Questions on gpuRelu

[SkyTu]

Hello, I am currently reviewing the function gpuRelu in /GPU-MPC/fss/gpu_relu.cu and I am confused about its implementation. Specifically, I am trying to understand why the function tries to compute Relu(x - p) + q.

In /GPU-MPC/tests/fss/relu.cu, it appears that p and q are both set to (1<<16) - 1.

// Relu(x-p) + q, where x-p is guaranteed to be small
template <typename TIn, typename TOut, u64 p, u64 q, bool flipDRelu>
TOut *gpuRelu(SigmaPeer *peer, int party, GPUReluKey<TOut> &k, TIn *d_I, AESGlobalContext *gaes, Stats *s)
{   
    // std::cout << "calling gpuRelu" << std::endl;
    auto &dreluKey = k.dreluKey;
    std::vector<u32 *> h_mask({dreluKey.mask});
    auto d_drelu = gpuDcf<TIn, 1, dReluPrologue<p>, dReluEpilogue<p, flipDRelu>>(dreluKey.dpfKey, party, d_I, gaes, s, &h_mask);
    peer->reconstructInPlace(d_drelu, 1, k.numRelus, s);
    auto d_relu = gpuSelect<TIn, TOut, p, q>(peer, party, k.bout, k.selectKey, (u32 *)d_drelu, d_I, s);
    return d_relu;
}