Add improved avg profiling method

1 files changed, 104 insertions, 25 deletions

diff --git a/src/main.c b/src/main.c
index b6a979f..b1562ca 100644
--- a/src/main.c
+++ b/src/main.c
@@ -39,49 +39,59 @@
 
 #define PROF_ENABLE 2
 
-#ifdef PROF_ENABLE
-#if PROF_ENABLE == 0
-#define TEXT_ENABLE 1
-#define PROF(F,VAR) (profile_start(),(F),(VAR) = profile_stop())
-#elif PROF_ENABLE == 1
+#if PROF_ENABLE > 0 && PROF_ENABLE < 3
+
+#ifndef PROF_N_FRAMES
+#define PROF_N_FRAMES 15
+#endif
+
+#if PROF_ENABLE == 1
+// Profile method 1: Average per N frames.
 #define TEXT_ENABLE 1
-#define PROF(F,VAR) (profile_start(),(F),(VAR) = MAX(profile_stop(), (VAR)))
-#elif PROF_ENABLE == 2
 #define PROF(F,VAR) \
     do { \
         u32 __tmp_prof = profile_measure();\
         (F);\
-        (VAR) = profile_measure() - __tmp_prof;\
+        (VAR) += profile_measure() - __tmp_prof;\
     } while (0)
+
+#elif PROF_ENABLE == 2
+// Profile method 2: Maximum in N frames.
 #define TEXT_ENABLE 1
-// TODO: allow measuring inside a PROF function using profile_measure to store
-// temporary values and calcualting the difference.
-// TODO: allow per frame or per 60 second averaging of cycle measurement.
-// TODO: calculate fps based on frame timing.
+#define PROF(F,VAR) \
+    do { \
+        u32 __tmp_prof = profile_measure();\
+        (F);\
+        (VAR) = MAX(profile_measure() - __tmp_prof, (VAR));\
+    } while (0)
 #endif
+
 #ifndef PROF_SHOW_X
 #define PROF_SHOW_X 0
 #endif
 #ifndef PROF_SHOW_Y
 #define PROF_SHOW_Y 0
 #endif
+
 #define PROF_SHOW() \
     do { \
         txt_position((PROF_SHOW_X), (PROF_SHOW_Y));\
-        txt_printf("INPUT  %.8lu\n", input_cycles);\
-        txt_printf("EVAL   %.8lu\n", eval_cycles);\
+        txt_printf("INPUT  %.8lu\n", avg_input_cycles);\
+        txt_printf("EVAL   %.8lu\n", avg_eval_cycles);\
         txt_printf("VIDEO\n");\
-        txt_printf(">PIX   %.8lu\n", ppu_pixel_cycles);\
-        txt_printf(">FILL  %.8lu\n", ppu_fill_cycles);\
-        txt_printf(">1BPP  %.8lu\n", ppu_icn_cycles);\
-        txt_printf(">2BPP  %.8lu\n", ppu_chr_cycles);\
-        txt_printf(">FLIP  %.8lu\n", flip_cycles);\
-        txt_printf("AUDIO  %.8lu\n", mix_cycles);\
-        txt_printf("FRAME  %.8lu\n", frame_cycles);\
+        txt_printf(">PIX   %.8lu\n", avg_ppu_pixel_cycles);\
+        txt_printf(">FILL  %.8lu\n", avg_ppu_fill_cycles);\
+        txt_printf(">1BPP  %.8lu\n", avg_ppu_icn_cycles);\
+        txt_printf(">2BPP  %.8lu\n", avg_ppu_chr_cycles);\
+        txt_printf(">FLIP  %.8lu\n", avg_flip_cycles);\
+        txt_printf("AUDIO  %.8lu\n", avg_mix_cycles);\
+        txt_printf("TOTAL  %.8lu\n", avg_frame_cycles);\
         screen_fill(BG_BACK, 0, 0, 8 * 16, 8 * 10, 0);\
         flipbuf();\
     } while (0)
 
+static u32 prof_frame_counter = 0;
+
 static u32 frame_cycles = 0;
 static u32 ppu_pixel_cycles = 0;
 static u32 ppu_fill_cycles = 0;
@@ -91,15 +101,84 @@ static u32 flip_cycles = 0;
 static u32 eval_cycles = 0;
 static u32 input_cycles = 0;
 static u32 mix_cycles = 0;
+
+static u32 avg_ppu_pixel_cycles = 0;
+static u32 avg_ppu_fill_cycles = 0;
+static u32 avg_ppu_chr_cycles = 0;
+static u32 avg_ppu_icn_cycles = 0;
+static u32 avg_flip_cycles = 0;
+static u32 avg_eval_cycles = 0;
+static u32 avg_input_cycles = 0;
+static u32 avg_mix_cycles = 0;
+static u32 avg_frame_cycles = 0;
+
+#if PROF_ENABLE == 1
 #define FRAME_START()\
     do { \
-        frame_cycles = 0;\
+        if (prof_frame_counter == PROF_N_FRAMES) {\
+            avg_ppu_pixel_cycles = ppu_pixel_cycles / prof_frame_counter;\
+            avg_ppu_fill_cycles  = ppu_fill_cycles / prof_frame_counter;\
+            avg_ppu_chr_cycles   = ppu_chr_cycles / prof_frame_counter;\
+            avg_ppu_icn_cycles   = ppu_icn_cycles / prof_frame_counter;\
+            avg_flip_cycles      = flip_cycles / prof_frame_counter;\
+            avg_eval_cycles      = eval_cycles / prof_frame_counter;\
+            avg_input_cycles     = input_cycles / prof_frame_counter;\
+            avg_mix_cycles       = mix_cycles / prof_frame_counter;\
+            avg_frame_cycles     = frame_cycles / prof_frame_counter;\
+            prof_frame_counter   = 0;\
+            frame_cycles         = 0;\
+            ppu_pixel_cycles     = 0;\
+            ppu_fill_cycles      = 0;\
+            ppu_chr_cycles       = 0;\
+            ppu_icn_cycles       = 0;\
+            flip_cycles          = 0;\
+            eval_cycles          = 0;\
+            input_cycles         = 0;\
+            mix_cycles           = 0;\
+        }\
         profile_start();\
     } while (0)
-#define FRAME_END() (frame_cycles = profile_stop())
+#elif PROF_ENABLE == 2
+#define FRAME_START()\
+    do { \
+        if (prof_frame_counter == PROF_N_FRAMES) {\
+            avg_ppu_pixel_cycles = ppu_pixel_cycles;\
+            avg_ppu_fill_cycles  = ppu_fill_cycles;\
+            avg_ppu_chr_cycles   = ppu_chr_cycles;\
+            avg_ppu_icn_cycles   = ppu_icn_cycles;\
+            avg_flip_cycles      = flip_cycles;\
+            avg_eval_cycles      = eval_cycles;\
+            avg_input_cycles     = input_cycles;\
+            avg_mix_cycles       = mix_cycles;\
+            avg_frame_cycles     = frame_cycles / prof_frame_counter;\
+            prof_frame_counter   = 0;\
+            frame_cycles         = 0;\
+            ppu_pixel_cycles     = 0;\
+            ppu_fill_cycles      = 0;\
+            ppu_chr_cycles       = 0;\
+            ppu_icn_cycles       = 0;\
+            flip_cycles          = 0;\
+            eval_cycles          = 0;\
+            input_cycles         = 0;\
+            mix_cycles           = 0;\
+        }\
+        profile_start();\
+    } while (0)
+#endif
+
+#define FRAME_END() \
+    do { \
+        prof_frame_counter++;\
+        frame_cycles += profile_stop();\
+    } while (0)
+
 #else
+
+// No profiling.
 #define PROF(F,VAR) (F)
 #define PROF_SHOW()
+#define FRAME_START()
+#define FRAME_END()
 #endif
 
 static time_t seconds = 0;
@@ -646,8 +725,9 @@ main(void) {
     // NOTE: A VBLANK is 83776 cycles, anything other than that will make it so
     // we fail to render at 60FPS.
     while(true) {
-        FRAME_START();
+        PROF_SHOW();
         bios_vblank_wait();
+        FRAME_START();
         PROF(handle_input(&u), input_cycles);
         PROF(uxn_eval(&u, PEEK2(&u.dev[0x20])), eval_cycles);
         PROF(sound_mix(), mix_cycles);
@@ -660,7 +740,6 @@ main(void) {
             frame_counter = 0;
         }
         FRAME_END();
-        PROF_SHOW();
     }
 
     return 0;