Add new profiling macros

4 files changed, 184 insertions, 50 deletions

diff --git a/src/gba/gba.h b/src/gba/gba.h
index 99eb1f7..27a6a9a 100644
--- a/src/gba/gba.h
+++ b/src/gba/gba.h
@@ -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");
 }
 
 //
diff --git a/src/main.c b/src/main.c
index 993ca37..14e04e9 100644
--- a/src/main.c
+++ b/src/main.c
@@ -11,45 +11,14 @@ WITH REGARD TO THIS SOFTWARE.
 
 #include "gba/gba.h"
 
-#include "renderer_m0.c"
+#include "renderer_m4.c"
 
 //
 // Config parameters.
 //
 
-#if PROF_ENABLE == 0
-#define PROF(F,VAR) (profile_start(),(F),(VAR) = profile_stop())
-#else
-#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_printf("RECT:  %.9lu\n", test_rect_cycles);\
-        txt_printf("FRECT: %.9lu\n", test_fill_rect_cycles);\
-        txt_printf("CLEAR: %.9lu\n", test_clear_cycles);\
-        txt_printf("FLIP:  %.9lu\n", flip_cycles);\
-        txt_printf("CHR:   %.9lu\n", test_chr_cycles);\
-        txt_printf("ICN:   %.9lu\n", test_icn_cycles);\
-        txt_printf("LINE:  %.9lu\n", test_lines_cycles);\
-        txt_render();\
-        txt_clear();\
-        frame_counter++;\
-    } while (0)
-#define PROF_INIT() \
-    u32 frame_counter = 0;\
-    u32 test_fill_rect_cycles = 0;\
-    u32 test_rect_cycles = 0;\
-    u32 test_clear_cycles = 0;\
-    u32 test_chr_cycles = 0;\
-    u32 test_icn_cycles = 0;\
-    u32 test_lines_cycles = 0;\
-    u32 flip_cycles = 0;
+#define PROF_ENABLE 1
+#include "profiling.c"
 
 void
 test_clear(void) {
@@ -116,8 +85,15 @@ test_moving_line(void) {
     int inc_x = 1;
     int inc_y = 0;
     while (true) {
-        screen_fill(0);
-        draw_line(x, y, 239 - x, 159 - y, 2);
+        poll_keys();
+        if (key_tap(KEY_A)) {
+            break;
+        }
+        bios_vblank_wait();
+        FRAME_START();
+        PROF(flip_buffer(), flip_cycles);
+        PROF(screen_fill(3), clear_cycles);
+        PROF(draw_line(x, y, 239 - x, 159 - y, 2), line_cycles);
         x += inc_x;
         y += inc_y;
         if (x == 239 && inc_x == 1) {
@@ -130,8 +106,29 @@ test_moving_line(void) {
             inc_x = 1;
             inc_y = 0;
         }
+        FRAME_END();
+        PROF_SHOW();
+    }
+}
+
+void
+test_all_static(void) {
+    while (true) {
+        poll_keys();
+        if (key_tap(KEY_A)) {
+            break;
+        }
         bios_vblank_wait();
-        flip_buffer();
+        FRAME_START();
+        PROF(flip_buffer(), flip_cycles);
+        PROF(test_clear(), clear_cycles);
+        PROF(test_lines(), line_cycles);
+        PROF(test_rect(), rect_cycles);
+        PROF(test_fill_rect(), fill_rect_cycles);
+        PROF(test_chr(), chr_cycles);
+        PROF(test_icn(), icn_cycles);
+        FRAME_END();
+        PROF_SHOW();
     }
 }
 
@@ -146,19 +143,12 @@ int main(void) {
     irq_init();
     irs_set(IRQ_VBLANK, irs_stub);
 
-    // Main loop.
-    PROF_INIT();
+    // TODO: Test sprites in movement.
+    // TODO: Verify all works well, there were some modifications and
+    // bugfixes on the ppu of the uxngba branch.
     while (true) {
-        PROF(test_clear(), test_clear_cycles);
-        PROF(test_lines(), test_lines_cycles);
-        PROF(test_rect(), test_rect_cycles);
-        PROF(test_fill_rect(), test_fill_rect_cycles);
-        PROF(test_chr(), test_chr_cycles);
-        PROF(test_icn(), test_icn_cycles);
-        draw_filled_rect(0, 0, 140, 60, 0);
-        PROF_SHOW();
-        bios_vblank_wait();
-        PROF(flip_buffer(), flip_cycles);
+        test_moving_line();
+        test_all_static();
     }
 
     return 0;
diff --git a/src/profiling.c b/src/profiling.c
new file mode 100644
index 0000000..89e2ce7
--- /dev/null
+++ b/src/profiling.c
@@ -0,0 +1,140 @@
+//
+// 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 5
+#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
+
+#define PROF_SHOW() \
+    do { \
+        txt_position((PROF_SHOW_X), (PROF_SHOW_Y));\
+        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("TOTAL  %.8lu\n", avg_frame_cycles);\
+        txt_render();\
+    } 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 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;
+
+#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;\
+            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;\
+            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;\
+            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;\
+            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/renderer_m0.c b/src/renderer_m0.c
index 35e9ceb..eaf4658 100644
--- a/src/renderer_m0.c
+++ b/src/renderer_m0.c
@@ -12,7 +12,7 @@
 // Parameters.
 //
 
-#define SUBPIXEL_LINES 0
+#define SUBPIXEL_LINES 1
 #define DEC_BIG_LUT 1
 
 // Front/back buffers for double buffering.