batmat 0.0.17
Batched linear algebra routines
Loading...
Searching...
No Matches
Low-level operations

Detailed Description

Low-level operations such as gathers, transposes, lane-wise rotations, inverse square roots, conditional negations, etc.

Conditional negation of floating point numbers

template<class T>
T batmat::ops::detail::cneg (T x, T signs)
 Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.
template<class T, class Abi>
T batmat::ops::detail::cneg (T x, T signs)
 Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.
template<class T, class Abi>
datapar::simd< T, Abi > batmat::ops::detail::cneg (datapar::simd< T, Abi > x, datapar::simd< T, Abi > signs)
 Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.

Gathering elements from memory

template<class T, class AbiT, class I, class AbiI, class M>
datapar::simd< T, AbiT > batmat::ops::gather (const T *p, datapar::simd< I, AbiI > idx, M mask)
 Gathers elements from memory at the addresses specified by idx, which should be an integer SIMD vector, and returns them in a SIMD vector of type datapar::simd<T, AbiT>.

Lane-wise rotations of SIMD vectors

template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::rotl (datapar::simd< F, Abi > x)
 Rotates the elements of x by s positions to the left.
template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::rotr (datapar::simd< F, Abi > x)
 Rotate the elements of x to the right by S positions.
template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::shiftl (datapar::simd< F, Abi > x)
 Shift the elements of x to the left by S positions, shifting in zeros.
template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::shiftr (datapar::simd< F, Abi > x)
 Shift the elements of x to the right by S positions, shifting in zeros.
template<class F, class Abi>
datapar::simd< F, Abi > batmat::ops::rot (datapar::simd< F, Abi > x, int s)
 Rotate the elements of x to the right by s positions.

Inverse square root

template<std::floating_point T>
T batmat::ops::rsqrt (T x)
 Inverse square root.
template<class T, class Abi>
datapar::simd< T, Abi > batmat::ops::rsqrt (datapar::simd< T, Abi > x)
 Inverse square root.

Transposition

template<index_t R, index_t C, class T>
void batmat::ops::transpose_dyn (const T *pa, index_t lda, T *pb, index_t ldb, index_t d=R)
 Transposes the R × C matrix at pa with leading dimension lda, writing the result to pb with leading dimension ldb, writing only the d first columns of the result.
template<index_t R, index_t C, class T>
void batmat::ops::transpose (const T *pa, index_t lda, T *pb, index_t ldb)
 Transposes the R × C matrix at pa with leading dimension lda, writing the result to pb with leading dimension ldb.

Function Documentation

◆ cneg() [1/3]

template<class T>
T batmat::ops::detail::cneg ( T x,
T signs )

#include <batmat/ops/cneg.hpp>

Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.

only the sign bit of an IEEE-754 floating point number).

Definition at line 42 of file cneg.hpp.

◆ cneg() [2/3]

template<class T, class Abi>
T batmat::ops::detail::cneg ( T x,
T signs )
inline

#include <batmat/ops/cneg.hpp>

Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.

only the sign bit of an IEEE-754 floating point number).

Definition at line 51 of file cneg.hpp.

◆ cneg() [3/3]

template<class T, class Abi>
datapar::simd< T, Abi > batmat::ops::detail::cneg ( datapar::simd< T, Abi > x,
datapar::simd< T, Abi > signs )
inline

#include <batmat/ops/cneg.hpp>

Conditionally negates the sign bit of x, depending on signs, which should contain only ±0 (i.e.

only the sign bit of an IEEE-754 floating point number).

Definition at line 63 of file cneg.hpp.

◆ gather()

template<class T, class AbiT, class I, class AbiI, class M>
datapar::simd< T, AbiT > batmat::ops::gather ( const T * p,
datapar::simd< I, AbiI > idx,
M mask )
inline

#include <batmat/ops/gather.hpp>

Gathers elements from memory at the addresses specified by idx, which should be an integer SIMD vector, and returns them in a SIMD vector of type datapar::simd<T, AbiT>.

The elements are gathered relative to the base address p. The gathering is masked by mask,

Definition at line 56 of file gather.hpp.

◆ rotl()

template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::rotl ( datapar::simd< F, Abi > x)
inline

#include <batmat/ops/rotate.hpp>

Rotates the elements of x by s positions to the left.

For example, rotl<1>([x0, x1, x2, x3]) == [x1, x2, x3, x0] and rotl<-1>([x0, x1, x2, x3]) == [x3, x0, x1, x2].

Definition at line 226 of file rotate.hpp.

◆ rotr()

template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::rotr ( datapar::simd< F, Abi > x)
inline

#include <batmat/ops/rotate.hpp>

Rotate the elements of x to the right by S positions.

For example, rotr<1>([x0, x1, x2, x3]) == [x3, x0, x1, x2] and rotr<-1>([x0, x1, x2, x3]) == [x1, x2, x3, x0].

Definition at line 239 of file rotate.hpp.

◆ shiftl()

template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::shiftl ( datapar::simd< F, Abi > x)
inline

#include <batmat/ops/rotate.hpp>

Shift the elements of x to the left by S positions, shifting in zeros.

For example, shiftl<1>([x0, x1, x2, x3]) == [x1, x2, x3, 0] and shiftl<-1>([x0, x1, x2, x3]) == [0, x0, x1, x2].

Definition at line 252 of file rotate.hpp.

◆ shiftr()

template<int S, class F, class Abi>
datapar::simd< F, Abi > batmat::ops::shiftr ( datapar::simd< F, Abi > x)
inline

#include <batmat/ops/rotate.hpp>

Shift the elements of x to the right by S positions, shifting in zeros.

For example, shiftr<1>([x0, x1, x2, x3]) == [0, x0, x1, x2] and shiftr<-1>([x0, x1, x2, x3]) == [x1, x2, x3, 0].

Definition at line 267 of file rotate.hpp.

◆ rsqrt() [1/2]

template<std::floating_point T>
T batmat::ops::rsqrt ( T x)

#include <batmat/ops/rsqrt.hpp>

Inverse square root.

Definition at line 15 of file rsqrt.hpp.

◆ rsqrt() [2/2]

template<class T, class Abi>
datapar::simd< T, Abi > batmat::ops::rsqrt ( datapar::simd< T, Abi > x)

#include <batmat/ops/rsqrt.hpp>

Inverse square root.

May be implemented using an rsqrt instruction followed by Newton iterations for better performance, depending on the SIMD ABI. This allows it to be performed in parallel with a normal square root instruction, enabling better performance of the Cholesky micro-kernels.

Definition at line 25 of file rsqrt.hpp.

◆ transpose_dyn()

template<index_t R, index_t C, class T>
void batmat::ops::transpose_dyn ( const T * pa,
index_t lda,
T * pb,
index_t ldb,
index_t d = R )
inline

#include <batmat/ops/transpose.hpp>

Transposes the R × C matrix at pa with leading dimension lda, writing the result to pb with leading dimension ldb, writing only the d first columns of the result.

Definition at line 20 of file transpose.hpp.

◆ transpose()

template<index_t R, index_t C, class T>
void batmat::ops::transpose ( const T * pa,
index_t lda,
T * pb,
index_t ldb )
inline

#include <batmat/ops/transpose.hpp>

Transposes the R × C matrix at pa with leading dimension lda, writing the result to pb with leading dimension ldb.

Definition at line 63 of file transpose.hpp.

◆ rot()

template<class F, class Abi>
datapar::simd< F, Abi > batmat::ops::detail::rot ( datapar::simd< F, Abi > x,
int s )
inline

#include <batmat/ops/rotate.hpp>

Rotate the elements of x to the right by s positions.

For example, rotr<1>([x0, x1, x2, x3]) == [x3, x0, x1, x2] and rotr<-1>([x0, x1, x2, x3]) == [x1, x2, x3, x0].

Definition at line 18 of file rotate.hpp.