I am trying to run my metal program on my iPhone SE.
I tried many numbers for threadsPerThreadGroup and threadsPerGrid sizes and all of them gave me this error: TLValidateFeatureSupport:3539: failed assertion `Dispatch Threads with Non-Uniform Threadgroup Size is only supported on MTLGPUFamilyApple4 and later.'
Here is my code.
var threadsPerThreadGroup: MTLSize
var threadsPerGrid: MTLSize
computeCommandEncoder.setComputePipelineState(updateShader)
let w = updateShader.threadExecutionWidth
threadsPerThreadGroup = MTLSize(width: w, height: 1, depth: 1)
threadsPerGrid = MTLSize(width: Int(constants.bufferLength), height: 1, depth: 1)
if(frames % 2 == 0) {
computeCommandEncoder.setBuffer(buffer1, offset: 0, index: 0)
computeCommandEncoder.setBuffer(buffer2, offset: 0, index: 1)
} else {
computeCommandEncoder.setBuffer(buffer2, offset: 0, index: 0)
computeCommandEncoder.setBuffer(buffer1, offset: 0, index: 1)
}
computeCommandEncoder.setBytes(&constants, length: MemoryLayout<MyConstants>.stride, index: 2)
computeCommandEncoder.dispatchThreads(threadsPerGrid, threadsPerThreadgroup: threadsPerThreadGroup)
frames += 1
I am using iOS 13.4 and XCode 11.4.
threadExecutionWidth evaluates to 32 and constants.bufferLength is 512.
Use [dispatchThreads] only if the device supports non-uniform threadgroup sizes.
That is not worded as clearly as it could be. It means that dispatchThreads
does not work on pre-A11 GPUs.
If you want a solution that works on all devices, you have to calculate how many threadgroups go into a grid yourself, and use dispatchThreadgroups.
If you want to have both methods in your code, you can detect the device's feature set at runtime.