Replace puticn with optimized version

1 files changed, 158 insertions, 19 deletions

diff --git a/src/uxn/devices/ppu.c b/src/uxn/devices/ppu.c
index d9208a5..3865e95 100644
--- a/src/uxn/devices/ppu.c
+++ b/src/uxn/devices/ppu.c
@@ -32,6 +32,116 @@ static Uint8 font[][8] = {
     {0x00, 0x7c, 0x82, 0x80, 0xf0, 0x80, 0x82, 0x7c},
     {0x00, 0x7c, 0x82, 0x80, 0xf0, 0x80, 0x80, 0x80}};
 
+static Uint32 unpack_icon_lut[256] = {
+    0x00000000, 0x00000001, 0x00000010, 0x00000011, 0x00000100,
+    0x00000101, 0x00000110, 0x00000111, 0x00001000, 0x00001001,
+    0x00001010, 0x00001011, 0x00001100, 0x00001101, 0x00001110,
+    0x00001111, 0x00010000, 0x00010001, 0x00010010, 0x00010011,
+    0x00010100, 0x00010101, 0x00010110, 0x00010111, 0x00011000,
+    0x00011001, 0x00011010, 0x00011011, 0x00011100, 0x00011101,
+    0x00011110, 0x00011111, 0x00100000, 0x00100001, 0x00100010,
+    0x00100011, 0x00100100, 0x00100101, 0x00100110, 0x00100111,
+    0x00101000, 0x00101001, 0x00101010, 0x00101011, 0x00101100,
+    0x00101101, 0x00101110, 0x00101111, 0x00110000, 0x00110001,
+    0x00110010, 0x00110011, 0x00110100, 0x00110101, 0x00110110,
+    0x00110111, 0x00111000, 0x00111001, 0x00111010, 0x00111011,
+    0x00111100, 0x00111101, 0x00111110, 0x00111111, 0x01000000,
+    0x01000001, 0x01000010, 0x01000011, 0x01000100, 0x01000101,
+    0x01000110, 0x01000111, 0x01001000, 0x01001001, 0x01001010,
+    0x01001011, 0x01001100, 0x01001101, 0x01001110, 0x01001111,
+    0x01010000, 0x01010001, 0x01010010, 0x01010011, 0x01010100,
+    0x01010101, 0x01010110, 0x01010111, 0x01011000, 0x01011001,
+    0x01011010, 0x01011011, 0x01011100, 0x01011101, 0x01011110,
+    0x01011111, 0x01100000, 0x01100001, 0x01100010, 0x01100011,
+    0x01100100, 0x01100101, 0x01100110, 0x01100111, 0x01101000,
+    0x01101001, 0x01101010, 0x01101011, 0x01101100, 0x01101101,
+    0x01101110, 0x01101111, 0x01110000, 0x01110001, 0x01110010,
+    0x01110011, 0x01110100, 0x01110101, 0x01110110, 0x01110111,
+    0x01111000, 0x01111001, 0x01111010, 0x01111011, 0x01111100,
+    0x01111101, 0x01111110, 0x01111111, 0x10000000, 0x10000001,
+    0x10000010, 0x10000011, 0x10000100, 0x10000101, 0x10000110,
+    0x10000111, 0x10001000, 0x10001001, 0x10001010, 0x10001011,
+    0x10001100, 0x10001101, 0x10001110, 0x10001111, 0x10010000,
+    0x10010001, 0x10010010, 0x10010011, 0x10010100, 0x10010101,
+    0x10010110, 0x10010111, 0x10011000, 0x10011001, 0x10011010,
+    0x10011011, 0x10011100, 0x10011101, 0x10011110, 0x10011111,
+    0x10100000, 0x10100001, 0x10100010, 0x10100011, 0x10100100,
+    0x10100101, 0x10100110, 0x10100111, 0x10101000, 0x10101001,
+    0x10101010, 0x10101011, 0x10101100, 0x10101101, 0x10101110,
+    0x10101111, 0x10110000, 0x10110001, 0x10110010, 0x10110011,
+    0x10110100, 0x10110101, 0x10110110, 0x10110111, 0x10111000,
+    0x10111001, 0x10111010, 0x10111011, 0x10111100, 0x10111101,
+    0x10111110, 0x10111111, 0x11000000, 0x11000001, 0x11000010,
+    0x11000011, 0x11000100, 0x11000101, 0x11000110, 0x11000111,
+    0x11001000, 0x11001001, 0x11001010, 0x11001011, 0x11001100,
+    0x11001101, 0x11001110, 0x11001111, 0x11010000, 0x11010001,
+    0x11010010, 0x11010011, 0x11010100, 0x11010101, 0x11010110,
+    0x11010111, 0x11011000, 0x11011001, 0x11011010, 0x11011011,
+    0x11011100, 0x11011101, 0x11011110, 0x11011111, 0x11100000,
+    0x11100001, 0x11100010, 0x11100011, 0x11100100, 0x11100101,
+    0x11100110, 0x11100111, 0x11101000, 0x11101001, 0x11101010,
+    0x11101011, 0x11101100, 0x11101101, 0x11101110, 0x11101111,
+    0x11110000, 0x11110001, 0x11110010, 0x11110011, 0x11110100,
+    0x11110101, 0x11110110, 0x11110111, 0x11111000, 0x11111001,
+    0x11111010, 0x11111011, 0x11111100, 0x11111101, 0x11111110,
+    0x11111111
+};
+
+static Uint32 unpack_icon_lut_flipx[256] = {
+    0x00000000, 0x10000000, 0x01000000, 0x11000000, 0x00100000,
+    0x10100000, 0x01100000, 0x11100000, 0x00010000, 0x10010000,
+    0x01010000, 0x11010000, 0x00110000, 0x10110000, 0x01110000,
+    0x11110000, 0x00001000, 0x10001000, 0x01001000, 0x11001000,
+    0x00101000, 0x10101000, 0x01101000, 0x11101000, 0x00011000,
+    0x10011000, 0x01011000, 0x11011000, 0x00111000, 0x10111000,
+    0x01111000, 0x11111000, 0x00000100, 0x10000100, 0x01000100,
+    0x11000100, 0x00100100, 0x10100100, 0x01100100, 0x11100100,
+    0x00010100, 0x10010100, 0x01010100, 0x11010100, 0x00110100,
+    0x10110100, 0x01110100, 0x11110100, 0x00001100, 0x10001100,
+    0x01001100, 0x11001100, 0x00101100, 0x10101100, 0x01101100,
+    0x11101100, 0x00011100, 0x10011100, 0x01011100, 0x11011100,
+    0x00111100, 0x10111100, 0x01111100, 0x11111100, 0x00000010,
+    0x10000010, 0x01000010, 0x11000010, 0x00100010, 0x10100010,
+    0x01100010, 0x11100010, 0x00010010, 0x10010010, 0x01010010,
+    0x11010010, 0x00110010, 0x10110010, 0x01110010, 0x11110010,
+    0x00001010, 0x10001010, 0x01001010, 0x11001010, 0x00101010,
+    0x10101010, 0x01101010, 0x11101010, 0x00011010, 0x10011010,
+    0x01011010, 0x11011010, 0x00111010, 0x10111010, 0x01111010,
+    0x11111010, 0x00000110, 0x10000110, 0x01000110, 0x11000110,
+    0x00100110, 0x10100110, 0x01100110, 0x11100110, 0x00010110,
+    0x10010110, 0x01010110, 0x11010110, 0x00110110, 0x10110110,
+    0x01110110, 0x11110110, 0x00001110, 0x10001110, 0x01001110,
+    0x11001110, 0x00101110, 0x10101110, 0x01101110, 0x11101110,
+    0x00011110, 0x10011110, 0x01011110, 0x11011110, 0x00111110,
+    0x10111110, 0x01111110, 0x11111110, 0x00000001, 0x10000001,
+    0x01000001, 0x11000001, 0x00100001, 0x10100001, 0x01100001,
+    0x11100001, 0x00010001, 0x10010001, 0x01010001, 0x11010001,
+    0x00110001, 0x10110001, 0x01110001, 0x11110001, 0x00001001,
+    0x10001001, 0x01001001, 0x11001001, 0x00101001, 0x10101001,
+    0x01101001, 0x11101001, 0x00011001, 0x10011001, 0x01011001,
+    0x11011001, 0x00111001, 0x10111001, 0x01111001, 0x11111001,
+    0x00000101, 0x10000101, 0x01000101, 0x11000101, 0x00100101,
+    0x10100101, 0x01100101, 0x11100101, 0x00010101, 0x10010101,
+    0x01010101, 0x11010101, 0x00110101, 0x10110101, 0x01110101,
+    0x11110101, 0x00001101, 0x10001101, 0x01001101, 0x11001101,
+    0x00101101, 0x10101101, 0x01101101, 0x11101101, 0x00011101,
+    0x10011101, 0x01011101, 0x11011101, 0x00111101, 0x10111101,
+    0x01111101, 0x11111101, 0x00000011, 0x10000011, 0x01000011,
+    0x11000011, 0x00100011, 0x10100011, 0x01100011, 0x11100011,
+    0x00010011, 0x10010011, 0x01010011, 0x11010011, 0x00110011,
+    0x10110011, 0x01110011, 0x11110011, 0x00001011, 0x10001011,
+    0x01001011, 0x11001011, 0x00101011, 0x10101011, 0x01101011,
+    0x11101011, 0x00011011, 0x10011011, 0x01011011, 0x11011011,
+    0x00111011, 0x10111011, 0x01111011, 0x11111011, 0x00000111,
+    0x10000111, 0x01000111, 0x11000111, 0x00100111, 0x10100111,
+    0x01100111, 0x11100111, 0x00010111, 0x10010111, 0x01010111,
+    0x11010111, 0x00110111, 0x10110111, 0x01110111, 0x11110111,
+    0x00001111, 0x10001111, 0x01001111, 0x11001111, 0x00101111,
+    0x10101111, 0x01101111, 0x11101111, 0x00011111, 0x10011111,
+    0x01011111, 0x11011111, 0x00111111, 0x10111111, 0x01111111,
+    0x11111111
+};
+
 u32 *backbuffer_bg0;
 u32 *backbuffer_bg1;
 static u8 dirty_tiles[20][30] = {0};
@@ -53,32 +163,61 @@ putcolors(Ppu *p, Uint8 *addr) {
 
 void
 putpixel(Ppu *p, Uint32 *layer, Uint16 x, Uint16 y, Uint8 color) {
-    Uint32 pos = ((y & 7) + (((x >> 3) + (y >> 3) * 30) * 8));
-    Uint32 shift = (x & 7) << 2;
+    size_t tile_x = x / 8;
+    size_t tile_y = y / 8;
+    size_t start_col = x % 8;
+    size_t start_row = y % 8;
+    Uint32 pos = (start_row + ((tile_x + tile_y * 30) * 8));
+    Uint32 shift = start_col * 4;
     layer[pos] = (layer[pos] & (~(0xF << shift))) | (color << shift);
-    dirty_tiles[y >> 3][x >> 3] = 1;
+    dirty_tiles[tile_y][tile_x] = 1;
 }
 
 void
 puticn(Ppu *p, Uint32 *layer, Uint16 x, Uint16 y, Uint8 *sprite, Uint8 color,
         Uint8 flipx, Uint8 flipy) {
-    Uint16 v, h;
-    for(v = 0; v < 8; v++) {
-        for(h = 0; h < 8; h++) {
-            Uint8 ch1 = ((sprite[v] >> (7 - h)) & 0x1);
-            // if(ch1 == 1 || (color != 0x05 && color != 0x0a && color != 0x0f))
-                // putpixel(p,
-                //     layer,
-                //     x + (flipx ? 7 - h : h),
-                //     y + (flipy ? 7 - v : v),
-                //     ch1 ? color % 4 : color / 4);
-            // TODO: Add flip options directly to the screenblock?
-            putpixel(p,
-                layer,
-                x + h,
-                y + v,
-                ch1 ? color & 0x3 : color >> 2);
+    size_t tile_x = x / 8;
+    size_t tile_y = y / 8;
+    size_t start_col = x % 8;
+    size_t start_row = y % 8;
+    Uint32 pos = (start_row + ((tile_x + tile_y * 30) * 8));
+    Uint32 shift_left = start_col * 4;
+    Uint32 shift_right = (8 - start_col) * 4;
+        Uint32 *layer_ptr = &layer[pos];
+        if (flipy) flipy = 7;
+
+        Uint32 *lut = flipx ? unpack_icon_lut : unpack_icon_lut_flipx;
+
+    // There are 4 possible cases:
+    // 1. The sprite is exactly at the tile boundary. We can just copy the
+    //    tile directly to memory.
+    // 2. The sprite covers 2 tiles horizontally.
+    // 3. The sprite covers 2 tiles vertically.
+    // 4. The sprite covers 4 tiles.
+    if (start_row == 0 && start_col == 0) {
+        for (size_t i = 0; i < 8; ++i) {
+            layer_ptr[i] = lut[sprite[i ^ flipy]] * color;
+        }
+        dirty_tiles[tile_y][tile_x] = 1;
+    } else {
+        u32 row_mask = 0xFFFFFFFF << shift_left;
+        for (size_t i = 0; i < (8 - start_row); ++i) {
+            u32 sprite_row = lut[sprite[i ^ flipy]] * color;
+            layer_ptr[0] = (layer_ptr[0] & ~row_mask) | (sprite_row << shift_left);
+            layer_ptr[8] = (layer_ptr[8] & row_mask) | (sprite_row >> shift_right);
+            layer_ptr++;
+        }
+        layer_ptr += 8 * 29;
+        for (size_t i = (8 - start_row); i < 8; ++i) {
+            u32 sprite_row = lut[sprite[i ^ flipy]] * color;
+            layer_ptr[0] = (layer_ptr[0] & ~row_mask) | (sprite_row << shift_left);
+            layer_ptr[8] = (layer_ptr[8] & row_mask) | (sprite_row >> shift_right);
+            layer_ptr++;
         }
+        dirty_tiles[tile_y][tile_x] = 1;
+        dirty_tiles[tile_y][tile_x + 1] = 1;
+        dirty_tiles[tile_y + 1][tile_x] = 1;
+        dirty_tiles[tile_y + 1][tile_x + 1] = 1;
     }
 }