TeamDS

Signed Distance Field Generator

Team Members

1. Dawood Shirzada - Developer

Email All

Progress

What is Signed Distance Field?

Signed Distance Field also know as SDF, is a technique developed by Valve company that uses low resolution textures to display extremely high resolution looking fonts and decals. Valve used SDF in their game engines that run such games as Half-Life 2, Counter-Strike 2 and etc. SDF is so effective that no matter how many times the font or decal is zoomed in, it will always look crisp and sharp while using very small textures. This allows fonts and decals in game to have much higher quality with low memory compare to using regular high resolution textures.

For more detailed information please read Valve's publication: http://www.valvesoftware.com/publications/2007/SIGGRAPH2007_AlphaTestedMagnification.pdf

Examples from Valve

Examples in action

https://youtu.be/CGZRHJvJYIg?t=40

How does Signed Distance Field work?

SDF ONLY works with monochromatic images such as decals and fonts. SDF takes the original texture as an input and creates a SDF version of that texture and saves it in an image format. The game engine then uses the low resolution SDF texture instead.

SDF version is actually much different than the original image. the SDF version no longer stores the pixel color intensity like normal images do, but instead stores the distance to nearest opposite color. For example, since monochromatic images only have black and white, for every white pixel, you look for the nearest black pixel and stores the distance between the two and vice versa.

To read the SDF version of image, we will need to use a custom shader that can understand the SDF version. With help of custom shaders, we can do many more effects such as edge glows, drop shadows, soft edges and etc. All these effects at virtually no additional rendering costs. SDF is a huge win when it comes to gaming performance!

How to convert image to SDF version?

It turns out to convert a image to SDF, it is very computationally expensive. There are however, many methods that approximates and are relatively fast, but the Brute-Force method produces the most accurate result and it is the method that Valve used for their textures. Therefore, I will be using this method as well.

Big-O Complexity

For every pixel in a image, we will need to test it against every other pixel. This makes its complexity O(n^2). For example a 256x256 has 65,536 pixel. Each pixel would have to be tested against 65536 pixels to find out the nearest corresponding pixel. So it needs 65,536 * 65,536 = 4,294,967,296 array element look ups!

Why I chose this

I chose SDF image conversion because it has lots of potential for parallelization. Since all the operations of a pixel are independent of each other and reads data from one single array, this allows for massive gains when using GPU multi-threading.

SDF Brute-Force Method

void Generate(const float src[], float dst[], int width, int height, int spread)
{
	
	int spreadSize = spread * spread;
	int size = width * height;
	for (int i = 0; i < size; i++)
	{
		Vector2 localVec(i - ((i / width) * width), i / width);
		float shortestDist = MAX_FLOAT_VALUE;
		float pixelVal = src[i];
		if (pixelVal > 0) // It's an inside pixel
		{
			// Find closest outside pixel
			for (int j = 0; j < size; j++)
			{
				float pixelVal2 = src[j];
				if (pixelVal2 == 0)// Outside pixel
				{
					// Calculate distance
					Vector2 targetVec(j - ((j / width) * width), j / width);
					float dist = localVec.CalcDistance(targetVec);
					if (dist < shortestDist) shortestDist = dist;
				}

			}

			float spread01 = (shortestDist / spread);
			if (spread01 > 1) spread01 = 1; // clamp it
			dst[i] = (spread01 * .5f) + 0.5f;
		}
		else // It's an outisde pixel
		{
			// Find closest inside pixel
			for (int j = 0; j < size; j++)
			{
				float pixelVal2 = src[j];
				if (pixelVal2 > 0)// Inside pixel
				{
					// Calculate distance
					Vector2 targetVec(j - ((j / width) * width), j / width);
					float dist = localVec.CalcDistance(targetVec);
					if (dist < shortestDist) shortestDist = dist;
				}
			}

			float spread01 = (shortestDist / spread);
			if (spread01 > 1) spread01 = 1; // clamp it
			dst[i] = (1 - spread01) *.5f;
		}
	}
}

Converting a 128x128 image

---

Converting a 256x256 image

---

Converting a 512x512 image

---

Analysis

As you can see, it gets much more expensive as the image size increases. This is due to the n^2 algorithm. In order to to be able to convert images to SDF for game engine in reasonable amount of time, we need to look into GPU parallelism.

Converting the code to Cuda

For now, this code will be a direct conversion to Cuda code without any special consideration for GPU related optimizations.

Launch Config

int main(int argc, char **argv)
{
	if (argc != 2)
	{
		cout << "Incorrect number of arg";
		return 0;
	}

	char* path = argv[1];
	BinaryBitmap* bitmap = LoadBitmap(path);
	if (bitmap == NULL)
		return 0;


	int d;
	cudaDeviceProp prop;
	cudaGetDevice(&d);
	cudaGetDeviceProperties(&prop, d);
	unsigned ntpb = prop.maxThreadsDim[0];

	int size = bitmap->GetSize();
	int width = bitmap->_width;
	int height = bitmap->_height;

	// Allocate GPU memory
	float* d_src;
	float * d_dst;
	cudaMalloc((void**)&d_src, size * sizeof(float));
	cudaMalloc((void**)&d_dst, size * sizeof(float));

	// Copy src to device src
	cudaMemcpy(d_src, bitmap->_pixels, size * sizeof(float), cudaMemcpyHostToDevice);

	// Prepare kernal launch
	// Calc how many blocks to launch
	int numOfBlocks =  ((size + ntpb - 1) / ntpb); //(size / ntpb) + 1;
	// Launch grid
	SDFGenerateCuda << <numOfBlocks, ntpb >> >(d_src, d_dst, width, height, 64);

	// Wait for kernel to finish before copying
	cudaDeviceSynchronize();

	// buffer array for SDF pixels
	float* dst = new float[bitmap->GetSize()];
	cudaMemcpy(dst, d_dst, size * sizeof(float), cudaMemcpyDeviceToHost);


	SaveBitmap(path, dst);

	Pause();

	// Free memory back
	cudaFree(d_src);
	cudaFree(d_dst);
	delete bitmap;
	delete dst;

	return 1;

}

The Kernal

 __global__ void SDFGenerateCuda(const float src[], float dst[], int width, int height, int spread)
{
	int size = width * height;
	int i = blockIdx.x * blockDim.x + threadIdx.x;
	if(i >= size)
		return;
	
	
	Vector2 localVec(i - ((i / width) * width), i / width);
	float shortestDist = MAX_FLOAT_VALUE;
	float pixelVal = src[i];
	if (pixelVal > 0) // It's an inside pixel
	{
		// Find closest outside pixel
		for (int j = 0; j < size; j++)
		{
			float pixelVal2 = src[j];
			if (pixelVal2 == 0)// Outside pixel
			{
				// Calculate distance
				Vector2 targetVec(j - ((j / width) * width), j / width);
				float dist = localVec.CalcDistance(targetVec);
				if (dist < shortestDist) shortestDist = dist;
			}

		}

		float spread01 = (shortestDist / spread);
		if (spread01 > 1) spread01 = 1; // clamp it
		dst[i] = (spread01 * .5f) + 0.5f;
	}
	else // It's an outisde pixel
	{
		// Find closest inside pixel
		for (int j = 0; j < size; j++)
		{
			float pixelVal2 = src[j];
			if (pixelVal2 > 0)// Inside pixel
			{
				// Calculate distance
				Vector2 targetVec(j - ((j / width) * width), j / width);
				float dist = localVec.CalcDistance(targetVec);
				if (dist < shortestDist) shortestDist = dist;
			}
		}

		float spread01 = (shortestDist / spread);
		if (spread01 > 1) spread01 = 1; // clamp it
		dst[i] = (1 - spread01) *.5f;
	}
	
}