基于FPGA的HLS CNN加速器

#include <string.h> #include "platform.h" #include "xparameters.h" #include<stdio.h> #include "xil_printf.h" #include "ff.h" #include "xdevcfg.h" #include "xtime_l.h" #include"xconv.h" #include"xil_cache.h" #pragma GCC optimize("O3") static FATFS fatfs; void print_float(float* buff,int length){ for(int i=0;i<length;i++) printf("%f\n",buff[i]); } void print_short(short* buff,int length){ for(int i=0;i<length;i++) printf("%d\n",buff[i]); } int SD_Init() { FRESULT rc; rc = f_mount(&fatfs,"",0); if(rc) { xil_printf("ERROR : f_mount returned %d\r\n",rc); return XST_FAILURE; } return XST_SUCCESS; } int SD_Transfer_read(char *FileName,u32 DestinationAddress,u32 ByteLength) { FIL fil; FRESULT rc; UINT br; rc = f_open(&fil,FileName,FA_READ); if(rc) { xil_printf("ERROR : f_open returned %d\r\n",rc); return XST_FAILURE; } rc = f_lseek(&fil, 0); if(rc) { xil_printf("ERROR : f_lseek returned %d\r\n",rc); return XST_FAILURE; } rc = f_read(&fil, (void*)DestinationAddress,ByteLength,&br); if(rc) { xil_printf("ERROR : f_read returned %d\r\n",rc); return XST_FAILURE; } rc = f_close(&fil); if(rc) { xil_printf(" ERROR : f_close returned %d\r\n", rc); return XST_FAILURE; } return XST_SUCCESS; } int SD_Transfer_write(char *FileName,u32 SourceAddress,u32 ByteLength) { FIL fil; FRESULT rc; UINT bw; rc = f_open(&fil,FileName,FA_CREATE_ALWAYS | FA_WRITE); if(rc) { xil_printf("ERROR : f_open returned %d\r\n",rc); return XST_FAILURE; } rc = f_lseek(&fil, 0); if(rc) { xil_printf("ERROR : f_lseek returned %d\r\n",rc); return XST_FAILURE; } rc = f_write(&fil,(void*) SourceAddress,ByteLength,&bw); if(rc) { xil_printf("ERROR : f_write returned %d\r\n", rc); return XST_FAILURE; } rc = f_close(&fil); if(rc){ xil_printf("ERROR : f_close returned %d\r\n",rc); return XST_FAILURE; } return XST_SUCCESS; } void float2short(short* buff1,float* buff2,int length){ for(int i=0;i<length;i++){ buff1[i]=(short)(buff2[i]*1024); } } struct param_short{ short images[10000][1][28][28]; short conv1_w[64][1][5][5]; short conv1_b[64]; short conv1_out[64][28][28]; short conv2_w[64][64][5][5]; short conv2_b[64]; short pool1_out[64][14][14]; short conv3_w[64][64][5][5]; short conv3_b[64]; short conv3_out[64][14][14]; short conv4_w[64][64][5][5]; short conv4_b[64]; short pool2_out[64][7][7]; }; struct param{ float conv1_w[64][1][5][5]; float conv1_b[64]; float conv1_out[64][28][28]; float conv2_w[64][64][5][5]; float conv2_b[64]; float conv2_out[64][28][28]; float pool1_out[64][14][14]; float conv3_w[64][64][5][5]; float conv3_b[64]; float conv3_out[64][14][14]; float conv4_w[64][64][5][5]; float conv4_b[64]; float conv4_out[64][14][14]; float pool2_out[64][7][7]; float fc1_w[10][64*7*7]; float fc1_b[10]; float fc1_out[10]; }; void pool(int h,int w,int ch,float* in,float* out){ int i,j,n,kx,ky; for(i=0;i<h/2;i++) for(j=0;j<w/2;j++) for(n=0;n<ch;n++){ //float tmp=in[n][2*i][2*j] float tmp1=*(in+n*h*w+2*i*w+2*j); float tmp2=*(in+n*h*w+2*i*w+2*j+1); float tmp3=*(in+n*h*w+(2*i+1)*w+2*j); float tmp4=*(in+n*h*w+(2*i+1)*w+2*j+1); float max1=(tmp1>tmp2)?tmp1:tmp2; float max2=(tmp3>tmp4)?tmp3:tmp4; *(out+n*h*w/4+i*w/2+j)=(max1>max2)?max1:max2; } } void conv(int ch_in,int ch_out,int pad,int stride,int k,int h,int w, float* in,float *weight,float *bias,float *out){ int h_o,w_o; int i,j,n,m; int kx,ky; h_o=(h-k+2*pad)/stride+1; w_o=(w-k+2*pad)/stride+1; for(i=0;i<h_o;i++) for(j=0;j<w_o;j++) for(m=0;m<ch_out;m++){ //计算out[m][i][j]的地址 int addr_o=m*h_o*w_o+i*w_o+j; *(out+addr_o)=bias[m]; for(n=0;n<ch_in;n++) for(kx=0;kx<k;kx++) for(ky=0;ky<k;ky++) { //out[m][i][j]+=in[n][i*stride+kx-pad][j*stride+ky-pad]*weight[m][n][kx][ky] int row=i*stride+kx-pad; int col=j*stride+ky-pad; int addr_i=n*h*w+row*w+col; // int addr_w=m*ch_in*k*k+n*k*k+kx*k+ky; if(row>=0&&row<h&&col>=0&&col<w) *(out+addr_o)+=(*(in+addr_i))*(*(weight+addr_w)); } *(out+addr_o)=(*(out+addr_o)>0.0)?(*(out+addr_o)):0.0; } } void fc(float *in,float *w,float *b,float *out,int nin,int nout){ //out=w*in+b int i,j; for(i=0;i<nout;i++){ float tmp=b[i]; for(j=0;j<nin;j++){ //w[i][j]*in[j] tmp+=w[i*nin+j]*in[j]; } out[i]=tmp; } } void conv_pl(XConv HlsXConv,short *IN,short *W,short *B,short *OUT,int N,int M,int SIZE,int pool){ XConv_Set_in_V(&HlsXConv,(u32)IN); XConv_Set_weight_V(&HlsXConv,(u32)W); XConv_Set_bias_V(&HlsXConv,(u32)B); XConv_Set_out_V(&HlsXConv,(u32)OUT); XConv_Set_N_V(&HlsXConv,N); XConv_Set_M_V(&HlsXConv,M); XConv_Set_SIZE_V(&HlsXConv,SIZE); XConv_Set_pool_V(&HlsXConv,pool); XConv_Start(&HlsXConv); while(XConv_IsDone(&HlsXConv)==0); return; } int main() { init_platform(); XTime tEnd,tCur; u32 tUsed; int labels[10000]; float images[10000][1][28][28]; struct param my_param; struct param_short my_param_short; SD_Init(); SD_Transfer_read("test_labels.bin",(u32)labels,10000*sizeof(int)); SD_Transfer_read("test_images.bin",(u32)images,10000*1*28*28*sizeof(float)); // SD_Transfer_read("conv1.weight.bin",(u32)my_param.conv1_w,64*1*5*5*sizeof(float)); SD_Transfer_read("conv1.bias.bin",(u32)my_param.conv1_b,64*sizeof(float)); SD_Transfer_read("conv2.weight.bin",(u32)my_param.conv2_w,64*64*5*5*sizeof(float)); SD_Transfer_read("conv2.bias.bin",(u32)my_param.conv2_b,64*sizeof(float)); SD_Transfer_read("conv3.weight.bin",(u32)my_param.conv3_w,64*64*5*5*sizeof(float)); SD_Transfer_read("conv3.bias.bin",(u32)my_param.conv3_b,64*sizeof(float)); SD_Transfer_read("conv4.weight.bin",(u32)my_param.conv4_w,64*64*5*5*sizeof(float)); SD_Transfer_read("conv4.bias.bin",(u32)my_param.conv4_b,64*sizeof(float)); SD_Transfer_read("fc1.bias.bin",(u32)my_param.fc1_b,10*sizeof(float)); SD_Transfer_read("fc1.weight.bin",(u32)my_param.fc1_w,10*64*7*7*sizeof(float)); print("SD read parameters over!\r\n"); float2short((short*)my_param_short.images,(float*)images,10000*1*28*28); float2short((short*)my_param_short.conv1_w,(float*)my_param.conv1_w,64*1*5*5); float2short((short*)my_param_short.conv1_b,(float*)my_param.conv1_b,64); float2short((short*)my_param_short.conv2_w,(float*)my_param.conv2_w,64*64*5*5); float2short((short*)my_param_short.conv2_b,(float*)my_param.conv2_b,64); float2short((short*)my_param_short.conv3_w,(float*)my_param.conv3_w,64*64*5*5); float2short((short*)my_param_short.conv3_b,(float*)my_param.conv3_b,64); float2short((short*)my_param_short.conv4_w,(float*)my_param.conv4_w,64*64*5*5); float2short((short*)my_param_short.conv4_b,(float*)my_param.conv4_b,64); Xil_DCacheFlushRange((u32)my_param_short.images,10000*1*28*28*sizeof(short)); Xil_DCacheFlushRange((u32)my_param_short.con