From 95432d94ee6d5cd0c1f71c36b78e195f1b96892f Mon Sep 17 00:00:00 2001
From: John Tsiombikas <nuclear@member.fsf.org>
Date: Sun, 11 Feb 2018 10:05:49 +0200
Subject: [PATCH] added perf_start/perf_end assembly macros

---
 GNUmakefile   |    6 +++---
 Makefile.dj   |    1 +
 src/gfxutil.c |    7 ++++---
 src/greets.c  |    8 +++++++-
 src/util.h    |   51 +++++++++++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 66 insertions(+), 7 deletions(-)

diff --git a/GNUmakefile b/GNUmakefile
index d6c67d1..830297d 100644
--- a/GNUmakefile
+++ b/GNUmakefile
@@ -1,5 +1,5 @@
 src = $(wildcard src/*.c) $(wildcard src/sdl/*.c)
-obj = $(src:.c=.o)
+obj = $(src:.c=.o) $(asmsrc:.asm=.o)
 dep = $(obj:.o=.d)
 bin = demo
 
@@ -29,8 +29,8 @@ mikmod:
 
 .PHONY: cleanlibs
 cleanlibs:
-	$(MAKE) -C libs/imago -f Makefile.dj clean
-	$(MAKE) -C libs/oldmik -f Makefile.dj clean
+	$(MAKE) -C libs/imago clean
+	$(MAKE) -C libs/mikmod clean
 
 .PHONY: clean
 clean:
diff --git a/Makefile.dj b/Makefile.dj
index 0f152b7..764ad9d 100644
--- a/Makefile.dj
+++ b/Makefile.dj
@@ -1,4 +1,5 @@
 src = $(wildcard src/*.c) $(wildcard src/dos/*.c)
+asmsrc = $(wildcard src/*.asm)
 obj = $(src:.c=.cof)
 dep = $(obj:.cof=.dep)
 bin = demo.exe
diff --git a/src/gfxutil.c b/src/gfxutil.c
index cb51780..b82a612 100644
--- a/src/gfxutil.c
+++ b/src/gfxutil.c
@@ -149,20 +149,21 @@ void draw_line(int x0, int y0, int x1, int y1, unsigned short color)
 		int sum = sptr[0] * (rad + 1); \
 		int count = (rad * 2 + 1) << 8; \
 		int midsize = w - rad * 2; \
+		int firstpix = sptr[0]; \
 		int lastpix = sptr[pstep * (w - 1)]; \
 		/* add up the contributions for the -1 pixel */ \
 		for(j=0; j<rad; j++) { \
 			sum += sptr[pstep * j]; \
 		} \
 		/* first part adding sptr[rad] and subtracting sptr[0] */ \
-		for(j=0; j<rad + 1; j++) { \
-			sum += (int)sptr[pstep * rad] - (int)sptr[0]; \
+		for(j=0; j<=rad; j++) { \
+			sum += (int)sptr[pstep * rad] - firstpix; \
 			sptr += pstep; \
 			*dptr = scale * sum / count; \
 			dptr += pstep; \
 		} \
 		/* middle part adding sptr[rad] and subtracting sptr[-(rad+1)] */ \
-		for(j=0; j<midsize - 1; j++) { \
+		for(j=1; j<midsize; j++) { \
 			sum += (int)sptr[pstep * rad] - (int)sptr[-(rad + 1) * pstep]; \
 			sptr += pstep; \
 			*dptr = scale * sum / count; \
diff --git a/src/greets.c b/src/greets.c
index 0ad90a1..57f85e8 100644
--- a/src/greets.c
+++ b/src/greets.c
@@ -25,6 +25,8 @@ void wait_vsync(void);
 
 #define RAND_FIELD_MAX	0.7
 
+#define BLUR_RAD	5
+
 #define PCOUNT		4000
 #define MAX_LIFE	7.0f
 #define PALPHA		1.0f
@@ -231,8 +233,12 @@ static void draw(void)
 		}
 	}
 
-#define BLUR_RAD	5
+	/*perf_start();*/
 	blur_grey_horiz(prev_smokebuf, cur_smokebuf, fb_width, fb_height, BLUR_RAD, 240);
+	/*
+	perf_end();
+	printf("blur perf: %lu\n", (unsigned long)perf_interval_count);
+	*/
 	blur_grey_vert(cur_smokebuf, prev_smokebuf, fb_width, fb_height, BLUR_RAD, 240);
 	swap_smoke_buffers();
 
diff --git a/src/util.h b/src/util.h
index eb3773d..f8ae175 100644
--- a/src/util.h
+++ b/src/util.h
@@ -24,4 +24,55 @@ static INLINE int32_t cround64(double val)
 	return *(int32_t*)&val;
 }
 
+uint32_t perf_start_count, perf_interval_count;
+
+#ifdef __WATCOMC__
+void perf_start(void);
+#pragma aux perf_start = \
+	"rdtsc" \
+	"mov [perf_start_count], eax" \
+	modify[eax edx];
+
+void perf_end(void);
+#pragma aux perf_end = \
+	"rdtsc" \
+	"sub eax, [perf_start_count]" \
+	"mov [perf_interval_count], eax" \
+	modify [eax edx];
+#endif
+
+#ifdef __GNUC__
+#define perf_start()  asm volatile ( \
+	"rdtsc\n" \
+	"mov %%eax, %0\n" \
+	: "=m"(perf_start_count) :: "%eax", "%edx")
+
+#define perf_end() asm volatile ( \
+	"rdtsc\n" \
+	"sub %1, %%eax\n" \
+	"mov %%eax, %0\n" \
+	: "=m"(perf_interval_count) \
+	: "m"(perf_start_count) \
+	: "%eax", "%edx")
+#endif
+
+#ifdef _MSC_VER
+#define perf_start() \
+	do { \
+		__asm { \
+			rdtsc \
+			mov [perf_start_count], eax \
+		} \
+	} while(0)
+
+#define perf_end() \
+	do { \
+		__asm { \
+			rdtsc \
+			sub eax, [perf_start_count] \
+			mov [perf_interval_count], eax \
+		} \
+	} while(0)
+#endif
+
 #endif	/* UTIL_H_ */
-- 
1.7.10.4