神经网络的Add+maxpool函数 HLS ip核设计

先定义一下四个值取最大函数

#define max(a,b) ((a>b)?a:b)
Dtype_16f maxfour(Dtype_16f a,Dtype_16f b,Dtype_16f c,Dtype_16f d){
	return max(max(a,b),max(c,d));
}

这个函数就是简单的复制，封装成函数可以并行，不封装的看分析好像不并行，我也不清楚

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));

}

Add 函数

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_16f winxhin){
		INT8 cc,hh,ww;
		Dtype_11f inchan=inchannel;
		int of;
		static Dtype_16f pool_bus1[Tww];
		static Dtype_16f pool_bus2[Tww];

	//	memcpy(pool_bus1, (Dtype_16f *)(In_ddr + of), Tw*sizeof(Dtype_16f));
	//	memcpy(pool_bus2, (Dtype_16f *)(In_ddr2 + of), Tw*sizeof(Dtype_16f));
		if(cz+inchannel>chin)
			inchan=chin-cz;
		for(cc=0;cc<inchan;cc++){
			Dtype_11f C=cc+cz;
			for(hh=0;hh<Th;hh++){
				Dtype_11f H=hh+hz;
				of=C*winxhin+H*win+wz;
				cpy_poolbuf(In_ddr, pool_bus1, Tw, of);
				cpy_poolbuf(In_ddr2, pool_bus2, Tw, of);
//				memcpy(pool_bus1, (Dtype_16f *)(In_ddr + of), Tw*sizeof(Dtype_16f));
//				memcpy(pool_bus2, (Dtype_16f *)(In_ddr2 + of), Tw*sizeof(Dtype_16f));
				for(ww=0;ww<Tw;ww++){
					pool_in[cc][hh][ww]=pool_bus1[ww]+pool_bus2[ww];
				}

			}

		}
}

这个函数是将add函数和pool函数写在一起的，主要是我的网络 YOLO-ghostnet（https://kns.cnki.net/kcms/detail/detail.aspx?dbcode=CAPJ&dbname=CAPJLAST&filename=JSJC20210318000&v=g9BjGJf5ZLX%25mmd2FERDnZ0yhaQE3qlTNYNvbEOMhXkvxwGY7cDtv9Pc4mCD4L1Cogb0u）的结构中都是 add后面跟pool

void pool(Dtype_16f* In_ddr,Dtype_16f* In_ddr2,Dtype_16f* pool_outddr,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_16f winxhin){
	INT8 cc,hh,ww;
	Dtype_11f Wout=win>>1;
//	load_poolin(In_ddr,In_ddr2, pool_in,Th, Tw,  hz,wz, cz, win, chin,  winxhin);
	INT8 th=Th>>1;
	INT8 tw=Tw>>1;
	Dtype_11f hzz=hz>>1;
	Dtype_11f wzz=wz>>1;
	static Dtype_16f pool_lin[tww];
	Dtype_11f inchan=inchannel;
	if(cz+inchannel>chin)
		inchan=chin-cz;
	Add(In_ddr,In_ddr2, pool_in,Th, Tw,  hz,wz, cz, win, chin,  winxhin);
	for(cc=0;cc<inchan;cc++){
		for(hh=0;hh<th;hh++){
			INT8 h2=hh<<1;
			for(ww=0;ww<tw;ww++){
//#pragma HLS PIPELINE
				INT8 w2=ww<<1;
				pool_lin[ww]=maxfour(
						pool_in[cc][h2][w2],
						pool_in[cc][h2+1][w2],
						pool_in[cc][h2][w2+1],
						pool_in[cc][h2+1][w2+1]
				);
			}
			Dtype_11f hof=hzz+hh;
			Dtype_11f cof=cz+cc;
			int off=cof*Wout*Wout+hof*Wout+wzz;
			memcpy((Dtype_16f *)(pool_outddr + off),pool_lin,tw*sizeof(Dtype_16f));
		}
	}
}