c - 如何将这个汇编代码转换为内部代码？

Question

下面看起来像是内在函数，但是，我对内在函数不熟悉。请帮我转换真实的代码。特别是， testFunc() 对我来说更加模糊。我想这也适用于两个浮点 vector 的点积，但是标签 Lrep 和 Lexit 让我感到困惑。请帮我弄清楚。移动处理器可以使用内在函数吗？

void testFunc(int M, int N, int K, float* A, float* B, float* C)
{
    float *a;
    float *b = new float[K*N];
    float *pointb = B;
    float *bb;
    float *answer = C;
    float c[8];

    for (int j = 0, k; j < K; j++) {
        bb = b + j;
        for (k = N / 8; k > 0; k--) {
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
            *bb = *pointb++; bb += K;
        }
        for (k = N / 8 * 8; k < N; k++) {
            *bb = *pointb++; bb += K;
        }
    }

    int K8 = K / 8 * 8;

    for (int i = 0; i < M; i++) for (int k = 0; k < N; k++) {
        a = A + i * K;
        bb = b + k * K;
        __asm {
            mov             esi, K8;
            sub             esi, 8;
            shl             esi, 2;
            xor             edi, edi;
            mov             edx, a;
            mov             ebx, bb;
            vxorps          ymm3, ymm3, ymm3;
        Lrep:
            cmp             edi, esi;
            jg              Lexit;
            vmovups         ymm0, ymmword ptr[edx + edi];
            vfmadd231ps     ymm3, ymm0, ymmword ptr[ebx + edi];
            add             edi, 32;
            jmp             Lrep;
        Lexit:
            vmovups         ymmword ptr[c], ymm3;
        }

        for (int j = K8; j < K; ) {
            *c += *(a + j) * *(bb + j); j++;
        }

        *answer = (c[0] + c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7]);
        answer++;
    }
}

和

pA = A;
for (k = 0; k < K; k++) {
    pC = C;
    for (i = 0; i < M; i++) {
        pA = A + i * K + k;
        pB = B + k * N;
        for (j = N / 32; j > 0; j--) {
            _asm {
                mov             eax, pC;
                mov             ebx, pA;
                mov             ecx, pB;
                vmovups         ymm0, ymmword ptr[eax];
                vmovss          xmm1, dword ptr[ebx];
                vbroadcastss    ymm4, xmm1;
                vmovups         ymm2, ymmword ptr[ecx];
                vfmadd231ps     ymm0, ymm4, ymm2;
                vmovups         ymmword ptr[eax], ymm0;
            }
            pC += 8; pB += 8;
            _asm {
                mov             eax, pC;
                mov             ebx, pA;
                mov             ecx, pB;
                vmovups         ymm0, ymmword ptr[eax];
                vmovss          xmm1, dword ptr[ebx];
                vbroadcastss    ymm4, xmm1;
                vmovups         ymm2, ymmword ptr[ecx];
                vfmadd231ps     ymm0, ymm4, ymm2;
                vmovups         ymmword ptr[eax], ymm0;
            }
            pC += 8; pB += 8;
            _asm {
                mov             eax, pC;
                mov             ebx, pA;
                mov             ecx, pB;
                vmovups         ymm0, ymmword ptr[eax];
                vmovss          xmm1, dword ptr[ebx];
                vbroadcastss    ymm4, xmm1;
                vmovups         ymm2, ymmword ptr[ecx];
                vfmadd231ps     ymm0, ymm4, ymm2;
                vmovups         ymmword ptr[eax], ymm0;
            }
            pC += 8; pB += 8;
            _asm {
                mov             eax, pC;
                mov             ebx, pA;
                mov             ecx, pB;
                vmovups         ymm0, ymmword ptr[eax];
                vmovss          xmm1, dword ptr[ebx];
                vbroadcastss    ymm4, xmm1;
                vmovups         ymm2, ymmword ptr[ecx];
                vfmadd231ps     ymm0, ymm4, ymm2;
                vmovups         ymmword ptr[eax], ymm0;
            }
            pC += 8; pB += 8;
        }
        for (j = N / 32 * 32; j < N; j++) {
            *pC += *pA * *pB;
            pC += 1; pB += 1;
        }
    }
}

Answer 1

在内联函数中，这段代码重复了 4 次。

{
// vmovups         ymm0, ymmword ptr[eax];
__m256 tempC = _mm256_loadu_ps((float*)pC);

// vmovss          xmm1, dword ptr[ebx];
// vbroadcastss    ymm4, xmm1;
__m256 tempA = _mm256_set1_ps(*pA);

// vmovups         ymm2, ymmword ptr[ecx];
__m256 tempB = _mm256_loadu_ps((float*)pB);

// vfmadd231ps     ymm0, ymm4, ymm2;
__m256 result = _mm256_fmadd_ps(tempA, tempB, tempC);

// vmovups         ymmword ptr[eax], ymm0;
_mm256_storeu_ps(pC, result);
}

pC += 8; pB += 8;

不过，不断地从 pA 广播相同的值似乎有点多余。