Using SurfaceFormat.Single and HLSL for GPGPU with XNA

Question

I'm trying to implement a so-called ping-pong technique in XNA; you basically have two RenderTarget2D A and B and at each iteration you use one as texture and the other as target - and vice versa - for a quad rendered through an HLSL pixel shader.

• step1: A--PS-->B
• step2: B--PS-->A
• step3: A--PS-->B

...

In my setup, both RenderTargets are SurfaceFormat.Single.

In my .fx file, I have a tachnique to do the update, and another to render the "current buffer" to the screen.

Before starting the "ping-pong", buffer A is filled with test data with SetData<float>(float[]) function: this seems to work properly, because if I render a quad on the screen through the "Draw" pixel shader, i do see the test data being correctly rendered.

However, if i do update buffer B, something does not function proerly and the next rendering to screen will be all black.

For debug purposes, i replaced the "Update" HLSL pixel shader with one that should simply copy buffer A into B (or B into A depending on which among "ping" and "pong" phases we are...). From some examples i found on the net, i see that in order to correctly fetch a float value from a texture sampler from HLSL code, i should only need to care for the red channel.

So, basically the debug "Update" HLSL function is:

float4 ComputePS(float2 inPos  : TEXCOORD0) : COLOR0
{
  float v1 = tex2D(bufSampler, inPos.xy).r;
  return float4(v1,0,0,1);
}

which still doesn't work and results in a all-zeroes ouput. Here's the "Draw" function that seems to properly display initial data:

float4 DrawPS(float2 inPos  : TEXCOORD0) : COLOR0
{
  float v1 = tex2D(bufSampler, inPos.xy).r;
  return float4(v1,v1,v1,1);
}

Now: playing around with HLSL doesn't change anything, so maybe I'm missing something on the c# side of this, so here's the infamous Update() function:

_effect.Parameters["bufTexture"].SetValue(buf[_currentBuf]);
_graphicsDevice.SetRenderTarget(buf[1 - _currentBuf]);
_graphicsDevice.Clear(Color.Black); // probably not needed since RenderTargetUsage is DiscardContents
_effect.CurrentTechnique = _computeTechnique;
_computeTechnique.Passes[0].Apply();
_quadRender.Render();
_graphicsDevice.SetRenderTarget(null);
_currentBuf = 1 - _currentBuf;

Any clue?

Answer 1

Just found the answer myself, and it was a newbie issue.
Turns out the vertex shader was shifting the wole rendering that in a couple of updates, quickly exited the area visible when drawing results on the screen.

To fix this and have every pixel of the texture buffer aligned with its matching pixel on the RenderTarget buffer, the drawn quad has to be from [-1,-1] to [1,1] (and not from [0,0] to [1,1] as one could be tricked to think) and following vertex transform is needed:

output.Pos = input.Pos + float4(-0.5f/renderTargetSize.x, 0.5f/renderTargetSize.y, 0, 0);

so that Pos points exactly to the middle of a pixel/texel.