DirectCompute Buffer

Started by
5 comments, last by Krohm 10 years, 1 month ago

Hm, when using a buffer of type int4 for my input to the compute shader, everything works fine. The shader resource view has its format set to R32G32B32A32_SInt.

But when I try to create a buffer of type int3 (R32G32B32_Sint), I get this error "Additional information: Attempted to read or write protected memory. This is often an indication that other memory is corrupt.".

So, is the int3 value somehow padded up to 16 bytes?

Or is there some other cause behind this?

Advertisement

Yep. Each element is separately padded up to 16 for c buffers
What kind of buffer are you using, (cbuffer, tbuffer, strctured buffer?), Because the padding rule is only applicable for constant buffers?

Yep. Each element is separately padded up to 16 for c buffers
What kind of buffer are you using, (cbuffer, tbuffer, strctured buffer?), Because the padding rule is only applicable for constant buffers?

hm, a t buffer, i think
Buffer<int3> input : register(t0);

No, that isn't a tbuffer - it looks like a structured buffer to me. If it is a structured buffer, then it isn't 16 byte aligned. Can you post the code for the buffer's description when you create it? That will clarify the type of the buffer, and also the size of the buffer which you can compare to the size of the CPU data that you are trying to load into it.

Also, do you know what operation you are performing when you get the error message?

//=============================================================================
// Buffer<int3> input : register(t0);

inputBuffer = new Buffer(device, new BufferDescription()
{
StructureByteStride = Int3.SizeInBytes,
SizeInBytes = N * Int3.SizeInBytes,
Usage = ResourceUsage.Dynamic, // the cpu can write to it.
BindFlags = BindFlags.ShaderResource, // will be read-only by the GPU.
CpuAccessFlags = CpuAccessFlags.Write,
OptionFlags = ResourceOptionFlags.None,
});

// srView to our unput buffer
var srViewDesc = new ShaderResourceViewDescription();
srViewDesc.Format = Format.R32G32B32_SInt;
srViewDesc.Dimension = ShaderResourceViewDimension.Buffer;
srViewDesc.Buffer.FirstElement = 0;
srViewDesc.Buffer.ElementCount = N;
ShaderResourceView srView = new ShaderResourceView(device, inputBuffer, srViewDesc);
//=============================================================================

The error appears the second time i call UnmapSubresource

(I have a makeshift 'loop', using a goto statement)

//=============================================================================

DataStream dataStream;
context.MapSubresource(inputBuffer, MapMode.WriteDiscard, SharpDX.Direct3D11.MapFlags.None, out dataStream);

for(int i = 0; i < N; i++)
dataStream.Write(input);

context.UnmapSubresource(inputBuffer, 0);
dataStream.Dispose();

//=============================================================================

It works with single integers, two integers, and 4. But with 3, it crashes on the second attempt to update the buffer.

p.s. yeah, this is in SharpDX, on a D3D10 card.

Ok, I did some tests.

So, there are two ways to define a "typed" buffer.

Buffer of a custom structure type:

C#

bufferDesc.OptionFlags = ResourceOptionFlags.BufferStructured;

srViewDesc.Format = Format.Unknown;

HLSL

StructuredBuffer<int3> input : register(t0);

Primitive type Buffer:

C#

bufferDesc.OptionsFlags = ResourceOptionFlags.None

srViewDesc.Format = Format.R32G32B32_SInt;

HLSL

Buffer<int3> input : register(t0);

The structured buffer works for every type I tried.

Using a primitive type buffer works for int, int2, int4, but not int3.

So my question is, why?

Also, is there any difference in the two types of buffers, other than the fact that the structured buffer lets you use custom types?


So my question is, why?
It is a good question! I have recently searched the net for a similar problem. The only rationale I've found which seemed fairly robust involved interactions with texturing hardware. Apparently, buffers are really meant to be "texture-like" and historically, 3-component textures never really made it to silicon.

I only report what I have read, but it does have indeed some sense.

Previously "Krohm"

This topic is closed to new replies.

Advertisement