Shader storage buffer bug in Catalyst 13.4
AMD added support for OpenGL 4.3 in their latest 13.4 driver but it seems they have a bug on their implementation of Shader Buffer Objects. I have a shader that increments an atomic counter and uses it as an index to write to a Shader Buffer Object buffer that hold points data. The buffer is big enough to old 100.000 points but the shader is only writting the first 1696 points.
I can see this because after running the shader I map the atomic counter and the point buffer to look at their contents. The point buffer is only filled to the first 1696 positions although the atomic counter was incremented beyond that value. It seems the shader is assuming that the array has only 1696 elements, this is confirmed by the bokehPoints.length() that I'm writting into the buffer which returns precisely 1696. The extension specification says that we can query the size of the attached shader buffer object with glGetInteger64i_v(GL_SHADER_STORAGE_BUFFER_SIZE, 1, &value) which in my case it returns the expected value of 3.200.000 which represents sizeof(Point) * 100000 and not sizeof(Point) * 1696.
So to sum up, the shader is assuming that the array is smaller than it actually is. Note that I have no problems running this stuff on NVidia cards so I'd like to know if anyone is experienced the same problems with AMD or if it's me who is doing something wrong.
This is how the buffer is created:
glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(Point) * 100000, NULL, GL_DYNAMIC_DRAW);
These are the rendering commands:
//Reset the value of the atomic counter and bind it.
GLuint resetValue = 0;
glNamedBufferSubDataEXT(depthOfField.getBokehPointsAtomicCounter(), 0, sizeof(GLuint), &resetValue);
glBindBufferBase(GL_ATOMIC_COUNTER_BUFFER, 0, 0);
//Bind the points buffer.
glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, bokehPointsBufferID);
//Draw full screen quad.
//NOTE: I know glBegin/glEnd is evil.
glVertex3f(0, 0, 0);
glVertex3f(1, 0, 0);
glVertex3f(1, 1, 0);
glVertex3f(0, 1, 0);
glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT | GL_ATOMIC_COUNTER_BARRIER_BIT);
Declaration of the atomic counter and shader buffer object in the shader:
layout (binding = 0, offset = 0) uniform atomic_uint bokehPointCounter;
layout (packed, binding = 1) buffer BokehPointBuffer
writeonly BokehPoint bokehPoints;
And this is how I write to the buffer in the shader:
uint index = atomicCounterIncrement(bokehPointCounter);
bokehPoints[index].position = vec4(texCoord, 0.0, float(bokehPoints.length()));
bokehPoints[index].colorAndRadius = vec4(color * scale, radius);