FPGA(HLS)实现神经网络的 Add函数、Maxpool函数、upsample2D函数

#include"addp_upool.h"
#include "string.h"

Dtype_16f maxfour(Dtype_16f a,Dtype_16f b,Dtype_16f c,Dtype_16f d){
	return max(max(a,b),max(c,d));
}

void cpy_poolbuf(Dtype_16f* In_ddr,Dtype_16f pool_bus1[Tww],INT8 Tw,int of){
	memcpy(pool_bus1, (Dtype_16f *)(In_ddr + of), Tw*sizeof(Dtype_16f));
}
void addcpdr2_lin(Dtype_16f* In_ddr,Dtype_16f* In_ddr2,Dtype_16f pool_bus1[Tww],Dtype_16f pool_bus2[Tww],Dtype_11f hz,Dtype_11f wz,Dtype_11f cz,Dtype_11f win,
		Dtype_19f winxhin,Dtype_16f t,Dtype_11f inchan,INT8 Tw,Dtype_16f inxTh){
	if(t>=inxTh)
		return;
	Dtype_11f t1=t%inchan;
	Dtype_11f t2=t/inchan;
	Dtype_11f c=t1+cz;
	Dtype_11f h=t2+hz;
	int of=c*winxhin+h*win+wz;
	cpy_poolbuf(In_ddr, pool_bus1, Tw, of);
	cpy_poolbuf(In_ddr2, pool_bus2, Tw, of);
}

void addcplin2_buf(Dtype_16f pool_in[inchannel][Thh][Tww],Dtype_16f pool_bus1[Tww],Dtype_16f pool_bus2[Tww],
			Dtype_16f t,Dtype_11f inchan,INT8 Tw){
	if(t<0)
		return;
	INT8 t3;
	Dtype_11f t1=t%inchan;
	Dtype_11f t2=t/inchan;
	for(t3=0;t3<Tw;t3++){
#pragma HLS PIPELINE II=1
		pool_in[t1][t2][t3]=pool_bus1[t3]+pool_bus2[t3];
	}

}

void Add(Dtype_16f* In_ddr,Dtype_16f* In_ddr2,Dtype_16f pool_in[inchannel][Thh][Tww],
		INT8 Th,INT8 Tw, Dtype_11f hz,
Dtype_11f wz,Dtype_11f cz,Dtype_11f win,Dtype_11f chin, Dtype_19f winxhin){
		if(cz>=chin)
			return;
		INT8 cc,hh,ww;
		Dtype_11f inchan=inchannel;
//		int of;
		Dtype_16f t;
		static Dtype_16f pool_bus1[Tww];
		static Dtype_16f pool_bus2[Tww];
		static Dtype_16f pool_busb1[Tww];
		static Dtype_16f pool_busb2[Tww];
		if(cz+inchannel>chin)
			inchan=chin-cz;
		Dtype_16f inxTh=inchan*Th;
		bool pp=true;
		for(t=0;t<inxTh+1;t++){
			if(pp)
			{
				addcpdr2_lin( In_ddr, In_ddr2, pool_bus1, pool_bus2, hz, wz, cz, win,
						 winxhin, t, inchan, Tw,inxTh);
				addcplin2_buf( pool_in, pool_busb1, pool_busb2,
							 t-1, inchan, Tw);
				pp=false;

			}
			else