如何使用 Altivec 将向量存储到内存中未对齐的位置
Posted
技术标签:
【中文标题】如何使用 Altivec 将向量存储到内存中未对齐的位置【英文标题】:How to store a vector to an unaligned location in memory with Altivec 【发布时间】:2016-02-10 14:04:45 【问题描述】:我从tutorial 知道,未对齐的加载和存储可能如下所示:
//Load a vector from an unaligned location in memory
__vector unsigned char LoadUnaligned(const unsigned char * src )
__vector unsigned char permuteVector = vec_lvsl(0, src);
__vector unsigned char low = vec_ld( 0, src);
__vector unsigned char high = vec_ld( 16, src);
return vec_perm( low, high, permuteVector);
//Store a vector to an unaligned location in memory
void StoreUnaligned(__vector unsigned char v, __vector unsigned char * dst)
//Load the surrounding area
__vector unsigned char low = vec_ld( 0, dst);
__vector unsigned char high = vec_ld( 16, dst);
//Prepare the constants that we need
__vector unsigned char permuteVector = vec_lvsr( 0, (int*) dst);
__vector signed char oxFF = vec_splat_s8( -1 );
__vector signed char ox00 = vec_splat_s8( 0 );
//Make a mask for which parts of the vectors to swap out
__vector unsigned char mask = vec_perm( ox00, oxFF, permuteVector );
//Right rotate our input data
v = vec_perm( v, v, permuteVector );
//Insert our data into the low and high vectors
low = vec_sel( v, low, mask );
high = vec_sel( high, v, mask );
//Store the two aligned result vectors
vec_st( low, 0, dst);
vec_st( high, 16, dst);
看起来很糟糕。为了存储一个向量需要做大量的工作! 并且有适当的性能损失。
void SomeFuncA(const unsigned char * src, size_t size, unsigned char * dst)
for(size_t i = 0; i < size; i += 16)
__vector unsigned char a = vec_ld(0, src + i);
//simple work
vec_st(a, 0, dst + i);
void SomeFuncU(const unsigned char * src, size_t size, unsigned char * dst)
for(size_t i = 0; i < size; i += 16)
__vector unsigned char a = LoadUnaligned(src + i);
//simple work
StoreUnaligned(dst + i, a);
第二个函数的工作速度比第一个慢 3-4 倍。 由于我无法控制输入和输出内存的对齐,所以我必须实现这两个版本。 如何最大程度地减少未对齐大小写的性能损失?
【问题讨论】:
【参考方案1】:首先我想提一下,如果您多次将 Altivec 向量保存到未对齐的内存中,则不需要仅在开始和结束时将先前的内存状态保存在数组中间。 所以Simd Library中有一个有用的函数和类,实现了这个功能:
typedef __vector uint8_t v128_u8;
const v128_u8 K8_00 = vec_splat_u8(0x00);
const v128_u8 K8_FF = vec_splat_u8(0xFF);
template <bool align> inline v128_u8 Load(const uint8_t * p);
template <> inline v128_u8 Load<false>(const uint8_t * p)
v128_u8 lo = vec_ld(0, p);
v128_u8 hi = vec_ld(16, p);
return vec_perm(lo, hi, vec_lvsl(0, p));
template <> inline v128_u8 Load<true>(const uint8_t * p)
return vec_ld(0, p);
template <bool align> struct Storer;
template <> struct Storer<true>
template <class T> Storer(T * ptr)
:_ptr((uint8_t*)ptr)
template <class T> inline void First(T value)
vec_st((v128_u8)value, 0, _ptr);
template <class T> inline void Next(T value)
_ptr += 16;
vec_st((v128_u8)value, 0, _ptr);
inline void Flush()
private:
uint8_t * _ptr;
;
template <> struct Storer<false>
template <class T> inline Storer(T * ptr)
:_ptr((uint8_t*)ptr)
_perm = vec_lvsr(0, _ptr);
_mask = vec_perm(K8_00, K8_FF, _perm);
template <class T> inline void First(T value)
_last = (v128_u8)value;
v128_u8 background = vec_ld(0, _ptr);
v128_u8 foreground = vec_perm(_last, _last, _perm);
vec_st(vec_sel(background, foreground, _mask), 0, _ptr);
template <class T> inline void Next(T value)
_ptr += 16;
vec_st(vec_perm(_last, (v128_u8)value, _perm), 0, _ptr);
_last = (v128_u8)value;
inline void Flush()
v128_u8 background = vec_ld(16, _ptr);
v128_u8 foreground = vec_perm(_last, _last, _perm);
vec_st(vec_sel(foreground, background, _mask), 16, _ptr);
private:
uint8_t * _ptr;
v128_u8 _perm;
v128_u8 _mask;
v128_u8 _last;
;
它的使用方式如下:
template<bool align> void SomeFunc(const unsigned char * src, size_t size, unsigned char * dst)
Storer<align> _dst(dst);
__vector unsigned char a = Load<align>(src);
//simple work
_dst.First(a);// save first block
for(size_t i = 16; i < size; i += 16)
__vector unsigned char a = Load<align>(src + i);
//simple work
_dst.Next(a);// save body
_dst.Flush(); // save tail
与对齐版本相比,性能损失将是 30-40%。 这当然令人不快,但可以容忍。
另一个好处是减少了代码 - 所有函数(对齐和未对齐)都具有相同的实现。
【讨论】:
以上是关于如何使用 Altivec 将向量存储到内存中未对齐的位置的主要内容,如果未能解决你的问题,请参考以下文章