What's the meaning of `Shared Memory executed` in nsys

Developer Tools Nsight Systems Profiling Linux Targets
1

Hi,

Here is a result in nsys gui when I click one kernel:

Begins: 8.28491s
Ends: 8.35452s (+69.615 ms)
grid: <<<64, 64, 1>>>
block: <<<16, 16, 1>>>
Launch Type: Regular
Static Shared Memory: 8,192 bytes
Dynamic Shared Memory: 0 bytes
Registers Per Thread: 127
Local Memory Per Thread: 0 bytes
Local Memory Total: 205,258,752 bytes
Shared Memory executed: 65,536 bytes
Shared Memory Bank Size: 4 B
Theoretical occupancy: 25 %
Launched from thread: 1363913
Latency: ←642.541 ms
Correlation ID: 263
Stream: Default stream 7

What’s the meaning of Shared Memory executed and what’s the difference between it and Static Shared Memory?

2

@skottapalli

3

“Static Shared Memory” refers to the amount of shared memory that is statically allocated for a kernel during compilation.

“Shared Memory executed” is Shared memory size set by the driver.

Nsight Systems gets the information about kernels from CUPTI. Please see 6.46. CUpti_ActivityKernel8 — Cupti 13.0 documentation for more details on the information reported by CUPTI for each kernel.

4

This is the shared memory carve out size executed by the driver. This may differ from the requested size. This is not the thread block shared memory size.

5

Thanks! @skottapalli @Greg

Yes, I found that the ‘Static Shared Memory’ metric is consistent with the size I allocated in my code, while ‘Shared Memory executed’ is much larger.