c - 通过使用 AVX 内在函数重写来提高 math.h 函数的性能

Question

我有一个简单的数学库，它链接到一个在模拟器硬件(32 位 RTOS)上运行的项目，并且编译器工具链基于 GCC 5.5 的变体。主要项目代码位于 Matlab 中，但核心数学运算(数组数据上的 cmath 函数)是用 C 重写的以提高性能。查看 Compiler Explorer，对于 GCC 5.5 32 bit 来说，优化代码的质量似乎不太好。 (供引用:Clang trunk 32bit)。据我了解，Clang 在优化循环方面做得更好。一个示例代码片段:

...

void cfunctionsLog10(unsigned int n, const double* x, double* y) {
    int i;
    for (i = 0; i < n; i++) {
        y[i] = log10(x[i]);
    }
}

以及GCC 5.5生成的相应程序集

cfunctionsLog10(unsigned int, double const*, double*):
        push    ebp
        push    edi
        push    esi
        push    ebx
        sub     esp, 12
        mov     esi, DWORD PTR [esp+32]
        mov     ebp, DWORD PTR [esp+36]
        mov     edi, DWORD PTR [esp+40]
        test    esi, esi
        je      .L28
        xor     ebx, ebx
.L27:
        sub     esp, 8
        push    DWORD PTR [ebp+4+ebx*8]
        push    DWORD PTR [ebp+0+ebx*8]
        call    __log10_finite
        fstp    QWORD PTR [edi+ebx*8]
        add     ebx, 1
        add     esp, 16
        cmp     ebx, esi
        jne     .L27
.L28:
        add     esp, 12
        pop     ebx
        pop     esi
        pop     edi
        pop     ebp
        ret

Clang 的生产地:

cfunctionsLog10(unsigned int, double const*, double*):              # @cfunctionsLog10(unsigned int, double const*, double*)
        push    ebp
        push    ebx
        push    edi
        push    esi
        sub     esp, 76
        mov     esi, dword ptr [esp + 96]
        test    esi, esi
        je      .LBB2_8
        mov     edi, dword ptr [esp + 104]
        mov     ebx, dword ptr [esp + 100]
        xor     ebp, ebp
        cmp     esi, 4
        jb      .LBB2_7
        lea     eax, [ebx + 8*esi]
        cmp     eax, edi
        jbe     .LBB2_4
        lea     eax, [edi + 8*esi]
        cmp     eax, ebx
        ja      .LBB2_7
.LBB2_4:
        mov     ebp, esi
        xor     esi, esi
        and     ebp, -4
.LBB2_5:                                # =>This Inner Loop Header: Depth=1
        vmovsd  xmm0, qword ptr [ebx + 8*esi + 16] # xmm0 = mem[0],zero
        vmovsd  qword ptr [esp], xmm0
        vmovsd  xmm0, qword ptr [ebx + 8*esi] # xmm0 = mem[0],zero
        vmovsd  xmm1, qword ptr [ebx + 8*esi + 8] # xmm1 = mem[0],zero
        vmovsd  qword ptr [esp + 8], xmm0 # 8-byte Spill
        vmovsd  qword ptr [esp + 16], xmm1 # 8-byte Spill
        call    log10
        fstp    tbyte ptr [esp + 64]    # 10-byte Folded Spill
        vmovsd  xmm0, qword ptr [esp + 16] # 8-byte Reload
        vmovsd  qword ptr [esp], xmm0
        call    log10
        fstp    tbyte ptr [esp + 16]    # 10-byte Folded Spill
        vmovsd  xmm0, qword ptr [esp + 8] # 8-byte Reload
        vmovsd  qword ptr [esp], xmm0
        vmovsd  xmm0, qword ptr [ebx + 8*esi + 24] # xmm0 = mem[0],zero
        vmovsd  qword ptr [esp + 8], xmm0 # 8-byte Spill
        call    log10
        vmovsd  xmm0, qword ptr [esp + 8] # 8-byte Reload
        vmovsd  qword ptr [esp], xmm0
        fstp    qword ptr [esp + 56]
        fld     tbyte ptr [esp + 16]    # 10-byte Folded Reload
        fstp    qword ptr [esp + 48]
        fld     tbyte ptr [esp + 64]    # 10-byte Folded Reload
        fstp    qword ptr [esp + 40]
        call    log10
        fstp    qword ptr [esp + 32]
        vmovsd  xmm0, qword ptr [esp + 56] # xmm0 = mem[0],zero
        vmovsd  xmm1, qword ptr [esp + 40] # xmm1 = mem[0],zero
        vmovhps xmm0, xmm0, qword ptr [esp + 48] # xmm0 = xmm0[0,1],mem[0,1]
        vmovhps xmm1, xmm1, qword ptr [esp + 32] # xmm1 = xmm1[0,1],mem[0,1]
        vmovups xmmword ptr [edi + 8*esi + 16], xmm1
        vmovups xmmword ptr [edi + 8*esi], xmm0
        add     esi, 4
        cmp     ebp, esi
        jne     .LBB2_5
        mov     esi, dword ptr [esp + 96]
        cmp     ebp, esi
        je      .LBB2_8
.LBB2_7:                                # =>This Inner Loop Header: Depth=1
        vmovsd  xmm0, qword ptr [ebx + 8*ebp] # xmm0 = mem[0],zero
        vmovsd  qword ptr [esp], xmm0
        call    log10
        fstp    qword ptr [edi + 8*ebp]
        inc     ebp
        cmp     esi, ebp
        jne     .LBB2_7
.LBB2_8:
        add     esp, 76
        pop     esi
        pop     edi
        pop     ebx
        pop     ebp
        ret

由于我无法直接使用 Clang，因此使用 AVX 内在函数重写 C 源代码是否有任何值(value)。我认为大部分性能成本来自 cmath 函数调用，其中大多数没有内在实现。

---

编辑: 使用 vectorclass library 重新实现:

void vclfunctionsTanh(unsigned int n, const double* x, double* y) 
{
    const int N = n;
    const int VectorSize = 4;
    const int FirstPass = N & (-VectorSize);

    int i = 0;    
    for (; i < FirstPass; i+= 4)
    {
        Vec4d data = Vec4d.load(x[i]);
        Vec4d ans = tanh(data);
        ans.store(y+i);
    }

    
    for (;i < N; ++i)
        y[i]=std::tanh(x[i]);
}

Answer 1

vector 类库具有常见数学函数的内联 vector 版本，包括 log10。

https://github.com/vectorclass/