Add an improved profiling method for more granularity

1 files changed, 49 insertions, 17 deletions

diff --git a/src/main.c b/src/main.c
index 86a9376..b6a979f 100644
--- a/src/main.c
+++ b/src/main.c
@@ -37,7 +37,7 @@
 #define CONTROL_METHODS CONTROL_CONTROLLER,CONTROL_MOUSE,CONTROL_KEYBOARD
 #endif
 
-#define PROF_ENABLE 1
+#define PROF_ENABLE 2
 
 #ifdef PROF_ENABLE
 #if PROF_ENABLE == 0
@@ -47,6 +47,12 @@
 #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;\
+    } while (0)
 #define TEXT_ENABLE 1
 // TODO: allow measuring inside a PROF function using profile_measure to store
 // temporary values and calcualting the difference.
@@ -62,23 +68,35 @@
 #define PROF_SHOW() \
     do { \
         txt_position((PROF_SHOW_X), (PROF_SHOW_Y));\
-        txt_printf("FLIP: %.8lu\n", flip_cycles);\
-        txt_printf("EVAL: %.8lu\n", eval_cycles);\
-        screen_fill(BG_BACK, 0, 0, 15 * 8, 2 * 8, 0);\
+        txt_printf("INPUT  %.8lu\n", input_cycles);\
+        txt_printf("EVAL   %.8lu\n", 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);\
+        screen_fill(BG_BACK, 0, 0, 8 * 16, 8 * 10, 0);\
         flipbuf();\
     } while (0)
 
-
-// static u32 ppu_pixel_cycles = 0;
-// static u32 ppu_fill_cycles = 0;
-// static u32 ppu_chr_cycles = 0;
-// static u32 ppu_icn_cycles = 0;
+static u32 frame_cycles = 0;
+static u32 ppu_pixel_cycles = 0;
+static u32 ppu_fill_cycles = 0;
+static u32 ppu_chr_cycles = 0;
+static u32 ppu_icn_cycles = 0;
 static u32 flip_cycles = 0;
 static u32 eval_cycles = 0;
 static u32 input_cycles = 0;
 static u32 mix_cycles = 0;
-#define FRAME_START() ; // TODO
-#define FRAME_END()   ; // TODO
+#define FRAME_START()\
+    do { \
+        frame_cycles = 0;\
+        profile_start();\
+    } while (0)
+#define FRAME_END() (frame_cycles = profile_stop())
 #else
 #define PROF(F,VAR) (F)
 #define PROF_SHOW()
@@ -147,9 +165,9 @@ screen_deo(u8 *ram, u8 *d, u8 port) {
                 u16 y1 = SCREEN_HEIGHT - 1;
                 if(ctrl & 0x10) x1 = x0, x0 = 0;
                 if(ctrl & 0x20) y1 = y0, y0 = 0;
-                screen_fill(layer, x0, y0, x1, y1, color);
+                PROF(screen_fill(layer, x0, y0, x1, y1, color), ppu_fill_cycles);
             } else {
-                ppu_pixel(layer, x0, y0, color);
+                PROF(ppu_pixel(layer, x0, y0, color), ppu_pixel_cycles);
                 if(d[0x6] & 0x1) POKE2(d + 0x8, x0 + 1); /* auto x+1 */
                 if(d[0x6] & 0x2) POKE2(d + 0xa, y0 + 1); /* auto y+1 */
             }
@@ -174,9 +192,19 @@ screen_deo(u8 *ram, u8 *d, u8 port) {
             for(size_t i = 0; i <= n; i++) {
                 u8 *sprite = &ram[addr];
                 if (twobpp) {
-                    ppu_2bpp(layer, x + dy * i, y + dx * i, sprite, color, flipx, flipy);
+                    PROF(ppu_2bpp(layer,
+                                x + dy * i,
+                                y + dx * i,
+                                sprite,
+                                color,
+                                flipx, flipy), ppu_chr_cycles);
                 } else {
-                    ppu_1bpp(layer, x + dy * i, y + dx * i, sprite, color, flipx, flipy);
+                    PROF(ppu_1bpp(layer,
+                                x + dy * i,
+                                y + dx * i,
+                                sprite,
+                                color,
+                                flipx, flipy), ppu_icn_cycles);
                 }
                 addr += (d[0x6] & 0x04) << (1 + twobpp);
             }
@@ -615,20 +643,24 @@ main(void) {
     // Main loop.
     uxn_eval(&u, PAGE_PROGRAM);
     u8 frame_counter = 0;
+    // NOTE: A VBLANK is 83776 cycles, anything other than that will make it so
+    // we fail to render at 60FPS.
     while(true) {
         FRAME_START();
+        bios_vblank_wait();
         PROF(handle_input(&u), input_cycles);
         PROF(uxn_eval(&u, PEEK2(&u.dev[0x20])), eval_cycles);
         PROF(sound_mix(), mix_cycles);
-        bios_vblank_wait();
+        // TODO: allow configuration to do VSYNC at 15 or 30 fps to avoid too
+        // much memory copying on demanding uxn roms.
         PROF(flipbuf(), flip_cycles);
         frame_counter++;
         if (frame_counter == 60) {
             seconds++;
             frame_counter = 0;
         }
-        PROF_SHOW();
         FRAME_END();
+        PROF_SHOW();
     }
 
     return 0;