如何将两个双 (m128d) 转换/合并为一个 (m128)？

问题描述

我正在尝试转换此标量代码：

struct CICDecimator {
    static constexpr int64_t scale = ((int64_t)1) << 32;
    int mStages;
    int mFactor;
    float mGainCorrection;
    int64_t *pIntegrators;
    int64_t *pCombs;

    CICDecimator(int stages = 4,int factor = 8) {
        assert(stages > 0);
        assert(factor > 0);

        mStages = stages;
        pIntegrators = new int64_t[mStages + 1]{};
        pCombs = new int64_t[mStages]{};

        if (mFactor != factor) {
            mFactor = factor;
            mGainCorrection = 1.0f / (float)(pow(mFactor,mStages));
        }
    }
    ~CICDecimator() {
        delete[] pIntegrators;
        delete[] pCombs;
    }

    float process(const float *buffer) {
        for (int i = 0; i < mFactor; i++) {
            pIntegrators[0] = buffer[i] * scale;

            for (int j = 1; j <= mStages; j++) {
                pIntegrators[j] += pIntegrators[j - 1];
            }
        }

        int64_t s = pIntegrators[mStages];
        for (int i = 0; i < mStages; i++) {
            int64_t t = s;
            s -= pCombs[i];
            pCombs[i] = t;
        }

        return mGainCorrection * (s / (float)scale);
    }
};

到 SSE/SIMD 版本（使用 -march=nocona 构建），尝试在 process() 函数（在音频应用程序中称为重载）上获得性能。

主要问题是缓冲区的输入将是 m128 (4 x float) 值。并且似乎 CICDecimator 需要 64 位整数才能使精度正常工作（因此，只有两个插槽）。

这是我的实际代码：

struct CICDecimatorV4 {
    const __m128 scale4 = _mm_set1_ps((float)(((int64_t)1) << 32));
    int mStages;
    int mFactor;
    __m128 mGainCorrection4;
    __m128d *pIntegrators4_1;
    __m128d *pIntegrators4_2;
    __m128d *pCombs4_1;
    __m128d *pCombs4_2;

    CICDecimatorV4(int stages = 4,int factor = 8) {
        assert(stages > 0);
        assert(factor > 0);

        mStages = stages;
        pIntegrators4_1 = new __m128d[mStages + 1]{};
        pIntegrators4_2 = new __m128d[mStages + 1]{};
        pCombs4_1 = new __m128d[mStages]{};
        pCombs4_2 = new __m128d[mStages]{};

        if (mFactor != factor) {
            mFactor = factor;
            mGainCorrection4 = _mm_div_ps(_mm_set1_ps(1.0f),_mm_set1_ps((float)pow(mFactor,mStages)));
        }
    }
    ~CICDecimatorV4() {
        delete[] pIntegrators4_1;
        delete[] pIntegrators4_2;
        delete[] pCombs4_1;
        delete[] pCombs4_2;
    }

    __m128 process(const __m128 *buffer) {
        for (int i = 0; i < mFactor; i++) {
            // float to double
            __m128i bufferInt = _mm_cvttps_epi32(_mm_mul_ps(buffer[i],scale4));
            pIntegrators4_1[0] = _mm_cvtepi32_pd(bufferInt);
            pIntegrators4_2[0] = _mm_cvtepi32_pd(_mm_srli_si128(bufferInt,8));

            for (int j = 1; j <= mStages; j++) {
                pIntegrators4_1[j] += pIntegrators4_1[j - 1];
                pIntegrators4_2[j] += pIntegrators4_2[j - 1];
            }
        }

        __m128d sD1 = pIntegrators4_1[mStages];
        __m128d sD2 = pIntegrators4_2[mStages];
        for (int i = 0; i < mStages; i++) {
            __m128d t1 = sD1;
            __m128d t2 = sD2;

            sD1 -= pCombs4_1[i];
            sD2 -= pCombs4_2[i];

            pCombs4_1[i] = t1;
            pCombs4_2[i] = t2;
        }

        // double to float
        __m128 sF1 = _mm_cvtpd_ps(sD1);
        __m128 sF2 = _mm_cvtpd_ps(sD2);
        __m128 sF; // <<<< TODO: i need to merge the two float on a single one

        return _mm_mul_ps(mGainCorrection4,_mm_div_ps(sF,scale4));
    }
};

我将 int64_t 转换为 m128d（两个插槽），将输入 m128 拆分为两个 m128d 并复制每个计算以进行保持向量和双精度计算。

问题：

1. 这个方法正确吗？
2. 如何将两个 m128d 转换为单个 m128？ （首先在下部，其次在上部）（即 <<<< TODO part）
3. 可以整体改进吗？ (:P)

谢谢

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）