8 KERNEL PROC Changes

Changes: Eltwise Add Software layer 8 Kernel Proc functionality fixes
Xilinx · Aug 16, 2018 · 2ac2f2f · 2ac2f2f
1 parent 84318fb
commit 2ac2f2f
Show file tree

Hide file tree

Showing 39 changed files with 1,135 additions and 107 deletions.
diff --git a/SD_Card/zcu102/alexnet.elf b/SD_Card/zcu102/alexnet.elf
diff --git a/SD_Card/zcu102/alexnetfcn.elf b/SD_Card/zcu102/alexnetfcn.elf
diff --git a/SD_Card/zcu102/eval.elf b/SD_Card/zcu102/eval.elf
diff --git a/SD_Card/zcu102/googlenet.elf b/SD_Card/zcu102/googlenet.elf
diff --git a/SD_Card/zcu102/lib/libparser_arm.so b/SD_Card/zcu102/lib/libparser_arm.so
diff --git a/SD_Card/zcu102/lib/libxstack.so b/SD_Card/zcu102/lib/libxstack.so
diff --git a/SD_Card/zcu102/resnet50.elf b/SD_Card/zcu102/resnet50.elf
diff --git a/SD_Card/zcu102/ssd.elf b/SD_Card/zcu102/ssd.elf
diff --git a/SD_Card/zcu102/vgg.elf b/SD_Card/zcu102/vgg.elf
diff --git a/SD_Card/zcu104/alexnet.elf b/SD_Card/zcu104/alexnet.elf
diff --git a/SD_Card/zcu104/alexnetfcn.elf b/SD_Card/zcu104/alexnetfcn.elf
diff --git a/SD_Card/zcu104/eval.elf b/SD_Card/zcu104/eval.elf
diff --git a/SD_Card/zcu104/googlenet.elf b/SD_Card/zcu104/googlenet.elf
diff --git a/SD_Card/zcu104/lib/libparser_arm.so b/SD_Card/zcu104/lib/libparser_arm.so
diff --git a/SD_Card/zcu104/lib/libxstack.so b/SD_Card/zcu104/lib/libxstack.so
diff --git a/SD_Card/zcu104/resnet50.elf b/SD_Card/zcu104/resnet50.elf
diff --git a/SD_Card/zcu104/ssd.elf b/SD_Card/zcu104/ssd.elf
diff --git a/SD_Card/zcu104/vgg.elf b/SD_Card/zcu104/vgg.elf
diff --git a/software/bufmgmt/xi_buf_mgmt_utils.hpp b/software/bufmgmt/xi_buf_mgmt_utils.hpp
@@ -394,12 +394,14 @@ int convWeightsSize(int out_depth, int in_depth, int kernel_h, int kernel_w, int
 	int size_of_weights;
 	int no_kernals,no_planes;
 
-	remainder=no_kernals_1%KER_PROC;
+	//remainder=no_kernals_1%KER_PROC;
+	remainder=no_kernals_1%CONV_IO_PACK_ELEMS;
 
 	if(remainder==0)
 		no_kernals=no_kernals_1;
 	else
-		no_kernals=no_kernals_1+KER_PROC-remainder;
+		//no_kernals=no_kernals_1+KER_PROC-remainder;
+		no_kernals=no_kernals_1+CONV_IO_PACK_ELEMS-remainder;
 
 	if(no_kernals<4)
 		no_kernals=4; //TODO: make it as a argument minimum number of kernels are processing

diff --git a/software/checkers/checkers.cpp b/software/checkers/checkers.cpp
@@ -124,7 +124,7 @@ int errorCheck(
 		return -1;
 	}
 
-	if(inLayer.kernType == CONV)
+	if((inLayer.kernType == CONV) || (inLayer.kernType == ELTWISEADD))
 	{
 #if PACKED_INOUT
 		status = cpCheck_packed(inLayer);

diff --git a/software/checkers/checkers_utils.hpp b/software/checkers/checkers_utils.hpp
@@ -19,7 +19,6 @@ limitations under the License.
 
 #include "ap_int.h"
 #include "../common/kernelinfo_class.h"
-#include "../swkernels/xi_format_converter.hpp"
 
 //# Checks Convolution/Pool funtionality
 int cpCheck(
@@ -191,6 +190,12 @@ int cpCheck(
 
 		relu   = params[10];
 		en_batch_size_one = params[126];//58];
+
+		int pool_split_cnt = params[110];
+		if(pool_split_cnt>1)
+		{
+			indepth = params[109];
+		}
 	}
 	else if(inLayer.kernType == NORM)
 	{
@@ -200,6 +205,14 @@ int cpCheck(
 		en_batch_size_one = params[8];
 		offline_quant_mode = params[9];
 	}
+	else if(inLayer.kernType == ELTWISEADD)
+	{
+		height = params[4];
+		width  = params[5];
+		indepth = params[3];
+		en_batch_size_one = 0;//params[8];
+		offline_quant_mode = 1;//params[9];
+	}
 	else //POOL
 	{
 		height = params[2];
@@ -690,7 +703,7 @@ int cpCheck(
 		return -1;
 	else
 		return 0;
-#endif
+#endif  //if LAYERWISE_OUTPUT_WRITE
 
 }
 // cpCheck
@@ -3019,6 +3032,11 @@ void LoadInputData(xChangeLayer inLayer, char *lay_path)
 		planes	= params[5];
 		height	= params[0];
 		width	= params[1];
+		int pool_split_cnt = params[110];
+		if(pool_split_cnt>1)
+		{
+			planes = params[109];
+		}
 	}
 
 	if(inLayer.kernType == POOL)

diff --git a/software/common/kernelinfo_class.cpp b/software/common/kernelinfo_class.cpp
@@ -198,12 +198,14 @@ void update_conv_module_opcode_table(Opcode_info & obj_module_opcode,map<opcode_
 	/****** OPCODE_CRELU ***********/
 	obj_module_opcode.op_code = OPCODE_CRELU;
 	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
-	/****** OPCODE_BN ***********/
+	/****** OPCODE_BN + SCALE***********/
 	obj_module_opcode.op_code = OPCODE_BN;
 	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
 	/****** OPCODE_ELTWISE ***********/
+#if XI_DIET_CHAI_Z==0
 	obj_module_opcode.op_code = OPCODE_ELTWISE;
 	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
+#endif
 	/****** OPCODE_FUSE_BN_CONV ***********/
 	obj_module_opcode.op_code = OPCODE_FUSE_BN_CONV;
 	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
@@ -1283,6 +1285,163 @@ void update_permute_xunpack_module_opcode_table(Opcode_info & obj_module_opcode,
 	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
 
 }
+void update_opcde_sw_eltwise_module_opcode_table(Opcode_info & obj_module_opcode,map<opcode_num_e, Opcode_info>  & opcode_info_map,int batch_size){
+
+	Module_info module_default;
+	module_default.module_target_info = CPU;
+	module_default.module_type = SW_ELTWISE_ADD;
+	module_default.module_name = "Eltwise"; /// TODO abid needs to update the layer information based on the enum instead of layer name
+	module_default.num_total_ports=11; // TODO need to check with buffer management & init module
+	obj_module_opcode.mega_module = module_default;
+
+	/*** xcustom mega module ports  & these port order should be same as deploy file*********/
+
+	Port_info port_defualt;
+	port_defualt.port_width=XI_128BIT_WIDTH;
+	port_defualt.data_type = XI_INT8;
+	port_defualt.size_datatype = XI_INT8_S;
+	port_defualt.memory_align_required=32; // TODO need to check with buffer management & init module
+	port_defualt.memory_map_info = __MALLOC_TYPE;
+	port_defualt.port_enum=XI_INPUT;
+	port_defualt.qant_scheme_type = g_quant_value;
+
+
+	packinfo input_pack;
+
+	input_pack.dim = XI_CHANNEL;
+	input_pack.mem_contiguous = true;
+	input_pack.mod_div_factor = (port_defualt.port_width/port_defualt.size_datatype)/batch_size;
+	input_pack.sep=XI_DIVISION;
+
+	port_defualt.pack_info_vec.push_back(input_pack);
+
+	input_pack.dim = XI_HEIGHT;
+	input_pack.mem_contiguous = true;
+	input_pack.mod_div_factor = 0;
+	input_pack.sep=XI_NONE;
+
+	port_defualt.pack_info_vec.push_back(input_pack);
+
+	input_pack.dim = XI_WIDTH;
+	input_pack.mem_contiguous = true;
+	input_pack.mod_div_factor = 0;
+	input_pack.sep=XI_NONE;
+
+	port_defualt.pack_info_vec.push_back(input_pack);
+
+	input_pack.dim = XI_INTERLEAVED;
+	input_pack.mem_contiguous = true;
+	input_pack.mod_div_factor = (port_defualt.port_width/port_defualt.size_datatype)/batch_size;
+	input_pack.sep=XI_MODULO;
+	port_defualt.pack_info_vec.push_back(input_pack);
+
+	input_pack.dim = XI_BATCH;
+	input_pack.mem_contiguous = true;
+	input_pack.mod_div_factor = 0;
+	input_pack.sep=XI_NONE;
+
+	port_defualt.pack_info_vec.push_back(input_pack);
+
+	obj_module_opcode.port_vec.push_back(port_defualt);
+
+
+	packinfo input_pack1;
+	Port_info port_defualt2;
+	port_defualt2.port_width=XI_128BIT_WIDTH;
+	port_defualt2.data_type = XI_INT8;
+	port_defualt2.size_datatype = XI_INT8_S;
+	port_defualt2.memory_align_required=32; // TODO need to check with buffer management & init module
+	port_defualt2.memory_map_info = __MALLOC_TYPE;
+	port_defualt2.port_enum=XI_INPUT;
+	port_defualt2.qant_scheme_type = g_quant_value;
+
+	input_pack1.dim = XI_CHANNEL;
+	input_pack1.mem_contiguous = true;
+	input_pack1.mod_div_factor = (port_defualt2.port_width/port_defualt2.size_datatype)/batch_size;
+	input_pack1.sep=XI_DIVISION;
+
+	port_defualt2.pack_info_vec.push_back(input_pack1);
+
+	input_pack1.dim = XI_HEIGHT;
+	input_pack1.mem_contiguous = true;
+	input_pack1.mod_div_factor = 0;
+	input_pack1.sep=XI_NONE;
+
+	port_defualt2.pack_info_vec.push_back(input_pack1);
+
+	input_pack1.dim = XI_WIDTH;
+	input_pack1.mem_contiguous = true;
+	input_pack1.mod_div_factor = 0;
+	input_pack1.sep=XI_NONE;
+
+	port_defualt2.pack_info_vec.push_back(input_pack1);
+
+	input_pack1.dim = XI_INTERLEAVED;
+	input_pack1.mem_contiguous = true;
+	input_pack1.mod_div_factor = (port_defualt2.port_width/port_defualt2.size_datatype)/batch_size;
+	input_pack1.sep=XI_MODULO;
+	port_defualt2.pack_info_vec.push_back(input_pack1);
+
+	input_pack1.dim = XI_BATCH;
+	input_pack1.mem_contiguous = true;
+	input_pack1.mod_div_factor = 0;
+	input_pack1.sep=XI_NONE;
+
+	port_defualt2.pack_info_vec.push_back(input_pack1);
+	obj_module_opcode.port_vec.push_back(port_defualt2);
+
+	Port_info port_defualt1;
+	port_defualt1.port_width=XI_128BIT_WIDTH;
+	port_defualt1.data_type = XI_INT8;
+	port_defualt1.size_datatype = XI_INT8_S;
+	port_defualt1.memory_align_required=32; // TODO need to check with buffer management & init module
+	port_defualt1.memory_map_info = __MALLOC_TYPE;
+	port_defualt1.port_enum=XI_OUTPUT;
+	port_defualt1.qant_scheme_type = g_quant_value;
+
+	packinfo output_pack0;
+
+	output_pack0.dim = XI_CHANNEL;
+	output_pack0.mem_contiguous = true;
+	output_pack0.mod_div_factor = (port_defualt1.port_width/port_defualt1.size_datatype)/batch_size;
+	output_pack0.sep=XI_DIVISION;
+
+	port_defualt1.pack_info_vec.push_back(output_pack0);
+
+	output_pack0.dim = XI_HEIGHT;
+	output_pack0.mem_contiguous = true;
+	output_pack0.mod_div_factor = 0;
+	output_pack0.sep=XI_NONE;
+
+	port_defualt1.pack_info_vec.push_back(output_pack0);
+
+	output_pack0.dim = XI_WIDTH;
+	output_pack0.mem_contiguous = true;
+	output_pack0.mod_div_factor = 0;
+	output_pack0.sep=XI_NONE;
+
+	port_defualt1.pack_info_vec.push_back(output_pack0);
+
+	output_pack0.dim = XI_INTERLEAVED;
+	output_pack0.mem_contiguous = true;
+	output_pack0.mod_div_factor = (port_defualt1.port_width/port_defualt1.size_datatype)/batch_size;
+	output_pack0.sep=XI_MODULO;
+
+	port_defualt1.pack_info_vec.push_back(output_pack0);
+
+	output_pack0.dim = XI_BATCH;
+	output_pack0.mem_contiguous = true;
+	output_pack0.mod_div_factor = 0;
+	output_pack0.sep=XI_NONE;
+
+	port_defualt1.pack_info_vec.push_back(output_pack0);
+	obj_module_opcode.port_vec.push_back(port_defualt1);
+
+
+	/****** OPCODE_SW_ELTWISE_ADD  *******/
+	obj_module_opcode.op_code = OPCODE_SW_ELTWISE_ADD;
+	opcode_info_map[obj_module_opcode.op_code]=obj_module_opcode;
+}
 void update_sf_xpack_module_opcode_table(Opcode_info & obj_module_opcode,map<opcode_num_e, Opcode_info>  & opcode_info_map,int batch_size){
 
 	Module_info module_default;
@@ -1823,12 +1982,16 @@ kernelInfo::kernelInfo()
 	update_l2normalize_module_opcode_table(opcde_l2normalize,opcode_info_map,batch_size);
 
 
+
 	Opcode_info opcde_sf_xpack;
 	update_sf_xpack_module_opcode_table(opcde_sf_xpack,opcode_info_map,batch_size);
 
 	Opcode_info permute_xunpack;
 	update_permute_xunpack_module_opcode_table(permute_xunpack,opcode_info_map,batch_size);
-
+#if XI_DIET_CHAI_Z
+	Opcode_info opcde_sw_eltwise;
+	update_opcde_sw_eltwise_module_opcode_table(opcde_sw_eltwise,opcode_info_map,batch_size);
+#endif
 	//	opcode_info_info.insert()
 	//
 	//	kernelinfo_func_keys.insert(pair<int, kernelfunction>(int(HW_CONV), &kernelInfo::kernelinfo_function));
@@ -1913,7 +2076,10 @@ kernelInfo::kernelInfo(string &quant_schem)
 
 	Opcode_info permute_xunpack;
 	update_permute_xunpack_module_opcode_table(permute_xunpack,opcode_info_map,batch_size);
-
+#if XI_DIET_CHAI_Z
+	Opcode_info opcde_sw_eltwise;
+	update_opcde_sw_eltwise_module_opcode_table(opcde_sw_eltwise,opcode_info_map,batch_size);
+#endif
 	//	opcode_info_info.insert()
 	//
 	//	kernelinfo_func_keys.insert(pair<int, kernelfunction>(int(HW_CONV), &kernelInfo::kernelinfo_function));

diff --git a/software/common/kernelinfo_class.h b/software/common/kernelinfo_class.h
@@ -94,7 +94,8 @@ enum moudle_type {
 	SW_XPACK=104,
 	SW_XUNPACK=105,
 	SW_PERMUTE=106,
-	SW_L2NORMALIZE=107
+	SW_L2NORMALIZE=107,
+	SW_ELTWISE_ADD=108
 };
 typedef enum moudle_type _moudle_type_e;
 

diff --git a/software/common/xi_kernels.cpp b/software/common/xi_kernels.cpp
@@ -158,7 +158,7 @@ void ConvolutionForward(
 			istg_out1    = istg_out1 + scalar_conv_args[112];
 #if !SINGLE_IO_PORT
 			input_other2 = input_other2 + scalar_conv_args[111];
-			istg_out2    = output2 + scalar_conv_args[112];
+			istg_out2    = istg_out2+ scalar_conv_args[112];
 #endif
 			/*
 			if(loop_iter==(pool_split_cnt-2))
@@ -173,11 +173,12 @@ void ConvolutionForward(
 				scalar_conv_args[14] = scalar_conv_args[117];
 				scalar_conv_args[15] = scalar_conv_args[118];
 			}*/
-
+/*
 			if(loop_iter==(pool_split_cnt-1))
 			{
 				scalar_conv_args[4] = scalar_conv_args[109];
 			}
+			*/
 		}
 
 #ifdef __SDSOC
@@ -506,5 +507,12 @@ void XpackForward(void *input1, void *input2, void *output1, void *output2, stri
 	XpackWrapper(input1, input2, output1, output2, quant, params, float_params);
 }
 
+void EltwiseaddWrapper(IO_DATA_TYPE *input1, IO_DATA_TYPE *input2, IO_DATA_TYPE *input3, IO_DATA_TYPE *input4, IO_DATA_TYPE *output1, IO_DATA_TYPE *output2, int *params);
+
+void EltwiseaddForward(void *input1, void *input2, void *input3, void *input4, void *output1, void *output2, int *params)
+{
+	EltwiseaddWrapper((IO_DATA_TYPE *)input1, (IO_DATA_TYPE *)input2, (IO_DATA_TYPE *)input3, (IO_DATA_TYPE *)input4, (IO_DATA_TYPE *)output1, (IO_DATA_TYPE *)output2, params);
+}
+
 
 #endif//_XI_KERNELS_HPP_
diff --git a/software/common/xi_kernels.h b/software/common/xi_kernels.h
@@ -110,5 +110,6 @@ void CropForward(int *input, int *output, int *scalars);
 
 void XpackForward(void *input1, void *input2, void *output1, void *output2, string quant, int *params, float *float_params);
 
+void EltwiseaddForward(void *input1, void *input2, void *input3, void *input4, void *output1, void *output2, int *params);
 
 #endif//_XI_KERNELS_H_