How many 32-bit variables can I use concurrently in a single thread of a Metal compute kernel without worrying about the variables getting spilled into the device memory? Alternatively: how many 32-bit registers does a single thread have available for itself?
Let's say that each thread of my compute kernel needs to store and work with its own array of N float variables, where N can be 128, 256, 512 or more. To achieve maximum possible performance, I do not want to the local thread variables to get spilled into the slow device memory. I want all N variables to be stored "on-chip", in the thread memory space.
To make my question more concrete, let's say there is an array thread float localArray[N]. Assuming an unrealistic hypothetical scenario where localArray is the only variable in the whole kernel, what is the maximum value of N for which no portion of localArray would get spilled into the device memory?
I searched in the Metal feature set tables, but I could not find any details.
Hello,
You'll want to consider maxThreadgroupMemoryLength:
"The maximum threadgroup memory available to a compute kernel, in bytes."
Likewise recommendedMaxWorkingSetSize.
Spilled bytes are shown in a GPU trace with the Metal Debugger.
We recommend Metal Compute on MacBook Pro to learn more.
