h48

commit ca68b59475f6b1773fa62e6dbc4b68349602b441
parent 762d285f3b914c877887469de06a70cee9028159
Author: Sebastiano Tronto <sebastiano@tronto.net>
Date:   Thu, 12 Sep 2024 16:55:28 +0200

Update build files to enable multithreading

Diffstat:
M
Makefile
 | 
16
++++++++--------
M
configure.sh
 | 
73
++++++++++++++++++++++++++++++++++++++++++++++++++++---------------------
M
tools/010_stats_tables_h48/stats_tables_h48.c
 | 
13
++++++-------

3 files changed, 66 insertions(+), 36 deletions(-)
diff --git a/Makefile b/Makefile
@@ -3,34 +3,34 @@ include config.mk
 all: cube.o debugcube.o
 
 cube.s: clean
-	${CC} -D${ARCH} ${CFLAGS} -c -S -o cube.s src/nissy.c
+	${CC} ${MACROS} ${CFLAGS} -c -S -o cube.s src/nissy.c
 
 cube.o: clean
-	${CC} -D${ARCH} ${CFLAGS} -c -o cube.o src/nissy.c
+	${CC} ${MACROS} ${CFLAGS} -c -o cube.o src/nissy.c
 
 debugcube.o: clean
-	${CC} -D${ARCH} ${DBGFLAGS} -c -o debugcube.o src/nissy.c
+	${CC} ${MACROS} ${DBGFLAGS} -c -o debugcube.o src/nissy.c
 
 clean:
 	rm -rf *.o run
 
 test: debugcube.o
-	CC="${CC} -D${ARCH} ${DBGFLAGS}" ./test/test.sh
+	CC="${CC} ${MACROS} ${DBGFLAGS}" ./test/test.sh
 
 tool: cube.o
 	mkdir -p tools/results
-	CC="${CC} ${CFLAGS}" CUBEOBJ=cube.o ./tools/run_tool.sh
+	CC="${CC} ${MACROS} ${CFLAGS}" CUBEOBJ=cube.o ./tools/run_tool.sh
 
 debugtool: debugcube.o
 	mkdir -p tools/results
-	CC="${CC} ${DBGFLAGS}" CUBEOBJ=debugcube.o ./tools/run_tool.sh
+	CC="${CC} ${MACROS} ${DBGFLAGS}" CUBEOBJ=debugcube.o ./tools/run_tool.sh
 
 shell: cube.o
 	mkdir -p tables
-	${CC} ${CFLAGS} -o run cube.o shell.c
+	${CC} ${MACROS} ${CFLAGS} -o run cube.o shell.c
 
 debugshell: debugcube.o
 	mkdir -p tables
-	${CC} ${DBGFLAGS} -o run debugcube.o shell.c
+	${CC} ${MACROS} ${DBGFLAGS} -o run debugcube.o shell.c
 
 .PHONY: all clean test tool debugtool shell debugshell
diff --git a/configure.sh b/configure.sh
@@ -1,51 +1,82 @@
 #!/bin/sh
 
-cc=${CC:-cc}
+greparch() {
+	$CC -march=native -dM -E - </dev/null 2>/dev/null | grep "$1"
+}
+
+grepsan() {
+	$CC -fsanitize=$1 -dM -E -x c - </dev/null 2>/dev/null | grep "SANITIZE"
+}
+
+detectthreads() {
+	echo 16 # TODO: choose based on system
+}
 
 detectarch() {
-	${cc} -march=native -dM -E - </dev/null 2>/dev/null | grep "$1"
+	[ -n "$(greparch __AVX2__)" ] && detected="AVX2"
+	[ -n "$(greparch __ARM_NEON)" ] && detected="NEON"
+	[ -z "$detected" ] && detected="PORTABLE"
+
+	echo "$detected"
 }
-detectsan() {
-	${cc} -fsanitize=$1 -dM -E -x c - </dev/null 2>/dev/null \
-		| grep "SANITIZE"
+
+validatecc() {
+	if ! (which "$CC" >/dev/null 2>&1) ; then
+		echo "Error: compiler '$CC' not found"
+		exit 1
+	fi
 }
 
-[ -n "$(detectarch __AVX2__)" ] && detected="AVX2"
-[ -n "$(detectarch __ARM_NEON)" ] && detected="NEON"
-[ -z "$detected" ] && detected="PORTABLE"
+validatethreads() {
+	if [ "$THREADS" -le 0 ] || [ "$THREADS" -gt 128 ]; then
+		echo "Error: number of threads must be between 1 and 128"
+		exit 1
+	fi
+}
+
+validatearch() {
+	case "$ARCH" in
+	AVX2|NEON|PORTABLE)
+		;;
+	*)
+		echo "Error: architecture '$ARCH' not supported"
+		exit 1
+		;;
+	esac
+}
 
-ARCH=${ARCH-"$detected"}
+CC=${CC:-cc}
+THREADS=${THREADS-"$(detectthreads)"}
+ARCH=${ARCH-"$(detectarch)"}
 
-case "$ARCH" in
-AVX2|NEON|PORTABLE)
-	;;
-*)
-	echo "Error: architecture $ARCH not supported"
-	exit 1
-	;;
-esac
+validatecc
+validatethreads
+validatearch
 
 STD="-std=c99"
 WFLAGS="-pedantic -Wall -Wextra"
 WNOFLAGS="-Wno-unused-parameter -Wno-unused-function -Wno-unknown-pragmas"
 
 [ "$ARCH" = "AVX2" ] && AVX="-mavx2"
-[ -n "$(detectsan address)" ] && ADDR="-fsanitize=address"
-[ -n "$(detectsan undefined)" ] && UNDEF="-fsanitize=undefined"
+[ -n "$(grepsan address)" ] && ADDR="-fsanitize=address"
+[ -n "$(grepsan undefined)" ] && UNDEF="-fsanitize=undefined"
 SAN="$ADDR $UNDEF"
 LIBS="-lpthread"
 
 CFLAGS="$STD $LIBS $WFLAGS $WNOFLAGS $AVX -O3"
 DBGFLAGS="$STD $LIBS $WFLAGS $WNOFLAGS $SAN $AVX -g3 -DDEBUG"
+MACROS="-DTHREADS=$THREADS -D$ARCH"
 
 echo "Selected architecture: $ARCH"
-echo "Compiler: ${CC:-cc}"
+echo "Number of threads: $THREADS"
+echo "Compiler: $CC"
 
 {
 echo "ARCH = $ARCH";
 echo "";
 echo "CFLAGS = $CFLAGS";
 echo "DBGFLAGS = $DBGFLAGS";
+echo "MACROS = $MACROS"
 echo "";
-echo "CC = ${cc}"
+echo "CC = $CC"
 } > config.mk
diff --git a/tools/010_stats_tables_h48/stats_tables_h48.c b/tools/010_stats_tables_h48/stats_tables_h48.c
@@ -3,7 +3,6 @@
 #include "../tool.h"
 
 #define MAXMOVES 20
-#define NTHREADS 32
 #define NCUBES_PER_THREAD 10000
 #define LOG_EVERY (NCUBES_PER_THREAD / 10)
 
@@ -57,10 +56,10 @@ run_thread(void *arg)
 void run(void) {
 	int64_t i, j, k, tot;
 	double avg;
-	pthread_t thread[NTHREADS];
-	thread_arg_t arg[NTHREADS];
+	pthread_t thread[THREADS];
+	thread_arg_t arg[THREADS];
 
-	for (i = 0; i < NTHREADS; i++) {
+	for (i = 0; i < THREADS; i++) {
 		arg[i] = (thread_arg_t) {
 			.thread_id = i,
 			.n = NCUBES_PER_THREAD,
@@ -69,18 +68,18 @@ void run(void) {
 		pthread_create(&thread[i], NULL, run_thread, &arg[i]);
 	}
 
-	for (i = 0; i < NTHREADS; i++)
+	for (i = 0; i < THREADS; i++)
 		pthread_join(thread[i], NULL);
 
 	for (j = 0; j < 12; j++) {
 		printf("Data for h=%" PRId64 "\n", j);
 		for (k = 0, avg = 0.0; k <= 16; k++) {
-			for (i = 0, tot = 0; i < NTHREADS; i++)
+			for (i = 0, tot = 0; i < THREADS; i++)
 				tot += arg[i].v[j][k];
 			printf("%" PRId64 "\t%" PRId64 "\n", k, tot);
 			avg += tot * k;
 		}
-		avg /= (double)(NCUBES_PER_THREAD * NTHREADS);
+		avg /= (double)(NCUBES_PER_THREAD * THREADS);
 		printf("Average: %.4lf\n", avg);
 		printf("\n");
 	}