bones-compiler 1.3.1 → 1.6.0

data/examples/neighbourhood/example02.c

@@ -8,16 +8,16 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........neighbourhood/example2.c
+// Filename...........neighbourhood/example02.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 #define A 256
 #define B 512
 
-// This is 'example2', demonstrating a 2D array, a 2D neighbourhood and a for-loop-less notation of the neighbourhood accesses
+// This is 'example02', demonstrating a 2D array, a 2D neighbourhood and a for-loop-less notation of the neighbourhood accesses
 int main(void) {
 	int i,j;
 	
@@ -33,7 +33,8 @@ int main(void) {
 	}
 	
 	// Perform the computation
-	#pragma species kernel 0:255,0:511|neighbourhood(-1:1,-1:1) -> 0:255,0:511|element
+	#pragma scop
+	#pragma species kernel in[0:255,0:511]|neighbourhood(-1:1,-1:1) -> out[0:255,0:511]|element
 	for(i=0;i<A;i++) {
 		for(j=0;j<B;j++) {
 			if (i >= 1 && j >= 1 && i < (A-1) && j < (B-1)) {
@@ -47,6 +48,7 @@ int main(void) {
 		}
 	}
 	#pragma species endkernel example2
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/examples/neighbourhood/example03.c

@@ -8,9 +8,9 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........neighbourhood/example3.c
+// Filename...........neighbourhood/example03.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
@@ -23,7 +23,7 @@
 float ** alloc_2D(int size1, int size2);
 void free_2D(float ** array_2D);
 
-// This is 'example3', demonstrating a neighbourhood with only some values used (a cross) and a math.h square root function call
+// This is 'example03', demonstrating a neighbourhood with only some values used (a cross) and a math.h square root function call
 int main(void) {
 	int i,j;
 	int sizea = A;
@@ -41,7 +41,8 @@ int main(void) {
 	}
 	
 	// Perform the computation
-	#pragma species kernel 0:sizea-1,0:sizeb-1|neighbourhood(-1:1,-1:1) -> 0:sizea-1,0:sizeb-1|element
+	#pragma scop
+	#pragma species kernel in[0:sizea-1,0:sizeb-1]|neighbourhood(-1:1,-1:1) -> out[0:sizea-1,0:sizeb-1]|element
 	for(i=0;i<sizea;i++) {
 		for(j=0;j<sizeb;j++) {
 			if (i >= 1 && j >= 1 && i < (sizea-1) && j < (sizeb-1)) {
@@ -55,6 +56,7 @@ int main(void) {
 		}
 	}
 	#pragma species endkernel example3
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	free_2D(in);

data/examples/neighbourhood/example04.c

@@ -8,14 +8,14 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........neighbourhood/example4.c
+// Filename...........neighbourhood/example04.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 
-// This is 'example4', demonstrating naming (optional) in the classification to distingish the two input arrays
+// This is 'example04', demonstrating naming (optional) in the classification to distingish the two input arrays
 int main(void) {
 	int i;
 	float factor;
@@ -33,6 +33,7 @@ int main(void) {
 	}
 	
 	// Perform the computation
+	#pragma scop
 	#pragma species kernel B[0:size-1]|neighbourhood(-1:1) ^ A[0:size-1]|element -> C[0:size-1]|element
 	for(i=0;i<size;i++) {
 		factor = A[i]/100.0;
@@ -44,6 +45,7 @@ int main(void) {
 		}
 	}
 	#pragma species endkernel example4
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/examples/neighbourhood/example05.c

@@ -10,7 +10,7 @@
 // == File information
 // Filename...........neighbourhood/example05.c
 // Author.............Cedric Nugteren
-// Last modified on...07-May-2013
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
@@ -31,11 +31,13 @@ int main(void) {
 	}
 	
 	// Perform the computation
+	#pragma scop
 	#pragma species kernel A[2:N]|neighbourhood(0:1) -> B[2:N-1]|element
 	for (i=2; i<N; i++) {
 		B[i] = A[i] + A[i+1];
 	}
 	#pragma species endkernel example05
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/examples/shared/example01.c

@@ -8,16 +8,16 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........shared/example1.c
+// Filename...........shared/example01.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 #include <stdlib.h>
 #define SIZE 512*1024
 
-// This is 'example1', a basic associative and commutative reduction to scalar
+// This is 'example01', a basic associative and commutative reduction to scalar
 int main(void) {
 	int i;
 	
@@ -32,11 +32,13 @@ int main(void) {
 	
 	// Perform the computation
 	B[0] = 0;
-	#pragma species kernel 0:SIZE-1|element -> 0:0|shared
+	#pragma scop
+	#pragma species kernel A[0:SIZE-1]|element -> B[0:0]|shared
 	for(i=0;i<SIZE;i++) {
 		B[0] = B[0] + A[i];
 	}
 	#pragma species endkernel example1
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/examples/shared/example02.c

@@ -8,14 +8,14 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........shared/example2.c
+// Filename...........shared/example02.c
 // Author.............Cedric Nugteren
-// Last modified on...09-May-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 
-// This is 'example2', demonstrating a 3D input reduction to scalar
+// This is 'example02', demonstrating a 3D input reduction to scalar
 int main(void) {
 	int a,b,c;
 	
@@ -34,7 +34,8 @@ int main(void) {
 	
 	// Perform the computation
 	out[0] = 0;
-	#pragma species kernel 0:7,0:15,0:31|element -> 0:0|shared
+	#pragma scop
+	#pragma species kernel in[0:7,0:15,0:31]|element -> out[0:0]|shared
 	for(a=0;a<8;a++) {
 		for(b=0;b<16;b++) {
 			for(c=0;c<32;c++) {
@@ -43,6 +44,7 @@ int main(void) {
 		}
 	}
 	#pragma species endkernel example2
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/examples/shared/example03.c

@@ -8,16 +8,16 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........shared/example3.c
+// Filename...........shared/example03.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 #include <stdlib.h>
 #define SIZE 1024
 
-// This is 'example3', demonstrating a reduction to a 2D array
+// This is 'example03', demonstrating a reduction to a 2D array
 int main(void) {
 	int i,p,q;
 	int index1,index2;
@@ -39,13 +39,15 @@ int main(void) {
 	}
 	
 	// Perform the computation
-	#pragma species kernel 0:SIZE-1|element -> 0:19,0:9|shared
+	#pragma scop
+	#pragma species kernel in[0:SIZE-1]|element -> B[0:19,0:9]|shared
 	for(i=0;i<SIZE;i++) {
 		index1 = in[i]%20;
 		index2 = in[i]%10;
 		B[index1][index2] = B[index1][index2] + 1;
 	}
 	#pragma species endkernel example3
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	free(in);

data/examples/shared/example04.c

@@ -8,16 +8,16 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........shared/example4.c
+// Filename...........shared/example04.c
 // Author.............Cedric Nugteren
-// Last modified on...16-April-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 #include <stdlib.h>
 #define SIZE 1024*1024
 
-// This is 'example4', demonstrating a basic 256-bin histogram computation
+// This is 'example04', demonstrating a basic 256-bin histogram computation
 int main(void) {
 	int i;
 	unsigned char index;
@@ -37,12 +37,14 @@ int main(void) {
 	}
 	
 	// Perform the computation
-	#pragma species kernel 0:SIZE-1|element -> 0:255|shared
+	#pragma scop
+	#pragma species kernel A[0:SIZE-1]|element -> B[0:255]|shared
 	for(i=0;i<SIZE;i++) {
 		index = A[i];
 		B[index]++;
 	}
 	#pragma species endkernel example4
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	free(A);

data/examples/shared/example05.c

@@ -8,14 +8,14 @@
 // Web address........http://parse.ele.tue.nl/bones/
 //
 // == File information
-// Filename...........shared/example5.c
+// Filename...........shared/example05.c
 // Author.............Cedric Nugteren
-// Last modified on...07-May-2012
+// Last modified on...10-October-2014
 //
 
 #include <stdio.h>
 
-// This is 'example5', demonstrating an inner-loop only classification of a reduction to scalar
+// This is 'example05', demonstrating an inner-loop only classification of a reduction to scalar
 int main(void) {
 	int a,b,c;
 	
@@ -32,14 +32,16 @@ int main(void) {
 	}
 	
 	// Perform the computation
+	#pragma scop
 	for(a=0;a<16;a++) {
-		#pragma species kernel a:a,0:a|element -> 0:0|shared
+		#pragma species kernel in[a:a,0:a]|element -> out[0:0]|shared
 		for(b=0;b<=a;b++) {
 			out[0] = out[0] - in[a][b]*in[a][b];
 		}
 		#pragma species endkernel example5
 		out[0] = 1.002;
 	}
+	#pragma endscop
 	
 	// Clean-up and exit the function
 	fflush(stdout);

data/lib/adarwin/engine.rb

@@ -83,16 +83,27 @@ module Adarwin
 			# Parse the original source code into AST form (using CAST)
 			original_ast = parser.parse(preprocessor.parsed_code)
 			
+			# Process every SCoP, one by one
+			@id = 0
+			@result[:species_code] = preprocessor.target_code
+			preprocessor.scop_code.each do |scop_code|
+				process_scop(scop_code)
+			end
+		end
+		
+		def process_scop(scop_code)
 			# Create an AST of the SCoP (using CAST) and save a backup
-			scop_ast = C::Block.parse('{'+preprocessor.scop_code+'}')
+			scop_ast = C::Block.parse('{'+scop_code+'}')
 			original_scop_ast = scop_ast.clone
 			
 			# Process the scop to identify the loop nests of interest and to find the
 			# corresponding species. This is the method performing most of the work.
 			@nests = []
-			@id = 0
 			populate_nests(scop_ast)
 			
+			# return if no loop nests are found in the code
+			return unless @nests.length > 0
+			
 			# Remove inner-loop (nested) species. This removes all species that are
 			# found within another species. For completeness, this might be desired in
 			# some cases.
@@ -141,7 +152,7 @@ module Adarwin
 			puts modified_scop if !@options[:silent]
 			
 			# Store the result
-			@result[:species_code] = preprocessor.target_code.gsub(preprocessor.scop_code,modified_scop)
+			@result[:species_code].gsub!(scop_code,modified_scop)
 		end
 		
 		# This method writes the output code to a file.
@@ -149,7 +160,7 @@ module Adarwin
 			
 			# Populate the species file
 			# TODO: The filename is fixed, make this an optional argument
-			File.open(File.join(@options[:application].split('.').first+'_species'+'.c'),'w') do |target|
+			File.open(File.join(@options[:application].rpartition('.').first+'_species'+'.c'),'w') do |target|
 				target.puts @result[:species_code]
 			end
 		end
@@ -172,9 +183,9 @@ module Adarwin
 						# Only continue if the nest is an actual loop nest
 						if nest.for_statement?
 							@nests.push(Nest.new(new_level,nest,@id,@basename,!@options[:silent]))
+							@id += 1
 						end
 					end
-					@id += 1
 				end
 				
 				# Proceed to the next depth level.

data/lib/adarwin/memorycopies.rb

@@ -2,14 +2,16 @@
 
 # Recursive copy optimisations
 def recursive_copy_optimisations(nests,options)
-	perform_copy_optimisations1(nests,options)
-	perform_copy_optimisations2(nests,options)
-	nests.each do |nest|
-		children = get_children(nest)
-		recursive_copy_optimisations(children,options) if !children.empty?
+	2.times do
+		perform_copy_optimisations1(nests,options)
+		perform_copy_optimisations2(nests,options)
+		nests.each do |nest|
+			children = get_children(nest)
+			recursive_copy_optimisations(children,options) if !children.empty?
+		end
+		perform_copy_optimisations3(nests,options)
+		perform_copy_optimisations3(nests,options)
 	end
-	perform_copy_optimisations3(nests,options)
-	perform_copy_optimisations3(nests,options)
 end
 
 # First set of copyin/copyout optimisations (recursive)
@@ -134,8 +136,18 @@ def perform_copy_optimisations3(nests,options)
 				# Move copyins to outer loops
 				children.first.copyins.each do |copyin|
 					to_outer_loop = true
-					nest.outer_loops.map{ |l| l[:var] }.each do |var|
-						to_outer_loop = false if copyin.depends_on?(var)
+					nest.outer_loops.map{ |l| l[:var] }.each_with_index do |var,lindex|
+						if copyin.depends_on?(var)
+							to_outer_loop = false
+							if copyin.tD[0].a == var && copyin.tD[0].b == var
+								loopinfo = nest.outer_loops[lindex]
+								if loopinfo[:step] == "1"
+									copyin.tD[0].a = loopinfo[:min]
+									copyin.tD[0].b = loopinfo[:max]
+									to_outer_loop = true
+								end
+							end
+						end
 					end
 					children.drop(1).each do |child|
 						to_outer_loop = false if child.copyins.map{ |c| c.tN }.include?(copyin.tN)

data/lib/adarwin/nest.rb

@@ -47,12 +47,15 @@ module Adarwin
 			@all_loops = @code.get_all_loops()
 			@outer_loops = @code.get_direct_loops()
 			@inner_loops = @all_loops - @outer_loops
+
+			# Get all local variable declarations
+			@var_declarations = @code.get_var_declarations()
 			
 			# Process the read/write nodes in the loop body to obtain the array
 			# reference characterisations. The references also need to be aware of all
 			# loop data and of any if-statements in the loop body.
 			@references = @code.clone.get_accesses().map do |reference|
-				Reference.new(reference,@id,@inner_loops,@outer_loops,@verbose)
+				Reference.new(reference,@id,@inner_loops,@outer_loops,@var_declarations,@verbose)
 			end
 		
 			# Perform the dependence test. The result can be either true or false.
@@ -121,6 +124,19 @@ module Adarwin
 			# Else, set the species for the individual accesses.
 			read_names = (@reads.empty?) ? ['0:0|void'] : @reads.map{ |r| r.to_species }
 			write_names = (@writes.empty?) ? ['0:0|void'] : @writes.map{ |r| r.to_species }
+
+			# Remove a 'full' access pattern in case there is a same 'shared' write pattern
+			write_names.each do |write_name|
+				write_parts = write_name.split(PIPE)
+				if write_parts.last == 'shared'
+					read_names.each do |read_name|
+						read_parts = read_name.split(PIPE)
+						if read_parts.last == 'full' && read_parts.first == write_parts.first
+							read_names.delete(read_name)
+						end
+					end
+				end
+			end
 			
 			# Combine the descriptions (using Reference's +to_s+ method) into species
 			species_in = read_names.uniq.join(' '+WEDGE+' ')
@@ -174,6 +190,7 @@ module Adarwin
 			return false if @removed
 			return false if @has_dependences
 			return false if @species == ''
+			return false if (@writes) && (@writes.select{ |a| a.pattern == 'shared' }.length > 3)
 			only_full = (@reads) ? @reads.select{ |a| a.pattern != 'full' }.empty? : false
 			only_shared = (@writes) ? @writes.select{ |a| a.pattern != 'shared' }.empty? : false
 			return !(only_full && only_shared)