#ifndef GELU_KERNEL_CUH_
#define GELU_KERNEL_CUH_

#include <cuda_runtime.h>
#include "../utils/cuda_utils.cuh"

#define GELU_SCALING_FACTOR sqrtf(2.0f / M_PI)

__global__ void gelu_forward_kernel(float* __restrict__ out, const float* __restrict__ inp, int N) {
    // Implement this

    int x = blockDim.x * blockIdx.x + threadIdx.x;

    if(x < N){
        float val = inp[x];
        float cube = 0.044715f * val * val * val;
        out[x] = 0.5f * val * (1.0f + tanhf(GELU_SCALING_FACTOR * (val + cube)));
    }

}

// Launch kernel here
void gelu_forward(float* out, const float* inp, int N) {
    // Implement this
    dim3 block(256,1,1);
    dim3 grid((N + block.x -1) / block.x);

    gelu_forward_kernel<<<grid, block>>>(out, inp, N);

}

#endif // GELU_KERNEL_CUH_