- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
When profiling the PlaidML benchmarks, I found that the same kernel was appearing many times, but each time followed by a unique suffix in the form of __k### where ### is a number. The same kernel would appear with a different __k suffix, but when examining the ISA code generated by the compiler, only one version of the kernel is compiled. For example in densenet169, kernel_c149_sdk_7 appears with 82 different names. Each name has kernel_c149_sdk_7 in it, but is followed by __k### with different numbers.
What is the significance of the __k## suffix? Are these kernels the same or is there different assembly generated for them?
Here is the list of kernels called by densenet169 for reference: https://docs.google.com/spreadsheets/d/1F0qq--nHB4sI3_sk77J3sfjkgYoHuFXX0XELrNImSWA/edit#gid=990823576
I simply counted the occurrences of each unique kernel name, starting at row 398 you can see the kernel_c149_sdk_7 I am referring to.
Thank you for your time.
- Tags:
- Intel® Advanced Vector Extensions (Intel® AVX)
- Intel® Streaming SIMD Extensions
- Parallel Computing
Link Copied
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page