeof (vSIMDType);
/** The number of elements that this vector can hold. */
static constexpr size_t SIMDNumElements = SIMDRegisterSize / sizeof (ElementType);
vSIMDType value;
//==============================================================================
/** Returns the number of elements in this vector. */
static constexpr size_t size() noexcept { return SIMDNumElements; }
//==============================================================================
/** Creates a new SIMDRegister from the corresponding scalar primitive.
The scalar is extended to all elements of the vector. */
inline static SIMDRegister JUCE_VECTOR_CALLTYPE expand (ElementType s) noexcept { return {CmplxOps::expand (s)}; }
/** Creates a new SIMDRegister from the internal SIMD type (for example
__mm128 for single-precision floating point on SSE architectures). */
inline static SIMDRegister JUCE_VECTOR_CALLTYPE fromNative (vSIMDType a) noexcept { return {a}; }
/** Creates a new SIMDRegister from the first SIMDNumElements of a scalar array. */
inline static SIMDRegister JUCE_VECTOR_CALLTYPE fromRawArray (const ElementType* a) noexcept
{
jassert (isSIMDAligned (a));
return {CmplxOps::load (a)};
}
/** Copies the elements of the SIMDRegister to a scalar array in memory. */
inline void JUCE_VECTOR_CALLTYPE copyToRawArray (ElementType* a) const noexcept
{
jassert (isSIMDAligned (a));
CmplxOps::store (value, a);
}
//==============================================================================
/** Returns the idx-th element of the receiver. Note that this does not check if idx
is larger than the native register size. */
inline ElementType JUCE_VECTOR_CALLTYPE operator[] (size_t idx) const noexcept
{
jassert (idx < SIMDNumElements);
return reinterpret_cast<const ElementType*> (&value) [idx];
}
/** Returns the idx-th element of the receiver. Note that this does not check if idx
is larger than the native register size. */
inline ElementType& JUCE_VECTOR_CALLTYPE operator[] (size_t idx) noexcept
{
jassert (idx < SIMDNumElements);
return reinterpret_cast<ElementType*> (&value) [idx];
}
//==============================================================================
/** Adds another SIMDRegister to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator+= (SIMDRegister v) noexcept { value = NativeOps::add (value, v.value); return *this; }
/** Subtracts another SIMDRegister to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator-= (SIMDRegister v) noexcept { value = NativeOps::sub (value, v.value); return *this; }
/** Subtracts another SIMDRegister to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator*= (SIMDRegister v) noexcept { value = CmplxOps::mul (value, v.value); return *this; }
//==============================================================================
/** Broadcasts the scalar to all elements of the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator= (ElementType s) noexcept { value = CmplxOps::expand (s); return *this; }
/** Adds a scalar to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator+= (ElementType s) noexcept { value = NativeOps::add (value, CmplxOps::expand (s)); return *this; }
/** Subtracts a scalar to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator-= (ElementType s) noexcept { value = NativeOps::sub (value, CmplxOps::expand (s)); return *this; }
/** Multiplies a scalar to the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator*= (ElementType s) noexcept { value = CmplxOps::mul (value, CmplxOps::expand (s)); return *this; }
//==============================================================================
/** Bit-and the reciver with SIMDRegister v and store the result in the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator&= (vMaskType v) noexcept { value = NativeOps::bit_and (value, toVecType (v.value)); return *this; }
/** Bit-or the reciver with SIMDRegister v and store the result in the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator|= (vMaskType v) noexcept { value = NativeOps::bit_or (value, toVecType (v.value)); return *this; }
/** Bit-xor the reciver with SIMDRegister v and store the result in the receiver. */
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator^= (vMaskType v) noexcept { value = NativeOps::bit_xor (value, toVecType (v.value)); return *this; }
//==============================================================================
/** Bit-and each element of the reciver with the scalar s and store the result in the receiver.*/
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator&= (MaskType s) noexcept { value = NativeOps::bit_and (value, toVecType (s)); return *this; }
/** Bit-or each element of the reciver with the scalar s and store the result in the receiver.*/
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator|= (MaskType s) noexcept { value = NativeOps::bit_or (value, toVecType (s)); return *this; }
/** Bit-xor each element of the reciver with the scalar s and store the result in the receiver.*/
inline SIMDRegister& JUCE_VECTOR_CALLTYPE operator^= (MaskType s) noexcept { value = NativeOps::bit_xor (value, toVecType (s)); return *this; }
//==============================================================================
/** Returns the sum of the receiver and v.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator+ (SIMDRegister v) const noexcept { return { NativeOps::add (value, v.value) }; }
/** Returns the difference of the receiver and v.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator- (SIMDRegister v) const noexcept { return { NativeOps::sub (value, v.value) }; }
/** Returns the product of the receiver and v.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator* (SIMDRegister v) const noexcept { return { CmplxOps::mul (value, v.value) }; }
//==============================================================================
/** Returns a vector where each element is the sum of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator+ (ElementType s) const noexcept { return { NativeOps::add (value, CmplxOps::expand (s)) }; }
/** Returns a vector where each element is the difference of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator- (ElementType s) const noexcept { return { NativeOps::sub (value, CmplxOps::expand (s)) }; }
/** Returns a vector where each element is the difference of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator* (ElementType s) const noexcept { return { CmplxOps::mul (value, CmplxOps::expand (s)) }; }
//==============================================================================
/** Returns the bit-and of the receiver and v. */
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator& (vMaskType v) const noexcept { return { NativeOps::bit_and (value, toVecType (v.value)) }; }
/** Returns the bit-or of the receiver and v. */
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator| (vMaskType v) const noexcept { return { NativeOps::bit_or (value, toVecType (v.value)) }; }
/** Returns the bit-xor of the receiver and v. */
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator^ (vMaskType v) const noexcept { return { NativeOps::bit_xor (value, toVecType (v.value)) }; }
/** Returns a vector where each element is the bit-inverted value of the corresponding element in the receiver.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator~ () const noexcept { return { NativeOps::bit_not (value) }; }
//==============================================================================
/** Returns a vector where each element is the bit-and'd value of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator& (MaskType s) const noexcept { return { NativeOps::bit_and (value, toVecType (s)) }; }
/** Returns a vector where each element is the bit-or'd value of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator| (MaskType s) const noexcept { return { NativeOps::bit_or (value, toVecType (s)) }; }
/** Returns a vector where each element is the bit-xor'd value of the corresponding element in the receiver and the scalar s.*/
inline SIMDRegister JUCE_VECTOR_CALLTYPE operator^ (MaskType s) const noexcept { return { NativeOps::bit_xor (value, toVecType (s)) }; }
//==============================================================================
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is equal to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE equal (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::equal (a.value, b.value)); }
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is not equal to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE notEqual (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::notEqual (a.value, b.value)); }
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is less than to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE lessThan (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::greaterThan (b.value, a.value)); }
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is than or equal to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE lessThanOrEqual (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::greaterThanOrEqual (b.value, a.value)); }
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is greater than to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE greaterThan (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::greaterThan (a.value, b.value)); }
/** Returns a SIMDRegister of the corresponding integral type where each element has each bit set
if the corresponding element of a is greater than or equal to the corresponding element of b, or zero otherwise.
The result can then be used in bit operations defined above to avoid branches in vector SIMD code. */
static inline vMaskType JUCE_VECTOR_CALLTYPE greaterThanOrEqual (SIMDRegister a, SIMDRegister b) noexcept { return toMaskType (NativeOps::greaterThanOrEqual (a.value, b.value)); }
//==============================================================================
/** Returns a new vector where each element is the minimum of the corresponding element of a and b. */
static inline SIMDRegister JUCE_VECTOR_CALLTYPE min (SIMDRegister a, SIMDRegister b) noexcept { return { NativeOps::min (a.value, b.value) }; }
/** Returns a new vector where each element is the maximum of the corresponding element of a and b. */
static inline SIMDRegister JUCE_VECTOR_CALLTYPE max (SIMDRegister a, SIMDRegister b) noexcept { return { NativeOps::max (a.value, b.value) }; }
//==============================================================================
/** Multiplies a and b and adds the result to c. */
static inline SIMDRegister JUCE_VECTOR_CALLTYPE multiplyAdd (SIMDRegister a, const SIMDRegister b, SIMDRegister c) noexcept
{
return { CmplxOps::muladd (a.value, b.value, c.value) };
}
//==============================================================================
/** Returns a scalar which is the sum of all elements of the receiver. */
inline ElementType sum() const noexcept { return CmplxOps::sum (value); }
//==============================================================================
/** Checks if the given pointer is suffeciently aligned for using SIMD operations. */
static inline bool isSIMDAligned (const ElementType* ptr) noexcept
{
uintptr_t bitmask = SIMDRegisterSize - 1;
return (reinterpret_cast<uintptr_t> (ptr) & bitmask) == 0;
}
/** Returns the next position in memory where isSIMDAligned returns true.
If the current position in memory is already aligned then this method
will simply return the pointer.
*/
static inline ElementType* getNextSIMDAlignedPtr (ElementType* ptr) noexcept
{
return snapPointerToAlignment (ptr, SIMDRegisterSize);
}
#ifndef DOXYGEN
static inline const ElementType* getNextSIMDAlignedPtr (const ElementType* ptr) noexcept
{
return snapPointerToAlignment (ptr, SIMDRegisterSize);
}
#endif
private:
static inline vMaskType JUCE_VECTOR_CALLTYPE toMaskType (vSIMDType a) noexcept
{
union
{
vSIMDType in;
vMaskSIMDType out;
} u;
u.in = a;
return vMaskType::fromNative (u.out);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE toVecType (vMaskSIMDType a) noexcept
{
union
{
vMaskSIMDType in;
vSIMDType out;
} u;
u.in = a;
return u.out;
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE toVecType (MaskType a) noexcept
{
union
{
vMaskSIMDType in;
vSIMDType out;
} u;
u.in = CmplxSIMDOps<MaskType>::expand (a);
return u.out;
}
};
//==============================================================================
/* This class is used internally by SIMDRegister to abstract away differences
in operations which are different for complex and pure floating point types. */
// the pure floating-point version
template <typename Scalar>
struct CmplxSIMDOps
{
typedef typename SIMDNativeOps<Scalar>::vSIMDType vSIMDType;
static inline vSIMDType JUCE_VECTOR_CALLTYPE load (const Scalar* a) noexcept
{
return SIMDNativeOps<Scalar>::load (a);
}
static inline void JUCE_VECTOR_CALLTYPE store (vSIMDType value, Scalar* dest) noexcept
{
SIMDNativeOps<Scalar>::store (value, dest);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE expand (Scalar s) noexcept
{
return SIMDNativeOps<Scalar>::expand (s);
}
static inline Scalar JUCE_VECTOR_CALLTYPE sum (vSIMDType a) noexcept
{
return SIMDNativeOps<Scalar>::sum (a);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE mul (vSIMDType a, vSIMDType b) noexcept
{
return SIMDNativeOps<Scalar>::mul (a, b);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE muladd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept
{
return SIMDNativeOps<Scalar>::multiplyAdd (a, b, c);
}
};
// The pure complex version
template <typename Scalar>
struct CmplxSIMDOps<std::complex<Scalar>>
{
typedef typename SIMDNativeOps<Scalar>::vSIMDType vSIMDType;
static inline vSIMDType JUCE_VECTOR_CALLTYPE load (const std::complex<Scalar>* a) noexcept
{
return SIMDNativeOps<Scalar>::load (reinterpret_cast<const Scalar*> (a));
}
static inline void JUCE_VECTOR_CALLTYPE store (vSIMDType value, std::complex<Scalar>* dest) noexcept
{
SIMDNativeOps<Scalar>::store (value, reinterpret_cast<Scalar*> (dest));
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE expand (std::complex<Scalar> s) noexcept
{
const int n = sizeof (vSIMDType) / sizeof (Scalar);
union
{
vSIMDType v;
Scalar floats[n];
} u;
for (int i = 0; i < n; ++i)
u.floats[i] = (i & 1) == 0 ? s.real() : s.imag();
return u.v;
}
static inline std::complex<Scalar> JUCE_VECTOR_CALLTYPE sum (vSIMDType a) noexcept
{
vSIMDType result = SIMDNativeOps<Scalar>::oddevensum (a);
auto* ptr = reinterpret_cast<const Scalar*> (&result);
return std::complex<Scalar> (ptr[0], ptr[1]);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE mul (vSIMDType a, vSIMDType b) noexcept
{
return SIMDNativeOps<Scalar>::cmplxmul (a, b);
}
static inline vSIMDType JUCE_VECTOR_CALLTYPE muladd (vSIMDType a, vSIMDType b, vSIMDType c) noexcept
{
return SIMDNativeOps<Scalar>::add (a, SIMDNativeOps<Scalar>::cmplxmul (b, c));
}
};
//==============================================================================
#ifndef DOXYGEN
namespace util
{
template <typename Type>
inline void snapToZero (SIMDRegister<Type>&) noexcept {}
}
#endif
} // namespace dsp
// Extend some common used global functions to SIMDRegister types
template <typename Type>
inline dsp::SIMDRegister<Type> JUCE_VECTOR_CALLTYPE jmin (dsp::SIMDRegister<Type> a, dsp::SIMDRegister<Type> b) { return dsp::SIMDRegister<Type>::min (a, b); }
template <typename Type>
inline dsp::SIMDRegister<Type> JUCE_VECTOR_CALLTYPE jmax (dsp::SIMDRegister<Type> a, dsp::SIMDRegister<Type> b) { return dsp::SIMDRegister<Type>::max (a, b); }
} // namespace juce