issue #35: merged E+G + updated fire & adadelta kernels

device/calcenergy.cl

@@ -23,7 +23,10 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
 
 //#define DEBUG_ENERGY_KERNEL
 
+// No needed to be included as all kernel sources are stringified
+#if 0
 #include "calcenergy_basic.h"
+#endif
 
 typedef struct
 {
@@ -689,6 +692,8 @@ if (get_local_id (0) == 0) {
 
 }
 
+// No needed to be included as all kernel sources are stringified
+#if 0
 #include "kernel1.cl"
 #include "kernel2.cl"
 #include "auxiliary_genetic.cl"
@@ -698,3 +703,6 @@ if (get_local_id (0) == 0) {
 #include "kernel_sd.cl"
 #include "kernel_fire.cl"
 #include "kernel_ad.cl"
+#include "calcEnerGrad.cl"
+#endif
+

device/kernel_ad.cl

@@ -312,6 +312,12 @@ gradient_minAD(
 		}
 		#endif
 
+// Replacing separate gradient and energy 
+// calculations with a single & unified
+// gpu_calc_energrad() function
+// IMPORTANT: be careful with input/output (RE) assignment
+// of genotypes, energy, and gradients
+#if 0
 		// =============================================================
 		// Calculating gradient
 		barrier(CLK_LOCAL_MEM_FENCE);
@@ -437,20 +443,92 @@ gradient_minAD(
 				calc_coords_y,
 				calc_coords_z,
 				partial_energies,
-				#if defined (DEBUG_ENERGY_KERNEL)
+			#if defined (DEBUG_ENERGY_KERNEL)
 				partial_interE,
 				partial_intraE,
-				#endif
-#if 0
+			#endif
+			#if 0
 				true,
+			#endif
+				kerconst_interintra,
+				kerconst_intracontrib,
+				kerconst_intra,
+				kerconst_rotlist,
+				kerconst_conform
+				);
+		// =============================================================
 #endif
+
+		// =============================================================
+		// =============================================================
+		// =============================================================
+		// Calculating energy & gradient
+		barrier(CLK_LOCAL_MEM_FENCE);
+
+		gpu_calc_energrad(
+				dockpars_rotbondlist_length,
+				dockpars_num_of_atoms,
+				dockpars_gridsize_x,
+				dockpars_gridsize_y,
+				dockpars_gridsize_z,
+								    	// g1 = gridsize_x
+				dockpars_gridsize_x_times_y, 		// g2 = gridsize_x * gridsize_y
+				dockpars_gridsize_x_times_y_times_z,	// g3 = gridsize_x * gridsize_y * gridsize_z
+				dockpars_fgrids,
+				dockpars_num_of_atypes,
+				dockpars_num_of_intraE_contributors,
+				dockpars_grid_spacing,
+				dockpars_coeff_elec,
+				dockpars_qasp,
+				dockpars_coeff_desolv,
+				dockpars_smooth,
+
+				// Some OpenCL compilers don't allow declaring 
+				// local variables within non-kernel functions.
+				// These local variables must be declared in a kernel, 
+				// and then passed to non-kernel functions.
+				genotype,
+				&energy,
+				&run_id,
+
+				calc_coords_x,
+				calc_coords_y,
+				calc_coords_z,
+				partial_energies,
+				#if defined (DEBUG_ENERGY_KERNEL)
+				partial_interE,
+				partial_intraE,
+				#endif
+
 				kerconst_interintra,
 				kerconst_intracontrib,
 				kerconst_intra,
 				kerconst_rotlist,
 				kerconst_conform
+				,
+				rotbonds_const,
+				rotbonds_atoms_const,
+				num_rotating_atoms_per_rotbond_const
+				,
+	     			angle_const,
+	     			dependence_on_theta_const,
+	     			dependence_on_rotangle_const
+			 	// Gradient-related arguments
+			 	// Calculate gradients (forces) for intermolecular energy
+			 	// Derived from autodockdev/maps.py
+				,
+				dockpars_num_of_genes,
+				gradient_inter_x,
+				gradient_inter_y,
+				gradient_inter_z,
+				gradient_intra_x,
+				gradient_intra_y,
+				gradient_intra_z,
+				gradient
 				);
 		// =============================================================
+		// =============================================================
+		// =============================================================
 
 		#if defined (DEBUG_ENERGY_ADADELTA)
 		if (/*(get_group_id(0) == 0) &&*/ (get_local_id(0) == 0)) {

device/kernel_fire.cl

@@ -440,11 +440,17 @@ gradient_minFire(
 			candidate_genotype [gene_counter] = genotype [gene_counter] + dt * velocity [gene_counter];	
 		}
 
+// Replacing separate gradient and energy 
+// calculations with a single & unified
+// gpu_calc_energrad() function
+// IMPORTANT: be careful with input/output (RE) assignment
+// of genotypes, energy, and gradients
+#if 0
+		// =============================================================
 		// Calculating (candidate) gradient
 		// from "candidate_genotype"
 		barrier(CLK_LOCAL_MEM_FENCE);
 
-		// =============================================================
 		gpu_calc_gradient(
 				dockpars_rotbondlist_length,
 				dockpars_num_of_atoms,
@@ -536,20 +542,92 @@ gradient_minFire(
 				calc_coords_y,
 				calc_coords_z,
 				partial_energies,
-				#if defined (DEBUG_ENERGY_KERNEL)
+			#if defined (DEBUG_ENERGY_KERNEL)
 				partial_interE,
 				partial_intraE,
-				#endif
-#if 0
+			#endif
+			#if 0
 				true,
+			#endif
+				kerconst_interintra,
+				kerconst_intracontrib,
+				kerconst_intra,
+				kerconst_rotlist,
+				kerconst_conform
+				);
+		// =============================================================
 #endif
+
+		// =============================================================
+		// =============================================================
+		// =============================================================
+		// Calculating energy & gradient
+		barrier(CLK_LOCAL_MEM_FENCE);
+
+		gpu_calc_energrad(
+				dockpars_rotbondlist_length,
+				dockpars_num_of_atoms,
+				dockpars_gridsize_x,
+				dockpars_gridsize_y,
+				dockpars_gridsize_z,
+								    	// g1 = gridsize_x
+				dockpars_gridsize_x_times_y, 		// g2 = gridsize_x * gridsize_y
+				dockpars_gridsize_x_times_y_times_z,	// g3 = gridsize_x * gridsize_y * gridsize_z
+				dockpars_fgrids,
+				dockpars_num_of_atypes,
+				dockpars_num_of_intraE_contributors,
+				dockpars_grid_spacing,
+				dockpars_coeff_elec,
+				dockpars_qasp,
+				dockpars_coeff_desolv,
+				dockpars_smooth,
+
+				// Some OpenCL compilers don't allow declaring 
+				// local variables within non-kernel functions.
+				// These local variables must be declared in a kernel, 
+				// and then passed to non-kernel functions.
+				candidate_genotype,
+				&candidate_energy,
+				&run_id,
+
+				calc_coords_x,
+				calc_coords_y,
+				calc_coords_z,
+				partial_energies,
+				#if defined (DEBUG_ENERGY_KERNEL)
+				partial_interE,
+				partial_intraE,
+				#endif
+
 				kerconst_interintra,
 				kerconst_intracontrib,
 				kerconst_intra,
 				kerconst_rotlist,
 				kerconst_conform
+				,
+				rotbonds_const,
+				rotbonds_atoms_const,
+				num_rotating_atoms_per_rotbond_const
+				,
+	     			angle_const,
+	     			dependence_on_theta_const,
+	     			dependence_on_rotangle_const
+			 	// Gradient-related arguments
+			 	// Calculate gradients (forces) for intermolecular energy
+			 	// Derived from autodockdev/maps.py
+				,
+				dockpars_num_of_genes,
+				gradient_inter_x,
+				gradient_inter_y,
+				gradient_inter_z,
+				gradient_intra_x,
+				gradient_intra_y,
+				gradient_intra_z,
+				candidate_gradient
 				);
 		// =============================================================
+		// =============================================================
+		// =============================================================
 
 		// Calculating power
         	// power is force * velocity.

stringify_ocl_krnls.sh

@@ -6,35 +6,53 @@ IN_HEADER1=$HEADER_DIR/"defines.h"
 IN_HEADER2=$HEADER_DIR/"calcenergy_basic.h"
 
 echo " "
-echo "Stringified input header-files: "
+echo "Stringified input header files: "
 echo $IN_HEADER1
 echo $IN_HEADER2
 
-# kernel files
+# device source-code folder
 KERNEL_DIR="./device"
-IN_KERNELm=$KERNEL_DIR/"calcenergy.cl"
-IN_KERNEL1=$KERNEL_DIR/"kernel1.cl"
-IN_KERNEL2=$KERNEL_DIR/"kernel2.cl"
-IN_KERNELa=$KERNEL_DIR/"auxiliary_genetic.cl"
-IN_KERNEL3=$KERNEL_DIR/"kernel3.cl"
-IN_KERNEL4=$KERNEL_DIR/"kernel4.cl"
-IN_KERNELb=$KERNEL_DIR/"calcgradient.cl"
-IN_KERNEL5=$KERNEL_DIR/"kernel_sd.cl"
-IN_KERNEL6=$KERNEL_DIR/"kernel_fire.cl"
-IN_KERNEL7=$KERNEL_DIR/"kernel_ad.cl"
+
+# non-kernel function files
+## AUXI:   auxiliary
+## ENER:   energy
+## GRAD:   gradient
+## ENEGRA: merged energy & gradient
+IN_FUNCTION_AUXI=$KERNEL_DIR/"auxiliary_genetic.cl"
+IN_FUNCTION_ENER=$KERNEL_DIR/"calcenergy.cl"
+IN_FUNCTION_GRAD=$KERNEL_DIR/"calcgradient.cl"
+IN_FUNCTION_MERGED_EG=$KERNEL_DIR/"calcMergedEneGra.cl"
+
+# kernel files
+## INIT: energy calculation of initial population
+## EVAL: evaluation count 
+## GS_ORIG_GA: Global Search Original Genetic Algorithm
+## LS_ORIG_SW: Local Search Original Solis-Wets
+## LS_GRAD_{SD | FI | AD}: Local Search Gradient-based {Steepest-Descent | Fire | Ada-Delta}
+IN_KERNEL_INIT=$KERNEL_DIR/"kernel1.cl"
+IN_KERNEL_EVAL=$KERNEL_DIR/"kernel2.cl"
+IN_KERNEL_GS_ORIG_GA=$KERNEL_DIR/"kernel4.cl"
+IN_KERNEL_LS_ORIG_SW=$KERNEL_DIR/"kernel3.cl"
+IN_KERNEL_LS_GRAD_SD=$KERNEL_DIR/"kernel_sd.cl"
+IN_KERNEL_LS_GRAD_FI=$KERNEL_DIR/"kernel_fire.cl"
+IN_KERNEL_LS_GRAD_AD=$KERNEL_DIR/"kernel_ad.cl"
+
+echo " "
+echo "Stringified input non-kernel files: "
+echo $IN_FUNCTION_AUXI
+echo $IN_FUNCTION_ENER
+echo $IN_FUNCTION_GRAD
+echo $IN_FUNCTION_MERGED_EG
 
 echo " "
 echo "Stringified input kernel-files: "
-echo $IN_KERNELm
-echo $IN_KERNEL1
-echo $IN_KERNEL2
-echo $IN_KERNELa
-echo $IN_KERNEL3
-echo $IN_KERNEL4
-echo $IN_KERNELb
-echo $IN_KERNEL5
-echo $IN_KERNEL6
-echo $IN_KERNEL7
+echo $IN_KERNEL_INIT
+echo $IN_KERNEL_EVAL
+echo $IN_KERNEL_GS_ORIG_GA
+echo $IN_KERNEL_LS_ORIG_SW
+echo $IN_KERNEL_LS_GRAD_SD
+echo $IN_KERNEL_LS_GRAD_FI
+echo $IN_KERNEL_LS_GRAD_AD
 
 # output file
 OUT=host/inc/stringify.h
@@ -47,23 +65,24 @@ echo " "
 # temporal file
 TMP=$KERNEL_DIR/stringify_tmp
 
-echo "// OCLADOCK: AUTOMATICALLY GENERATED FILE, DO NOT EDIT" >$TMP
+echo "// OCLADock: AUTOMATICALLY GENERATED FILE, DO NOT EDIT." >$TMP
 echo "#ifndef STRINGIFY_H" >>$TMP
 echo "#define STRINGIFY_H" >>$TMP
 
 echo "const char *calcenergy_ocl =" >>$TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_HEADER1 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_HEADER2 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNELm >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL1 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL2 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNELa >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL3 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL4 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNELb >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL5 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL6 >> $TMP
-sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL7 >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_HEADER1            >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_HEADER2            >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_FUNCTION_AUXI      >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_FUNCTION_ENER      >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_FUNCTION_GRAD      >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_FUNCTION_MERGED_EG >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_INIT        >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_EVAL        >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_GS_ORIG_GA  >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_LS_ORIG_SW  >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_LS_GRAD_SD  >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_LS_GRAD_FI  >> $TMP
+sed 's/\\/\\\\/g;s/"/\\"/g;s/^/"/;s/$/\\n"/' $IN_KERNEL_LS_GRAD_AD  >> $TMP
 echo ";" >>$TMP
 
 echo "#endif // End of STRINGIFY_H" >>$TMP