Compute

Compute pipelines process data outside a render pass. Record compute work on the compute or graphics queue according to the surrounding workload.

Create a Compute Pipeline

Define a Slang compute entry point and its thread-group size:

[shader("compute")]
[numthreads(16, 16, 1)]
void CSMain(uint3 dispatchThreadID : SV_DispatchThreadID)
{
    // Process one element or pixel.
}

Compile the entry point and create its Shader as described in Shaders, then create the pipeline:

ComputePipeline pipeline = context.CreateComputePipeline(new() { ComputeShader = shader });

Dispatch Work

Calculate group counts from the workload dimensions and the shader's [numthreads] values, rounding each dimension up to a complete group.

Bind the pipeline and constants, then dispatch:

CommandBuffer commandBuffer = context.ComputeQueue.CommandBuffer();

commandBuffer.SetPipeline(pipeline);
commandBuffer.SetConstantBuffer(buffer, 0);
commandBuffer.Dispatch(groupCountX, groupCountY, 1);

commandBuffer.Submit();

Guard out-of-range threads in the shader when group counts round up the workload dimensions.

Dispatch Indirectly

DispatchIndirect reads GroupCountX, GroupCountY, and GroupCountZ from an IndirectDispatchArgs record.

Create the argument buffer with BufferUsages.Indirect. If earlier GPU work writes the arguments, also add the appropriate storage usage and record a barrier before dispatch.

See Bindless Resources for passing buffers and textures to the shader. See Synchronization when a dispatch consumes or produces data for other GPU work.