GB Video: Draw SGB border pieces that overlap GB graphics (fixes #1339)

CHANGES

@@ -14,6 +14,7 @@ Features:
 Emulation fixes:
  - ARM7: Fix unsigned multiply timing
  - GB Memory: Add cursory cartridge open bus emulation (fixes mgba.io/i/2032)
+ - GB Video: Draw SGB border pieces that overlap GB graphics (fixes mgba.io/i/1339)
  - GBA: Improve timing when not booting from BIOS
  - GBA BIOS: Work around IRQ handling hiccup in Mario & Luigi (fixes mgba.io/i/1059)
  - GBA I/O: Redo internal key input, enabling edge-based key IRQs

include/mgba/internal/gb/renderers/software.h

@@ -60,6 +60,7 @@ struct GBVideoSoftwareRenderer {
 	uint8_t sgbPacket[128];
 	uint8_t sgbCommandHeader;
 	bool sgbBorders;
+	uint32_t sgbBorderMask[18];
 
 	uint8_t lastHighlightAmount;
 };

src/gb/renderers/software.c

@@ -64,37 +64,58 @@ static void _regenerateSGBBorder(struct GBVideoSoftwareRenderer* renderer) {
 	}
 	int x, y;
 	for (y = 0; y < 224; ++y) {
+		int localY = y & 0x7;
+		if (!localY && y >= 40 && y < 184) {
+			renderer->sgbBorderMask[(y - 40) >> 3] = 0;
+		}
 		for (x = 0; x < 256; x += 8) {
-			if (x >= 48 && x < 208 && y >= 40 && y < 184) {
-				continue;
-			}
 			uint16_t mapData;
 			LOAD_16LE(mapData, (x >> 2) + (y & ~7) * 8, renderer->d.sgbMapRam);
 			if (UNLIKELY(SGBBgAttributesGetTile(mapData) >= 0x100)) {
 				continue;
 			}
 
-			int localY = y & 0x7;
+			if (x >= 48 && x < 208 && y >= 40 && y < 184) {
-			if (SGBBgAttributesIsYFlip(mapData)) {
+				if (!localY) {
-				localY = 7 - localY;
+					unsigned tileBase = SGBBgAttributesGetTile(mapData) * 8;
+					uint32_t bits = 0;
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 0];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 1];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 2];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 3];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 4];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 5];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 6];
+					bits |= ((uint32_t*) renderer->d.sgbCharRam)[tileBase + 7];
+					if (bits) {
+						renderer->sgbBorderMask[(y - 40) >> 3] |= 1 << ((x - 48) >> 3);
+					}
+				}
+				continue;
 			}
+
+			int yFlip = 0;
+			if (SGBBgAttributesIsYFlip(mapData)) {
+				yFlip = 7;
+			}
+			unsigned tileBase = (SGBBgAttributesGetTile(mapData) * 16 + (localY ^ yFlip)) * 2;
 			uint8_t tileData[4];
-			tileData[0] = renderer->d.sgbCharRam[(SGBBgAttributesGetTile(mapData) * 16 + localY) * 2 + 0x00];
+			tileData[0] = renderer->d.sgbCharRam[tileBase + 0x00];
-			tileData[1] = renderer->d.sgbCharRam[(SGBBgAttributesGetTile(mapData) * 16 + localY) * 2 + 0x01];
+			tileData[1] = renderer->d.sgbCharRam[tileBase + 0x01];
-			tileData[2] = renderer->d.sgbCharRam[(SGBBgAttributesGetTile(mapData) * 16 + localY) * 2 + 0x10];
+			tileData[2] = renderer->d.sgbCharRam[tileBase + 0x10];
-			tileData[3] = renderer->d.sgbCharRam[(SGBBgAttributesGetTile(mapData) * 16 + localY) * 2 + 0x11];
+			tileData[3] = renderer->d.sgbCharRam[tileBase + 0x11];
 
 			size_t base = y * renderer->outputBufferStride + x;
 			int paletteBase = SGBBgAttributesGetPalette(mapData) * 0x10;
 			int colorSelector;
 
-			int flip = 0;
+			int xFlip = 0;
 			if (SGBBgAttributesIsXFlip(mapData)) {
-				flip = 7;
+				xFlip = 7;
 			}
 			for (i = 7; i >= 0; --i) {
 				colorSelector = (tileData[0] >> i & 0x1) << 0 | (tileData[1] >> i & 0x1) << 1 | (tileData[2] >> i & 0x1) << 2 | (tileData[3] >> i & 0x1) << 3;
-				renderer->outputBuffer[(base + 7 - i) ^ flip] = renderer->palette[paletteBase | colorSelector];
+				renderer->outputBuffer[(base + 7 - i) ^ xFlip] = renderer->palette[paletteBase | colorSelector];
 			}
 		}
 	}
@@ -233,6 +254,7 @@ static void GBVideoSoftwareRendererInit(struct GBVideoRenderer* renderer, enum G
 	}
 
 	memset(softwareRenderer->palette, 0, sizeof(softwareRenderer->palette));
+	memset(softwareRenderer->sgbBorderMask, 0, sizeof(softwareRenderer->sgbBorderMask));
 
 	softwareRenderer->lastHighlightAmount = 0;
 }
@@ -667,6 +689,46 @@ static void GBVideoSoftwareRendererDrawRange(struct GBVideoRenderer* renderer, i
 		for (; x < endX; ++x) {
 			row[x] = softwareRenderer->palette[p | softwareRenderer->lookup[softwareRenderer->row[x] & OBJ_PRIO_MASK]];
 		}
+		if (softwareRenderer->sgbBorderMask[y >> 3]) {
+			uint32_t borderMask = softwareRenderer->sgbBorderMask[y >> 3];
+			int localY = y & 0x7;
+			for (x = startX; x < endX; x += 8) {
+				if (!(borderMask & (1 << (x >> 3)))) {
+					continue;
+				}
+				uint16_t mapData;
+				LOAD_16LE(mapData, (x >> 2) + 12 + (y & ~7) * 8 + 320, softwareRenderer->d.sgbMapRam);
+				if (UNLIKELY(SGBBgAttributesGetTile(mapData) >= 0x100)) {
+					continue;
+				}
+
+				int yFlip = 0;
+				if (SGBBgAttributesIsYFlip(mapData)) {
+					yFlip = 7;
+				}
+				unsigned tileBase = (SGBBgAttributesGetTile(mapData) * 16 + (localY ^ yFlip)) * 2;
+				uint8_t tileData[4];
+				tileData[0] = softwareRenderer->d.sgbCharRam[tileBase + 0x00];
+				tileData[1] = softwareRenderer->d.sgbCharRam[tileBase + 0x01];
+				tileData[2] = softwareRenderer->d.sgbCharRam[tileBase + 0x10];
+				tileData[3] = softwareRenderer->d.sgbCharRam[tileBase + 0x11];
+
+				int paletteBase = SGBBgAttributesGetPalette(mapData) * 0x10;
+				int colorSelector;
+
+				int flip = 0;
+				if (SGBBgAttributesIsXFlip(mapData)) {
+					flip = 7;
+				}
+				int i;
+				for (i = 7; i >= 0; --i) {
+					colorSelector = (tileData[0] >> i & 0x1) << 0 | (tileData[1] >> i & 0x1) << 1 | (tileData[2] >> i & 0x1) << 2 | (tileData[3] >> i & 0x1) << 3;
+					if (colorSelector) {
+						row[(x + 7 - i) ^ flip] = softwareRenderer->palette[paletteBase | colorSelector];
+					}
+				}
+			}
+		}
 		break;
 	case 1:
 		break;