custom YUY2 -> IYUV conversion code

improves from 1+60ms to 3ms

src/debug_tox_call.cpp

@@ -45,6 +45,61 @@ static bool isFormatPlanar(SDL_PixelFormat f) {
 	;
 }
 
+static SDL_Surface* convertYUY2_IYUV(SDL_Surface* surf) {
+	if (surf->format != SDL_PIXELFORMAT_YUY2) {
+		return nullptr;
+	}
+	if ((surf->w % 2) != 0) {
+		// hmmm, we dont handle odd widths
+		return nullptr;
+	}
+
+	SDL_LockSurface(surf);
+
+	SDL_Surface* conv_surf = SDL_CreateSurface(surf->w, surf->h, SDL_PIXELFORMAT_IYUV);
+	SDL_LockSurface(conv_surf);
+
+	// YUY2 is 4:2:2 packed
+	// Y is simple, we just copy it over
+	// U V are double the resolution (vertically), so we avg both
+	// Packed mode: Y0+U0+Y1+V0 (1 plane)
+
+	uint8_t* y_plane = static_cast<uint8_t*>(conv_surf->pixels);
+	uint8_t* u_plane = static_cast<uint8_t*>(conv_surf->pixels) + conv_surf->w*conv_surf->h;
+	uint8_t* v_plane = static_cast<uint8_t*>(conv_surf->pixels) + conv_surf->w*conv_surf->h + (conv_surf->w/2)*(conv_surf->h/2);
+
+	const uint8_t* yuy2_data = static_cast<const uint8_t*>(surf->pixels);
+
+	for (int y = 0; y < surf->h; y++) {
+		for (int x = 0; x < surf->w; x += 2) {
+			// every pixel uses 2 bytes
+			const uint8_t* yuy2_curser = yuy2_data + y*surf->w*2 + x*2;
+			uint8_t src_y0 = yuy2_curser[0];
+			uint8_t src_u = yuy2_curser[1];
+			uint8_t src_y1 = yuy2_curser[2];
+			uint8_t src_v = yuy2_curser[3];
+
+			y_plane[y*conv_surf->w + x] = src_y0;
+			y_plane[y*conv_surf->w + x+1] = src_y1;
+
+			size_t uv_index = (y/2) * (conv_surf->w/2) + x/2;
+			if (y % 2 == 0) {
+				// first write
+				u_plane[uv_index] = src_u;
+				v_plane[uv_index] = src_v;
+			} else {
+				// second write, mix with existing value
+				u_plane[uv_index] = (int(u_plane[uv_index]) + int(src_u)) / 2;
+				v_plane[uv_index] = (int(v_plane[uv_index]) + int(src_v)) / 2;
+			}
+		}
+	}
+
+	SDL_UnlockSurface(conv_surf);
+	SDL_UnlockSurface(surf);
+	return conv_surf;
+}
+
 struct PushConversionQueuedVideoStream : public QueuedFrameStream2<SDLVideoFrame> {
 	SDL_PixelFormat _forced_format {SDL_PIXELFORMAT_IYUV};
 
@@ -55,19 +110,29 @@ struct PushConversionQueuedVideoStream : public QueuedFrameStream2<SDLVideoFrame
 		SDL_Surface* converted_surf = value.surface.get();
 		if (converted_surf->format != SDL_PIXELFORMAT_IYUV) {
 			//std::cerr << "DTC: need to convert from " << SDL_GetPixelFormatName(converted_surf->format) << " to SDL_PIXELFORMAT_IYUV\n";
-			if (isFormatPlanar(converted_surf->format)) {
+			if (converted_surf->format == SDL_PIXELFORMAT_YUY2) {
+				// optimized custom impl
+
+				//auto start = Message::getTimeMS();
+				converted_surf = convertYUY2_IYUV(converted_surf);
+				//auto end = Message::getTimeMS();
+				// 3ms
+				//std::cerr << "DTC: timing " << SDL_GetPixelFormatName(converted_surf->format) << "->SDL_PIXELFORMAT_IYUV: " << end-start << "ms\n";
+			} else if (isFormatPlanar(converted_surf->format)) {
 				// meh, need to convert to rgb as a stopgap
 
 				//auto start = Message::getTimeMS();
 				//SDL_Surface* tmp_conv_surf = SDL_ConvertSurfaceAndColorspace(converted_surf, SDL_PIXELFORMAT_RGBA32, nullptr, SDL_COLORSPACE_RGB_DEFAULT, 0);
 				SDL_Surface* tmp_conv_surf = SDL_ConvertSurfaceAndColorspace(converted_surf, SDL_PIXELFORMAT_RGB24, nullptr, SDL_COLORSPACE_RGB_DEFAULT, 0);
 				//auto end = Message::getTimeMS();
+				// 1ms
 				//std::cerr << "DTC: timing " << SDL_GetPixelFormatName(converted_surf->format) << "->SDL_PIXELFORMAT_RGB24: " << end-start << "ms\n";
 
 				// TODO: fix sdl rgb->yuv conversion resulting in too dark (colorspace) issues
 				//start = Message::getTimeMS();
 				converted_surf = SDL_ConvertSurfaceAndColorspace(tmp_conv_surf, SDL_PIXELFORMAT_IYUV, nullptr, SDL_COLORSPACE_YUV_DEFAULT, 0);
 				//end = Message::getTimeMS();
+				// 60ms
 				//std::cerr << "DTC: timing SDL_PIXELFORMAT_RGB24->SDL_PIXELFORMAT_IYUV: " << end-start << "ms\n";
 
 				SDL_DestroySurface(tmp_conv_surf);