#ifndef __ATTENTION_CUH__
#define __ATTENTION_CUH__

// WINDOW = 129 because the CPU reference uses an inclusive range [q_t - 128, q_t],
// which covers 129 positions (128 look-back steps + the current token itself).
#define WINDOW 129

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

__global__ void permute_kernel(float* q, float* k, float* v, const float* inp, int B, int N, int NH, int d) {
    // Implement this

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

    if( x < B*NH*N*d){
        int b = (x / d / N / NH);
        int nh = (x / d / N) % NH;
        int t = (x / d) % N;
        int d_ = x % d;

        int inp_idx = (b * N * 3 * NH * d) + (t * 3 * NH * d) + (0 * NH * d) + (nh * d) + d_;
        q[b * NH * N * d + nh * N * d + t * d + d_] = inp[inp_idx];
        k[b * NH * N * d + nh * N * d + t * d + d_] = inp[inp_idx + NH * d];
        v[b * NH * N * d + nh * N * d + t * d + d_] = inp[inp_idx + 2 * (NH * d)];
    }
}


__global__ void unpermute_kernel(float* inp, float *out, int B, int N, int NH, int d) {
    // Implement this
    int x = blockDim.x * blockIdx.x + threadIdx.x;

    if( x < B*NH*N*d){
        int b = (x / d / N / NH);
        int nh = (x / d / N) % NH;
        int t = (x / d) % N;
        int d_ = x % d;

        out[b*N*NH*d + t*NH*d + nh*d + d_] = inp[b*NH*N*d + nh*N*d + t*d + d_];
    }

}

__global__ void preattention_scores_kernel(float* q, float* k, int B, int T, int NH, int HS, float* preatt){
    int x = blockDim.x * blockIdx.x + threadIdx.x;

    //Make int terms of WINDOW instead of full T, so its T * WINDOW
    if(x < B* T* NH* WINDOW){
        int b = (x / WINDOW / T / NH);
        int nh = (x / WINDOW / T) % NH;
        int t1 = (x / WINDOW) % T;
        int w = x % WINDOW;

        //Key start point based on offset
        int t2 = t1 - w;

        float sum = 0.0f;

        //Need to check to not look too far back and stay in window range
        //First check if t2 (Key Step) is behind or equal to t1 (Query Step)
        // int lowest_margin = t1 - WINDOW;
        if (t2 >= 0) {
            for (int i = 0; i < HS; i++) {
                sum += q[b*NH*T*HS + nh*T*HS + t1*HS + i] * k[b*NH*T*HS + nh*T*HS + t2*HS + i];
            }
        }
        else {
            sum = -FLT_MAX;
        }
        preatt[b * NH * T * WINDOW + nh * T * WINDOW + t1 * WINDOW + w] = sum;
        //preatt[x] = sum;
    }   
}

__global__ void attention_v_kernel(float* vaccum, float* att, float* v, int B, int T, int NH, int HS){
    int x = blockDim.x * blockIdx.x + threadIdx.x;

    if(x < B*NH*T*HS){
        int b   = (x / HS / T / NH);
        int nh  = (x / HS / T) % NH;
        int t1   = (x / HS) % T;
        int hs  = x % HS;
        float sum = 0.0f;
        for (int w = 0; w < WINDOW; w++) {
            int t2 = t1 - w;
            if (t2 >= 0) {
                sum += att[b*NH*T*WINDOW + nh*T*WINDOW + t1*WINDOW + w] * v[b*NH*T*HS + nh*T*HS + t2*HS + hs];
            }  
        }

        vaccum[b * NH * T * HS + nh * T * HS + t1 * HS + hs] = sum;
        //vaccum[x] = sum;
    }
}

// Launch all kernels related to attention here 
void local_attention_forward_gpu(float* out, float* qkvr, float* att, float* inp, int B, int T, int NH, int HS) {
    // Implement this

    float *q, *k, *v;
    q = qkvr + 0 * B * T * NH * HS;
    k = qkvr + 1 * B * T * NH * HS;
    v = qkvr + 2 * B * T * NH * HS;

    dim3 block(256,1,1);
    dim3 grid((B * NH * T * HS + block.x-1) / block.x,1,1);

    permute_kernel<<<grid, block>>>(q, k, v, inp, B, T, NH, HS);

    float* preatt;
    cudaMalloc(&preatt, B * NH * T * WINDOW * sizeof(float));

    dim3 grid_preatt((B * NH * T * WINDOW + block.x-1) / block.x,1,1);
    preattention_scores_kernel<<<grid_preatt, block>>>(q, k, B, T, NH, HS, preatt);

    float* packed_att;
    cudaMalloc(&packed_att, B * NH * T * WINDOW * sizeof(float));

    float scale = 1.0 / sqrtf(HS);
    dim3 grid_softmax (( B * NH * T + block.x -1 )/ block.x,1,1);
    local_softmax_forward_kernel<<<grid_softmax, block>>>(packed_att, scale, preatt, B * NH, T);

    float* vaccum;
    cudaMalloc(&vaccum, B * NH * T * HS * sizeof(float));

    attention_v_kernel<<<grid,block>>>(vaccum, packed_att, v, B, T, NH, HS);

    unpermute_kernel<<<grid, block>>>(vaccum, out, B, T, NH, HS);

    cudaFree(preatt);
    cudaFree(vaccum);
    cudaFree(packed_att);
}

#endif // __ATTENTION_CUH__