[GPU] Logging cleanup

[echeresh]

Jira: https://jira.devtools.intel.com/browse/MFDNN-11400.

List of changes:

• Moved logging functionality to src/gpu/intel/logging.hpp with some slight refactoring (constant -> enum class)
  • Renamed ir_{warning,perf,trace,...} to gpu-prefixed versions
• Reworked logging format: see src/gpu/intel/logging.hpp
  • Example: [TRACE][ codegen.cpp:95] codegen:bind h_34 -> r18 - r19
  • Added level tag and file name/line number information
• Simplified logging implementation, mostly minor things:
  • Removed ir_check related not-relly-important features as base_logger_t, fatal and dynamic log levels, IR check guard
  • Added automatic new-line after every log message
• Dropped ir_assert (and similar macros), renamed usages to the existing gpu_assert, etc

Things left for the future:

• Documentation
• New log level: debug (or maybe refactored info/trace combination). Now "trace" is used for both very detailed tracing and some useful/more high-level logging. There should be more granularity, e.g.:
  • Trace: max detail level, logging after every IR pass and even from inside of some passes
  • Debug/info: less details but enough to understand what kernel looks like, e.g. kernel parameters + final IR

[rjoursler]

Simplifying log levels, e.g. drop perf-specific levels as they are not really used AFAIK (?)

I have often used this when trying to optimize IR creation time. The main issue is that the trace level introduces overhead (~30% from what I remember) and this overhead was not evenly distributed. This makes is so that performance in trace mode only loosely corresponds to performance in release mode when prioritizing optimizations. Could we add perf information (perhaps in a more compressed format) to the info level, analogous to how oneDNN verbose maps into spdlog levels?

[echeresh]

Simplifying log levels, e.g. drop perf-specific levels as they are not really used AFAIK (?)

I have often used this when trying to optimize IR creation time. The main issue is that the trace level introduces overhead (~30% from what I remember) and this overhead was not evenly distributed. This makes is so that performance in trace mode only loosely corresponds to performance in release mode when prioritizing optimizations. Could we add perf information (perhaps in a more compressed format) to the info level, analogous to how oneDNN verbose maps into spdlog levels?

Thanks. Yes, combining it with info might be a good option. Let's keep it as is for now. I plan to open one more PR, will think about this later.

[echeresh]

make test
set test_scope=NIGHTLY
disable test_device_cpu
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lp
enable arch_gpu_xe-lpg
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg
enable arch_gpu_xe3-lpg