bones-compiler 1.3.1 → 1.6.0

data/lib/bones/variablelist.rb

@@ -15,7 +15,7 @@ module Bones
 		def select(direction)
 			array = Variablelist.new()
 			self.each do |element|
-				array.push(element) if ((direction == INPUT) && (element.input?)) || ((direction == OUTPUT) && (element.output?))
+				array.push(element) if ((direction == INPUT) && (element.input?)) || ((direction == OUTPUT) && (element.output?)) || (element.direction == INOUT)
 			end
 			return array
 		end

data/lib/cast.rb

@@ -0,0 +1,11 @@
+require "cast/cast"
+require "cast/tempfile"
+require "cast/preprocessor"
+require "cast/node"
+require "cast/node_list"
+require "cast/c_nodes"
+require "cast/c.tab"
+require "cast/parse"
+require "cast/to_s"
+require "cast/inspect"
+require "cast/version"

data/lib/castaddon.rb

@@ -12,12 +12,10 @@ require 'cast'
 # for Bones and A-Darwin.
 require 'castaddon/node_common.rb'
 require 'castaddon/node_adarwin.rb'
-if File.exists?('lib/bones.rb')
-	require 'castaddon/node_bones.rb'
-	require 'castaddon/transformations.rb'
-	require 'castaddon/type.rb'
-	require 'castaddon/index.rb'
-end
+require 'castaddon/node_bones.rb'
+require 'castaddon/transformations.rb'
+require 'castaddon/type.rb'
+require 'castaddon/index.rb'
 
 # Modify the NodeArray and NodeChain lists to output correct
 # code when printed to a file.
@@ -30,3 +28,22 @@ class C::NodeList
 	end
 end
 
+# update the FloatLiteral and IntLiteral classes in CAST to print suffixes
+class C::FloatLiteral
+	def to_s
+		if suffix.nil?
+			val.to_s
+		else
+			val.to_s + suffix
+		end
+	end
+end
+class C::IntLiteral
+	def to_s
+		if suffix.nil?
+			val.to_s
+		else
+			val.to_s + suffix
+		end
+	end
+end

data/lib/castaddon/node_adarwin.rb

@@ -146,6 +146,10 @@ class C::Node
 		elsif less?
 			results << ['',simplify("#{expr1}")+'='+simplify("(#{expr2})-1")]
 			
+		# Equal (==)
+		elsif equality?
+			results << ['','']#[simplify("#{expr1}"),simplify("(#{expr2})")]
+
 		# Unsupported conditions
 		else
 			raise_error("Unsupported if-condition: #{self.to_s}")
@@ -234,6 +238,19 @@ class C::Node
 		end
 	end
 	
+	# This method retrieves all variable declarations
+	def get_var_declarations()
+		vars = []
+		self.preorder do |node|
+			if node.declaration?
+				node.declarators.each do |decl|
+					vars << decl.name
+				end
+			end
+		end
+		return vars
+	end
+	
 	# This method retrieves the value from the current node. The value can be an
 	# integer (in case of a constant) or a string (in case of a variable).
 	def get_value()

data/lib/castaddon/node_common.rb

@@ -152,6 +152,12 @@ class C::Node
 		return (self.class == C::Or)
 	end
 	
+	# This method returns 'true' if the node is of the 'Equal'
+	# class. Otherwise, it returns 'false'.
+	def equality?
+		return (self.class == C::Equal)
+	end
+	
 	# This method returns 'true' if the node is of the 'Less'
 	# class. Otherwise, it returns 'false'.
 	def less?

data/lib/castaddon/transformations.rb

@@ -155,11 +155,13 @@ class C::Node
 	# Accepted inputs at this point: 2, 3 and 4 (CUDA/OPENCL)
 	# Also accepted input: 8 (CUDA), 9 (OPENCL) (to create an atomic version of the code)
 	# TODO: Complete the atomic support, e.g. add support for multiplications and ternary operators
-	def transform_reduction(input_variable,output_variable,id)
+	def transform_reduction(input_variable,shared_variable,id)
 		
 		# Pre-process assign-add type constructions
-		if self.stmts[0].expr.addassign?
-			self.stmts[0].expr.replace_with(C::Assign.parse(self.stmts[0].expr.lval.to_s+'='+self.stmts[0].expr.lval.to_s+'+'+self.stmts[0].expr.rval.to_s))
+		self.preorder do |node|
+			if node.addassign?
+				node.replace_with(C::Assign.parse(node.lval.to_s+'='+node.lval.to_s+'+'+node.rval.to_s))
+			end
 		end
 		
 		# Create atomic code
@@ -167,19 +169,21 @@ class C::Node
 			function_name = (id == 8) ? 'atomicAdd' : 'atomic_add'
 			self.preorder do |node|
 				if node.assign?
-					if node.lval.index? && node.lval.variable_name == output_variable.name
+					if node.lval.index? && node.lval.variable_name == shared_variable.name
 						if node.rval.add?
-							if node.rval.expr1.variable_name == output_variable.name
+							if node.rval.expr1.variable_name == shared_variable.name
 								node.replace_with(C::Call.parse(function_name+'(&'+node.rval.expr1.to_s+','+node.rval.expr2.to_s+')'))
-							elsif node.rval.expr2.variable_name == output_variable.name
+							elsif node.rval.expr2.variable_name == shared_variable.name
 								node.replace_with(C::Call.parse(function_name+'(&'+node.rval.expr2.to_s+','+node.rval.expr1.to_s+')'))
 							end
 						elsif node.rval.subtract?
-							if node.rval.expr1.variable_name == output_variable.name
+							if node.rval.expr1.variable_name == shared_variable.name
 								node.replace_with(C::Call.parse(function_name+'(&'+node.rval.expr1.to_s+',-'+node.rval.expr2.to_s+')'))
-							elsif node.rval.expr2.variable_name == output_variable.name
+							elsif node.rval.expr2.variable_name == shared_variable.name
 								node.replace_with(C::Call.parse(function_name+'(&'+node.rval.expr2.to_s+',-'+node.rval.expr1.to_s+')'))
 							end
+						elsif node.assign?
+							
 						else
 							raise_error('Unsupported atomic reduction operator: '+node.rval.type.inspect)
 						end
@@ -196,7 +200,7 @@ class C::Node
 				nodes.preorder do |node|
 					if (node.index?)
 						results[0] = nodes if node.variable_name == input_variable.name
-						results[1] = nodes if node.variable_name == output_variable.name
+						results[1] = nodes if node.variable_name == shared_variable.name
 					end
 				end
 			end

data/skeletons/CPU-C/common/timer_globals.c

@@ -2,6 +2,11 @@
 //////////// Timers ////////////////////
 ////////////////////////////////////////
 
+// Includes
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
 // Timer
 struct timeval bones_start_time1;
 

data/skeletons/CPU-OPENCL-INTEL/common/timer_globals.c

@@ -3,6 +3,11 @@
 //////////// Timers ////////////////////
 ////////////////////////////////////////
 
+// Includes
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/time.h>
+
 // Timer
 struct timeval bones_start_time1;
 

data/skeletons/CPU-OPENMP/common/timer_globals.c

@@ -4,6 +4,8 @@
 
 // Includes
 #include <stdio.h>
+#include <stdlib.h>
+#include <sys/time.h>
 
 // Timer
 struct timeval bones_start_time1;

data/skeletons/CPU-OPENMP/kernel/D-element-to-1-shared.kernel.c

@@ -1,10 +1,10 @@
 /* STARTDEF
-void bones_kernel_<algorithm_name>_0(int bones_thread_id, int bones_thread_count, int bones_size, <devicedefinitions>, <argument_definition>);
+void bones_kernel_<algorithm_name>_0(int bones_thread_id, int bones_thread_count, int bones_size, <devicedefinitions>, <in0_type>* bones_partial_result, <argument_definition>);
 void bones_kernel_<algorithm_name>_1(int bones_size, <devicedefinitions>, <argument_definition>);
 void bones_kernel_<algorithm_name>_2(<out0_type> bones_initial_value, <out0_type><out0_devicepointer> <out0_name>, <argument_definition>);
 ENDDEF */
 // Start of the <algorithm_name> kernel (main part)
-void bones_kernel_<algorithm_name>_0(int bones_thread_id, int bones_thread_count, int bones_size, <devicedefinitions>, <argument_definition>) {
+void bones_kernel_<algorithm_name>_0(int bones_thread_id, int bones_thread_count, int bones_size, <devicedefinitions>, <in0_type>* bones_partial_result, <argument_definition>) {
   const int bones_work = DIV_CEIL(bones_size,bones_thread_count);
   const int bones_global_id = bones_thread_id;
   <ids>
@@ -22,7 +22,7 @@ void bones_kernel_<algorithm_name>_0(int bones_thread_id, int bones_thread_count
   }
   
   // Store the result
-  <out0_name>[bones_thread_id] = bones_private_memory;
+  bones_partial_result[bones_thread_id] = bones_private_memory;
 }
 
 // Start of the <algorithm_name> kernel (secondary part)

data/skeletons/GPU-CUDA/common/scheduler.c

@@ -47,7 +47,7 @@ void bones_memcpy(void *dst, void *src, int size, enum cudaMemcpyKind direction,
 
 // Perform a task (CUDA memory copy)
 void bones_scheduler_copy(Task current_task) {
-  usleep(400);
+  //usleep(400);
   cudaMemcpyAsync(current_task.dst, current_task.src, current_task.size, current_task.direction, memory_stream);
   cudaStreamSynchronize(memory_stream);
 }
@@ -77,7 +77,7 @@ void* bones_scheduler(void* ptr) {
     
     // Perform the found task
     if (found_task != LARGE_INT) {
-      printf("Performing task %d, dl %d [scheduler]\n",found_task,tasks[found_task].deadline);
+      //printf("Performing task %d, dl %d [scheduler]\n",found_task,tasks[found_task].deadline);
       bones_scheduler_copy(tasks[found_task]);
       tasks[found_task].status = 2;
     }