Add icn and chr drawing functions

2 files changed, 114 insertions, 116 deletions

diff --git a/src/main.c b/src/main.c
index c6f3c49..129493b 100644
--- a/src/main.c
+++ b/src/main.c
@@ -120,8 +120,6 @@ int main(void) {
     irq_init();
     irs_set(IRQ_VBLANK, irs_stub);
 
-    // draw_filled_rect(0, 0, 7, 7, 1);
-    // flip_buffer();
     // draw_filled_rect(0, 0, 7, 7, 2);
     // draw_filled_rect(8, 0, 15, 7, 3);
     // BG_H_SCROLL_0 = 240;
@@ -133,28 +131,27 @@ int main(void) {
     int inc = 1;
     while (true) {
         bios_vblank_wait();
-        flip_buffer();
 
         PROF(test_clear(), test_clear_cycles);
         // PROF(test_lines(), test_lines_cycles);
 
         // ANIMATION EXAMPLE
-        draw_rect(0 + x, 0, 100 + x, 100, 2);
-        x += inc;
-        if (x >= 50) {
-            inc = -1;
-        } else if (x <= 0) {
-            inc = 1;
-        }
+        // draw_rect(0 + x, 0, 100 + x, 100, 2);
+        // x += inc;
+        // if (x >= 50) {
+        //     inc = -1;
+        // } else if (x <= 0) {
+        //     inc = 1;
+        // }
 
         // draw_rect(50, 10, 200, 110, 1);
-        PROF(test_rect(), test_rect_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();
-        // PROF(flip_buffer(), flip_cycles);
+        draw_filled_rect(0, 0, 140, 60, 0);
+        PROF_SHOW();
+        PROF(flip_buffer(), flip_cycles);
     }
 
     return 0;
diff --git a/src/renderer_m0.c b/src/renderer_m0.c
index b187903..3d1a4d0 100644
--- a/src/renderer_m0.c
+++ b/src/renderer_m0.c
@@ -75,14 +75,15 @@ static u32 *backbuf = BUF_0;
 
 IWRAM_CODE
 void screen_fill(u8 clr)  {
-    // TODO: revise this with updated renderer.
-#if 1
+    // We have to make sure we leave the last tile blank to use as alpha channel
+    // when moving the BG during double buffering.
+#if 0
     u32 *dst = backbuf;
-    for(int i = 0; i < KB(20) / 4; i++) {
+    for(int i = 0; i < KB(20) / 4 - 8; i++) {
         *dst++ = 0x11111111 * clr;
     }
 #else
-    dma_fill(backbuf, 0x11111111 * clr, KB(20), 3);
+    dma_fill(backbuf, 0x11111111 * clr, KB(20) - 32, 3);
 #endif
 }
 
@@ -475,65 +476,65 @@ flip_buffer(void) {
     // }
 }
 
-static u32 dec_nibble[] = {
-    0x00000000, 0x01000000, 0x00010000, 0x01010000,
-    0x00000100, 0x01000100, 0x00010100, 0x01010100,
-    0x00000001, 0x01000001, 0x00010001, 0x01010001,
-    0x00000101, 0x01000101, 0x00010101, 0x01010101,
+static u16 dec_nibble[] = {
+    0x0000, 0x1000, 0x0100, 0x1100,
+    0x0010, 0x1010, 0x0110, 0x1110,
+    0x0001, 0x1001, 0x0101, 0x1101,
+    0x0011, 0x1011, 0x0111, 0x1111,
 };
 
-static u32 dec_nibble_flip_x[] = {
-    0x00000000, 0x00000001, 0x00000100, 0x00000101,
-    0x00010000, 0x00010001, 0x00010100, 0x00010101,
-    0x01000000, 0x01000001, 0x01000100, 0x01000101,
-    0x01010000, 0x01010001, 0x01010100, 0x01010101,
+static u16 dec_nibble_flip_x[] = {
+    0x0000, 0x0001, 0x0010, 0x0011,
+    0x0100, 0x0101, 0x0110, 0x0111,
+    0x1000, 0x1001, 0x1010, 0x1011,
+    0x1100, 0x1101, 0x1110, 0x1111,
 };
 
 IWRAM_CODE
 static inline
-u64
+u32
 decode_1bpp(u8 row, u8 flip_x) {
     if (flip_x) {
-        u32 *lut = dec_nibble_flip_x;
-        return (u64)lut[(row >> 4) & 0xF] << 32 | (u64)lut[(row >> 0) & 0xF];
+        u16 *lut = dec_nibble_flip_x;
+        return (u32)lut[(row >> 4) & 0xF] << 16 | (u32)lut[(row >> 0) & 0xF];
     }
-    u32 *lut = dec_nibble;
-    return (u64)lut[(row >> 0) & 0xF] << 32 | (u64)lut[(row >> 4) & 0xF];
+    u16 *lut = dec_nibble;
+    return (u32)lut[(row >> 0) & 0xF] << 16 | (u32)lut[(row >> 4) & 0xF];
 }
 
 IWRAM_CODE
 static inline
 void
 draw_2bpp_row(size_t x, size_t y, u8 a, u8 b, u8 flip_x) {
-    // BOUNDCHECK_SCREEN(x, y);
-
-    // size_t tile_x = x / 8;
-    // size_t start_col = x % 8;
-    // size_t shift_left = start_col * 8;
-    // size_t shift_right = (8 - start_col) * 8;
-
-    // u64 *dst = &backbuf[(y * 30 + tile_x) * 8 / 2];
-    // if (start_col == 0) {
-    //     u64 clr_a = decode_1bpp(a, flip_x);
-    //     u64 clr_b = decode_1bpp(b, flip_x);
-    //     u64 mask_a = (clr_a * 0xF);
-    //     u64 mask_b = (clr_b * 0xF);
-    //     u64 mask = (mask_a | mask_b);
-    //     u64 color = clr_a + (clr_b << 1);
-    //     dst[0] = (dst[0] & ~mask) | color;
-    // } else {
-    //     u64 clr_a = decode_1bpp(a, flip_x);
-    //     u64 clr_b = decode_1bpp(b, flip_x);
-    //     u64 mask_a = (clr_a * 0xF);
-    //     u64 mask_b = (clr_b * 0xF);
-    //     u64 mask = (mask_a | mask_b);
-    //     u64 color = clr_a + (clr_b << 1);
-    //     dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-    //     if ((x + 7) > (SCREEN_WIDTH)) {
-    //         return;
-    //     }
-    //     dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
-    // }
+    BOUNDCHECK_SCREEN(x, y);
+
+    size_t tile_x = x / 8;
+    size_t tile_y = y / 8;
+    size_t start_col = x % 8;
+    size_t start_row = y % 8;
+    size_t shift_left = start_col * 4;
+    size_t shift_right = (8 - start_col) * 4;
+
+    u32 *dst = &backbuf[tile_x * 8 + tile_y * 8 * 32];
+    dst += start_row;
+    if (start_col == 0) {
+        u32 clr_a = decode_1bpp(a, flip_x);
+        u32 clr_b = decode_1bpp(b, flip_x);
+        u32 mask_a = (clr_a * 0xF);
+        u32 mask_b = (clr_b * 0xF);
+        u32 mask = (mask_a | mask_b);
+        u32 color = clr_a + (clr_b << 1);
+        dst[0] = (dst[0] & ~mask) | color;
+    } else {
+        u32 clr_a = decode_1bpp(a, flip_x);
+        u32 clr_b = decode_1bpp(b, flip_x);
+        u32 mask_a = (clr_a * 0xF);
+        u32 mask_b = (clr_b * 0xF);
+        u32 mask = (mask_a | mask_b);
+        u32 color = clr_a + (clr_b << 1);
+        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
+        dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+    }
 
     // TODO: different blend modes?
 }
@@ -542,29 +543,29 @@ IWRAM_CODE
 static inline
 void
 draw_1bpp_row(size_t x, size_t y, u8 a, u8 clr, u8 flip_x) {
-    // BOUNDCHECK_SCREEN(x, y);
-
-    // size_t tile_x = x / 8;
-    // size_t start_col = x % 8;
-    // size_t shift_left = start_col * 8;
-    // size_t shift_right = (8 - start_col) * 8;
-
-    // u64 *dst = &backbuf[(y * 30 + tile_x) * 8 / 2];
-    // if (start_col == 0) {
-    //     u64 color = decode_1bpp(a, flip_x);
-    //     u64 mask = (color * 0xF);
-    //     color *= clr;
-    //     dst[0] = (dst[0] & ~mask) | color;
-    // } else {
-    //     u64 color = decode_1bpp(a, flip_x);
-    //     u64 mask = (color * 0xF);
-    //     color *= clr;
-    //     dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
-    //     if ((x + 7) > (SCREEN_WIDTH)) {
-    //         return;
-    //     }
-    //     dst[1] = (dst[1] & ~(mask >> shift_right)) | (color >> shift_right);
-    // }
+    BOUNDCHECK_SCREEN(x, y);
+
+    size_t tile_x = x / 8;
+    size_t tile_y = y / 8;
+    size_t start_col = x % 8;
+    size_t start_row = y % 8;
+    size_t shift_left = start_col * 4;
+    size_t shift_right = (8 - start_col) * 4;
+
+    u32 *dst = &backbuf[tile_x * 8 + tile_y * 8 * 32];
+    dst += start_row;
+    if (start_col == 0) {
+        u32 color = decode_1bpp(a, flip_x);
+        u32 mask = (color * 0xF);
+        color *= clr;
+        dst[0] = (dst[0] & ~mask) | color;
+    } else {
+        u32 color = decode_1bpp(a, flip_x);
+        u32 mask = (color * 0xF);
+        color *= clr;
+        dst[0] = (dst[0] & ~(mask << shift_left))  | (color << shift_left);
+        dst[8] = (dst[8] & ~(mask >> shift_right)) | (color >> shift_right);
+    }
 
     // TODO: different blend modes?
 }
@@ -572,41 +573,41 @@ draw_1bpp_row(size_t x, size_t y, u8 a, u8 clr, u8 flip_x) {
 IWRAM_CODE
 void
 draw_chr(size_t x, size_t y, u8 *sprite, u8 flip_x, u8 flip_y) {
-    // BOUNDCHECK_SCREEN(x, y);
-    // if (!flip_y) {
-    //     for(size_t v = 0; v < 8; v++) {
-    //         if ((y + v) >= SCREEN_HEIGHT) break;
-    //         u8 ch1 = sprite[v + 0];
-    //         u8 ch2 = sprite[v + 8];
-    //         draw_2bpp_row(x, y + v, ch1, ch2, flip_x);
-    //     }
-    // } else {
-    //     for(size_t v = 0; v < 8; v++) {
-    //         if ((y + v) >= SCREEN_HEIGHT) break;
-    //         u8 ch1 = sprite[(7 - v) + 0];
-    //         u8 ch2 = sprite[(7 - v) + 8];
-    //         draw_2bpp_row(x, y + v, ch1, ch2, flip_x);
-    //     }
-    // }
+    BOUNDCHECK_SCREEN(x, y);
+    if (!flip_y) {
+        for(size_t v = 0; v < 8; v++) {
+            if ((y + v) >= SCREEN_HEIGHT) break;
+            u8 ch1 = sprite[v + 0];
+            u8 ch2 = sprite[v + 8];
+            draw_2bpp_row(x, y + v, ch1, ch2, flip_x);
+        }
+    } else {
+        for(size_t v = 0; v < 8; v++) {
+            if ((y + v) >= SCREEN_HEIGHT) break;
+            u8 ch1 = sprite[(7 - v) + 0];
+            u8 ch2 = sprite[(7 - v) + 8];
+            draw_2bpp_row(x, y + v, ch1, ch2, flip_x);
+        }
+    }
 }
 
 IWRAM_CODE
 void
 draw_icn(size_t x, size_t y, u8 *sprite, u8 clr, u8 flip_x, u8 flip_y) {
-    // BOUNDCHECK_SCREEN(x, y);
-    // if (!flip_y) {
-    //     for(size_t v = 0; v < 8; v++) {
-    //         if ((y + v) >= SCREEN_HEIGHT) break;
-    //         u8 ch1 = sprite[v];
-    //         draw_1bpp_row(x, y + v, ch1, clr, flip_x);
-    //     }
-    // } else {
-    //     for(size_t v = 0; v < 8; v++) {
-    //         if ((y + v) >= SCREEN_HEIGHT) break;
-    //         u8 ch1 = sprite[(7 - v)];
-    //         draw_1bpp_row(x, y + v, ch1, clr, flip_x);
-    //     }
-    // }
+    BOUNDCHECK_SCREEN(x, y);
+    if (!flip_y) {
+        for(size_t v = 0; v < 8; v++) {
+            if ((y + v) >= SCREEN_HEIGHT) break;
+            u8 ch1 = sprite[v];
+            draw_1bpp_row(x, y + v, ch1, clr, flip_x);
+        }
+    } else {
+        for(size_t v = 0; v < 8; v++) {
+            if ((y + v) >= SCREEN_HEIGHT) break;
+            u8 ch1 = sprite[(7 - v)];
+            draw_1bpp_row(x, y + v, ch1, clr, flip_x);
+        }
+    }
 }
 
 #include "font.h"
@@ -649,7 +650,7 @@ renderer_init(void) {
 
     // Setup initial background state.
     BG_H_SCROLL_0 = 240;
-    BG_H_SCROLL_1 = 0;
+    BG_H_SCROLL_1 = 240;
 
     // FIXME: clean this up.
     // u32 *clr_tile =((u32*)(MEM_VRAM) + 0xC000 / 4);