InfiniTensor/include/cuda/cuda_softmax.h

9 lines
325 B
C++

#pragma once
#include "utils/small_array.h"
namespace infini {
void softmax_kernel(int num_blocks, float *input, float *output, int size,
int dimsize, int stride);
void softmax_kernel(int num_blocks, half *input, half *output, int size,
int dimsize, int stride);
} // namespace infini