// !!! This is a file automatically generated by hipify!!!
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <stdio.h>
#include <stdlib.h>

#ifdef __HIP_PLATFORM_HCC__
#include <hip/hip_cooperative_groups.h>
#else
#include <cooperative_groups.h>
#endif
#include <hiprand/hiprand_kernel.h>

#include "context_hip.h"
#include "cublas_wrappers_hip.h"

#define THREADS 256
#define TILE_DIM 32

#define minus_infinity -1 * std::numeric_limits<float>::infinity()

#define FINAL_MASK 0xffffffff

template <typename T>
void launch_fused_add2(T* out,
                       const T* inp1,
                       const T* inp2,
                       int batch_size,
                       int seq_length,
                       int hidden_size,
                       hipStream_t& stream);

template <typename T>
void launch_fused_add4(T* out,
                       const T* inp1,
                       const T* inp2,
                       const T* inp3,
                       const T* inp4,
                       int batch_size,
                       int seq_length,
                       int hidden_size,
                       hipStream_t& stream);

template <typename T>
void launch_fused_add3(T* out,
                       const T* inp1,
                       const T* inp2,
                       const T* inp3,
                       int batch_size,
                       int seq_length,
                       int hidden_size,
                       hipStream_t& stream);