printf update

runtime/include/vx_print.h

@@ -7,10 +7,6 @@
 extern "C" {
 #endif
 
-void vx_prints(const char * str);
-void vx_printx(unsigned value);
-void vx_printv(const char * str, unsigned value);
-
 int vx_vprintf(const char* format, va_list va);
 int vx_printf(const char * format, ...);
 int vx_putchar(int c);

runtime/src/vx_print.c

@@ -1,4 +1,5 @@
 #include <vx_print.h>
+#include <vx_intrinsics.h>
 #include <stdlib.h>
 #include <stdbool.h>
 #include <stdio.h>
@@ -7,159 +8,46 @@
 extern "C" {
 #endif
 
-static const char* skip_flags(const char* format) {
+int __attribute__((noinline)) __vprintf(int index, int tid, const char* format, va_list va) {
-	for (;;) {
+	__if (index == tid) {		
-		int c = *format++;
+		return vprintf(format, va);
-		switch (c) {
+	}__endif
-			case '-':
+	return 0;
-			case '+':
-			case ' ':
-			case '#': break;
-			default : {
-				return --format;
-			}
-		}
-	}
-	return NULL;	
-}
-
-static const char* skip_width(const char* format) {
-	if (*format == '*') {
-		++format;
-	} else {
-		char *endptr;
-		strtol(format, &endptr, 10);
-		format = endptr;
-	}
-	return format;
-}
-
-static const char* skip_precision(const char* format) {
-	if (*format == '.') {
-		++format;
-		if (*format == '*') {
-			++format;
-		} else {
-			char *endptr;
-			strtol(format, &endptr, 10);
-			format = endptr;
-		}
-	}
-	return format;
-}
-
-static const char* skip_modifier(const char* format) {
-	switch (*format) {
-	case 'h':
-		format++;
-		if (*format == 'h') {
-			format++;
-		}
-		break;
-	case 'l':
-		++format;
-		if (*format == 'l') {
-			++format;
-		}
-		break;
-	case 'j':
-	case 'z':
-    case 't':
-	case 'L':
-		++format;
-		break;
-	default:
-		break;
-	}
-	return format;
-}
-
-static const char* parse_format(const char* format, va_list va) {
-	char buffer[64];
-	char fmt[64];
-
-	const char* p = format;
-	p = skip_flags(p);	
-	p = skip_width(p);
-	p = skip_precision(p);
-	p = skip_modifier(p);
-	++p;
-
-	int i;
-
-	fmt[0] = '%';	
-	for (i = 0; i < (p - format); ++i) {
-		fmt[i+1] = format[i];
-	}
-	fmt[i+1] = 0;
-
-	int len = vsnprintf(buffer, 256, fmt, va);
-
-	for (i = 0; i < len; ++i) {
-		vx_putchar(buffer[i]);
-	}
-
-	return p;
 }
 
 int vx_vprintf(const char* format, va_list va) {
-	if (format == NULL)
+	int ret = 0;
-		return -1;
 
-	const char* p = format;
+	// need to execute single-threaded due to potential thread-data dependency
-	int c = *p++;
+	// use manual goto loop to disable compiler optimizations affceting split/join placement
-	while (c) {		
+
-  	if (c == '%') {			
+	volatile int nt = vx_num_threads();	
-			p = parse_format(p, va);
+	int tid = vx_thread_id();
-			c = *p++;	
+
-		} else {
+	for (int i = 0; i < nt; ++i) {
-			vx_putchar(c);
+		ret |= __vprintf(i, tid, format, va);
-			c = *p++;
-		}		
 	}
-
+		
-	return (int)(p - format);
-}
-
-int vx_printf(const char * format, ...) {
-	va_list va;
-  	va_start(va, format);
-	int ret = vx_vprintf(format, va);
-  	va_end(va);
   	return ret;
 }
 
-static const char hextoa[] = "0123456789abcdef";
+int vx_printf(const char * format, ...) {
+	int ret = 0;
 
-void vx_prints(const char * str) {
+	// need to execute single-threaded due to potential thread-data dependency
-	int c = *str++;
+	// use manual goto loop to disable compiler optimizations affceting split/join placement
-	while (c) {
+
-		vx_putchar(c);
+	volatile int nt = vx_num_threads();	
-		c = *str++;
+	int tid = vx_thread_id();
+
+	va_list va;
+	va_start(va, format);
+	for (int i = 0; i < nt; ++i) {
+		ret |= __vprintf(i, tid, format, va);
 	}
-}
+	va_end(va);
-
+		
-void vx_printx(unsigned value) {
+  	return ret;
-	if (value < 16) {
-		vx_putchar(hextoa[value]);
-	} else {
-		int i = 32;
-		bool start = false;
-		do {
-			int temp = (value >> (i - 4)) & 0xf;
-			if (temp != 0) 
-				start = true;
-			if (start) 
-				vx_putchar(hextoa[temp]);
-			i-= 4;
-		} while (i != 0);	
-	}
-	vx_putchar('\n');
-}
-
-void vx_printv(const char * str, unsigned value) {
-	vx_prints(str);
-	vx_printx(value);
 }
 
 #ifdef __cplusplus

tests/runtime/dev/vx_dev_main.c

@@ -79,7 +79,7 @@ int main() {
 	vx_spawn_tasks(arguments.numRows * arguments.numColums, mat_add_kernel, &arguments);
 	vx_print_mat(z, arguments.numRows, arguments.numColums);
 	
-	vx_prints("Passed!\n");
+	vx_printf("Passed!\n");
 
 	return 0;
 }

tests/runtime/nlTest/vx_nl_main.c

@@ -16,7 +16,7 @@ int main() {
 	float isq = 1.0f / sqrt(fNum);
 	vx_printf("fibonacci(%d) = %d\n", Num, fib);
 	vx_printf("invAqrt(%f) = %f\n", fNum, isq);
-	vx_prints("Passed!\n");
+	vx_printf("Passed!\n");
 	return 0;
 }
 

tests/runtime/simple/main.c

@@ -105,8 +105,8 @@ int main() {
 			unsigned index = (i * arguments.numColums) + j;
 			vx_printf("0x%x ", z[index]);
 		}
-		vx_prints("\n");
+		vx_printf("\n");
 	}
-	vx_prints("Passed!\n");
+	vx_printf("Passed!\n");
 	return 0;
 }

tests/runtime/simple/tests.c

@@ -17,10 +17,10 @@ void test_tmc() {
 	test_tmc_impl();
 	vx_tmc(1);
 
-	vx_printx(tmc_array[0]);
+	vx_printf("%x", tmc_array[0]);
-	vx_printx(tmc_array[1]);
+	vx_printf("%x", tmc_array[1]);
-	vx_printx(tmc_array[2]);
+	vx_printf("%x", tmc_array[2]);
-	vx_printx(tmc_array[3]);
+	vx_printf("%x", tmc_array[3]);
 
 	return;
 }
@@ -57,10 +57,10 @@ void test_divergence() {
 
 	vx_tmc(1);
 
-	vx_printx(div_arr[0]);
+	vx_printf("%x", div_arr[0]);
-	vx_printx(div_arr[1]);
+	vx_printf("%x", div_arr[1]);
-	vx_printx(div_arr[2]);
+	vx_printf("%x", div_arr[2]);
-	vx_printx(div_arr[3]);
+	vx_printf("%x", div_arr[3]);
 }
 
 unsigned wsapwn_arr[4];
@@ -76,8 +76,8 @@ void simple_kernel() {
 void test_wsapwn() {
 	vx_wspawn(4, (unsigned)simple_kernel);
 	simple_kernel();
-	vx_printx(wsapwn_arr[0]);
+	vx_printf("%x", wsapwn_arr[0]);
-	vx_printx(wsapwn_arr[1]);
+	vx_printf("%x", wsapwn_arr[1]);
-	vx_printx(wsapwn_arr[2]);
+	vx_printf("%x", wsapwn_arr[2]);
-	vx_printx(wsapwn_arr[3]);
+	vx_printf("%x", wsapwn_arr[3]);
 }