Skip to content
Snippets Groups Projects
Normal view Permalink
wino_cell.cpp 27 KiB
Newer Older
  • Learn to ignore specific revisions
    • xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    1 2 3 4 
    #ifndef _WINO_CELL_HPP_
#define _WINO_CELL_HPP_

#include "wino_macro.h"
    xliu79's avatar
    C testbench version work
    xliu79 committed
    5 6 
    #include <ap_int.h>
#include <hls_stream.h>
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    7
    xliu79's avatar
    add some pragma  
    xliu79 committed
    8 
    #include "wino_buffer.cpp"
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    9 
    #include "wino_transform.cpp"
    xliu79's avatar
    C testbench version work
    xliu79 committed
    10
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    11 
    #include <dsp_builtins.h>
    xliu79's avatar
    C testbench version work
    xliu79 committed
    12 13
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    14 
    #define OUT_WIDTH 18
    xliu79's avatar
    C testbench version work
    xliu79 committed
    15 16 17 
    
#define B_WIDTH_IN 8
#define B_WIDTH_OUT 12
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    18 
    #define BTB_WIDTH_IN 12
    xliu79's avatar
    C testbench version work
    xliu79 committed
    19 20 21 22 23 24 25 26 27 28 29 30 31 32 
    #define BT_WIDTH_OUT 16

#define GTGG_WIDTH_IN 16
#define GTGG_WIDTH_W 16
#define GTGG_WIDTH_OUT 24

#define A_WIDTH_IN 24
#define A_WIDTH_OUT 28

#define AT_WIDTH_IN 28
#define AT_WIDTH_OUT 32
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 
    void load_input_tile(
	ap_int<BTB_WIDTH> input_tile[WINO_WIDTH][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE],
	ap_uint<BTB_WIDTH*BATCH_SIZE*WINO_DOMAIN_SIZE_SQUARE> stream_temp_reg[WINO_WIDTH][INDEPTH_MINITILE_SIZE]
)
{
	#pragma HLS pipeline
	#pragma HLS array_partition variable = input_tile complete dim=5
	#pragma HLS array_partition variable = input_tile complete dim=4
	#pragma HLS array_partition variable = input_tile complete dim=3
	#pragma HLS array_partition variable = input_tile complete dim=2
	#pragma HLS array_partition variable = input_tile complete dim=1

	#pragma HLS array_partition variable = stream_temp_reg0 complete dim=1
	#pragma HLS array_partition variable = stream_temp_reg0 complete dim=2 

	for(int w=0;w<WINO_WIDTH;w++){
	#pragma HLS unroll
		for(int iid=0;iid<INDEPTH_MINITILE_SIZE;iid++){
			#pragma HLS unroll
			for(int i=0;i<WINO_DOMAIN_SIZE;i++){
				#pragma HLS unroll
				for(int j=0;j<WINO_DOMAIN_SIZE;j++){
					#pragma HLS unroll
					for(int k=0;k<BATCH_SIZE;k++){
					#pragma HLS unroll
						//if(stream_pingpong_flag)
							input_tile[w][iid][i][j][k]=
							stream_temp_reg[w][iid].range(  ((i*WINO_DOMAIN_SIZE+j)*BATCH_SIZE+k)*BTB_WIDTH+7, ((i*WINO_DOMAIN_SIZE+j)*BATCH_SIZE+k)*BTB_WIDTH);
					}
				}
			}
		}
	}
}

void load_weight_tile(
		ap_int<W_WIDTH> weight_tile[WINO_HEIGHT][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE],
		ap_uint<W_WIDTH*INDEPTH_MINITILE_SIZE*WINO_DOMAIN_SIZE_SQUARE> weight_value_temp[WINO_HEIGHT]
)
{
		#pragma HLS array_partition variable = weight_tile complete dim=4
		#pragma HLS array_partition variable = weight_tile complete dim=3
		#pragma HLS array_partition variable = weight_tile complete dim=2
		#pragma HLS array_partition variable = weight_tile complete dim=1

		#pragma HLS array_partition variable=weight_value_temp complete 

		for(int wh=0;wh<WINO_HEIGHT;wh++)
		{
			#pragma HLS unroll
			for(int id=0;id<INDEPTH_MINITILE_SIZE;id++)
			{
				#pragma HLS unroll
				for(int i=0;i<WINO_DOMAIN_SIZE;i++)
				{
					#pragma HLS unroll
					for(int j=0;j<WINO_DOMAIN_SIZE;j++)
					{
						#pragma HLS unroll
						weight_tile[wh][id][i][j]=weight_value_temp[wh].range(
							(id*36+i*6+j)*W_WIDTH+W_WIDTH-1,
							(id*36+i*6+j)*W_WIDTH);
					}
				}
			}
		}
}

template<int dummy>
void element_wise_mult_6x6(
		ap_int<UV_MUL_WIDTH> UV_MUL_TILE[INDEPTH_MINITILE_SIZE/2][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE],
		ap_int<BTB_WIDTH> input_tile[INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE],
		ap_int<W_WIDTH> weight_tile[INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE],
		ap_int<1> ap_clk_div2
)
{
	#pragma HLS pipeline

	#pragma HLS array_partition variable=UV_MUL_TILE complete dim=1
	#pragma HLS array_partition variable=UV_MUL_TILE complete dim=2
	#pragma HLS array_partition variable=UV_MUL_TILE complete dim=3
	#pragma HLS array_partition variable=UV_MUL_TILE complete dim=4



	#pragma HLS array_partition variable = input_tile complete dim=4
	#pragma HLS array_partition variable = input_tile complete dim=3
	#pragma HLS array_partition variable = input_tile complete dim=2
	#pragma HLS array_partition variable = input_tile complete dim=1

	#pragma HLS array_partition variable = weight_tile complete dim=3
	#pragma HLS array_partition variable = weight_tile complete dim=2
	#pragma HLS array_partition variable = weight_tile complete dim=1

	for(int wr=0;wr<WINO_DOMAIN_SIZE;wr++)
	{
	#pragma HLS unroll
		for(int wc=0; wc<WINO_DOMAIN_SIZE;wc++)
		{
		#pragma HLS unroll
			for(int id2=0,id=0;id2<INDEPTH_MINITILE_SIZE/2;id2++,id+=2)
			{
			#pragma HLS unroll
				for(int b=0;b<BATCH_SIZE;b++)
				{
				#pragma HLS unroll
					ap_int<BTB_WIDTH> input_val0=input_tile[id][wr][wc][b];
					ap_int<BTB_WIDTH> input_val1=input_tile[id+1][wr][wc][b];
					UV_MUL_TILE[id2][wr][wc][b]=__builtin_mac16x2(
							input_tile[id][wr][wc][b],
							input_tile[id+1][wr][wc][b],
							weight_tile[id][wr][wc],
							weight_tile[id+1][wr][wc],
							0,1,ap_clk_div2);
				}
			}
		}
	}
}

void element_wise_mult(
		ap_int<UV_MUL_WIDTH> UV_MUL[OUTDEPTH_MINITILE_SIZE][WINO_WIDTH][INDEPTH_MINITILE_SIZE/2][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE],
		ap_int<BTB_WIDTH> input_tile[WINO_WIDTH][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE],
		ap_int<W_WIDTH> weight_tile[OUTDEPTH_MINITILE_SIZE][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE],
		ap_int<1> ap_clk_div2
)
{
	#pragma HLS pipeline

	#pragma HLS array_partition variable=UV_MUL complete dim=1
	#pragma HLS array_partition variable=UV_MUL complete dim=2
	#pragma HLS array_partition variable=UV_MUL complete dim=3
	#pragma HLS array_partition variable=UV_MUL complete dim=4
	#pragma HLS array_partition variable=UV_MUL complete dim=5
	#pragma HLS array_partition variable=UV_MUL complete dim=6

	#pragma HLS array_partition variable = input_tile complete dim=5
	#pragma HLS array_partition variable = input_tile complete dim=4
	#pragma HLS array_partition variable = input_tile complete dim=3
	#pragma HLS array_partition variable = input_tile complete dim=2
	#pragma HLS array_partition variable = input_tile complete dim=1

	#pragma HLS array_partition variable = weight_tile complete dim=4
	#pragma HLS array_partition variable = weight_tile complete dim=3
	#pragma HLS array_partition variable = weight_tile complete dim=2
	#pragma HLS array_partition variable = weight_tile complete dim=1

	for(int wr=0;wr<WINO_DOMAIN_SIZE;wr++)
	{
	#pragma HLS unroll
		for(int wc=0; wc<WINO_DOMAIN_SIZE;wc++)
		{
			for(int od=0;od<OUTDEPTH_MINITILE_SIZE;od++)
			{
				for(int id2=0,id=0;id2<INDEPTH_MINITILE_SIZE/2;id2++,id+=2)
				{
					ap_int<W_WIDTH> wino_val_d0 = weight_tile[od][id][wr][wc];
					ap_int<W_WIDTH> wino_val_d1 = weight_tile[od][id+1][wr][wc];
					for(int ww=0;ww<WINO_WIDTH;ww++)
					{
						for(int b=0;b<BATCH_SIZE;b++)
						{
							ap_int<BTB_WIDTH_IN> input_val0=input_tile[ww][id][wr][wc][b];
							ap_int<BTB_WIDTH_IN> input_val1=input_tile[ww][id+1][wr][wc][b];
							UV_MUL[od][ww][id2][wr][wc][b]=__builtin_mac16x2(input_val0,input_val1,wino_val_d0,wino_val_d1,0,1,ap_clk_div2);
						}
					}
				}
			}
		}
	}
}


template<class T, int dim1, int dim2, int dim3, int dim4> 
void load_reg_tile4(	T reg[dim1][dim2][dim3][dim4],
				T val[dim1][dim2][dim3][dim4])
{
	#pragma HLS pipeline
	#pragma HLS array_partition variable = reg complete dim=4
	#pragma HLS array_partition variable = reg complete dim=3
	#pragma HLS array_partition variable = reg complete dim=2
	#pragma HLS array_partition variable = reg complete dim=1

	#pragma HLS array_partition variable = val complete dim=4
	#pragma HLS array_partition variable = val complete dim=3
	#pragma HLS array_partition variable = val complete dim=2
	#pragma HLS array_partition variable = val complete dim=1

	for(int d4=0;d4<dim4;d4++){
	#pragma HLS unroll	
	for(int d3=0;d3<dim3;d3++){
	#pragma HLS unroll	
	for(int d2=0;d2<dim2;d2++){
	#pragma HLS unroll	
	for(int d1=0;d1<dim1;d1++){
	#pragma HLS unroll	
		reg[d1][d2][d3][d4]=val[d1][d2][d3][d4];
	}}}}

}


template<class T, int dim1, int dim2, int dim3> 
void load_reg_tile3(	T reg[dim1][dim2][dim3],
				T val[dim1][dim2][dim3])
{
	#pragma HLS pipeline
	#pragma HLS array_partition variable = reg complete dim=3
	#pragma HLS array_partition variable = reg complete dim=2
	#pragma HLS array_partition variable = reg complete dim=1

	#pragma HLS array_partition variable = val complete dim=3
	#pragma HLS array_partition variable = val complete dim=2
	#pragma HLS array_partition variable = val complete dim=1

	
	for(int d3=0;d3<dim3;d3++){
	#pragma HLS unroll	
	for(int d2=0;d2<dim2;d2++){
	#pragma HLS unroll	
	for(int d1=0;d1<dim1;d1++){
	#pragma HLS unroll	
		reg[d1][d2][d3]=val[d1][d2][d3];
	}}}

}

template<int dummy> 
void DSP_LOADER(ap_int<32> &rst, ap_int<16> a0, ap_int<16> a1, ap_int<16> b0, ap_int<16> b1, ap_int<32> accum, ap_int<1> clear, ap_int<1> ap_clk_div2)
{
	rst=__builtin_mac16x2(a0,a1,b0,b1,accum,clear,ap_clk_div2);
}


void wino_stream_block(
		hls::stream< ap_uint<BTB_WIDTH*BATCH_SIZE*WINO_DOMAIN_SIZE_SQUARE> > top_stream_in[WINO_WIDTH],
		hls::stream< ap_uint<W_WIDTH*INDEPTH_MINITILE_SIZE*WINO_DOMAIN_SIZE_SQUARE> > left_stream_in[4][WEIGHT_FEED_NUMBER_PER_PORT],
		ap_uint<OUT_WIDTH*2> out_buffer[OUTDEPTH_MINITILE_SIZE][WINO_OUT_SIZE_SQUARE][OUTPUT_BUFFER_DEPTH],
		ap_int<16> weightbuffer_load_indepth_number,
		ap_int<16> weightbuffer_load_outdepth_number,
		ap_int<16> weightbuffer_indepth_minitile_number,
		ap_int<16> weightbuffer_outdepth_minitile_number,
		ap_int<16> wino_tile_number_in_outwidth,
		ap_uint<24> total_input_stream_tile,
		ap_uint<16> loop_omini_base_reset_cycle,
		ap_uint<32> loop_wino_cell_bound,
		ap_uint<1> wino5x5_flag
		#if DEBUG_FILE_PRINT
		,ConvDesc_t conv_desc
		#endif
		,ap_uint<1> ap_clk_div2
		)
{
	#pragma HLS ap_stable port=ap_clk_div2

	#pragma HLS array_partition variable = out_buffer dim=1 complete

	ap_uint<BTB_WIDTH*BATCH_SIZE*WINO_DOMAIN_SIZE_SQUARE> stream_temp_reg0[WINO_WIDTH][INDEPTH_MINITILE_SIZE];
	#pragma HLS array_partition variable = stream_temp_reg0 complete dim=1
	#pragma HLS array_partition variable = stream_temp_reg0 complete dim=2 

	ap_uint<BTB_WIDTH*BATCH_SIZE*WINO_DOMAIN_SIZE_SQUARE> stream_temp_reg1[WINO_WIDTH][INDEPTH_MINITILE_SIZE];
	#pragma HLS array_partition variable = stream_temp_reg1 complete dim=1
	#pragma HLS array_partition variable = stream_temp_reg1 complete dim=2 

	ap_int<BTB_WIDTH> input_tile[WINO_WIDTH][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE];
	#pragma HLS array_partition variable = input_tile complete dim=5
	#pragma HLS array_partition variable = input_tile complete dim=4
	#pragma HLS array_partition variable = input_tile complete dim=3
	#pragma HLS array_partition variable = input_tile complete dim=2
	#pragma HLS array_partition variable = input_tile complete dim=1


	ap_int<BTB_WIDTH> input_tile_reg[INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE];
	#pragma HLS array_partition variable = input_tile complete dim=4
	#pragma HLS array_partition variable = input_tile complete dim=3
	#pragma HLS array_partition variable = input_tile complete dim=2
	#pragma HLS array_partition variable = input_tile complete dim=1



	#if 0
	memset(stream_temp_reg0,0xAA,2*2*36*2);
	memset(stream_temp_reg1,0xAA,2*2*36*2);
	#endif



	for(int i=0;i<INDEPTH_MINITILE_SIZE;i++)
	{
		#pragma hls pipeline
		for(int wino_array_col=0;wino_array_col<WINO_WIDTH;wino_array_col++)
		{
			#pragma HLS unroll
			top_stream_in[wino_array_col]>>stream_temp_reg0[wino_array_col][3];
			for(int imini_idx=0;imini_idx<INDEPTH_MINITILE_SIZE-1;imini_idx++)
			{
				#pragma HLS unroll
				stream_temp_reg0[wino_array_col][imini_idx]=stream_temp_reg0[wino_array_col][imini_idx+1];
			}
		}
	}



	



	ap_uint<1> load_input_flag=1;
	
	ap_uint<1> stream_pingpong_flag=1;


	ap_uint<24> loaded_input_stream_tile_number=1;
	#if DEBUG_FILE_PRINT
	total_input_stream_tile=conv_desc.weightbuffer_load_outdepth_number*conv_desc.weightbuffer_load_indepth_number*conv_desc.wino_tile_number_in_outwidth;
	#endif



	ap_uint<16> loop_omini_base_cnt=1;
	#if DEBUG_FILE_PRINT
	loop_omini_base_reset_cycle = conv_desc.weightbuffer_outdepth_minitile_number>INDEPTH_MINITILE_SIZE? conv_desc.weightbuffer_outdepth_minitile_number:INDEPTH_MINITILE_SIZE;
	#endif


	ap_uint<OUTPUT_BUFFER_DEPTH_BITWIDTH> out_buffer_address_outer=0;
	ap_uint<OUTPUT_BUFFER_DEPTH_BITWIDTH> out_buffer_address_inner=0;

	#if DEBUG_FILE_PRINT
	for(int oload_idx=0;oload_idx<weightbuffer_load_outdepth_number;oload_idx++)
	for(int iload_idx=0;iload_idx<weightbuffer_load_indepth_number;iload_idx++)
	for(int imini_base_idx=0;imini_base_idx<weightbuffer_indepth_minitile_number;imini_base_idx++)
	// for(int wino_tile_row_idx=0;wino_tile_row_idx<row_step;wino_tile_row_idx++)
	for(int wino_tile_idx=0;wino_tile_idx<wino_tile_number_in_outwidth;wino_tile_idx++)
	for(int omini_base_idx=0;omini_base_idx<loop_omini_base_reset_cycle ;omini_base_idx++)
	#else
	for(int cycle=0;cycle < loop_wino_cell_bound; cycle++)
	#endif
	{

		#pragma HLS pipeline
		#pragma HLS dependence variable=out_buffer inter false
		#pragma HLS dependence variable=out_buffer intra false



		ap_uint<1> load_input_flag_reg = (load_input_flag  && loaded_input_stream_tile_number !=  total_input_stream_tile);
		ap_uint<1> increment_out_inner_address_flag = (loop_omini_base_cnt < weightbuffer_outdepth_minitile_number);

		if(stream_pingpong_flag)
			load_input_tile(input_tile,stream_temp_reg0);
		else
			load_input_tile(input_tile,stream_temp_reg1);
		
		for(int wino_array_col=0;wino_array_col<WINO_WIDTH;wino_array_col++)
		{
			#pragma HLS unroll
			if(stream_pingpong_flag && load_input_flag_reg)
			{
				top_stream_in[wino_array_col]>>stream_temp_reg1[wino_array_col][3];
				for(int imini_idx=0;imini_idx<INDEPTH_MINITILE_SIZE-1;imini_idx++)
				{
					#pragma HLS unroll
					stream_temp_reg1[wino_array_col][imini_idx]=stream_temp_reg1[wino_array_col][imini_idx+1];
				}
			}
			else if(load_input_flag_reg)
			{
				top_stream_in[wino_array_col]>>stream_temp_reg0[wino_array_col][3];
				for(int imini_idx=0;imini_idx<INDEPTH_MINITILE_SIZE-1;imini_idx++)
				{
					#pragma HLS unroll
					stream_temp_reg0[wino_array_col][imini_idx]=stream_temp_reg0[wino_array_col][imini_idx+1];
				}
			}
		}



		ap_uint<W_WIDTH*INDEPTH_MINITILE_SIZE*WINO_DOMAIN_SIZE_SQUARE> weight_value_temp[WINO_HEIGHT];
		#pragma HLS array_partition variable=weight_value_temp complete 

		if(loop_omini_base_cnt <= weightbuffer_outdepth_minitile_number)
		{
			for(int i=0;i<4;i++)
			{
				#pragma HLS unroll
				for(int j=0;j<WEIGHT_FEED_NUMBER_PER_PORT;j++)
				{
					#pragma HLS unroll
					left_stream_in[i][j]>>weight_value_temp[i*WEIGHT_FEED_NUMBER_PER_PORT+j];
				}
			}
		}

		ap_int<W_WIDTH> weight_tile[WINO_HEIGHT][INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE];
		#pragma HLS array_partition variable = weight_tile complete dim=4
		#pragma HLS array_partition variable = weight_tile complete dim=3
		#pragma HLS array_partition variable = weight_tile complete dim=2
		#pragma HLS array_partition variable = weight_tile complete dim=1

		load_weight_tile(weight_tile,weight_value_temp);

		for(int wino_array_row=0;wino_array_row<WINO_HEIGHT;wino_array_row)
		{
			#pragma HLS unroll
			for(int wino_array_col=0;wino_array_col<WINO_WIDTH;wino_array_col++)
			{
				#pragma HLS unroll
				ap_int<BTB_WIDTH> input_tile_reg[INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE];
				#pragma HLS array_partition variable = input_tile_reg complete dim=4
				#pragma HLS array_partition variable = input_tile_reg complete dim=3
				#pragma HLS array_partition variable = input_tile_reg complete dim=2
				#pragma HLS array_partition variable = input_tile_reg complete dim=1

				ap_int<W_WIDTH> weight_tile_reg[INDEPTH_MINITILE_SIZE][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE];
				#pragma HLS array_partition variable = weight_tile_reg complete dim=3
				#pragma HLS array_partition variable = weight_tile_reg complete dim=2
				#pragma HLS array_partition variable = weight_tile_reg complete dim=1

				ap_int<UV_MUL_WIDTH> UV_MUL_TILE[INDEPTH_MINITILE_SIZE/2][WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE];
				#pragma HLS array_partition variable=UV_MUL_TILE complete dim=1
				#pragma HLS array_partition variable=UV_MUL_TILE complete dim=2
				#pragma HLS array_partition variable=UV_MUL_TILE complete dim=3
				#pragma HLS array_partition variable=UV_MUL_TILE complete dim=4

				load_reg_tile4<ap_int<BTB_WIDTH>,INDEPTH_MINITILE_SIZE,WINO_DOMAIN_SIZE,WINO_DOMAIN_SIZE,BATCH_SIZE>(input_tile_reg, input_tile[wino_array_col] );
				load_reg_tile3<ap_int<W_WIDTH>,INDEPTH_MINITILE_SIZE,WINO_DOMAIN_SIZE,WINO_DOMAIN_SIZE>(weight_tile_reg, weight_tile[wino_array_row]);
				element_wise_mult_6x6<0>(UV_MUL_TILE,input_tile_reg,weight_tile_reg, ap_clk_div2 );

				ap_int<UV_WIDTH> UV[WINO_DOMAIN_SIZE][WINO_DOMAIN_SIZE][BATCH_SIZE];
				#pragma HLS array_partition variable=UV complete dim=1
				#pragma HLS array_partition variable=UV complete dim=2
				#pragma HLS array_partition variable=UV complete dim=3

				for(int wino_row=0;wino_row<WINO_DOMAIN_SIZE;wino_row++)
				{
					#pragma HLS unroll
					for(int wino_col=0;wino_col<WINO_DOMAIN_SIZE;wino_col++)
					{
						#pragma HLS unroll
						for(int b=0;b<BATCH_SIZE;b++)
						{
							ap_int<UV_MUL_WIDTH> temp=0;
							for(int id2=0;id2<INDEPTH_MINITILE_SIZE/2;id2++)
							{
								#pragma HLS unroll
								temp+=UV_MUL_TILE[id2][wino_row][wino_col];
							}
							UV[wino_row][wino_col][b]=temp>>UV_QUANT_BIT;
						}
					}
				}


				ap_int<UVA_WIDTH> UVA[WINO_DOMAIN_SIZE][WINO_OUT_SIZE][BATCH_SIZE];
				#pragma HLS array_partition variable=UVA complete dim=1
				#pragma HLS array_partition variable=UVA complete dim=2
				#pragma HLS array_partition variable=UVA complete dim=3

		
				for(int ridx=0;ridx<WINO_DOMAIN_SIZE;ridx++)
				{
					#pragma HLS unroll
					for(int bidx=0;bidx<BATCH_SIZE;bidx++)
					{
						#pragma HLS unroll
						UVA_row(UVA,UV,ridx,bidx,wino5x5_flag);
					}
				}

				ap_int<ATA_WIDTH> ATA[WINO_OUT_SIZE][WINO_OUT_SIZE][BATCH_SIZE];
				#pragma HLS array_partition variable=ATA complete dim=1
				#pragma HLS array_partition variable=ATA complete dim=2
				#pragma HLS array_partition variable=ATA complete dim=3
				for(int cidx=0;cidx<WINO_OUT_SIZE;cidx++)
				{
					#pragma HLS unroll
					for(int bidx=0;bidx<BATCH_SIZE;bidx++)
					{
						#pragma HLS unroll
						ATA_col(ATA,UVA,cidx,bidx,wino5x5_flag);
					}
				}


				

			}
		}
		
		


		
		// element_wise_mult(UV_MUL,input_tile,weight_tile,ap_clk_div2);
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    532 533
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    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 
    		
		
		stream_pingpong_flag=~stream_pingpong_flag;
		

		if(loop_omini_base_cnt==loop_omini_base_reset_cycle ) 
		{
			load_input_flag = 1;
		}
		else if(loop_omini_base_cnt==INDEPTH_MINITILE_SIZE)
		{
			load_input_flag = 0;
		}

		if(loop_omini_base_cnt==loop_omini_base_reset_cycle)
		{
			loop_omini_base_cnt=1;
			loaded_input_stream_tile_number++;
		}
		else
		{
			loop_omini_base_cnt++;
		}	
	}
}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    560 561 562 563 564 565 566 567 568 569 570 571 572 
    


void wino6x6_stream_cell(
		hls::stream< ap_uint<GTGG_WIDTH_IN*2*36> > & top_stream_in,
		hls::stream< ap_uint<GTGG_WIDTH_IN*2*36> > & bottom_stream_out,
		hls::stream< ap_uint<W_WIDTH*2*36> > &left_stream_in,
		hls::stream< ap_uint<W_WIDTH*2*36> > &right_stream_out,
		ap_uint<OUT_WIDTH*2> out_buffer[16][1024],
		ap_int<16> weight_indepth_load_number,
		ap_int<16> weight_outdepth_load_number,
		ap_int<16> weight_indepth_feed_size,
		ap_int<16> weight_outdepth_feed_size,
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    573 574 575 576 577 578 
    		ap_int<16> row_tile_number
		#if DEBUG_FILE_PRINT
		,int ROW_IDX, int COL_IDX
		#endif

		)
    xliu79's avatar
    C testbench version work
    xliu79 committed
    579 580 
    {
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    581 
    #pragma HLS array_partition variable = out_buffer dim=1 complete
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    582 583 
    	ap_uint<GTGG_WIDTH_IN*2*36> stream_temp_reg0[2];
	ap_uint<GTGG_WIDTH_IN*2*36> stream_temp_reg1[2];
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    584 585 586 
    #pragma HLS array_partition variable = stream_temp_reg0 complete
#pragma HLS array_partition variable = stream_temp_reg1 complete
	#if DEBUG_FILE_PRINT
    xliu79's avatar
    input correct  
    xliu79 committed
    587 588 
    	memset(stream_temp_reg0,0xAA,2*2*36*2);
	memset(stream_temp_reg1,0xAA,2*2*36*2);
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    589 
    	#endif
    xliu79's avatar
    C testbench version work
    xliu79 committed
    590
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    591 
    	ap_int<W_WIDTH> G1[36][2];
    xliu79's avatar
    C testbench version work
    xliu79 committed
    592 593 
    	#pragma HLS array_partition variable=G1
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    594 
    	ap_uint<1> stream_pingpong_flag=0;
    xliu79's avatar
    C testbench version work
    xliu79 committed
    595
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    596 597 598 599 
    	int write_idx=0;
	int weight_stream_idx=0;
	int input_stream_idx =0;
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    600 
    	#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    601 602 603 604 605 606 
    	char filename[100];
	sprintf(filename,"wino_cell_output_%d_%d.txt",ROW_IDX, COL_IDX);
	char input_stream_filename[100];
	sprintf(input_stream_filename,"wino_cell_input_%d_%d.txt",ROW_IDX, COL_IDX);
	char weight_stream_filename[100];
	sprintf(weight_stream_filename,"wino_cell_weight_%d_%d.txt",ROW_IDX, COL_IDX);
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    607 
    	#endif
    xliu79's avatar
    C testbench version work
    xliu79 committed
    608
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    609 610 
    	top_stream_in>>stream_temp_reg0[0];
	bottom_stream_out<<stream_temp_reg0[0];
    xliu79's avatar
    add some pragma  
    xliu79 committed
    611 
    	#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    612 
    	wino_cell_stream_input<0>(stream_temp_reg0[0],input_stream_idx,input_stream_filename);
    xliu79's avatar
    add some pragma  
    xliu79 committed
    613 
    	#endif
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    614 615 
    	top_stream_in>>stream_temp_reg0[1];
	bottom_stream_out<<stream_temp_reg0[1];
    xliu79's avatar
    add some pragma  
    xliu79 committed
    616 
    	#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    617 
    	wino_cell_stream_input<0>(stream_temp_reg0[1],input_stream_idx,input_stream_filename);
    xliu79's avatar
    add some pragma  
    xliu79 committed
    618 
    	#endif
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    619 620
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    621 622
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    623 624 625 626 627 628 
    	ITER0:for(int iter0=0;iter0<weight_indepth_load_number;iter0++)
	ITER1:for(int iter1=0;iter1<weight_outdepth_load_number;iter1++)
	ITER2:for(int iter2=0;iter2<row_tile_number;iter2++)
	ITER3:for(int iter3=0;iter3<weight_indepth_feed_size/2;iter3++)
	{
	ITER4:for(int iter4=0;iter4<weight_outdepth_feed_size;iter4++)
    xliu79's avatar
    input correct  
    xliu79 committed
    629 
    	{
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    630
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    631 632 633 
    		#pragma HLS pipeline
#pragma HLS dependence variable=out_buffer inter false
#pragma HLS dependence variable=out_buffer intra false
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    634 635 
    			ap_int<GTGG_WIDTH_IN> input_tile[2][36][2];
			#pragma HLS array_partition variable = input_tile complete
    xliu79's avatar
    input correct  
    xliu79 committed
    636 
    			for(int i=0;i<36;i++)
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    637 
    			{
    xliu79's avatar
    input correct  
    xliu79 committed
    638 639 640 
    			#pragma HLS unroll
				if(stream_pingpong_flag)
				{
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    641 642 643 644 
    					input_tile[0][i][0]=stream_temp_reg1[0].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN-1,i*GTGG_WIDTH_IN*2);
					input_tile[0][i][1]=stream_temp_reg1[0].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN*2-1,i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN);
					input_tile[1][i][0]=stream_temp_reg1[1].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN-1,i*GTGG_WIDTH_IN*2);
					input_tile[1][i][1]=stream_temp_reg1[1].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN*2-1,i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN);
    xliu79's avatar
    input correct  
    xliu79 committed
    645 646 647 
    				}
				else
				{
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    648 649 650 651 
    					input_tile[0][i][0]=stream_temp_reg0[0].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN-1,i*GTGG_WIDTH_IN*2);
					input_tile[0][i][1]=stream_temp_reg0[0].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN*2-1,i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN);
					input_tile[1][i][0]=stream_temp_reg0[1].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN-1,i*GTGG_WIDTH_IN*2);
					input_tile[1][i][1]=stream_temp_reg0[1].range(i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN*2-1,i*GTGG_WIDTH_IN*2+GTGG_WIDTH_IN);
    xliu79's avatar
    input correct  
    xliu79 committed
    652 
    				}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    653 
    			}
    xliu79's avatar
    input correct  
    xliu79 committed
    654 655 
    
			if(iter4<2 )
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    656 
    			{
    xliu79's avatar
    input correct  
    xliu79 committed
    657 658 659 660 
    				if(stream_pingpong_flag)
				{
					top_stream_in>>stream_temp_reg0[iter4];
					bottom_stream_out<<stream_temp_reg0[iter4];
    xliu79's avatar
    add some pragma  
    xliu79 committed
    661 
    					#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    662 
    					wino_cell_stream_input<0>(stream_temp_reg0[iter4],input_stream_idx,input_stream_filename);
    xliu79's avatar
    add some pragma  
    xliu79 committed
    663 
    					#endif
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    664
    xliu79's avatar
    input correct  
    xliu79 committed
    665 666 667 668 669 
    				}
				else
				{
					top_stream_in>>stream_temp_reg1[iter4];
					bottom_stream_out<<stream_temp_reg1[iter4];
    xliu79's avatar
    add some pragma  
    xliu79 committed
    670 
    					#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    671 
    					wino_cell_stream_input<0>(stream_temp_reg1[iter4],input_stream_idx,input_stream_filename);
    xliu79's avatar
    add some pragma  
    xliu79 committed
    672 
    					#endif
    xliu79's avatar
    input correct  
    xliu79 committed
    673 
    				}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    674 
    			}
    xliu79's avatar
    C testbench version work
    xliu79 committed
    675 676
    xliu79's avatar
    input correct  
    xliu79 committed
    677 678 679 
    			ap_uint<W_WIDTH*2*36> weight_value_temp;
			left_stream_in>>weight_value_temp;
			right_stream_out<<weight_value_temp;
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    680
    xliu79's avatar
    C testbench version work
    xliu79 committed
    681
    xliu79's avatar
    input correct  
    xliu79 committed
    682 683 684 685 686 687 
    			for(int i=0;i<36;i++)
			{
			#pragma HLS unroll
				G1[i][0]=weight_value_temp.range(i*W_WIDTH+W_WIDTH-1,i*W_WIDTH);
				G1[i][1]=weight_value_temp.range(i*W_WIDTH+W_WIDTH-1+288,i*W_WIDTH +288 );
			}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    688
    xliu79's avatar
    input correct  
    xliu79 committed
    689 690 691 692 693 694 
    			ap_int<GTGG_WIDTH_OUT> UV[36][2];
			#pragma HLS array_partition variable=UV complete
			
			for(int i=0;i<36;i++)
			{
			#pragma HLS unroll
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    695 696 
    				UV[i][0] = G1[i][0]*input_tile[0][i][0] + G1[i][1]*input_tile[1][i][0];
				UV[i][1] = G1[i][0]*input_tile[0][i][1] + G1[i][1]*input_tile[1][i][1];
    xliu79's avatar
    input correct  
    xliu79 committed
    697 
    			}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    698
    xliu79's avatar
    input correct  
    xliu79 committed
    699 700 
    			ap_int<A_WIDTH_OUT>  UVA[6][4][2];
			#pragma HLS array_partition variable=UVA complete
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    701
    xliu79's avatar
    input correct  
    xliu79 committed
    702 
    			for(int b_idx=0;b_idx<2;b_idx++)
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    703 
    			{
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    704 
    				#pragma HLS unroll
    xliu79's avatar
    input correct  
    xliu79 committed
    705 706 
    				for(int i=0;i<6;i++)
				{
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    707 
    					#pragma HLS unroll
    xliu79's avatar
    input correct  
    xliu79 committed
    708 709 710 711 712 
    					UVA[i][0][b_idx]=(ap_int<A_WIDTH_OUT>) UV[i*6+0][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+1][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+2][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+3][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+4][b_idx];
					UVA[i][1][b_idx]=(ap_int<A_WIDTH_OUT>) UV[i*6+1][b_idx]-(ap_int<A_WIDTH_OUT>) UV[i*6+2][b_idx]+(((ap_int<A_WIDTH_OUT>) UV[i*6+3][b_idx]-(ap_int<A_WIDTH_OUT>) UV[i*6+4][b_idx])<<1);
					UVA[i][2][b_idx]=(ap_int<A_WIDTH_OUT>) UV[i*6+1][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+2][b_idx]+(((ap_int<A_WIDTH_OUT>) UV[i*6+3][b_idx]+(ap_int<A_WIDTH_OUT>) UV[i*6+4][b_idx])<<2);
					UVA[i][3][b_idx]=(ap_int<A_WIDTH_OUT>) UV[i*6+1][b_idx]-(ap_int<A_WIDTH_OUT>) UV[i*6+2][b_idx]+(((ap_int<A_WIDTH_OUT>) UV[i*6+3][b_idx]-(ap_int<A_WIDTH_OUT>) UV[i*6+4][b_idx])<<3)+(ap_int<A_WIDTH_OUT>) UV[i*6+5][b_idx];
				}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    713 
    			}
    xliu79's avatar
    input correct  
    xliu79 committed
    714 715 716 717 718 
    

			int address = iter1*row_tile_number* weight_outdepth_feed_size + iter4*row_tile_number + iter2;
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    719 
    			ap_int<36> outbuffer_value[16];
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    720 
    			#pragma HLS array_partition variable= outbuffer_value complete
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    721 722 
    			for(int i=0;i<16;i++)
			{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    723 
    				#pragma HLS unroll
    xliu79's avatar
    input correct  
    xliu79 committed
    724 
    				outbuffer_value[i]=out_buffer[i][address];
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    725 726 
    			}
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 
    

			ap_int<AT_WIDTH_OUT> ATUVA[16][2];
			#pragma HLS array_partition variable=ATUVA complete


			for(int b_idx=0;b_idx<2;b_idx++)
			{
			#pragma HLS unroll
				for(int i=0;i<4;i++)
				{
				#pragma HLS unroll
					ATUVA[0+i][b_idx]=(ap_int<AT_WIDTH_OUT>) UVA[0][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[1][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[2][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[3][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[4][i][b_idx];
					ATUVA[4+i][b_idx]=(ap_int<AT_WIDTH_OUT>) UVA[1][i][b_idx]-(ap_int<AT_WIDTH_OUT>) UVA[2][i][b_idx]+(((ap_int<AT_WIDTH_OUT>) UVA[3][i][b_idx]-(ap_int<AT_WIDTH_OUT>) UVA[4][i][b_idx])<<1);
					ATUVA[8+i][b_idx]=(ap_int<AT_WIDTH_OUT>) UVA[1][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[2][i][b_idx]+(((ap_int<AT_WIDTH_OUT>) UVA[3][i][b_idx]+(ap_int<AT_WIDTH_OUT>) UVA[4][i][b_idx])<<2);
					ATUVA[12+i][b_idx]=(ap_int<AT_WIDTH_OUT>) UVA[1][i][b_idx]-(ap_int<AT_WIDTH_OUT>) UVA[2][i][b_idx]+(((ap_int<AT_WIDTH_OUT>) UVA[3][i][b_idx]-(ap_int<AT_WIDTH_OUT>) UVA[4][i][b_idx])<<3) + (ap_int<AT_WIDTH_OUT>) UVA[5][i][b_idx];
				}
			}
    xliu79's avatar
    add some pragma  
    xliu79 committed
    745 
    			#if DEBUG_FILE_PRINT
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    746 
    			attach_output_vector<0>(ATUVA,write_idx,filename);
    xliu79's avatar
    add some pragma  
    xliu79 committed
    747 
    			#endif
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    748 749 
    
			ap_int<18> out_residual[16][2];
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    750
    xliu79's avatar
    input correct  
    xliu79 committed
    751 752 753 754 
    			if(iter0 ==0 && iter3 ==0)
			{
				for(int i=0;i<16;i++)
				{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    755 
    				#pragma HLS unroll
    xliu79's avatar
    input correct  
    xliu79 committed
    756 757 758 759 760 761 762 763 
    					out_residual[i][0]=0;
					out_residual[i][1]=0;
				}
			}
			else
			{
				for(int i=0;i<16;i++)
				{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    764 
    				#pragma HLS unroll
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    765 766 
    					out_residual[i][0]=outbuffer_value[i].range(17,0);
					out_residual[i][1]=outbuffer_value[i].range(35,18);
    xliu79's avatar
    input correct  
    xliu79 committed
    767 768 
    				}
			}
    xliu79's avatar
    C testbench version work
    xliu79 committed
    769
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    770
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    771 
    			ap_int<18> outbuffer_writeback_value[16][2];
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    772 773 
    			#pragma HLS array_partition variable = outbuffer_writeback_value complete dim=1
			#pragma HLS array_partition variable = outbuffer_writeback_value complete dim=2
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    774 
    			for(int i=0;i<16;i++)
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    775 
    			{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    776 
    			#pragma HLS unroll
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    777 778 
    				outbuffer_writeback_value[i][0]=ATUVA[i][0]+out_residual[i][0];
				outbuffer_writeback_value[i][1]=ATUVA[i][1]+out_residual[i][1];
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    779 
    			}
    xliu79's avatar
    C testbench version work
    xliu79 committed
    780
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    781 782 
    
			ap_int<18*2> outbuffer_writeback_value_batch[16];
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    783 
    			#pragma HLS array_partition variable = outbuffer_writeback_value_batch complete
    xliu79's avatar
    input correct  
    xliu79 committed
    784 785 
    			for(int i=0;i<16;i++)
			{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    786 
    				#pragma HLS unroll
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    787 
    				outbuffer_writeback_value_batch[i]=(outbuffer_writeback_value[i][1],outbuffer_writeback_value[i][0]  );
    xliu79's avatar
    input correct  
    xliu79 committed
    788 
    			}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    789
    xliu79's avatar
    input correct  
    xliu79 committed
    790 791 
    			for(int i=0;i<16;i++)
			{
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    792 
    #pragma HLS unroll
    xliu79's avatar
    still have some bug in output write  
    xliu79 committed
    793 
    				out_buffer[i][address]=outbuffer_writeback_value_batch[i];
    xliu79's avatar
    input correct  
    xliu79 committed
    794 
    			}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    795 796 
    		}
		stream_pingpong_flag=~stream_pingpong_flag;
    xliu79's avatar
    C testbench version work
    xliu79 committed
    797 
    	}
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    798
    xliu79's avatar
    C testbench version work
    xliu79 committed
    799 800 801 
    }
    xliu79's avatar
    mid check point of changing input load and buffering part  
    xliu79 committed
    802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 
    // //template<int dummy>
// void wino6x6_stream_bottomend(
// 	hls::stream< ap_uint<GTGG_WIDTH_IN*2*36> > & top_stream_in,
// 	ap_int<16> weight_indepth_load_number,
// 	ap_int<16> weight_outdepth_load_number,
// 	ap_int<16> weight_indepth_feed_size,
// 	ap_int<16> row_tile_number
// )
// {
// 	ap_uint<GTGG_WIDTH_IN*2*36> dummy_temp;
// 	for(int i=0;i<weight_indepth_load_number*weight_outdepth_load_number*weight_indepth_feed_size*row_tile_number+2;i++)
// 	{
// 		top_stream_in.read(dummy_temp);
// 	}
// }
    xliu79's avatar
    done golden input and weight feed  
    xliu79 committed
    817 818 819 820
    xliu79's avatar
    C testbench version work
    xliu79 committed
    821 822 823 824 825 826 827 828
    xliu79's avatar
    new version, update input/output structure, update input reading module and validated  
    xliu79 committed
    829 
    #endif