#24, added support for multiple DDR banks in host and krnls: cannot build due...

#24, added support for multiple DDR banks in host and krnls: cannot build due to insufficient resources

ofdock_taskpar_xl/Makefile

@@ -96,10 +96,13 @@ Krnl_GA_SRCS=./device/Krnl_GA.cl
 ## Choosing the all kernels in design for maximum memory port property 
 # Krnl_GA_CLFLAGS=--max_memory_ports all
 
-Krnl_GA_CLFLAGS=--max_memory_ports Krnl_GA \
-		--max_memory_ports Krnl_Conform \
-		--max_memory_ports Krnl_InterE \
-		--max_memory_ports Krnl_IntraE
+## Ideal configuration, but results in routing error
+#Krnl_GA_CLFLAGS=--max_memory_ports Krnl_GA \
+#		--max_memory_ports Krnl_Conform \
+#		--max_memory_ports Krnl_InterE \
+#		--max_memory_ports Krnl_IntraE
+
+#Krnl_GA_CLFLAGS=--max_memory_ports Krnl_GA
 
 # https://www.xilinx.com/support/documentation/sw_manuals/xilinx2017_4/ug1023-sdaccel-user-guide.pdf
 # Page 44
@@ -113,12 +116,37 @@ Krnl_GA_CLFLAGS=--max_memory_ports Krnl_GA \
 #		    --sp bandwidth_1.m_axi_gmem1:bank1 \
 #		    --sp bandwidth_1.m_axi_gmem2:bank2	
 
-Krnl_GA_LDCLFLAGS+=--sp Krnl_GA_1.m_axi_gmem0:bank0 \
-		   --sp Krnl_Conform_1.m_axi_gmem1:bank1 \
-		   --sp Krnl_InterE_1.m_axi_gmem2:bank2 \
-		   --sp Krnl_IntraE_1.m_axi_gmem3:bank3	
-
-
+## Ideal configuration, but results in routing error
+# First two args of IntraE are the 2nd and 3rd of InterE (atom_charges & atom_types),
+# Therefore they are mapped to the same DDR bank
+#Krnl_GA_LDCLFLAGS+=--sp Krnl_GA_1.m_axi_gmem0:bank0 \
+#		   --sp Krnl_GA_1.m_axi_gmem1:bank0 \
+#		   --sp Krnl_GA_1.m_axi_gmem2:bank0 \
+#		   --sp Krnl_GA_1.m_axi_gmem3:bank0 \
+#		   --sp Krnl_Conform_1.m_axi_gmem0:bank1 \
+#		   --sp Krnl_Conform_1.m_axi_gmem1:bank1 \
+#		   --sp Krnl_Conform_1.m_axi_gmem2:bank1 \
+#		   --sp Krnl_Conform_1.m_axi_gmem3:bank1 \
+#		   --sp Krnl_Conform_1.m_axi_gmem4:bank1 \
+#		   --sp Krnl_InterE_1.m_axi_gmem0:bank2 \
+#		   --sp Krnl_InterE_1.m_axi_gmem1:bank2 \
+#		   --sp Krnl_InterE_1.m_axi_gmem2:bank2 \
+#		   --sp Krnl_IntraE_1.m_axi_gmem0:bank2	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem1:bank2	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem2:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem3:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem4:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem5:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem6:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem7:bank3	\
+# 		   --sp Krnl_IntraE_1.m_axi_gmem8:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem9:bank3	\
+#		   --sp Krnl_IntraE_1.m_axi_gmem10:bank3
+
+Krnl_GA_LDCLFLAGS+=--sp Krnl_GA_1.m_axi_gmem:bank0      \
+		   --sp Krnl_Conform_1.m_axi_gmem:bank1 \
+		   --sp Krnl_InterE_1.m_axi_gmem:bank2  \
+		   --sp Krnl_IntraE_1.m_axi_gmem:bank3
 ##
 # https://www.xilinx.com/support/documentation/sw_manuals/xilinx2017_4/ug1281-sdaccel-debugging-guide.pdf
 # Page 32
@@ -139,9 +167,28 @@ ifneq ($(LAPC),none)
 ## # xbsak status --lapc
 ## Invalid codes read, skip decoding
 Krnl_GA_LDCLFLAGS+=--dk protocol:Krnl_GA_1:m_axi_gmem0 \
+		   --dk protocol:Krnl_GA_1:m_axi_gmem1 \
+		   --dk protocol:Krnl_GA_1:m_axi_gmem2 \
+		   --dk protocol:Krnl_GA_1:m_axi_gmem3 \
+		   --dk protocol:Krnl_Conform_1:m_axi_gmem0 \
 		   --dk protocol:Krnl_Conform_1:m_axi_gmem1 \
-                   --dk protocol:Krnl_InterE_1:m_axi_gmem2 \
-	           --dk protocol:Krnl_IntraE_1:m_axi_gmem3
+		   --dk protocol:Krnl_Conform_1:m_axi_gmem2 \
+		   --dk protocol:Krnl_Conform_1:m_axi_gmem3 \
+		   --dk protocol:Krnl_Conform_1:m_axi_gmem4 \
+                   --dk protocol:Krnl_InterE_1:m_axi_gmem0 \
+	 	   --dk protocol:Krnl_InterE_1:m_axi_gmem1 \
+		   --dk protocol:Krnl_InterE_1:m_axi_gmem2 \
+	           --dk protocol:Krnl_IntraE_1:m_axi_gmem0 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem1 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem2 \
+	           --dk protocol:Krnl_IntraE_1:m_axi_gmem3 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem4 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem5 \
+	           --dk protocol:Krnl_IntraE_1:m_axi_gmem6 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem7 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem8 \
+	           --dk protocol:Krnl_IntraE_1:m_axi_gmem9 \
+		   --dk protocol:Krnl_IntraE_1:m_axi_gmem10
 
 ## Commented because error was obtained:
 ## No interface pins matched 'get_bd_intf_pins -of /Krnl_Conform_1 -filter NAME=~m_axi_gmem'
@@ -160,7 +207,10 @@ SPM:=none
 ifneq ($(SPM),none)
 ##--profile_kernel data:krnl1:cu1:m_axi_gmem0
 ##--profile_kernel data:krnl2:cu2:m_axi_gmem
-Krnl_GA_LDCLFLAGS+=--profile_kernel Krnl_GA:Krnl_GA_1:all
+Krnl_GA_LDCLFLAGS+=--profile_kernel data:Krnl_GA:Krnl_GA_1:all \
+		   --profile_kernel data:Krnl_Conform:Krnl_Conform_1:all \
+		   --profile_kernel data:Krnl_InterE:Krnl_InterE_1:all \
+		   --profile_kernel data:Krnl_IntraE:Krnl_IntraE_1:all
 endif
 
 #-------------------------------

ofdock_taskpar_xl/host/src/performdocking.cpp

@@ -450,8 +450,12 @@ float* cpu_ref_ori_angles;
 #if defined (FIXED_POINT_INTERE)
 cl_mem mem_KerConstStatic_fixpt64_atom_charges_const;
 #endif
-cl_mem mem_KerConstStatic_atom_charges_const;
-cl_mem mem_KerConstStatic_atom_types_const;
+cl_mem mem_KerConstStatic_InterE_atom_charges_const;
+cl_mem mem_KerConstStatic_InterE_atom_types_const;
+
+cl_mem mem_KerConstStatic_IntraE_atom_charges_const;
+cl_mem mem_KerConstStatic_IntraE_atom_types_const;
+
 cl_mem mem_KerConstStatic_intraE_contributors_const;
 
 cl_mem mem_KerConstStatic_reqm_const;
@@ -541,8 +545,6 @@ cl_mem mem_dummy;
 
 
 
-
-
 
 
 
@@ -767,8 +769,38 @@ printf("%i %i\n", dockpars.num_of_intraE_contributors, myligand_reference.num_of
 	#else
 	size_floatgrids = (sizeof(float)) * (mygrid->num_of_atypes+2) * (mygrid->size_xyz[0]) * (mygrid->size_xyz[1]) * (mygrid->size_xyz[2]);
 	#endif
-#endif
-
+#endif
+
+
+
+// -----------------------------------------------------------------------------------------------------
+// Hardware specific
+// Specifiying exact memory bank from host code
+// Only valid if 4 banks are available (AWS)
+	cl_mem_ext_ptr_t d_bank0_ext; // Krnl_GA
+	cl_mem_ext_ptr_t d_bank1_ext; // Krnl_Conform
+	cl_mem_ext_ptr_t d_bank2_ext; // Krnl_InterE
+	cl_mem_ext_ptr_t d_bank3_ext; // Krnl_IntraE
+
+	d_bank0_ext.flags = XCL_MEM_DDR_BANK0;
+	d_bank0_ext.obj   = NULL;
+	d_bank0_ext.param = 0;
+
+	d_bank1_ext.flags = XCL_MEM_DDR_BANK1;
+	d_bank1_ext.obj   = NULL;
+	d_bank1_ext.param = 0;
+
+	d_bank2_ext.flags = XCL_MEM_DDR_BANK2;
+	d_bank2_ext.obj   = NULL;
+	d_bank2_ext.param = 0;
+
+	d_bank3_ext.flags = XCL_MEM_DDR_BANK3;
+	d_bank3_ext.obj   = NULL;
+	d_bank3_ext.param = 0;
+
+// Replacing common buffer creation with 
+// a Xilinx-specific where DDR banks can be specified
+#if 0
 #if defined (FIXED_POINT_INTERE)
 	mallocBufferObject(context,CL_MEM_READ_ONLY, MAX_NUM_OF_ATOMS*sizeof(fixedpt64),                &mem_KerConstStatic_fixpt64_atom_charges_const);
 #endif
@@ -826,7 +858,88 @@ printf("%i %i\n", dockpars.num_of_intraE_contributors, myligand_reference.num_of
 	mallocBufferObject(context,CL_MEM_WRITE_ONLY,size_evals_of_runs,  	&mem_gens_performed);
 #else
 	mallocBufferObject(context,CL_MEM_WRITE_ONLY,2*sizeof(unsigned int),  	&mem_evals_and_generations_performed);
+#endif
+#endif
+
+
+
+#if defined (FIXED_POINT_INTERE)
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(fixedpt64),                &d_bank2_ext, &mem_KerConstStatic_fixpt64_atom_charges_const);	
 #endif
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(float),                    &d_bank2_ext, &mem_KerConstStatic_InterE_atom_charges_const);	// InterE
+  	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(char),                     &d_bank2_ext, &mem_KerConstStatic_InterE_atom_types_const);	// InterE
+
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(float),                    &d_bank3_ext, &mem_KerConstStatic_IntraE_atom_charges_const);	// IntraE
+  	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(char),                     &d_bank3_ext, &mem_KerConstStatic_IntraE_atom_types_const);	// IntraE
+
+/*
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, 3*MAX_INTRAE_CONTRIBUTORS*sizeof(char),            &d_bank3_ext, &mem_KerConstStatic_intraE_contributors_const);	// IntraE
+*/
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_INTRAE_CONTRIBUTORS*sizeof(cl_char3),          &d_bank3_ext, &mem_KerConstStatic_intraE_contributors_const);	// IntraE
+
+        mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, ATYPE_NUM*sizeof(float),			        &d_bank3_ext, &mem_KerConstStatic_reqm_const);			// IntraE
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, ATYPE_NUM*sizeof(float),			        &d_bank3_ext, &mem_KerConstStatic_reqm_hbond_const);		// IntraE
+  	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, ATYPE_NUM*sizeof(unsigned int),		        &d_bank3_ext, &mem_KerConstStatic_atom1_types_reqm_const);	// IntraE
+  	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, ATYPE_NUM*sizeof(unsigned int),                    &d_bank3_ext, &mem_KerConstStatic_atom2_types_reqm_const);	// IntraE
+
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATYPES*MAX_NUM_OF_ATYPES*sizeof(float), &d_bank3_ext, &mem_KerConstStatic_VWpars_AC_const);		// IntraE
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATYPES*MAX_NUM_OF_ATYPES*sizeof(float), &d_bank3_ext, &mem_KerConstStatic_VWpars_BD_const);		// IntraE
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATYPES*sizeof(float), 		        &d_bank3_ext, &mem_KerConstStatic_dspars_S_const);		// IntraE
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATYPES*sizeof(float), 			&d_bank3_ext, &mem_KerConstStatic_dspars_V_const);		// IntraE
+	#if defined (FIXED_POINT_CONFORM)
+	// fixed-point
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTATIONS*sizeof(fixedpt), 		&d_bank1_ext, &mem_KerConstStatic_rotlist_const);			// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(cl_int3), 		        &d_bank1_ext, &mem_KerConstStatic_ref_coords_const);			// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTBONDS*sizeof(cl_int3), 		&d_bank1_ext, &mem_KerConstStatic_rotbonds_moving_vectors_const);	// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTBONDS*sizeof(cl_int3), 		&d_bank1_ext, &mem_KerConstStatic_rotbonds_unit_vectors_const);		// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_RUNS*sizeof(cl_int4),			&d_bank1_ext, &mem_KerConstStatic_ref_orientation_quats_const);		// Conform
+	#else
+	// floating-point (original)
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTATIONS*sizeof(int), 			&d_bank1_ext, &mem_KerConstStatic_rotlist_const);			// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ATOMS*sizeof(cl_float3), 		&d_bank1_ext, &mem_KerConstStatic_ref_coords_const); 			// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTBONDS*sizeof(cl_float3), 		&d_bank1_ext, &mem_KerConstStatic_rotbonds_moving_vectors_const);	// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_ROTBONDS*sizeof(cl_float3), 		&d_bank1_ext, &mem_KerConstStatic_rotbonds_unit_vectors_const);		// Conform
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX, MAX_NUM_OF_RUNS*sizeof(cl_float4),			&d_bank1_ext, &mem_KerConstStatic_ref_orientation_quats_const);		// Conform
+	#endif
+
+//#if defined (FIXED_POINT_INTERE)
+#if 0
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX,size_fixedpt64grids,   	&d_bank2_ext,	&mem_dockpars_fgrids);	// InterE
+#else
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX,size_floatgrids,   		&d_bank2_ext,	&mem_dockpars_fgrids);	// InterE
+
+	#if defined(SEPARATE_FGRID_INTERE)
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX,size_floatgrids2,   	&d_bank2_ext,	&mem_dockpars_fgrids2);	// InterE
+	mallocBufferObject(context,CL_MEM_READ_ONLY | CL_MEM_EXT_PTR_XILINX,size_floatgrids3,   	&d_bank2_ext,	&mem_dockpars_fgrids3);	// InterE
+	#endif
+#endif
+
+	mallocBufferObject(context,CL_MEM_READ_WRITE | CL_MEM_EXT_PTR_XILINX,size_populations,  	&d_bank0_ext,	&mem_dockpars_conformations_current);	// GA
+	mallocBufferObject(context,CL_MEM_READ_WRITE | CL_MEM_EXT_PTR_XILINX,size_energies,    		&d_bank0_ext,	&mem_dockpars_energies_current);	// GA
+
+#if defined(SINGLE_COPY_POP_ENE)
+	mallocBufferObject(context,CL_MEM_WRITE_ONLY | CL_MEM_EXT_PTR_XILINX,size_evals_of_runs,  	&d_bank0_ext,	&mem_evals_performed);			// GA
					
+	mallocBufferObject(context,CL_MEM_WRITE_ONLY | CL_MEM_EXT_PTR_XILINX,size_evals_of_runs,  	&d_bank0_ext,	&mem_gens_performed);			// GA
+#else
+	mallocBufferObject(context,CL_MEM_WRITE_ONLY | CL_MEM_EXT_PTR_XILINX,2*sizeof(unsigned int),  	&d_bank0_ext,	&mem_evals_and_generations_performed);	// GA
+#endif
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+// -----------------------------------------------------------------------------------------------------
+
 
 #if defined(SINGLE_COPY_POP_ENE)
 
@@ -840,9 +953,11 @@ printf("%i %i\n", dockpars.num_of_intraE_contributors, myligand_reference.num_of
 	#endif
 	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_fixpt64_atom_charges_const,            &KerConstStatic.fixpt64_atom_charges_const[0],            MAX_NUM_OF_ATOMS*sizeof(fixedpt64));
 #endif
-	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_atom_charges_const,            &KerConstStatic.atom_charges_const[0],            MAX_NUM_OF_ATOMS*sizeof(float));
-
-	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_atom_types_const,              &KerConstStatic.atom_types_const[0],              MAX_NUM_OF_ATOMS*sizeof(char));
+	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_InterE_atom_charges_const,            &KerConstStatic.atom_charges_const[0],            MAX_NUM_OF_ATOMS*sizeof(float));
+	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_InterE_atom_types_const,              &KerConstStatic.atom_types_const[0],              MAX_NUM_OF_ATOMS*sizeof(char));
+
+	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_IntraE_atom_charges_const,            &KerConstStatic.atom_charges_const[0],            MAX_NUM_OF_ATOMS*sizeof(float));
+	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_IntraE_atom_types_const,              &KerConstStatic.atom_types_const[0],              MAX_NUM_OF_ATOMS*sizeof(char));
 
 /*
 	memcopyBufferObjectToDevice(command_queue1,mem_KerConstStatic_intraE_contributors_const,     &KerConstStatic.intraE_contributors_const[0],     3*MAX_INTRAE_CONTRIBUTORS*sizeof(char));
@@ -1016,9 +1131,9 @@ printf("%i %i\n", dockpars.num_of_intraE_contributors, myligand_reference.num_of
 	#if defined (FIXED_POINT_INTERE)
 	setKernelArg(kernel3,1,  sizeof(mem_KerConstStatic_fixpt64_atom_charges_const),  &mem_KerConstStatic_fixpt64_atom_charges_const);
 	#else
-	setKernelArg(kernel3,1,  sizeof(mem_KerConstStatic_atom_charges_const),  &mem_KerConstStatic_atom_charges_const);
+	setKernelArg(kernel3,1,  sizeof(mem_KerConstStatic_InterE_atom_charges_const),  &mem_KerConstStatic_InterE_atom_charges_const);
 	#endif
-	setKernelArg(kernel3,2,  sizeof(mem_KerConstStatic_atom_types_const),    &mem_KerConstStatic_atom_types_const);
+	setKernelArg(kernel3,2,  sizeof(mem_KerConstStatic_InterE_atom_types_const),    &mem_KerConstStatic_InterE_atom_types_const);
 	setKernelArg(kernel3,3,  sizeof(unsigned char),                          &dockpars.g1);
 	setKernelArg(kernel3,4,  sizeof(unsigned int),                           &dockpars.g2);
 	setKernelArg(kernel3,5,  sizeof(unsigned int),                           &dockpars.g3);
@@ -1044,8 +1159,8 @@ printf("%i %i\n", dockpars.num_of_intraE_contributors, myligand_reference.num_of
 #endif // End of ENABLE_KERNEL3
 
 #ifdef ENABLE_KERNEL4 // Krnl_IntraE
-	setKernelArg(kernel4,0,  sizeof(mem_KerConstStatic_atom_charges_const),        &mem_KerConstStatic_atom_charges_const);
-	setKernelArg(kernel4,1,  sizeof(mem_KerConstStatic_atom_types_const),          &mem_KerConstStatic_atom_types_const);
+	setKernelArg(kernel4,0,  sizeof(mem_KerConstStatic_IntraE_atom_charges_const),        &mem_KerConstStatic_IntraE_atom_charges_const);
+	setKernelArg(kernel4,1,  sizeof(mem_KerConstStatic_IntraE_atom_types_const),          &mem_KerConstStatic_IntraE_atom_types_const);
 	setKernelArg(kernel4,2,  sizeof(mem_KerConstStatic_intraE_contributors_const), &mem_KerConstStatic_intraE_contributors_const);
 
   	setKernelArg(kernel4,3,  sizeof(dockpars.smooth),                              &dockpars.smooth);
@@ -3123,8 +3238,12 @@ void cleanup() {
 #if defined (FIXED_POINT_INTERE)
   if(mem_KerConstStatic_fixpt64_atom_charges_const)	  {clReleaseMemObject(mem_KerConstStatic_fixpt64_atom_charges_const);}
 #endif
-  if(mem_KerConstStatic_atom_charges_const)	  	  {clReleaseMemObject(mem_KerConstStatic_atom_charges_const);}
-  if(mem_KerConstStatic_atom_types_const)	   	  {clReleaseMemObject(mem_KerConstStatic_atom_types_const);}
+  if(mem_KerConstStatic_InterE_atom_charges_const)	  	  {clReleaseMemObject(mem_KerConstStatic_InterE_atom_charges_const);}
+  if(mem_KerConstStatic_InterE_atom_types_const)	   	  {clReleaseMemObject(mem_KerConstStatic_InterE_atom_types_const);}
+
+  if(mem_KerConstStatic_IntraE_atom_charges_const)	  	  {clReleaseMemObject(mem_KerConstStatic_IntraE_atom_charges_const);}
+  if(mem_KerConstStatic_IntraE_atom_types_const)	   	  {clReleaseMemObject(mem_KerConstStatic_IntraE_atom_types_const);}
+
   if(mem_KerConstStatic_intraE_contributors_const) 	  {clReleaseMemObject(mem_KerConstStatic_intraE_contributors_const);}
 
   if(mem_KerConstStatic_reqm_const) 	  		  {clReleaseMemObject(mem_KerConstStatic_reqm_const);}

ofdock_taskpar_xl/wrapcl/inc/BufferObjects.h

@@ -15,10 +15,19 @@
 /*
 
 */
+  #if 0
   int mallocBufferObject(cl_context   context,
 			 cl_mem_flags flags,
 			 size_t       size,
 			 cl_mem*      mem);
+  #endif
+
+  // Modified to support explicit host point
+  int mallocBufferObject(cl_context   context,
+			 cl_mem_flags flags,
+			 size_t       size,
+			 void*        host_pointer,
+			 cl_mem*      mem);
 
 /*
 

ofdock_taskpar_xl/wrapcl/src/BufferObjects.cpp

 #include "BufferObjects.h"
 
+#if 0
 int mallocBufferObject(cl_context   context,
                        cl_mem_flags flags,
                        size_t       size,
@@ -26,6 +27,36 @@ int mallocBufferObject(cl_context   context,
   *mem = local_mem;
   return CL_SUCCESS;
 }
+#endif
+
+int mallocBufferObject(cl_context   context,
+		       cl_mem_flags flags,
+		       size_t       size,
+		       void*        host_pointer,
+		       cl_mem*      mem){
+  cl_mem local_mem;
+
+  local_mem = clCreateBuffer(context, flags, size, host_pointer, NULL);
+  if (!local_mem){
+	printf("Error: clCreateBuffer()\n");
+	fflush(stdout);
+	return EXIT_FAILURE;
+  }
+
+#ifdef BUFFER_OBJECT_INFO_DISPLAY
+  cl_int err;
+  err = getBufferObjectInfo(local_mem);
+  if (err != CL_SUCCESS){
+	printf("Error: getBufferObjectInfo() %d\n",err);
+	fflush(stdout);
+	return EXIT_FAILURE;
+  }
+#endif
+
+  *mem = local_mem;
+  return CL_SUCCESS;
+}
+
 
 #ifdef BUFFER_OBJECT_INFO_DISPLAY
 int getBufferObjectInfo(cl_mem object){