• Advertisement
Sign in to follow this  

Use image as code array

This topic is 1269 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Hello again

 

I've got a scenario where I need a simple grayscale image to be represented as an array or similar construct within my HLSL pixel shader code. Each pixel of the image should be mapped to a value within the code construct in question.

 

In the particular case I don't have the means to pass the image as texture to the shader and sample it from there, nor could I load it from anywhere. It needs to be present as a (const / static) variable or #define construct as part of the source code. I neither have access to the DirectX / D3D API functions so everything needs to be pure HLSL.

 

There are tools that help converting the image into HLSL code but I'm not sure about the best storage construct.

 

Would it be recommendable to use e.g. an array of unsigned integers with each entry representing a grayscale / color value, or are there better ways of doing such in terms of performance and memory usage?

Share this post


Link to post
Share on other sites
Advertisement

Which shader model are you compiling your code to? SM4.0 and up have some (limited) support for indexing into immediate constant buffers, but if I recall correctly SM3.0 has no such functionality. 

Share this post


Link to post
Share on other sites

If the image represents a continuous function of some sort (such as a ramp or a wave), you could just procedurally evaluate it in the shader.

 

It is also possible to represent photographic images procedurally, if you store the image data's frequency coefficients over x and y (see Fast Fourier Transform). Depending on the image size, this can be very fast in the shader. You can even perform a principal component analysis before shader compilation to determine and select the dominant frequencies of the image, so as to save shader constants.

Edited by Nik02

Share this post


Link to post
Share on other sites

Which shader model are you compiling your code to? SM4.0 and up have some (limited) support for indexing into immediate constant buffers, but if I recall correctly SM3.0 has no such functionality. 

 

Generally SM 3.0, but SM 4.0 or 5.0 could be an option as well. Could you maybe provide an example or link about this "indexing into immediate constant buffers" thing?

 

 

 


If the image represents a continuous function of some sort (such as a ramp or a wave), you could just procedurally evaluate it in the shader.

It is also possible to represent photographic images procedurally, if you store the image data's frequency coefficients over x and y (see Fast Fourier Transform). Depending on the image size, this can be very fast in the shader. You can even perform a principal component analysis before shader compilation to determine and select the dominant frequencies of the image, so as to save shader constants.

 

I have cases of photographic images and others of procedural (such as perlin/simplex noise) which I need to "flatten" in order to gain performance.

 

As I said the transformation of the photo data into basically "numbers" is not the issue here as I already found tools for that, but my question was in what way / with what programming language constructs it would be best to save those numbers in my HLSL pixel shader code.

Edited by Meltac

Share this post


Link to post
Share on other sites

HLSL code generator

 

[source]

int numPixels = number of pixels in the original image data;
float pixels[numPixels]; // your original image data
string imageConsts = "float image[" + numPixels.toString() + "] = {";
for (int i = 0; i < numPixels;
++i)
{
  imageConsts += pixels.toString() + ", ";
}
imageConsts += "};";

[/source]

Edited by Nik02

Share this post


Link to post
Share on other sites

Thanks. So are you saying the best HLSL code structure for storing the image data would be a float array? Wouldn't that become quite expensive in terms of memory usage and compile time / run time performance?

Edited by Meltac

Share this post


Link to post
Share on other sites

Float is generally the most efficient data type for the shader units.

 

Of course, if you have SM4 or newer, you can get tighter packing with integers - but the unpacking logic itself uses some cycles. Unless my data is deprecated, integer ops get far less transistors than float ops.

 

If data bus bandwidth is your bottleneck, you can gain some performance using packed integers, though. It is best to profile your specific usage with these options.

Edited by Nik02

Share this post


Link to post
Share on other sites

If you want to see what the GPU does with the arrays to address them, you can inspect the compiled shader assembly by using fxc or shader reflection. This is useful when you want to compare (again at high level) the instruction cost of various approaches. That said, the driver can and will recompile the shader to GPU-specific optimized bytecode before it is actually run. Thus, fxc assembly can be seen as high-level metacode (even though it would be runnable as is).

 

Note that since the pixel shader is (likely) the most frequently run function in your program, you generally want to do as much work as possible at CPU or vertex level.

Edited by Nik02

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement