Omp offload

Makefile

@@ -25,21 +25,27 @@
 # Valid values: CPU, GPU, CUDA, OCLGPU
 
 ifeq ($(DEVICE), $(filter $(DEVICE),GPU CUDA))
-TEST_CUDA := $(shell ./test_cuda.sh nvcc "$(GPU_INCLUDE_PATH)" "$(GPU_LIBRARY_PATH)")
+  TEST_CUDA := $(shell ./test_cuda.sh nvcc "$(GPU_INCLUDE_PATH)" "$(GPU_LIBRARY_PATH)")
 # if user specifies DEVICE=CUDA it will be used (wether the test succeeds or not)
 # if user specifies DEVICE=GPU the test result determines wether CUDA will be used or not
-ifeq ($(DEVICE)$(TEST_CUDA),GPUyes)
-override DEVICE:=CUDA
-endif
+  ifeq ($(DEVICE)$(TEST_CUDA),GPUyes)
+    override DEVICE:=CUDA
+  endif
 endif
 ifeq ($(DEVICE),CUDA)
-override DEVICE:=GPU
-export
-include Makefile.Cuda
+  override DEVICE:=GPU
+  export
+  include Makefile.Cuda
+else
+ifeq ($(DEVICE), OMPGPU)
+  override DEVICE:=GPU
+  export
+  include Makefile.Ompt
 else
 ifeq ($(DEVICE),OCLGPU)
-override DEVICE:=GPU
-export
+  override DEVICE:=GPU
+  export
+  include Makefile.OpenCL
+endif
 endif
-include Makefile.OpenCL
 endif

Makefile.Ompt

@@ -0,0 +1,215 @@
+#
+# miniMDock OpenMP Target Offloading Makefile
+# ------------------------------------------------------
+# Note that environment variables must be defined
+# before compiling
+# DEVICE?
+
+UNAME := $(shell uname)
+
+#$(shell ./link_ompt.sh)
+#COMPILER=llvm
+
+#Using LLVM clang compiler
+ifeq ($(COMPILER), llvm)
+      CPP=clang++ -Ofast -std=c++17 -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda --cuda-path=$(GPU_PATH) -Xcuda-ptxas --maxrregcount=120
+##CPP=clang++ -Ofast -std=c++17 -fopenmp -lopmtarget -fopenmp-targets=nvptx64-nvidia-cuda --cuda-path=$(GPU_PATH)
+      COMP=llvm
+#Using IBM XL Compiler
+else ifeq ($(COMPILER), xl)
+      CPP=xlC -std=c++14 -qsmp=omp -qoffload -O3 -qstrict 
+      COMP=xl
+#Using NVIDIA nvc++ compiler
+else ifeq ($(COMPILER), nvhpc)
+      CPP=nvc++ -std=c++17 -mp=gpu -fast -gpu=cc70
+      COMP=nvhpc
+#Using gcc 
+else ifeq ($(COMPILER), gcc)
+      CPP=g++ -fopenmp -foffload=-lm  -O2 -std=c++17
+      COMP=gcc
+else ifeq ($(COMPILER), CC)
+      CPP=CC -fopenmp 
+      COMP=cce
+else ifeq ($(COMPILER), ROCM)
+      CPP=clang++  -fopenmp -target x86_64-pc-linux-gnu -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx908
+      COMP=rocm
+endif
+
+# ------------------------------------------------------
+# Project directories
+COMMON_DIR=./common
+HOST_INC_DIR=./host/inc
+HOST_SRC_DIR=./host/src
+KRNL_INC_DIR=./omp_offload/inc
+KRNL_SRC_DIR=./omp_offload/src
+BIN_DIR=./bin
+
+# Host sources
+HOST_SRC=$(wildcard $(HOST_SRC_DIR)/*.cpp)
+KRNL_SRC=$(wildcard $(KRNL_SRC_DIR)/*.cpp)
+SRC=$(HOST_SRC) $(KRNL_SRC)
+
+IFLAGS=-I$(COMMON_DIR) -I$(HOST_INC_DIR) -I$(KRNL_INC_DIR)
+#LFLAGS= -lomptarget 
+CFLAGS=$(IFLAGS) $(LFLAGS)
+
+OMPT_OPTS=-DUSE_OMPT
+TARGET := autodock_ompt
+
+ifeq ($(CARD), AMD)
+        CD=-DCARD_AMD
+        TARGET:=$(TARGET)_amd
+else
+        TARGET:=$(TARGET)_nv
+endif
+TARGET:=$(TARGET)_$(COMP)
+
+ifeq ($(OVERLAP), ON)
+	ifeq ($(DEVICE), GPU)
+		OMPT_OPTS+=-DUSE_OMP
+	endif
+endif
+
+ifeq ($(TASK2GPU), YES)
+                OMPT_OPTS+=-DTASKTOGPU
+endif
+
+BIN := $(wildcard $(TARGET)*)
+
+# ------------------------------------------------------
+# Number of work-items (wi)
+# Valid values: 32, 64, 128, 256
+NUMWI=
+
+ifeq ($(NUMWI), 32)
+        NWI=-DN32WI
+        TARGET:=$(TARGET)_32wi
+else ifeq ($(NUMWI), 64)
+        NWI=-DN64WI
+        TARGET:=$(TARGET)_64wi
+else ifeq ($(NUMWI), 128)
+        NWI=-DN128WI
+        TARGET:=$(TARGET)_128wi
+else ifeq ($(NUMWI), 256)
+                NWI=-DN256WI
+                TARGET:=$(TARGET)_256wi
+else
+        ifeq ($(DEVICE), CPU)
+                NWI=-DN16WI
+                TARGET:=$(TARGET)_16wi
+        else ifeq ($(DEVICE), GPU)
+                NWI=-DN64WI
+                TARGET:=$(TARGET)_64wi
+        endif
+endif
+
+# ------------------------------------------------------
+# Configuration
+# FDEBUG (full) : enables debugging on both host + device
+# LDEBUG (light): enables debugging on host
+# RELEASE
+CONFIG=RELEASE
+#CONFIG=FDEBUG
+
+ifeq ($(CONFIG),FDEBUG)
+	OPT =-O0 -g -Wall -DDOCK_DEBUG
+else ifeq ($(CONFIG),LDEBUG)
+	OPT =-O0 -g -Wall
+else ifeq ($(CONFIG),RELEASE)
+	OPT =-O3
+else
+	OPT =
+endif
+
+# ------------------------------------------------------
+# Reproduce results (remove randomness)
+REPRO=NO
+
+ifeq ($(REPRO),YES)
+	REP =-DREPRO
+else
+	REP =
+endif
+# ------------------------------------------------------
+
+all: odock
+
+check-env-dev:
+	@if test -z "$$DEVICE"; then \
+		echo "DEVICE is undefined"; \
+		exit 1; \
+	else \
+		if [ "$$DEVICE" = "GPU" ]; then \
+			echo "DEVICE is set to $$DEVICE"; \
+		else \
+			if [ "$$DEVICE" = "CPU" ]; then \
+				echo "DEVICE is set to $$DEVICE"; \
+			else \
+				if [ "$$DEVICE" = "SERIAL" ]; then \
+					echo "DEVICE is set to $$DEVICE"; \
+				else \
+					echo "DEVICE value is invalid. Set DEVICE to either CPU, GPU, or SERIAL (1 thread on CPU)"; \
+				fi; \
+			fi; \
+			if [ "$$OVERLAP" = "ON" ]; then \
+				echo "OVERLAP only works with the GPU version right now"; \
+				exit 1; \
+			fi; \
+		fi; \
+	fi; \
+	echo " "
+
+check-env-cpu:
+	@if test -z "$$CPU_INCLUDE_PATH"; then \
+		echo "CPU_INCLUDE_PATH is undefined"; \
+	else \
+		echo "CPU_INCLUDE_PATH is set to $$CPU_INCLUDE_PATH"; \
+	fi; \
+	if test -z "$$CPU_LIBRARY_PATH"; then \
+		echo "CPU_LIBRARY_PATH is undefined"; \
+	else \
+		echo "CPU_LIBRARY_PATH is set to $$CPU_LIBRARY_PATH"; \
+	fi; \
+	echo " "
+
+check-env-gpu:
+	@if test -z "$$GPU_INCLUDE_PATH"; then \
+		echo "GPU_INCLUDE_PATH is undefined"; \
+	else \
+		echo "GPU_INCLUDE_PATH is set to $$GPU_INCLUDE_PATH"; \
+	fi; \
+	if test -z "$$GPU_LIBRARY_PATH"; then \
+		echo "GPU_LIBRARY_PATH is undefined"; \
+	else \
+		echo "GPU_LIBRARY_PATH is set to $$GPU_LIBRARY_PATH"; \
+	fi; \
+	echo " "
+
+check-env-all: check-env-dev check-env-cpu check-env-gpu
+
+# ------------------------------------------------------
+# Priting out its git version hash
+
+GIT_VERSION := $(shell git describe --abbrev=40 --dirty --always --tags)
+
+CFLAGS+=-DVERSION=\"$(GIT_VERSION)\"
+
+# ------------------------------------------------------
+link-code:
+	ln -sf performdocking.h.Omp $(HOST_INC_DIR)/performdocking.h
+	ln -sf performdocking.cpp.Omp $(HOST_SRC_DIR)/performdocking.cpp
+
+unlink-code:
+	rm -f $(HOST_INC_DIR)/performdocking.h $(HOST_SRC_DIR)/performdocking.cpp
+
+#odock: check-env-all $(HOST_SRC) $(KRNL_SRC)
+odock: check-env-all $(SRC) link-code 
+	$(CPP) \
+	$(SRC) \
+	$(CFLAGS) \
+	-o $(BIN_DIR)/$(TARGET) \
+	$(NWI) $(OPT) $(DD) $(REP) $(CD) $(OMPT_OPTS)
+
+
+clean:
+	rm -f $(BIN_DIR)/* initpop.txt

common/calcenergy_basic.h

@@ -28,70 +28,97 @@ Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 
 #include "defines.h"
 
-#define RLIST_ATOMID_MASK       0x000000FF
-#define RLIST_RBONDID_MASK      0x0000FF00
-#define RLIST_RBONDID_SHIFT     8
-#define RLIST_FIRSTROT_MASK     0x00010000
-#define RLIST_GENROT_MASK       0x00020000
-#define RLIST_DUMMY_MASK        0x00040000
+#define RLIST_ATOMID_MASK 	0x000000FF
+#define RLIST_RBONDID_MASK 	0x0000FF00
+#define RLIST_RBONDID_SHIFT  	8
+#define RLIST_FIRSTROT_MASK  	0x00010000
+#define RLIST_GENROT_MASK 	0x00020000
+#define RLIST_DUMMY_MASK 	0x00040000
 
-#define DEG_TO_RAD              0.0174533f
+#define DEG_TO_RAD 		0.0174533f
 
 // LCG: linear congruential generator constants
-#define RAND_A                  1103515245u
-#define RAND_C                  12345u
+#define RAND_A 			1103515245u
+#define RAND_C 			12345u
 // WARNING: it is supposed that unsigned int is 32 bit long
-#define MAX_UINT                4294967296.0f
+#define MAX_UINT 		4294967296.0f
 
 // Sticking to array boundaries
 #define stick_to_bounds(x,a,b) x + (x <= a)*(a-x) + (x >= b)*(b-x)
 
 // e^2/4pie0 in kcal/mol
 #define ELEC_SCALE_FACTOR        332.06363f
 
-// Constants for dielelectric term of the
+// Macro for capturing grid values
+	// Original
+	#define GETGRIDVALUE(mempoi,gridsize_x,gridsize_y,gridsize_z,t,z,y,x)   *(mempoi + gridsize_x*(y + gridsize_y*(z + gridsize_z*t)) + x)
+
+	// Optimization 1
+	// #define GETGRIDVALUE_OPT(mempoi,gridsize_x,gridsize_y,mul_tmp,z,y,x)   *(mempoi + gridsize_x*(y + gridsize_y*(z + mul_tmp)) + x)
+
+	// Optimization 2
+	// Implemented direclty in the kernel code: calcenergy_fourkernels_intel.cl
+
+
+// Macro for trilinear interpolation
+#define TRILININTERPOL(cube, weights) (cube[idx_000]*weights[idx_000] + \
+				       cube[idx_010]*weights[idx_010] + \
+				       cube[idx_001]*weights[idx_001] + \
+				       cube[idx_011]*weights[idx_011] + \
+				       cube[idx_100]*weights[idx_100] + \
+				       cube[idx_110]*weights[idx_110] + \
+				       cube[idx_101]*weights[idx_101] + \
+				       cube[idx_111]*weights[idx_111])
+
+// Sticking to array boundaries
+#define stick_to_bounds(x,a,b) x + (x <= a)*(a-x) + (x >= b)*(b-x)
+
+// Constants for dielelectric term of the 
 // electrostatic component of the intramolecular energy/gradient
-#define DIEL_A                   -8.5525f
-#define DIEL_WAT                 78.4f
-#define DIEL_B                   (DIEL_WAT - DIEL_A)
-#define DIEL_LAMBDA              0.003627f
-#define DIEL_H                   DIEL_LAMBDA
-#define DIEL_K                   7.7839f
-#define DIEL_B_TIMES_H           (DIEL_B * DIEL_H)
-#define DIEL_B_TIMES_H_TIMES_K   (DIEL_B_TIMES_H * DIEL_K)
+#define DIEL_A 			-8.5525f
+#define DIEL_WAT 		78.4f
+#define DIEL_B 			(DIEL_WAT - DIEL_A)
+#define DIEL_LAMBDA		0.003627f
+#define DIEL_H			DIEL_LAMBDA
+#define DIEL_K			7.7839f
+#define DIEL_B_TIMES_H		(DIEL_B * DIEL_H)
+#define DIEL_B_TIMES_H_TIMES_K	(DIEL_B_TIMES_H * DIEL_K)
 
 // Used for Shoemake to quaternion transformation
 #if defined(M_PI)
-	#define PI_FLOAT (float)(M_PI)
+	#define PI_FLOAT		(float)(M_PI)
 #else
-	#define PI_FLOAT 3.14159265359f
+	#define PI_FLOAT		3.14159265359f
 #endif
-#define PI_TIMES_2       2.0f*PI_FLOAT
+#define PI_TIMES_2 		2.0f*PI_FLOAT
 
 // -------------------------------------------
 // Gradient-related defines
 // -------------------------------------------
 
-#define INFINITESIMAL_RADIAN            1E-3f
-#define HALF_INFINITESIMAL_RADIAN       (float)(0.5f * INFINITESIMAL_RADIAN)
-#define INV_INFINITESIMAL_RADIAN        (1.0f/INFINITESIMAL_RADIAN)
-#define COS_HALF_INFINITESIMAL_RADIAN   cos(HALF_INFINITESIMAL_RADIAN)
-#define SIN_HALF_INFINITESIMAL_RADIAN   sin(HALF_INFINITESIMAL_RADIAN)
-#define inv_angle_delta                 500.0f / PI_FLOAT
+#define INFINITESIMAL_RADIAN		1E-3
+#define HALF_INFINITESIMAL_RADIAN 	(float)(0.5f * INFINITESIMAL_RADIAN)
+#define INV_INFINITESIMAL_RADIAN	(1/INFINITESIMAL_RADIAN)
+#define COS_HALF_INFINITESIMAL_RADIAN	cos(HALF_INFINITESIMAL_RADIAN)
+#define SIN_HALF_INFINITESIMAL_RADIAN	sin(HALF_INFINITESIMAL_RADIAN)
+#define inv_angle_delta			500.0f / PI_FLOAT
 
 /*
 #define TRANGENE_ALPHA 1E-3
 #define ROTAGENE_ALPHA 1E-8
 #define TORSGENE_ALPHA 1E-13
 */
 
-#define STEP_INCREASE           1.2f
-#define STEP_DECREASE           0.2f
-#define STEP_START              1000.0f         // Starting step size. This might look gigantic but will cap
-#define MAX_DEV_TRANSLATION     2.0f            // 2 Angstrom, but must be divided by the gridspacing (store in variable)
-//#define MAX_DEV_ROTATION      0.2f            // Shoemake range [0, 1]
-#define MAX_DEV_ROTATION        0.5f/DEG_TO_RAD // 0.5f RAD
-#define MAX_DEV_TORSION         0.5f/DEG_TO_RAD // 0.5f RAD
+#define STEP_INCREASE 		1.2f
+#define STEP_DECREASE 		0.2f
+#define STEP_START		1E3		// Starting step size. This might look gigantic but will cap
+#define MAX_DEV_TRANSLATION	2.0f		// 2 Angstrom, but must be divided by the gridspacing (store in variable)
+//#define MAX_DEV_ROTATION	0.2f		// Shoemake range [0, 1]
+#define MAX_DEV_ROTATION	0.5f/DEG_TO_RAD	// 0.5f RAD
+#define MAX_DEV_TORSION		0.5f/DEG_TO_RAD	// 0.5f RAD
+
+
+
 
 
 #endif /* CALCENERGY_BASIC_H_ */

common/defines.h

@@ -98,7 +98,7 @@ 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 LS_CONT_FACTOR             0.5f
 
 // Improvements over Pechan's implementation
-#define MAPPED_COPY
+//#define MAPPED_COPY
 
 
 // Coefficients for CG-G0 pairs used in