Compute Shader - shared memory race condition when using optimization level 3

Started by
11 comments, last by Ronan Bel (Ubisoft) 12 years ago
Here's a very simple compute shader to illustrate the problem:

#define size 17
RWStructuredBuffer<float> OutputBuffer : register(u0);
groupshared float SharedPositions[size * size];
[numthreads(size, size, 1)]
void CS( uint3 GroupID : SV_GroupID, uint3 DispatchThreadID : SV_DispatchThreadID, uint3 GroupThreadID : SV_GroupThreadID, uint GroupIndex : SV_GroupIndex )
{
// Write Vertex Position to shared memory
SharedPositions[GroupIndex] = 1.0f;
// Wait for Shared Memory to be ready
GroupMemoryBarrierWithGroupSync();
OutputBuffer[GroupIndex] = SharedPositions[GroupIndex];
}


When compiling with D3D10_SHADER_OPTIMIZATION_LEVEL higher than 1, I get the following error:
error X3695: race condition writing to shared memory detected, consider making this write conditional.[/quote]

But since I write to "GroupIndex", no more than one thread will ever write to the same location.
What condition does the compiler require for this to compile properly?
Advertisement
*bump*

Can't seem to figure this one out.
GroupIndex is only unique inside the thread group. So if there are multiple groups you end up with a race condition for the output buffer.
Yes, but the groupshared memory is unique to each thread group, which is where the error is occurring. (or so I assume because it specifically says "race condition writing to shared memory")
My example is obviously not making much sense since each thread group would write to the same location of OutputBuffer,
I just quickly threw something simple together to illustrate my issue, my actual compute shader takes this into consideration.
Is the error on writing to the shared memory, or writing to the output resource?
It's definitely when writing to shared memory.
I completely removed any writing to output resources:

// Shared memory
groupshared float SharedPositions[1];
[numthreads(1, 1, 1)]
void CS( uint3 GroupID : SV_GroupID, uint3 DispatchThreadID : SV_DispatchThreadID, uint3 GroupThreadID : SV_GroupThreadID, uint GroupIndex : SV_GroupIndex )
{
// Write something to shared memory
SharedPositions[GroupIndex] = 1.0f;
// Wait for Shared Memory to be ready
GroupMemoryBarrierWithGroupSync();
}


The error still occurs:
ErrorTest.hlsl(2,19): error X3695: race condition writing to shared memory detected, consider making this write conditional.[/quote]

Line 2 is: groupshared float SharedPositions[1];
Character 19 is SharedPositions[1];
What the hell....
this errors as well:

groupshared float SharedPositions[1];
[numthreads(1, 1, 1)]
void CS( )
{
}


Basically, as soon as I define shared memory and use an optimization flag of 2 or higher when compiling, it errors.
Could this be a bug in the shader compiler?
I just tried compiling your shader with fxc.exe using /O3, and it compiles fine. Are you using cs_5_0?
I just found this link - would you happen to be using the app verifier???

I just tried compiling your shader with fxc.exe using /O3, and it compiles fine. Are you using cs_5_0?

It is working fine with fxc for me as well.
The problem only occurs with D3DX11CompileFromFile().
I am using cs_5_0.


I just found this link - would you happen to be using the app verifier???


This appear to be the exact same problem I'm experiencing.
However, I am not knowingly using app verifier.
It appears to be some sort of plugin for visual studio? If so I'm definitely not using it.

This topic is closed to new replies.

Advertisement