added unroll attributes

Former-commit-id: 754508be

ofdock_taskpar_xl/Makefile

@@ -634,7 +634,7 @@ include $(COMMON_REPO)/utility/rules.mk
 # Automate test
 # "exe" and "xclbin" are rules from "rules.mk"
 emu: 
-	XCL_EMULATION_MODE=sw_emu ./hostbin -ffile input/1stp/derived/1stp_protein.maps.fld -lfile input/1stp/derived/1stp_ligand.pdbqt -nrun 50 -nev 200000
+	XCL_EMULATION_MODE=sw_emu ./hostbin -ffile input/1stp/derived/1stp_protein.maps.fld -lfile input/1stp/derived/1stp_ligand.pdbqt -nrun 2 -nev 2000000
 	
 debug:
 	XCL_EMULATION_MODE=sw_emu xgdb -args ./hostbin -ffile input/1stp/derived/1stp_protein.maps.fld -lfile input/1stp/derived/1stp_ligand.pdbqt -nrun 1 -nev 100 -lsit 10 -psize 10

ofdock_taskpar_xl/device/Krnl_Conform.cl

@@ -481,7 +481,11 @@ while(active) {
 		float3 tmp_coords[2];
 		#endif
 
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_CONFORM_OUT:
 		for (uchar i=0; i<2; i++) {
 			tmp_coords[i] = loc_coords[pipe_cnt+i];
 		}

ofdock_taskpar_xl/device/Krnl_GA.cl

@@ -616,7 +616,11 @@ void Krnl_GA(
 		// Shift register to reduce II (initially II=6) of best entity for-loop 
 		float shift_reg[SHIFT_REG_SIZE];
 
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_GA_SHIFT: 
 		for (uchar i=0; i<SHIFT_REG_SIZE; i++) {
 			shift_reg[i] = 0.0f;
 		}
@@ -636,7 +640,11 @@ void Krnl_GA(
 			//if (pop_cnt>0) {
 			shift_reg[SHIFT_REG_SIZE_MINUS_ONE] = loc_energies[best_entity];
 
+/*
 			#pragma unroll
+*/
+			__attribute__((opencl_unroll_hint))
+			LOOP_GA_SHIFT_MINUS_ONE:
 			for (uchar j=0; j<SHIFT_REG_SIZE_MINUS_ONE; j++) {
 				shift_reg[j] = shift_reg[j+1];
 			}
@@ -655,7 +663,11 @@ void Krnl_GA(
 			// Elitism: copying the best entity to new population
 			// ---------------------------------------------------
 			if (new_pop_cnt == 1) {
+/*
 				#pragma unroll
+*/
+				__attribute__((opencl_unroll_hint))
+				LOOP_GA_ELITISM:
 				for (uchar gene_cnt=0; gene_cnt<ACTUAL_GENOTYPE_LENGTH; gene_cnt++) { 		
 					LocalPopNext[0][gene_cnt & MASK_GENOTYPE] = LocalPopCurr[best_entity][gene_cnt & MASK_GENOTYPE]; 	
 				} 		

ofdock_taskpar_xl/device/Krnl_IntraE.cl

@@ -103,7 +103,11 @@ while(active) {
 	//float shift_intraE[33];
 	fixedpt64 shift_intraE[33];
 
+/*
 	#pragma unroll
+*/
+	__attribute__((opencl_unroll_hint))
+	LOOP_INTRAE_SHIFT_INIT:
 	for (uchar i=0; i<33; i++) {
 		//shift_intraE[i] = 0.0f;
 		shift_intraE[i] = 0;
@@ -191,7 +195,11 @@ while(active) {
 						     fixedpt64_fromfloat(partialE3) + 
 						     fixedpt64_fromfloat(partialE4);
 
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_INTRAE_SHIFT:
 		for (uchar j=0; j<32; j++) {
 			shift_intraE[j] = shift_intraE[j+1];
 		}
@@ -204,7 +212,11 @@ while(active) {
 	#if defined (FIXED_POINT_INTRAE)
 	fixedpt64 fixpt_intraE = 0;
 
+/*
 	#pragma unroll
+*/
+	__attribute__((opencl_unroll_hint))
+	LOOP_INTRAE_SHIFT_RED:
 	for (uchar j=0; j<32; j++) {
 		//intraE += shift_intraE[j];
 		fixpt_intraE += shift_intraE[j];

ofdock_taskpar_xl/device/Krnl_PRNG.cl

@@ -32,7 +32,11 @@ void Krnl_Prng_BT_ushort_float(unsigned int Host_seed1,
 		uint   u_tmp[4]; // used as short in GA
 		float  f_tmp[4];	
 
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_PRNG_BT_USHORT_FLOAT:
 		for(uchar i=0; i<4; i++) {
 			uchar2 lsb;
 
@@ -98,7 +102,11 @@ void Krnl_Prng_GG_uchar(unsigned int  Host_seed,
 
 		uchar tmp[2];
 
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_PRNG_GG_UCHAR:
 		for(uchar i=0; i<2; i++) {
 			uchar lsb;
 			lsb = lfsr & 0x01u;
@@ -223,7 +231,11 @@ void Krnl_Prng_LS123_ushort(unsigned int Host_seed1,
 
 		ushort tmp[9];
 		
+/*
 		#pragma unroll
+*/
+		__attribute__((opencl_unroll_hint))
+		LOOP_PRNG_LS123_USHORT:
 		for (uint i=0; i<9; i++){
 			uchar  lsb[9];
 			lsb [i] = lfsr[i] & 0x01u;