faster rect fill and ttf glyph blit

2 files changed, 56 insertions, 8 deletions

diff --git a/kon/konfont.h b/kon/konfont.h
index 3cd1006..1e01b01 100644
--- a/kon/konfont.h
+++ b/kon/konfont.h
@@ -348,12 +348,35 @@ void kon_drawText(kon_framebuffer_t *fb, kon_font_t *font, int x, int y, const c
 		int glyphW = (int)(q.x1 - q.x0);
 		int glyphH = (int)(q.y1 - q.y0);
 
-		for (int gy = 0; gy < glyphH; gy++) {
-			for (int gx = 0; gx < glyphW; gx++) {
-				uint8_t alpha = font->atlasPixels[(src_y0 + gy) * font->atlasSize + (src_x0 + gx)];
+		int dstX = (int)q.x0;
+		int dstY = (int)q.y0;
+		int gx0 = 0, gy0 = 0;
+
+		if (dstX < 0) gx0 = -dstX;
+		if (dstY < 0) gy0 = -dstY;
+		int gx1 = glyphW, gy1 = glyphH;
+		if (dstX + gx1 > fb->width)  gx1 = fb->width  - dstX;
+		if (dstY + gy1 > fb->height) gy1 = fb->height - dstY;
+
+		for (int gy = gy0; gy < gy1; gy++) {
+			const uint8_t *in = font->atlasPixels + (size_t)(src_y0 + gy) * font->atlasSize + src_x0;
+			uint32_t *out = fb->data + (size_t)(dstY + gy) * fb->width + dstX;
+
+			for (int gx = gx0; gx < gx1; gx++) {
+				uint32_t alpha = in[gx];
 				if (alpha == 0) continue;
 
-				kon_putPixel(fb, (int)q.x0 + gx, (int)q.y0 + gy, ((uint32_t)alpha << 24) | baseColor);
+				if (alpha == 0xFF) {
+					out[gx] = 0xFF000000u | baseColor;
+					continue;
+				}
+
+				uint32_t dst = out[gx];
+				uint32_t inv = 255 - alpha;
+				uint32_t r = (((baseColor >> 16) & 0xFF) * alpha + ((dst >> 16) & 0xFF) * inv) / 255;
+				uint32_t g = (((baseColor >> 8)  & 0xFF) * alpha + ((dst >> 8)  & 0xFF) * inv) / 255;
+				uint32_t b = ((baseColor & 0xFF) * alpha + (dst & 0xFF) * inv) / 255;
+				out[gx] = (dst & 0xFF000000u) | (r << 16) | (g << 8) | b;
 			}
 		}
 	}
diff --git a/kon/konsofren.h b/kon/konsofren.h
index c45091b..e89915a 100644
--- a/kon/konsofren.h
+++ b/kon/konsofren.h
@@ -476,11 +476,36 @@ void kon_fillRectangle(kon_framebuffer_t *fb, int x, int y, int width, int heigh
 	if (x + width  > fb->width ) width  = fb->width  - x;
 	if (y + height > fb->height) height = fb->height - y;
 
-	for (int offset_y = 0; offset_y < height; offset_y++) {
-		for (int offset_x = 0; offset_x < width; offset_x++) {
-			uint32_t dst = fb->data[(offset_y + y) * fb->width + (offset_x + x)];
-			fb->data[(y + offset_y) * fb->width + (x + offset_x)] = kon_blendColor(dst, color);
+	if (width <= 0 || height <= 0) return;
+
+	uint32_t a = color >> 24;
+	if (a == 0) return;
+
+	if (a == 0xFF) {
+		for (int row = 0; row < height; row++) {
+			uint32_t *out = fb->data + (size_t)(y + row) * fb->width + x;
+			for (int col = 0; col < width; col++) out[col] = color;
+		}
+		return;
+	}
 
+	/* same result as kon_blendColor, with x / 255 done as shifts */
+	uint32_t inv = 255 - a;
+	uint32_t sr = ((color >> 16) & 0xFF) * a;
+	uint32_t sg = ((color >> 8)  & 0xFF) * a;
+	uint32_t sb = (color & 0xFF) * a;
+
+	for (int row = 0; row < height; row++) {
+		uint32_t *out = fb->data + (size_t)(y + row) * fb->width + x;
+		for (int col = 0; col < width; col++) {
+			uint32_t dst = out[col];
+			uint32_t r = sr + ((dst >> 16) & 0xFF) * inv;
+			uint32_t g = sg + ((dst >> 8)  & 0xFF) * inv;
+			uint32_t b = sb + (dst & 0xFF) * inv;
+			r = (r + 1 + (r >> 8)) >> 8;
+			g = (g + 1 + (g >> 8)) >> 8;
+			b = (b + 1 + (b >> 8)) >> 8;
+			out[col] = (dst & 0xFF000000u) | (r << 16) | (g << 8) | b;
 		}
 	}
 }