Function MNNExp

source/backend/cpu/arm/CommonOptFunctionNeon.cpp:478–575

void MNNExp(float* destPtr, const float* srcPtr, float* offset, size_t size) {
    float32x4_t maxVec = vdupq_n_f32(-offset[2]);
    float32x4_t sumVec0 = vdupq_n_f32(0);
    float32x4_t sumVec1 = vdupq_n_f32(0);
    if (offset[0] == 1.f && offset[1] == 0.f) {
        while (size >= 8) {
            float32x4_t srcVec0 = vld1q_f32(srcPtr);
            float32x4_t srcVec1 = vld1q_f32(srcPtr + 4);
            auto subVec0 = vsubq_f32(srcVec0, maxVec);
            auto subVec1 = vsubq_f32(srcVec1, maxVec);
            auto expVec0 = expApprox(subVec0);
            auto expVec1 = expApprox(subVec1);
            vst1q_f32(destPtr, expVec0);
            vst1q_f32(destPtr + 4, expVec1);
            sumVec0 = vaddq_f32(sumVec0, expVec0);
            sumVec1 = vaddq_f32(sumVec1, expVec1);
            srcPtr += 8;
            destPtr += 8;
            size -= 8;

        }
        while (size >= 4) {
            float32x4_t srcVec0 = vld1q_f32(srcPtr);
            auto subVec0 = vsubq_f32(srcVec0, maxVec);
            auto expVec0 = expApprox(subVec0);
            sumVec0 = vaddq_f32(sumVec0, expVec0);
            vst1q_f32(destPtr, expVec0);
            srcPtr += 4;
            destPtr += 4;
            size -= 4;
        }
        //merge
        sumVec0 = vaddq_f32(sumVec0, sumVec1);
        float32x2_t sumP = vpadd_f32(vget_low_f32(sumVec0), vget_high_f32(sumVec0));
        sumP = vpadd_f32(sumP, sumP);
        auto newSum = vget_lane_f32(sumP, 0);
        if (size > 0) {
            float tmp[4];
            memcpy(tmp, srcPtr, size * sizeof(float));
            float32x4_t srcVec0 = vld1q_f32(tmp);
            auto subVec0 = vsubq_f32(srcVec0, maxVec);
            auto expVec0 = expApprox(subVec0);
            vst1q_f32(tmp, expVec0);
            for (int i = 0; i < size; ++i) {
                newSum += tmp[i];
                destPtr[i] = tmp[i];
            }
        }
        offset[3] += newSum;
    } else {
        float32x4_t c0 = vdupq_n_f32(offset[0]);
        float32x4_t c1 = vdupq_n_f32(offset[1]);
        while (size >= 8) {
            float32x4_t srcVec0 = vld1q_f32(srcPtr);
            float32x4_t srcVec1 = vld1q_f32(srcPtr + 4);
            auto subVec0 = vsubq_f32(vmulq_f32(srcVec0, c0), maxVec);
            auto subVec1 = vsubq_f32(vmulq_f32(srcVec1, c0), maxVec);
            auto expVec0 = vaddq_f32(expApprox(subVec0), c1);