improved readability of gradient-norm calc

device/auxiliary_gradient.cl

@@ -66,10 +66,12 @@ bool is_gradDescent_enabled(
   	return *is_gradDescentEn;
 }
 
-float inner_product(__local float*	vector1,
-                    __local float*      vector2,
-                            uint 	inputSize,
-                    __local float*      init) {
+void gradient_norm(
+		   __local float*	vector1,
+                           uint 	inputSize,
+                   __local float*       init,
+		   __local float*       inner_product
+) {
 
 	float temp = 0.0f;
 
@@ -77,10 +79,11 @@ float inner_product(__local float*	vector1,
 	for(uint i = get_local_id(0); 
 		 i < inputSize; 
 		 i+= NUM_OF_THREADS_PER_BLOCK) {
-		init[i] = vector1[i] * vector2[i];
+		init[i] = vector1[i] * vector1[i];
 	}
 
-	// Accumulating dot product
+	// Accumulating dot product,
+	// and then getting the norm
 	if(get_local_id(0) == 0) {
 		for(uint i = 0; 
 			 i < inputSize; 
@@ -88,11 +91,7 @@ float inner_product(__local float*	vector1,
 			temp += init[i];
 		}
 
-		init [0] = temp;
+		*inner_product = native_sqrt(temp);
 	}
-	
-	barrier(CLK_LOCAL_MEM_FENCE);
-
-	return init[0];
 }
 

device/kernel_gradient.cl

@@ -168,7 +168,6 @@ gradient_minimizer(
 	__local float gradient_intra_z[MAX_NUM_OF_ATOMS];
 	__local float gradient_per_intracontributor[MAX_INTRAE_CONTRIBUTORS];
 	
-
 	// -------------------------------------------------------------------
 
 	// Variables to store partial energies
@@ -266,11 +265,7 @@ gradient_minimizer(
 		}
 
 	    	// Storing the norm of all gradients
-		local_gNorm = inner_product(local_gradient, local_gradient, dockpars_num_of_genes, dotProduct);
-
-		if (get_local_id(0) == 0) {
-			local_gNorm = native_sqrt(local_gNorm);
-		}
+		gradient_norm(local_gradient, dockpars_num_of_genes, dotProduct, &local_gNorm);
 
 		barrier(CLK_LOCAL_MEM_FENCE);
 
@@ -353,10 +348,12 @@ gradient_minimizer(
 
 	async_work_group_copy(dockpars_conformations_next+(run_id*dockpars_pop_size+entity_id)*GENOTYPE_LENGTH_IN_GLOBMEM,
 			      local_genotype,
-			      dockpars_num_of_genes,0);
+			      dockpars_num_of_genes, 0);
 
 
 	// FIXME: maybe not used outside?
   	// Copying final "local" results into "global" memory
+/*
   	local_nIter = local_nIter - 1;
+*/
 }