Increased maximum number of runs to 8192

Makefile.OpenCL

@@ -16,6 +16,7 @@ LIB_OPENCL = -lOpenCL
 UNAME := $(shell uname)
 
 ifeq ($(UNAME), Darwin)
+CPP = clang++
 # In case ScoreP (for profiling/tracing) is used,
 # need to link to a *.dylib for instrumentation
 ifneq (,$(findstring scorep,$(CPP)))

common/defines.h

@@ -87,7 +87,7 @@ enum {C=0,N=1,O=2,H=3,XX=4,P=5,S=6};  // see "bond_index" in the "AD4.1_bound.da
 #define MAX_INTRAE_CONTRIBUTORS    (MAX_NUM_OF_ATOMS * MAX_NUM_OF_ATOMS)
 #define MAX_NUM_OF_ROTATIONS       (MAX_NUM_OF_ATOMS * MAX_NUM_OF_ROTBONDS)
 #define MAX_POPSIZE                2048
-#define MAX_NUM_OF_RUNS            1000
+#define MAX_NUM_OF_RUNS            8192
 #define MAX_NUM_GRIDPOINTS         256
 
 // Must be larger than or equal to MAX_NUM_OF_ROTBONDS+6

host/inc/autostop.hpp

@@ -162,7 +162,7 @@ class AutoStop{
 	inline void print_intro(unsigned long num_of_generations, unsigned long num_of_energy_evals)
 	{
 		para_printf("\nExecuting docking runs, stopping automatically after either reaching %.2f kcal/mol standard deviation of\nthe best molecules of the last 4 * %u generations, %lu generations, or %lu evaluations:\n\n",stopstd,as_frequency,num_of_generations,num_of_energy_evals);
-		para_printf("Generations |  Evaluations |     Threshold    |  Average energy of best 10%%  | Samples |    Best energy\n");
+		para_printf("Generations |  Evaluations |     Threshold    |  Average energy of best 10%%  | Samples | Best Inter + Intra\n");
 		para_printf("------------+--------------+------------------+------------------------------+---------+-------------------\n");
 	}
 
@@ -221,7 +221,7 @@ class AutoStop{
 	inline void output_final_stddev(int generation_cnt, const float* energies, unsigned long total_evals){
 		if (autostopped){
 			para_printf("------------+--------------+------------------+------------------------------+---------+-------------------\n");
-			para_printf("\n%43s evaluation after reaching\n%40.2f +/-%8.2f kcal/mol combined.\n%34i samples, best energy %8.2f kcal/mol.\n","Finished",average(&average_sd2_N[0]),stddev(&average_sd2_N[0]),(unsigned int)average_sd2_N[2],overall_best_energy);
+			para_printf("\n%43s evaluation after reaching\n%40.2f +/-%8.2f kcal/mol combined.\n%31i samples, best inter + intra %8.2f kcal/mol.\n","Finished",average(&average_sd2_N[0]),stddev(&average_sd2_N[0]),(unsigned int)average_sd2_N[2],overall_best_energy);
 		} else {
 			// Stopped without autostop; output stddev statistics regardless
 
@@ -230,7 +230,7 @@ class AutoStop{
 
 			para_printf("%11u | %12lu |%8.2f kcal/mol |%8.2f +/-%8.2f kcal/mol |%8i |%8.2f kcal/mol\n",generation_cnt,total_evals/num_of_runs,threshold,curr_avg,curr_std,bestN,overall_best_energy);
 			para_printf("------------+--------------+------------------+------------------------------+---------+-------------------\n");
-			para_printf("\n%43s evaluation after reaching\n%33lu evaluations. Best energy %8.2f kcal/mol.\n","Finished",total_evals/num_of_runs,overall_best_energy);
+			para_printf("\n%43s evaluation after reaching\n%30lu evaluations. Best inter + intra %8.2f kcal/mol.\n","Finished",total_evals/num_of_runs,overall_best_energy);
 		}
 		fflush(stdout);
 	}

host/inc/getparameters.h

@@ -213,7 +213,8 @@ int get_commandpars(
                           char**,
                           double*,
                           Dockpars*,
-                    const bool late_call = true
+                    const bool late_call = true,
+                    const int batch_nr = -1
                    );
 
 std::vector<ReceptorAtom> read_receptor_atoms(

host/inc/miscellaneous.h

@@ -98,8 +98,12 @@ void vec_point2line(const double [], const double [], const double [], double []
 
 void rotate(double [], const double [], const double [], const double*, int);
 
+bool is_dirname(const char* filename);
+
 bool has_absolute_path(const char* filename);
 
+std::string get_base_filename(const char* filename);
+
 std::string get_filepath(const char* filename);
 
 // trim input string range -- careful: no range checking

host/src/getparameters.cpp

@@ -28,6 +28,7 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 #include <algorithm>
 #include <cctype>
 #include <locale>
+#include <sys/stat.h>
 
 #include "getparameters.h"
 
@@ -1333,7 +1334,8 @@ int get_commandpars(
                           char**    argv,
                           double*   spacing,
                           Dockpars* mypars,
-                    const bool      late_call
+                    const bool      late_call,
+                    const int       batch_nr
                    )
 // The function processes the command line arguments given with the argc and argv parameters,
 // and fills the proper fields of mypars according to that. If a parameter was not defined
@@ -1353,22 +1355,22 @@ int get_commandpars(
 		char* basefile = mypars->ligandfile;
 		if(!mypars->free_roaming_ligand) basefile = mypars->flexresfile;
 		if(mypars->xrayligandfile==NULL){
-			mypars->xrayligandfile      = strdup(basefile); // By default xray-ligand file is the same as the randomized input ligand
+			mypars->xrayligandfile = strdup(basefile); // By default xray-ligand file is the same as the randomized input ligand
 		}
 		if(mypars->xml2dlg){
 			if(strlen(mypars->load_xml)>4){ // .xml = 4 chars
 				i=strlen(mypars->load_xml)-4;
 				mypars->resname = (char*)malloc((i+1)*sizeof(char));
 				strncpy(mypars->resname,mypars->load_xml,i);    // Default is ligand file basename
-				mypars->resname[i]='\0';
+				mypars->resname[i] = '\0';
 			} else if(!mypars->resname) mypars->resname = strdup("docking"); // Fallback to old default
 		} else{
 			if(!mypars->resname){ // only need to set if it's not set yet
 				if(strlen(basefile)>6){ // .pdbqt = 6 chars
 					i=strlen(basefile)-6;
 					mypars->resname = (char*)malloc((i+1)*sizeof(char));
 					strncpy(mypars->resname,basefile,i);    // Default is ligand file basename
-					mypars->resname[i]='\0';
+					mypars->resname[i] = '\0';
 				} else mypars->resname = strdup("docking");               // Fallback to old default
 			}
 		}
@@ -1986,7 +1988,28 @@ int get_commandpars(
 		{
 			arg_recognized = 1;
 			free(mypars->resname); // as we assign a default value dynamically created to it
-			mypars->resname = strdup(argv [i+1]);
+			if(is_dirname(argv [i+1])){
+				struct stat res_stat;
+				int res_int = stat(argv [i+1], &res_stat);
+				if ((res_int != 0) || !(res_stat.st_mode & S_IFDIR)){
+					printf("\nError: Specified directory \"%s\" for output files with `--resnam` does not exist.\n",argv [i+1]);
+					exit(12);
+				}
+				char* basefile = mypars->ligandfile;
+				if(!mypars->free_roaming_ligand) basefile = mypars->flexresfile;
+				if(mypars->xml2dlg) basefile = mypars->load_xml;
+				std::string dir_res = get_filepath(argv [i+1]) + "/" + get_base_filename(basefile);
+				mypars->resname = strdup(dir_res.c_str());
+			} else{
+				if(batch_nr>0){
+					char* nrtmp = strdup(std::to_string(batch_nr).c_str());
+					mypars->resname = (char*)malloc((strlen(argv [i+1])+strlen(nrtmp)+2)*sizeof(char));
+					strcpy(mypars->resname, argv [i+1]);
+					strcat(mypars->resname,"_");
+					strcat(mypars->resname, nrtmp);
+					free(nrtmp);
+				} else mypars->resname = strdup(argv [i+1]);
+			}
 		}
 
 		// Argument: use modified QASP (from VirtualDrug) instead of original one used by AutoDock

host/src/main.cpp

@@ -236,7 +236,7 @@ int main(int argc, char* argv[])
 	}
 #endif
 	// Set up run profiles for timing
-	bool get_profiles = true; // hard-coded switch to use ALS's job profiler
+	bool get_profiles = ((filelist.filename!=NULL) || (initial_pars.dpffile!=NULL)); // hard-coded switch to use ALS's job profiler
 	Profiler profiler;
 	for (int i=0;i<n_files;i++){
 		profiler.p.push_back(Profile(i));
@@ -410,6 +410,16 @@ int main(int argc, char* argv[])
 						para_printf("    Ligand file: %s\n", mypars.ligandfile); fflush(stdout);
 					if(mypars.flexresfile)
 						para_printf("    Flexible residue: %s\n", mypars.flexresfile);
+					if(mypars.output_dlg || mypars.output_xml){
+						para_printf("    Output file: %s", mypars.resname);
+						if(mypars.output_dlg){
+							para_printf(".dlg");
+							if(mypars.output_xml) para_printf(" (+ xml)");
+						} else{
+							if(mypars.output_xml) para_printf(".xml");
+						}
+						para_printf("\n");
+					}
 					fflush(stdout);
 				} else para_printf("\n");
 				// End idling timer, start exec timer

host/src/miscellaneous.cpp

@@ -23,6 +23,8 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 */
 
 
+#include <sys/stat.h>
+
 #include "miscellaneous.h"
 
 float map2float(const char* c)
@@ -257,6 +259,15 @@ void rotate(double point [], const double movvec [], const double normvec [], co
 		        point [0], point [1], point [2]);
 }
 
+bool is_dirname(const char* filename)
+{
+	if((filename[strlen(filename)-1] == '/') ||
+	   (filename[strlen(filename)-1] == '\\')) return true;
+	struct stat res_stat;
+	int res_int = stat(filename, &res_stat);
+	return (res_stat.st_mode & S_IFDIR);
+}
+
 bool has_absolute_path(const char* filename)
 {
 	#ifndef _WIN32
@@ -269,19 +280,24 @@ bool has_absolute_path(const char* filename)
 	#endif
 }
 
+std::string get_base_filename(const char* filename)
+{
+	std::string result = filename;
+	std::size_t path   = result.find_last_of("/\\");
+	if(path == std::string::npos) path = 0; else path++;
+	result = result.substr(path);
+	return result.substr(0, result.find_last_of("."));
+}
+
 std::string get_filepath(const char* filename)
 {
-	#ifndef _WIN32
-	char* ts1 = strdup(filename);
-	std::string result = dirname(ts1);
-	free(ts1);
-	return result;
-	#else
-	char drive_tmp[_MAX_DRIVE];
-	char path_tmp[_MAX_DIR];
-	_splitpath(filename, drive_tmp, path_tmp, NULL, NULL);
-	return drive_tmp + path_tmp;
-	#endif
+	std::string result = filename;
+	std::size_t path   = result.find_last_of("/\\");
+	if(path == std::string::npos){
+		if(is_dirname(filename)) return filename;
+		return ".";
+	}
+	return result.substr(0, path);
 }
 
 #if 0

host/src/performdocking.cpp.Cuda

@@ -185,29 +185,6 @@ void setup_gpu_for_docking(
 		status = cudaMalloc((void**)&(tData.pMem_fgrids), cData.preallocated_gridsize*sizeof(float));
 		RTERROR(status, "pMem_fgrids: failed to allocate GPU memory.\n");
 	}
-	size_t size_populations = MAX_NUM_OF_RUNS * MAX_POPSIZE * GENOTYPE_LENGTH_IN_GLOBMEM*sizeof(float);
-	status = cudaMalloc((void**)&(tData.pMem_conformations1), size_populations);
-	RTERROR(status, "pMem_conformations1: failed to allocate GPU memory.\n");
-	status = cudaMalloc((void**)&(tData.pMem_conformations2), size_populations);
-	RTERROR(status, "pMem_conformations2: failed to allocate GPU memory.\n");
-	size_t size_energies = MAX_POPSIZE * MAX_NUM_OF_RUNS * sizeof(float);
-	status = cudaMalloc((void**)&(tData.pMem_energies1), size_energies);
-	RTERROR(status, "pMem_energies1: failed to allocate GPU memory.\n");
-	status = cudaMalloc((void**)&(tData.pMem_energies2), size_energies);
-	RTERROR(status, "pMem_energies2: failed to allocate GPU memory.\n");  
-	status = cudaMalloc((void**)&(tData.pMem_evals_of_new_entities), MAX_POPSIZE*MAX_NUM_OF_RUNS*sizeof(int));
-	RTERROR(status, "pMem_evals_of_new_Entities: failed to allocate GPU memory.\n");
-	size_t size_evals_of_runs = MAX_NUM_OF_RUNS*sizeof(int);
-#if defined (MAPPED_COPY)
-	status = cudaMallocManaged((void**)&(tData.pMem_gpu_evals_of_runs), size_evals_of_runs, cudaMemAttachGlobal);
-#else
-	status = cudaMalloc((void**)&(tData.pMem_gpu_evals_of_runs), size_evals_of_runs);
-#endif
-	RTERROR(status, "pMem_gpu_evals_of_runs: failed to allocate GPU memory.\n");
-	size_t blocksPerGridForEachEntity = MAX_POPSIZE * MAX_NUM_OF_RUNS;
-	size_t size_prng_seeds = blocksPerGridForEachEntity * NUM_OF_THREADS_PER_BLOCK * sizeof(unsigned int);
-	status = cudaMalloc((void**)&(tData.pMem_prng_states), size_prng_seeds);
-	RTERROR(status, "pMem_prng_states: failed to allocate GPU memory.\n");
 }
 void finish_gpu_from_docking(
                              GpuData& cData,
@@ -241,20 +218,6 @@ void finish_gpu_from_docking(
 		status = cudaFree(tData.pMem_fgrids);
 		RTERROR(status, "cudaFree: error freeing pMem_fgrids");
 	}
-	status = cudaFree(tData.pMem_conformations1);
-	RTERROR(status, "cudaFree: error freeing pMem_conformations1");
-	status = cudaFree(tData.pMem_conformations2);
-	RTERROR(status, "cudaFree: error freeing pMem_conformations2");
-	status = cudaFree(tData.pMem_energies1);
-	RTERROR(status, "cudaFree: error freeing pMem_energies1");
-	status = cudaFree(tData.pMem_energies2);
-	RTERROR(status, "cudaFree: error freeing pMem_energies2");
-	status = cudaFree(tData.pMem_evals_of_new_entities);
-	RTERROR(status, "cudaFree: error freeing pMem_evals_of_new_entities");
-	status = cudaFree(tData.pMem_gpu_evals_of_runs);
-	RTERROR(status, "cudaFree: error freeing pMem_gpu_evals_of_runs");
-	status = cudaFree(tData.pMem_prng_states);
-	RTERROR(status, "cudaFree: error freeing pMem_prng_states");
 	free(tData.device_name);
 }
 
@@ -357,6 +320,26 @@ parameters argc and argv:
 	size_evals_of_runs = mypars->num_of_runs*sizeof(int);
 	sim_state.cpu_evals_of_runs.resize(size_evals_of_runs);
 	memset(sim_state.cpu_evals_of_runs.data(), 0, size_evals_of_runs);
+
+	// allocating memory blocks for GPU
+	status = cudaMalloc((void**)&(tData.pMem_conformations1), size_populations);
+	RTERROR(status, "pMem_conformations1: failed to allocate GPU memory.\n");
+	status = cudaMalloc((void**)&(tData.pMem_conformations2), size_populations);
+	RTERROR(status, "pMem_conformations2: failed to allocate GPU memory.\n");
+	status = cudaMalloc((void**)&(tData.pMem_energies1), size_energies);
+	RTERROR(status, "pMem_energies1: failed to allocate GPU memory.\n");
+	status = cudaMalloc((void**)&(tData.pMem_energies2), size_energies);
+	RTERROR(status, "pMem_energies2: failed to allocate GPU memory.\n");  
+	status = cudaMalloc((void**)&(tData.pMem_evals_of_new_entities), MAX_POPSIZE*MAX_NUM_OF_RUNS*sizeof(int));
+	RTERROR(status, "pMem_evals_of_new_Entities: failed to allocate GPU memory.\n");
+#if defined (MAPPED_COPY)
+	status = cudaMallocManaged((void**)&(tData.pMem_gpu_evals_of_runs), size_evals_of_runs, cudaMemAttachGlobal);
+#else
+	status = cudaMalloc((void**)&(tData.pMem_gpu_evals_of_runs), size_evals_of_runs);
+#endif
+	RTERROR(status, "pMem_gpu_evals_of_runs: failed to allocate GPU memory.\n");
+	status = cudaMalloc((void**)&(tData.pMem_prng_states), size_prng_seeds);
+	RTERROR(status, "pMem_prng_states: failed to allocate GPU memory.\n");
 
 	// preparing the constant data fields for the GPU
 	kernelconstant_interintra*	KerConst_interintra = new kernelconstant_interintra;
@@ -893,6 +876,21 @@ parameters argc and argv:
 
 	free(cpu_prng_seeds);
 
+	status = cudaFree(tData.pMem_conformations1);
+	RTERROR(status, "cudaFree: error freeing pMem_conformations1");
+	status = cudaFree(tData.pMem_conformations2);
+	RTERROR(status, "cudaFree: error freeing pMem_conformations2");
+	status = cudaFree(tData.pMem_energies1);
+	RTERROR(status, "cudaFree: error freeing pMem_energies1");
+	status = cudaFree(tData.pMem_energies2);
+	RTERROR(status, "cudaFree: error freeing pMem_energies2");
+	status = cudaFree(tData.pMem_evals_of_new_entities);
+	RTERROR(status, "cudaFree: error freeing pMem_evals_of_new_entities");
+	status = cudaFree(tData.pMem_gpu_evals_of_runs);
+	RTERROR(status, "cudaFree: error freeing pMem_gpu_evals_of_runs");
+	status = cudaFree(tData.pMem_prng_states);
+	RTERROR(status, "cudaFree: error freeing pMem_prng_states");
+
 	delete KerConst_interintra;
 	delete KerConst_intracontrib;
 	delete KerConst_intra;

host/src/setup.cpp

@@ -26,7 +26,6 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 #include <stdio.h>
 #include <stdlib.h>
 #include <vector>
-#include <sys/stat.h>
 
 #include "filelist.hpp"
 #include "processgrid.h"
@@ -345,43 +344,9 @@ int setup(
 	//------------------------------------------------------------
 	// Capturing algorithm parameters (command line args)
 	//------------------------------------------------------------
-	char* orig_resname;
-	if(mypars->resname)
-		orig_resname = strdup(mypars->resname); // need to copy it since it might get free'd in the next line
-	else
-		orig_resname = strdup(""); // need an empty string for later if no resname has been specified
-
-	if(get_commandpars(&argc, argv, &(mygrid->spacing), mypars)<0)
+	if(get_commandpars(&argc, argv, &(mygrid->spacing), mypars, true, (filelist.nfiles>1)?i_file+1:0)<0)
 		return 1;
 
-	// command-line specified resname with more than one file
-	if (!mypars->xml2dlg){ // if the user specified an xml file, that's the one we want to use
-		if ((strcmp(orig_resname,mypars->resname)!=0) && (filelist.nfiles>1)){ // use resname as prefix
-			char* tmp = (char*)malloc(strlen(mypars->resname)+strlen(orig_resname)+1);
-			// take care of potential directory path
-			long long dir = strrchr(orig_resname,'/')-orig_resname+1;
-			if(dir>0){
-				strncpy(tmp, orig_resname, dir);
-				tmp[dir]='\0';
-				strcat(tmp, mypars->resname);
-				strcat(tmp, &orig_resname[dir]);
-			} else{
-				strcpy(tmp, mypars->resname);
-				strcat(tmp, orig_resname);
-			}
-			free(mypars->resname);
-			mypars->resname = tmp;
-		}
-	}
-	free(orig_resname);
-
-	struct stat res_stat;
-	int res_int = stat(get_filepath(mypars->resname).c_str(), &res_stat);
-	if ((res_int != 0) || !(res_stat.st_mode & S_IFDIR)){
-		printf("\nError: Specified directory \"%s\" for output files (e.g. with `--resnam`) does not exist.\n",get_filepath(mypars->resname).c_str());
-		exit(11);
-	}
-
 	Gridinfo mydummygrid;
 	// if -lxrayfile provided, then read xray ligand data
 	if (mypars->given_xrayligandfile)