Prepare profiling macros

4 files changed, 229 insertions, 49 deletions

diff --git a/src/gba/gba.h b/src/gba/gba.h
index b02d745..27a6a9a 100644
--- a/src/gba/gba.h
+++ b/src/gba/gba.h
@@ -98,18 +98,18 @@ typedef u16 Color;
 typedef Color Palette[16];
 
 // Inline function to calculate the 15 bit color value.
-#define RGB15(R,G,B) (u16)(((B) << 10) | ((G) << 5) | (R));
+#define RGB15(R,G,B) (u16)(((B) << 10) | ((G) << 5) | (R))
 
 // Some nice default colors.
-#define COLOR_BLACK  RGB15(0,  0, 0)
-#define COLOR_RED    RGB15(31, 0,10)
-#define COLOR_GREEN  RGB15(31, 0,10)
-#define COLOR_YELLOW RGB15(31, 0,10)
-#define COLOR_BLUE   RGB15(2, 17,31)
-#define COLOR_PURPLE RGB15(2, 17,31)
-#define COLOR_CYAN   RGB15(0, 27,30)
-#define COLOR_GREY   RGB15(16,17,19)
-#define COLOR_WHITE  RGB15(28,28,28)
+#define COLOR_BLACK  RGB15( 0,  0,  0)
+#define COLOR_RED    RGB15(31,  0, 10)
+#define COLOR_GREEN  RGB15( 0, 31, 18)
+#define COLOR_YELLOW RGB15(31, 31,  0)
+#define COLOR_BLUE   RGB15( 2, 17, 31)
+#define COLOR_PURPLE RGB15(15,  7, 31)
+#define COLOR_CYAN   RGB15( 0, 27, 30)
+#define COLOR_GREY   RGB15(16, 17, 19)
+#define COLOR_WHITE  RGB15(28, 28, 28)
 
 //
 // Tile memory access.
@@ -419,6 +419,8 @@ inline
 void
 dma_copy(void *dst, const void *src, u32 size, int channel) {
     dma_transfer_copy(dst, src, size / 4, channel, DMA_CHUNK_32 | DMA_ENABLE);
+    // Stall for 2 cycles in case we call this function more than once.
+    asm("nop"); asm("nop");
 }
 
 // Fill the dst location with the word set at src.
@@ -426,6 +428,8 @@ inline
 void
 dma_fill(void *dst, vu32 src, u32 size, int channel) {
     dma_transfer_fill(dst, src, size / 4, channel, DMA_CHUNK_32 | DMA_ENABLE);
+    // Stall for 2 cycles in case we call this function more than once.
+    asm("nop"); asm("nop");
 }
 
 //
@@ -661,6 +665,7 @@ wait_vsync(void) {
 #define LEN(ARR) (sizeof(ARR) / sizeof((ARR)[0]))
 
 // Fixed-point arithmetic for (i.P) numbers.
+#define FP_NUM(A,P)        ((A) << (P))
 #define FP_MUL(A,B,P)      (((A) * (B)) >> (P))
 #define FP_DIV(A,B,P)      (((A) << (P)) / (B))
 #define FP_LERP(Y0,Y1,X,P) ((Y0) + FP_MUL((X), ((Y1) - (Y0)), P))
@@ -683,4 +688,10 @@ memcpy32(u32 *dst, const u32 *src, u32 size) {
     }
 }
 
+//
+// Compiler hints.
+//
+
+#define UNROLL_LOOPS __attribute__((optimize("unroll-loops")))
+
 #endif // GBA_H
diff --git a/src/main.c b/src/main.c
index a322df2..e694057 100644
--- a/src/main.c
+++ b/src/main.c
@@ -15,45 +15,13 @@ WITH REGARD TO THIS SOFTWARE.
 #include "renderer.c"
 #include "sequencer.c"
 
+#define PROF_ENABLE 0
+#include "profiling.c"
+
 //
 // Config parameters.
 //
 
-#ifdef PROF_ENABLE
-#if PROF_ENABLE == 0
-#define PROF(F,VAR) (profile_start(),(F),(VAR) = profile_stop())
-#elif PROF_ENABLE == 1
-#define PROF(F,VAR) (profile_start(),(F),(VAR) = MAX(profile_stop(), (VAR)))
-#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("EVAL:  %lu    ", eval_cycles);\
-        txt_position((PROF_SHOW_X), (PROF_SHOW_Y)+1);\
-        txt_printf("FLIP:  %lu    ", flip_cycles);\
-        txt_position((PROF_SHOW_X), (PROF_SHOW_Y)+2);\
-        txt_printf("INPUT: %lu    ", input_cycles);\
-        txt_position((PROF_SHOW_X), (PROF_SHOW_Y)+3);\
-        txt_printf("FRAME: %lu    ", frame_counter);\
-        frame_counter++;\
-    } while (0)
-#define PROF_INIT() \
-    u32 frame_counter = 0;\
-    u32 input_cycles = 0;\
-    u32 eval_cycles = 0;\
-    u32 flip_cycles = 0;
-#else
-#define PROF(F,VAR) (F)
-#define PROF_SHOW()
-#define PROF_INIT()
-#endif
-
 int main(void) {
     // Adjust system wait times.
     SYSTEM_WAIT = SYSTEM_WAIT_CARTRIDGE;
@@ -72,12 +40,14 @@ int main(void) {
     sequencer_init();
 
     // Main loop.
-    PROF_INIT();
     while (true) {
+        poll_keys();
         bios_vblank_wait();
-        PROF(flip_buffer(), flip_cycles);
+        FRAME_START();
         PROF(handle_sequencer_input(), input_cycles);
         PROF_SHOW();
+        PROF(flip_buffer(), flip_cycles);
+        FRAME_END();
     }
 
     return 0;
diff --git a/src/profiling.c b/src/profiling.c
new file mode 100644
index 0000000..de969d2
--- /dev/null
+++ b/src/profiling.c
@@ -0,0 +1,201 @@
+//
+// Profiling macros.
+//
+
+#ifndef PROF_ENABLE
+#define PROF_ENABLE 0
+#endif
+
+#if PROF_ENABLE > 0 && PROF_ENABLE < 3
+
+#ifndef PROF_N_FRAMES
+#define PROF_N_FRAMES 30
+#endif
+
+// Profile method 1: Average per N frames.
+#if PROF_ENABLE == 1
+#define TEXT_ENABLE 1
+#define PROF(F,VAR) \
+    do { \
+        u32 __tmp_prof = profile_measure();\
+        F;\
+        (VAR) += profile_measure() - __tmp_prof;\
+    } while (0)
+
+// Profile method 2: Maximum in N frames.
+#elif PROF_ENABLE == 2
+#define TEXT_ENABLE 1
+#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
+
+static bool profile_show = true;
+static bool profile_bg_show = true;
+
+#define PROF_SHOW() \
+    do { \
+        if (key_tap(KEY_START)) {\
+            profile_show ^= 1;\
+        }\
+        if (key_tap(KEY_SELECT)) {\
+            profile_bg_show ^= 1;\
+        }\
+        if (profile_show) {\
+            txt_position((PROF_SHOW_X), (PROF_SHOW_Y));\
+            draw_filled_rect((PROF_SHOW_X), (PROF_SHOW_X), 8 * 18, 8 * 16, 0);\
+            txt_printf("VIDEO\n");\
+            txt_printf(">CLEAR  %.8lu\n", avg_clear_cycles);\
+            txt_printf(">LINES  %.8lu\n", avg_line_cycles);\
+            txt_printf(">RECT   %.8lu\n", avg_rect_cycles);\
+            txt_printf(">FRECT  %.8lu\n", avg_fill_rect_cycles);\
+            txt_printf(">1BPP   %.8lu\n", avg_icn_cycles);\
+            txt_printf(">2BPP   %.8lu\n", avg_chr_cycles);\
+            txt_printf(">FLIP   %.8lu\n", avg_flip_cycles);\
+            txt_printf("TEXT\n");\
+            txt_printf(">DRAWF  %.8lu\n", avg_txt_drawf_cycles);\
+            txt_printf(">PRINTF %.8lu\n", avg_txt_printf_cycles);\
+            txt_printf(">RENDER %.8lu\n", avg_txt_render_cycles);\
+            txt_printf(">CLEAR  %.8lu\n", avg_txt_clear_cycles);\
+            txt_printf("TOTAL   %.8lu\n", avg_frame_cycles);\
+        }\
+        if (profile_bg_show) {\
+            u32 frame_time =\
+                FP_DIV(\
+                    FP_NUM(avg_frame_cycles + 1, 2),\
+                    FP_NUM(2809, 2),\
+                    2) * 166;\
+            u32 fps =\
+                FP_DIV(\
+                    FP_NUM(280896 * 60, 2),\
+                    FP_NUM(avg_frame_cycles + 1, 2),\
+                    2);\
+            txt_printf("TIME    %.8lu\n", frame_time >> 2);\
+            txt_printf("FPS     %.8lu\n", (fps >> 2) + 1);\
+        }\
+    } while (0)
+
+static u32 prof_frame_counter = 0;
+
+static u32 frame_cycles      = 0;
+static u32 flip_cycles       = 0;
+static u32 clear_cycles      = 0;
+static u32 line_cycles       = 0;
+static u32 rect_cycles       = 0;
+static u32 fill_rect_cycles  = 0;
+static u32 chr_cycles        = 0;
+static u32 icn_cycles        = 0;
+static u32 txt_drawf_cycles  = 0;
+static u32 txt_printf_cycles = 0;
+static u32 txt_render_cycles = 0;
+static u32 txt_clear_cycles  = 0;
+static u32 input_cycles      = 0;
+
+static u32 avg_frame_cycles      = 0;
+static u32 avg_flip_cycles       = 0;
+static u32 avg_clear_cycles      = 0;
+static u32 avg_line_cycles       = 0;
+static u32 avg_rect_cycles       = 0;
+static u32 avg_fill_rect_cycles  = 0;
+static u32 avg_chr_cycles        = 0;
+static u32 avg_icn_cycles        = 0;
+static u32 avg_txt_drawf_cycles  = 0;
+static u32 avg_txt_printf_cycles = 0;
+static u32 avg_txt_render_cycles = 0;
+static u32 avg_txt_clear_cycles  = 0;
+static u32 avg_input_cycles      = 0;
+
+#if PROF_ENABLE == 1
+#define FRAME_START()\
+    do { \
+        if (prof_frame_counter == PROF_N_FRAMES) {\
+            avg_frame_cycles      = frame_cycles      / prof_frame_counter;\
+            avg_flip_cycles       = flip_cycles       / prof_frame_counter;\
+            avg_clear_cycles      = clear_cycles      / prof_frame_counter;\
+            avg_line_cycles       = line_cycles       / prof_frame_counter;\
+            avg_rect_cycles       = rect_cycles       / prof_frame_counter;\
+            avg_fill_rect_cycles  = fill_rect_cycles  / prof_frame_counter;\
+            avg_chr_cycles        = chr_cycles        / prof_frame_counter;\
+            avg_icn_cycles        = icn_cycles        / prof_frame_counter;\
+            avg_txt_drawf_cycles  = txt_drawf_cycles  / prof_frame_counter;\
+            avg_txt_printf_cycles = txt_printf_cycles / prof_frame_counter;\
+            avg_txt_render_cycles = txt_render_cycles / prof_frame_counter;\
+            avg_txt_clear_cycles  = txt_clear_cycles  / prof_frame_counter;\
+            avg_input_cycles      = input_cycles  / prof_frame_counter;\
+            frame_cycles          = 0;\
+            flip_cycles           = 0;\
+            clear_cycles          = 0;\
+            line_cycles           = 0;\
+            rect_cycles           = 0;\
+            fill_rect_cycles      = 0;\
+            chr_cycles            = 0;\
+            icn_cycles            = 0;\
+            txt_drawf_cycles      = 0;\
+            txt_printf_cycles     = 0;\
+            txt_render_cycles     = 0;\
+            txt_clear_cycles      = 0;\
+            input_cycles          = 0;\
+            prof_frame_counter    = 0;\
+        }\
+        profile_start();\
+    } while (0)
+#elif PROF_ENABLE == 2
+#define FRAME_START()\
+    do { \
+        if (prof_frame_counter == PROF_N_FRAMES) {\
+            avg_frame_cycles      = frame_cycles;\
+            avg_flip_cycles       = flip_cycles;\
+            avg_clear_cycles      = clear_cycles;\
+            avg_line_cycles       = line_cycles;\
+            avg_rect_cycles       = rect_cycles;\
+            avg_fill_rect_cycles  = fill_rect_cycles;\
+            avg_chr_cycles        = chr_cycles;\
+            avg_icn_cycles        = icn_cycles;\
+            avg_txt_drawf_cycles  = txt_drawf_cycles;\
+            avg_txt_printf_cycles = txt_printf_cycles;\
+            avg_txt_render_cycles = txt_render_cycles;\
+            avg_txt_clear_cycles  = txt_clear_cycles;\
+            avg_input_cycles      = input_cycles;\
+            frame_cycles          = 0;\
+            flip_cycles           = 0;\
+            clear_cycles          = 0;\
+            line_cycles           = 0;\
+            rect_cycles           = 0;\
+            fill_rect_cycles      = 0;\
+            chr_cycles            = 0;\
+            icn_cycles            = 0;\
+            txt_drawf_cycles      = 0;\
+            txt_printf_cycles     = 0;\
+            txt_render_cycles     = 0;\
+            txt_clear_cycles      = 0;\
+            input_cycles          = 0;\
+            prof_frame_counter    = 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
diff --git a/src/sequencer.c b/src/sequencer.c
index 2d2e2d3..b582cf4 100644
--- a/src/sequencer.c
+++ b/src/sequencer.c
@@ -2921,8 +2921,6 @@ clipboard_copy(void) {
 
 void
 handle_sequencer_input(void) {
-    poll_keys();
-
     if (key_tap(KEY_START)) {
         // Stop the sequencer or start playing from the beginning.
         toggle_playing();