Skip to content
Snippets Groups Projects
Normal view Permalink
wino_buffer.hpp 15.6 KiB
Newer Older
  • Learn to ignore specific revisions
    • xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 
    

#ifndef _WINO_BUFFER_HPP_
#define _WINO_BUFFER_HPP_

#include <ap_int.h>
#include <hls_stream.h>
#include "wino_macro.h"

#include "../testbench/debug.hpp"

template<int dummy>
void input_feed(
	ap_uint<16> input_buffer[8][16][INPUT_BUFFER_DEPTH],
	hls::stream< ap_uint<32*36> > &input_tile_stream1, 
	hls::stream< ap_uint<32*36> > &input_tile_stream2,
	ap_uint<16> start_row_idx,
	ap_uint<16> output_width,
	ap_uint<16> input_height,
	ap_uint<16> input_width,
	ap_uint<16> input_width_ceildiv_16,
	ap_uint<16> input_depth_aligned_8,
	ap_uint<16> pad_size,
    ap_uint<16> weight_outdepth_load_number,
	ap_uint<8> window_size,
	ap_uint<3> winograd_output_tile_size
)
{
	#pragma HLS array_partition variable=input_buffer dim=1 complete
	#pragma HLS array_partition variable=input_buffer dim=2 complete
	int tile_step =  winograd_output_tile_size*2;




    for(ap_uint<16> depth_tile_idx=0,depth_base_address=0;depth_tile_idx<input_depth_aligned_8/8;depth_tile_idx++,depth_base_address+=input_width_ceildiv_16)
	{
		for(int weight_load_outdepth_cnt=0;weight_load_outdepth_cnt<weight_outdepth_load_number;weight_load_outdepth_cnt++)
		{
			int start_col_idx =0;
			int depth_address =0;
			int row_tile_offset =0;


			for(int window_row_offset=0; window_row_offset<window_size; window_row_offset++)
			{
				for(int window_col_offset=0; window_col_offset<window_size; window_col_offset++)
				{
					ap_uint<1> row_legal_flag[6];
					#pragma HLS array_partition variable=row_legal_flag complete

					ap_uint<3> row_bank_idx[6];
					#pragma HLS array_partition variable=row_bank_idx complete

					ap_uint<1> row_addres_offset[8];
					#pragma HLS array_partition variable=row_addres_offset complete

					ap_uint<4> start_row_mod16 = start_row_idx-pad_size+window_row_offset;
					ap_uint<3> breakpoint = start_row_mod16.range(2,0);


					for(int i=0;i<8;i++)
					{
						#pragma HLS unroll
						if(i< breakpoint)
							row_addres_offset[i] = ~start_row_mod16[3];
						else
							row_addres_offset[i] = start_row_mod16[3];
					}			

					
					for(int i=0;i<6;i++)
					{
						#pragma HLS unroll
						row_bank_idx[i] = start_row_idx+i-pad_size+window_row_offset;
						row_legal_flag[i] = ( start_row_idx+i-pad_size >=0 && start_row_idx+i-pad_size < input_height);
					}


					
					ap_uint<10> input_col_idx[12];
	#pragma HLS array_partition variable=input_col_idx complete
					ap_int<16> input_head_col_idx=-pad_size;

					for(int i=0;i<6;i++)
					{
						#pragma HLS unroll
						input_col_idx[i]=i-pad_size+window_col_offset;
						input_col_idx[i+6]=i+winograd_output_tile_size-pad_size+window_col_offset;
					}

					for(ap_uint<10> out_col_idx=0;out_col_idx < output_width; out_col_idx+=tile_step)
					{		
					

						ap_uint<6> input_buffer_address_by8[16];
	#pragma HLS array_partition variable=input_buffer_address_by8 complete
						ap_uint<INPUT_BUFFER_DEPTH_BITWIDTH-4>  input_head_col_by16 = input_head_col_idx.range(10,4)+depth_base_address;
						ap_uint<4>  input_buffer_address_break_point = input_head_col_idx.range(3,0);
						
						ap_uint<1> col_legal_flag[12];
	#pragma HLS array_partition variable=col_legal_flag complete
						for(int i=0;i<12; i++)
						{
							#pragma HLS unroll
							col_legal_flag[i]= ( input_col_idx[i] >=0 && input_col_idx[i] < input_width);
						}
						for(int i=0;i<16;i++)
						{
							if(i>=input_buffer_address_break_point)
								input_buffer_address_by8[i] = input_head_col_by16;
							else
								input_buffer_address_by8[i] = input_head_col_by16+1;
						}

						DEPTH:for(ap_uint<4> depth_idx_in_tile=0;depth_idx_in_tile<8;depth_idx_in_tile++)
						{	

	#pragma HLS pipeline
							ap_uint<INPUT_BUFFER_DEPTH_BITWIDTH> input_buffer_address[8][16];
	#pragma HLS array_partition variable=input_buffer_address complete

	#pragma HLS array_partition variable=col_legal_flag complete
							ap_uint<16> input_buffer_val[8][16];
	#pragma HLS array_partition variable=input_buffer_val complete
							for(int j=0;j<8;j++)
							{
							#pragma HLS unroll
								for(int i=0;i<16;i++)
								{
									#pragma HLS unroll
									input_buffer_address[j][i].range(INPUT_BUFFER_DEPTH_BITWIDTH-1,0) = (row_addres_offset[j],( input_buffer_address_by8[i], (ap_uint<3>) depth_idx_in_tile.range(2,0) ));
								}
							}

							for(int i=0;i<8;i++)
							{
								#pragma HLS unroll
								for(int j=0;j<16;j++)
								{
									#pragma HLS unroll
									input_buffer_val[i][j]=input_buffer[i][j][ (ap_uint<10>) input_buffer_address[i][j] ];   
								}
							}
							// printf("depth %d\n", (int) depth_idx_in_tile);
							// for(int i=0;i<8;i++)
							// {
							// 	for(int j=0;j<16;j++)
							// 	{
							// 		printf("[%3d]", (unsigned char) input_buffer_val[i][j].range(7,0),  (unsigned char) input_buffer_val[i][j].range(15,8));
							// 	}
							// 	printf("\n");
							// }


							ap_uint<16> input_plane_tile_row[6][16];



	#pragma HLS array_partition variable=input_plane_tile_row dim=1 complete

	#pragma HLS array_partition variable=input_plane_tile_row dim=2 complete
							for(int j=0;j<16;j++)
							{
	#pragma HLS unroll
								for(int i=0;i<6;i++)
								{
	#pragma HLS unroll
									if(row_legal_flag[i])
									{
										input_plane_tile_row[i][j]=input_buffer_val[row_bank_idx[i]][j];
									}
									else
									{
										input_plane_tile_row[i][j]=0;
									}
								}
							}


							ap_uint<16> input_plane_tile[6][12];
							#pragma HLS array_partition variable=input_plane_tile complete
							for(int i=0;i<6;i++)
							{
							#pragma HLS unroll
								for(int j=0;j<12;j++)
								{
								#pragma HLS unroll
									if(col_legal_flag[j])
										input_plane_tile[i][j]=input_plane_tile_row[i][  (ap_uint<4>) input_col_idx[j].range(3,0) ];
									else
										input_plane_tile[i][j]=0;
								}
							}


							attach_streaming_content<0>(input_plane_tile,start_row_idx-pad_size,out_col_idx-pad_size,depth_tile_idx*8+depth_idx_in_tile,"input_stream_content.txt");


							ap_int<8> in[36][4];

		
							for(int k=0;k<6;k++)
							{
							#pragma HLS unroll
								for(int l=0;l<6;l++)
								{
								#pragma HLS unroll
									for(ap_uint<3> j=0;j<4;j++)
									{
									#pragma HLS unroll
    xliu79's avatar
    input correct  
    xliu79 committed
    212 
    										in[k*6+l][j].range(7,0)=  input_plane_tile[k][ j/2*6+l].range( j%2*8+7 ,j%2*8);
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    213 
    									}
    xliu79's avatar
    input correct  
    xliu79 committed
    214 
    									// printf("[%4x %4x]", (unsigned short) in[k*6+l][1], (unsigned short) in[k*6+l][3]);
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    215 
    								}
    xliu79's avatar
    input correct  
    xliu79 committed
    216 
    								// printf("\n");
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    217 
    							}
    xliu79's avatar
    input correct  
    xliu79 committed
    218 
    							// printf("\n");
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    219 220 221 222 223 224
    xliu79's avatar
    input correct  
    xliu79 committed
    225 
    							ap_int<16>  dB[6][6][4];
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    226 227 228 229 230 231 232 
    							#pragma HLS array_partition variable=dB complete
	
							for(int i=0;i<6;i++)
							{
							#pragma HLS unroll
								for(int j=0;j<4;j++)
								{
    xliu79's avatar
    input correct  
    xliu79 committed
    233 234 235 236 237 238 
    									dB[i][0][j]=(((ap_int<16>)in[i*6+0][j]-(ap_int<16>)in[i*6+2][j])<<2) - (ap_int<16>)in[i*6+2][j] + (ap_int<16>)in[i*6+4][j];
									dB[i][1][j]=(ap_int<16>)in[i*6+3][j]+(ap_int<16>)in[i*6+4][j]-(((ap_int<16>)in[i*6+1][j]+(ap_int<16>)in[i*6+2][j])<<2);
									dB[i][2][j]=(((ap_int<16>)in[i*6+1][j]-(ap_int<16>)in[i*6+2][j])<<2)+(ap_int<16>)in[i*6+4][j]-(ap_int<16>)in[i*6+3][j];
									dB[i][3][j]=(((ap_int<16>)in[i*6+3][j]-(ap_int<16>)in[i*6+1][j])<<1)+(ap_int<16>)in[i*6+4][j]-(ap_int<16>)in[i*6+2][j];
									dB[i][4][j]=(((ap_int<16>)in[i*6+1][j]-(ap_int<16>)in[i*6+3][j])<<1)+(ap_int<16>)in[i*6+4][j]-(ap_int<16>)in[i*6+2][j];
									dB[i][5][j]=(((ap_int<16>)in[i*6+1][j]-(ap_int<16>)in[i*6+3][j])<<2)+(ap_int<16>)in[i*6+5][j]-(ap_int<16>)in[i*6+3][j];
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    239 240 241 242 243 244 
    								}						
							}
    xliu79's avatar
    input correct  
    xliu79 committed
    245 246
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 
    							ap_int<16>  BTDB[6][6][4];
							#pragma HLS array_partition variable=BTDB complete
							for(int i=0;i<6;i++)
							{
							#pragma HLS unroll
								for(int j=0;j<4;j++)
								{
									BTDB[0][i][j]=((dB[0][i][j]-dB[2][i][j])<<2) - dB[2][i][j] + dB[4][i][j];
									BTDB[1][i][j]=dB[3][i][j]+dB[4][i][j]-((dB[1][i][j]+dB[2][i][j])<<2);
									BTDB[2][i][j]=((dB[1][i][j]-dB[2][i][j])<<2)+dB[4][i][j]-dB[3][i][j];
									BTDB[3][i][j]=((dB[3][i][j]-dB[1][i][j])<<1)+dB[4][i][j]-dB[2][i][j];
									BTDB[4][i][j]=((dB[1][i][j]-dB[3][i][j])<<1)+dB[4][i][j]-dB[2][i][j];
									BTDB[5][i][j]=((dB[1][i][j]-dB[3][i][j])<<2)+dB[5][i][j]-dB[3][i][j];
								}
							}

							ap_uint<32*36> stream_out1;
							ap_uint<32*36> stream_out2;
							for(int i=0;i<6;i++)
							{
							#pragma HLS unroll
								for(int j=0;j<6;j++)
								{
								#pragma HLS unroll
									stream_out1.range( (i*6+j)*32+31,(i*6+j)*32) =(BTDB[i][j][1],BTDB[i][j][0]);
									stream_out2.range( (i*6+j)*32+31,(i*6+j)*32) =(BTDB[i][j][3],BTDB[i][j][2]);
								}
							}
    xliu79's avatar
    input correct  
    xliu79 committed
    275 276 
    							
							attach_streaming_wino<0>(stream_out1,stream_out2,start_row_idx-pad_size,input_head_col_idx,depth_tile_idx*8+depth_idx_in_tile,"input_stream_wino.txt");
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 
    							input_tile_stream1<<stream_out1;
							input_tile_stream2<<stream_out2;
						}

						for(int i=0;i<12; i++)
						{
							#pragma HLS unroll
							input_col_idx[i]+=tile_step;
						}
						input_head_col_idx+=tile_step;
					}
				}//depth_tile1
			}// window_col
		}
	}//window_row

	ap_uint<32*36> stream_out1=0;
	ap_uint<32*36> stream_out2=0;
	input_tile_stream1<<stream_out1;
	input_tile_stream1<<stream_out1;
	input_tile_stream2<<stream_out2;
	input_tile_stream2<<stream_out2;
}



template<int dummy>
void load_weight_ddr_one_port(
	ap_uint<128>* DDR_interface,
	ap_uint<64> weight_buff[WEIGHT_FEED_NUMBER_PER_PORT][9][WEIGHT_BUFFER_DEPTH],
	int load_number,
	int port_load_number,
	ap_uint<1> pingpong)
{
	ap_uint<4> counter=0;

	ap_uint<9> port_load_cnt=0;

	ap_uint<9> buffer_address_offset=0;
	ap_uint<2> buffer_idx;
	

	for(int address = 0; address<load_number; address++)
	{
		ap_uint<128> temp128 = DDR_interface[address];
		ap_uint<64> temp64[2];
		#pragma HLS array_partition  variable = temp64 complete

		for(int i=0;i<2;i++)
		{
			#pragma HLS unroll
			temp64[i]=temp128.range(i*64+63,i*64);
		}

		ap_uint<10> buffer_address = (pingpong,buffer_address_offset);
		

		if(counter==0)
		{
			weight_buff[buffer_idx][0][buffer_address]=temp64[0];
			weight_buff[buffer_idx][1][buffer_address]=temp64[1];
		}
		else if(counter == 1)
		{
			weight_buff[buffer_idx][2][buffer_address]=temp64[0];
			weight_buff[buffer_idx][3][buffer_address]=temp64[1];
		}
		else if(counter == 2)
		{
			weight_buff[buffer_idx][4][buffer_address]=temp64[0];
			weight_buff[buffer_idx][5][buffer_address]=temp64[1];
		}
		else if(counter == 3)
		{
			weight_buff[buffer_idx][6][buffer_address]=temp64[0];
			weight_buff[buffer_idx][7][buffer_address]=temp64[1];
		}
		else if(counter == 4)
		{
			weight_buff[buffer_idx][8][buffer_address]=temp64[0];
		}

		if(port_load_cnt==port_load_number-1)
		{
			buffer_address_offset =0;
		}
		else if(counter==4 )
		{
			buffer_address_offset++;
		}

		if(port_load_cnt==port_load_number-1)
		{
			buffer_idx++;
			port_load_cnt=0;
		}
		else
		{
			port_load_cnt++;
		}

		if(counter==4)
		{
			counter=0;
		}
		else
		{
			counter++;
		}
	}
}

template<int dummy>
void load_weight_ddr(
	ap_uint<128>* weight_DDR0,
	ap_uint<128>* weight_DDR1,
	ap_uint<128>* weight_DDR2,
	ap_uint<128>* weight_DDR3,
	ap_uint<64> weight_buff[4][WEIGHT_FEED_NUMBER_PER_PORT][9][WEIGHT_BUFFER_DEPTH],
	int DDR_offset,
	int load_number,
	int port_load_number,
	ap_uint<1> skip_flag,
	ap_uint<1> pingpong)
{
	
	if(skip_flag)
		return;

	// printf("Load number %d, offset %d\n", load_number, DDR_offset);
	// fflush(stdout);

	load_weight_ddr_one_port<0>(
	weight_DDR0+DDR_offset,
	weight_buff[0],
	load_number,
	port_load_number,
	pingpong);

	load_weight_ddr_one_port<0>(
	weight_DDR1+DDR_offset,
	weight_buff[1],
	load_number,
	port_load_number,
	pingpong);

	load_weight_ddr_one_port<0>(
	weight_DDR2+DDR_offset,
	weight_buff[2],
	load_number,
	port_load_number,
	pingpong);

	load_weight_ddr_one_port<0>(
	weight_DDR3+DDR_offset,
	weight_buff[3],
	load_number,
	port_load_number,
	pingpong);
}

template<int dummy>
void weight_stream(
	ap_uint<64> weight_buff[WEIGHT_FEED_NUMBER_PER_PORT][9][WEIGHT_BUFFER_DEPTH],
	#if WEIGHT_FEED_NUMBER_PER_PORT == 2
	hls::stream<ap_uint<16*36> > & weight_stream0,
	hls::stream<ap_uint<16*36> > & weight_stream1,
	#endif
	int row_repeat_time, // output_row ceil_div out tiles
	int weight_feed_total_size,
	ap_int<1> pingpong
)
{
	int weight_feed_total_size_by2 = weight_feed_total_size/2;
	for(int i=0;i<row_repeat_time;i++)
	{
		for(ap_uint<9> buffer_addr_offset=0; buffer_addr_offset<weight_feed_total_size_by2; buffer_addr_offset++)
		{
			ap_uint<64> temp18[WEIGHT_FEED_NUMBER_PER_PORT][9];
			#pragma HLS array_partition variable = temp18 complete

			ap_uint<10> buffer_addr=(pingpong,buffer_addr_offset);

			for(int buffer_idx =0; buffer_idx< WEIGHT_FEED_NUMBER_PER_PORT; buffer_idx++)
			{
			#pragma HLS unroll
				for(int j18=0;j18<9;j18++)
				{
				#pragma HLS unroll
					temp18[buffer_idx][j18]=weight_buff[buffer_idx][j18][buffer_addr];
				}
			}
			ap_uint<16*36> temp16x36[WEIGHT_FEED_NUMBER_PER_PORT];
			
			for(int buffer_idx =0; buffer_idx< WEIGHT_FEED_NUMBER_PER_PORT; buffer_idx++)
			{
				for(int j18=0;j18<9;j18++)
				{
				#pragma HLS unroll
				temp16x36[buffer_idx].range(j18*64+63,j18*64)=temp18[buffer_idx][j18];
				}
			}
			#if WEIGHT_FEED_NUMBER_PER_PORT == 2
			weight_stream0<<temp16x36[0];
			weight_stream1<<temp16x36[1];
			#endif
		}
	}
}


template<int dummy>
void weight_feed(
	ap_uint<128>* weight_DDR0,
	ap_uint<128>* weight_DDR1,
	ap_uint<128>* weight_DDR2,
	ap_uint<128>* weight_DDR3,
	#if WEIGHT_FEED_NUMBER_PER_PORT == 2
	hls::stream<ap_uint<16*36> > & weight_stream0_0,
	hls::stream<ap_uint<16*36> > & weight_stream0_1,
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    497 
    	hls::stream<ap_uint<16*36> > & weight_stream1_0,
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    498 
    	hls::stream<ap_uint<16*36> > & weight_stream1_1,
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    499 
    	hls::stream<ap_uint<16*36> > & weight_stream2_0,
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    500 
    	hls::stream<ap_uint<16*36> > & weight_stream2_1,
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    501 
    	hls::stream<ap_uint<16*36> > & weight_stream3_0,
    xliu79's avatar
    done golden input and weight feed
    xliu79 committed
    502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 
    	hls::stream<ap_uint<16*36> > & weight_stream3_1,
	#endif
	ap_uint<16> weight_total_load_number,
	ap_uint<16> weight_total_feed_size,
	ap_uint<16> ddr_load_length,
	ap_uint<16> ddr_load_length_per_feed,
	ap_uint<16> row_repeat_times,
	ap_uint<16> first_flag,
	ap_uint<16> last_flag
)
{
	static ap_uint<16> DDR_offset;
	static ap_uint<16> DDR_load_cnt;
	static ap_uint<1> pingpong;

	static ap_uint<64> weight_buff[4][WEIGHT_FEED_NUMBER_PER_PORT][9][WEIGHT_BUFFER_DEPTH];

	if(first_flag)
		DDR_offset=ddr_load_length;
		DDR_load_cnt=1;
		pingpong = 0;

	load_weight_ddr<0>(
	weight_DDR0,
	weight_DDR1,
	weight_DDR2,
	weight_DDR3,
	weight_buff,
	0,
	ddr_load_length,
	ddr_load_length_per_feed,
	~first_flag,
	pingpong);

	for(int cnt=0;cnt<weight_total_load_number;cnt++)
	{

		// printf("load cnt%d\n", cnt);
		pingpong = ~pingpong;

		load_weight_ddr<0>(
		weight_DDR0,
		weight_DDR1,
		weight_DDR2,
		weight_DDR3,
		weight_buff,
		DDR_offset,
		ddr_load_length,
		ddr_load_length_per_feed,
		last_flag & (DDR_load_cnt==0) ,
		pingpong);

		weight_stream<0>(
		weight_buff[0],
		#if WEIGHT_FEED_NUMBER_PER_PORT == 2
		weight_stream0_0,
		weight_stream0_1,
		#endif
		row_repeat_times, // output_row ceil_div out tiles
		weight_total_feed_size,
		~pingpong);

		weight_stream<0>(
		weight_buff[1],
		#if WEIGHT_FEED_NUMBER_PER_PORT == 2
		weight_stream1_0,
		weight_stream1_1,
		#endif
		row_repeat_times, // output_row ceil_div out tiles
		weight_total_feed_size,
		~pingpong);

		weight_stream<0>(
		weight_buff[2],
		#if WEIGHT_FEED_NUMBER_PER_PORT == 2
		weight_stream2_0,
		weight_stream2_1,
		#endif
		row_repeat_times, // output_row ceil_div out tiles
		weight_total_feed_size,
		~pingpong);

		weight_stream<0>(
		weight_buff[3],
		#if WEIGHT_FEED_NUMBER_PER_PORT == 2
		weight_stream3_0,
		weight_stream3_1,
		#endif
		row_repeat_times, // output_row ceil_div out tiles
		weight_total_feed_size,
		~pingpong);

		if(DDR_load_cnt == weight_total_load_number-1)
		{
			DDR_load_cnt = 0;
			DDR_offset = 0;
		}
		else
		{
			DDR_load_cnt+=1;
			DDR_offset+=ddr_load_length;
		}
	}	
}

#endif