SIMD使用无符号乘法对64位* 64位到128位进行签名

时间:2015-03-02 10:20:29

标签: c x86 integer bit-manipulation sse

我创建了一个使用SIMD进行64位* 64位到128位的功能。目前我已经使用SSE2(acutally SSE4.1)实现了它。这意味着它可以同时运行两个64b * 64b到128b的产品。同样的想法可以扩展到AVX2或AVX512,同时提供四个或八个64b * 64到128b的产品。 我的算法基于http://www.hackersdelight.org/hdcodetxt/muldws.c.txt

该算法执行一次无符号乘法,一次有符号乘法和两次有符号*无符号乘法。使用_mm_mul_epi32_mm_mul_epu32可以轻松执行已签名的*已签名和无符号*无符号操作。但混合签名和未签名的产品给我带来了麻烦。 例如,考虑一下。

int32_t x = 0x80000000;
uint32_t y = 0x7fffffff;
int64_t z = (int64_t)x*y;

双字产品应为0xc000000080000000。但是如果你假设你的编译器知道如何处理混合类型,你怎么能得到这个呢?这就是我想出的:

int64_t sign = x<0; sign*=-1;        //get the sign and make it all ones
uint32_t t = abs(x);                 //if x<0 take two's complement again
uint64_t prod = (uint64_t)t*y;       //unsigned product
int64_t z = (prod ^ sign) - sign;    //take two's complement based on the sign

使用SSE可以像这样完成

__m128i xh;    //(xl2, xh2, xl1, xh1) high is signed, low unsigned
__m128i yl;    //(yh2, yl2, yh2, yl2)
__m128i xs     = _mm_cmpgt_epi32(_mm_setzero_si128(), xh); // get sign
        xs     = _mm_shuffle_epi32(xs, 0xA0);              // extend sign
__m128i t      = _mm_sign_epi32(xh,xh);                    // abs(xh)
__m128i prod   = _mm_mul_epu32(t, yl);                     // unsigned (xh2*yl2,xh1*yl1)
__m128i inv    = _mm_xor_si128(prod,xs);                   // invert bits if negative
__m128i z      = _mm_sub_epi64(inv,xs);                    // add 1 if negative

这给出了正确的结果。但是我必须做两次(平时一次)并且它现在是我功能的重要部分。使用SSE4.2,AVX2(四个128位产品),甚至AVX512(八个128位产品)是否有更有效的方法?

使用SIMD可能有更有效的方法吗?得到高位字的计算很多。

编辑:根据@ElderBug的评论,看起来这样做的方法不是使用SIMD,而是采用mul指令。对于它的价值,如果有人想要看到它有多复杂,这里就是完整的工作功能(我刚刚开始工作,所以我没有优化它,但我不认为它是值得的。)

void muldws1_sse(__m128i x, __m128i y, __m128i *lo, __m128i *hi) {
    __m128i lomask = _mm_set1_epi64x(0xffffffff);

    __m128i xh     = _mm_shuffle_epi32(x, 0xB1);    // x0l, x0h, x1l, x1h
    __m128i yh     = _mm_shuffle_epi32(y, 0xB1);    // y0l, y0h, y1l, y1h

    __m128i xs     = _mm_cmpgt_epi32(_mm_setzero_si128(), xh);
    __m128i ys     = _mm_cmpgt_epi32(_mm_setzero_si128(), yh);
            xs     = _mm_shuffle_epi32(xs, 0xA0);
            ys     = _mm_shuffle_epi32(ys, 0xA0);

    __m128i w0     = _mm_mul_epu32(x,  y);          // x0l*y0l, y0l*y0h
    __m128i w3     = _mm_mul_epi32(xh, yh);         // x0h*y0h, x1h*y1h
            xh     = _mm_sign_epi32(xh,xh);
            yh     = _mm_sign_epi32(yh,yh);

    __m128i w1     = _mm_mul_epu32(x,  yh);         // x0l*y0h, x1l*y1h
    __m128i w2     = _mm_mul_epu32(xh, y);          // x0h*y0l, x1h*y0l

    __m128i yinv   = _mm_xor_si128(w1,ys);          // invert bits if negative
            w1     = _mm_sub_epi64(yinv,ys);         // add 1
    __m128i xinv   = _mm_xor_si128(w2,xs);          // invert bits if negative
            w2     = _mm_sub_epi64(xinv,xs);         // add 1

    __m128i w0l    = _mm_and_si128(w0, lomask);
    __m128i w0h    = _mm_srli_epi64(w0, 32);

    __m128i s1     = _mm_add_epi64(w1, w0h);         // xl*yh + w0h;
    __m128i s1l    = _mm_and_si128(s1, lomask);      // lo(wl*yh + w0h);
    __m128i s1h    = _mm_srai_epi64(s1, 32);

    __m128i s2     = _mm_add_epi64(w2, s1l);         //xh*yl + s1l
    __m128i s2l    = _mm_slli_epi64(s2, 32);
    __m128i s2h    = _mm_srai_epi64(s2, 32);           //arithmetic shift right

    __m128i hi1    = _mm_add_epi64(w3, s1h);
            hi1    = _mm_add_epi64(hi1, s2h);

    __m128i lo1    = _mm_add_epi64(w0l, s2l);
    *hi = hi1;
    *lo = lo1;
}

情况变得更糟。在AVX512之前没有_mm_srai_epi64内在/指令,所以我必须自己制作。

static inline __m128i _mm_srai_epi64(__m128i a, int b) {
    __m128i sra = _mm_srai_epi32(a,32);
    __m128i srl = _mm_srli_epi64(a,32);
    __m128i mask = _mm_set_epi32(-1,0,-1,0);
    __m128i out = _mm_blendv_epi8(srl, sra, mask);
}

我在_mm_srai_epi64上面的实施不完整。我想我正在使用Agner Fog&#39; Vector Class Library。如果你查看文件vectori128.h,你会发现

static inline Vec2q operator >> (Vec2q const & a, int32_t b) {
    // instruction does not exist. Split into 32-bit shifts
    if (b <= 32) {
        __m128i bb   = _mm_cvtsi32_si128(b);               // b
        __m128i sra  = _mm_sra_epi32(a,bb);                // a >> b signed dwords
        __m128i srl  = _mm_srl_epi64(a,bb);                // a >> b unsigned qwords
        __m128i mask = _mm_setr_epi32(0,-1,0,-1);          // mask for signed high part
        return  selectb(mask,sra,srl);
    }
    else {  // b > 32
        __m128i bm32 = _mm_cvtsi32_si128(b-32);            // b - 32
        __m128i sign = _mm_srai_epi32(a,31);               // sign of a
        __m128i sra2 = _mm_sra_epi32(a,bm32);              // a >> (b-32) signed dwords
        __m128i sra3 = _mm_srli_epi64(sra2,32);            // a >> (b-32) >> 32 (second shift unsigned qword)
        __m128i mask = _mm_setr_epi32(0,-1,0,-1);          // mask for high part containing only sign
        return  selectb(mask,sign,sra3);
    }
}

2 个答案:

答案 0 :(得分:9)

我发现SIMD解决方案更简单,不需要signed*unsigned个产品。 我不再相信SIMD(至少使用AVX2和AV512)无法与mulx竞争。在某些情况下,SIMD可以与mulx竞争。我唯一知道的情况是FFT based multiplication of large numbers

诀窍是先做无符号乘法,然后再纠正。我从这个答案中学到了如何做到这一点32-bit-signed-multiplication-without-using-64-bit-data-type。对(hi,lo) = x*y进行无符号乘法进行校正很简单,然后像这样纠正hi

hi -= ((x<0) ? y : 0)  + ((y<0) ? x : 0)

这可以使用SSE4.2内在_mm_cmpgt_epi64

来完成
void muldws1_sse(__m128i x, __m128i y, __m128i *lo, __m128i *hi) {    
    muldwu1_sse(x,y,lo,hi);    
    //hi -= ((x<0) ? y : 0)  + ((y<0) ? x : 0);
    __m128i xs = _mm_cmpgt_epi64(_mm_setzero_si128(), x);
    __m128i ys = _mm_cmpgt_epi64(_mm_setzero_si128(), y);           
    __m128i t1 = _mm_and_si128(y,xs);
    __m128i t2 = _mm_and_si128(x,ys);
           *hi = _mm_sub_epi64(*hi,t1);
           *hi = _mm_sub_epi64(*hi,t2);
}

无符号乘法的代码更简单,因为它不需要混合signed*unsigned乘积。另外,由于它是无符号的,因此它不需要算术右移,它只有AVX512的指令。实际上以下功能只需要SSE2:

void muldwu1_sse(__m128i x, __m128i y, __m128i *lo, __m128i *hi) {    
    __m128i lomask = _mm_set1_epi64x(0xffffffff);

    __m128i xh     = _mm_shuffle_epi32(x, 0xB1);    // x0l, x0h, x1l, x1h
    __m128i yh     = _mm_shuffle_epi32(y, 0xB1);    // y0l, y0h, y1l, y1h

    __m128i w0     = _mm_mul_epu32(x,  y);          // x0l*y0l, x1l*y1l
    __m128i w1     = _mm_mul_epu32(x,  yh);         // x0l*y0h, x1l*y1h
    __m128i w2     = _mm_mul_epu32(xh, y);          // x0h*y0l, x1h*y0l
    __m128i w3     = _mm_mul_epu32(xh, yh);         // x0h*y0h, x1h*y1h

    __m128i w0l    = _mm_and_si128(w0, lomask);     //(*)
    __m128i w0h    = _mm_srli_epi64(w0, 32);

    __m128i s1     = _mm_add_epi64(w1, w0h);
    __m128i s1l    = _mm_and_si128(s1, lomask);
    __m128i s1h    = _mm_srli_epi64(s1, 32);

    __m128i s2     = _mm_add_epi64(w2, s1l);
    __m128i s2l    = _mm_slli_epi64(s2, 32);        //(*)
    __m128i s2h    = _mm_srli_epi64(s2, 32);

    __m128i hi1    = _mm_add_epi64(w3, s1h);
            hi1    = _mm_add_epi64(hi1, s2h);

    __m128i lo1    = _mm_add_epi64(w0l, s2l);       //(*)
    //__m128i lo1    = _mm_mullo_epi64(x,y);          //alternative

    *hi = hi1;
    *lo = lo1;
}

这使用

4x mul_epu32
5x add_epi64
2x shuffle_epi32
2x and
2x srli_epi64
1x slli_epi64
****************
16 instructions

AVX512具有_mm_mullo_epi64内在函数,可以使用一条指令计算lo。在这种情况下,可以使用替代方法(使用(*)注释注释行并取消注释替代行):

5x mul_epu32
4x add_epi64
2x shuffle_epi32
1x and
2x srli_epi64
****************
14 instructions

要更改全宽AVX2的代码,请将_mm替换为_mm256,将si128替换为si256,将__m128i替换为__m256i,以取消AVX512将其替换为_mm512si512__m512i

答案 1 :(得分:7)

使用各种指令考虑整数乘法的吞吐量限制的正确方法是根据多少&#34;产品位&#34;你可以按周期计算。

mulx生成一个64x64 - &gt;每个周期结果128;那个64x64 = 4096&#34;每个周期的产品位&#34;

如果你将SIMD上的乘数拼凑成32x32的指令 - &gt; 64位乘法,您需要能够在每个周期获得四个结果以匹配mulx(4x32x32 = 4096)。如果除了乘法之外没有算术,你只需在AVX2上收支平衡。不幸的是,正如您已经注意到的那样,除了乘法之外还有很多算术运算,所以这是当前一代硬件上的非启动算法。