Adding back memory check

[ivarusic-amd]

Added an option to disable the hipMemGetInfo memory check on Windows when building MGX. By default, the check is enabled, but it can be turned off if desired.

[pfultz2]

Whats the reason to disable this memory check?

[ivarusic-amd]

We had a similar check on Windows in the past, when we discovered that hipMemGetInfo returns ~70–80% of the available GPU memory. We believe there may be a regression in this function as few models start falling because of memory, and we also need to decide whether we want to enable it.

[pfultz2]

Dont add a cmake option. I dont think we need this. Just define this to true when windows is used:

#ifdef MIGRAPHX_DISABLE_AVAILABLE_GPU_MEMORY_CHECK
#ifdef _WIN32
#define MIGRAPHX_DISABLE_AVAILABLE_GPU_MEMORY_CHECK 1
#else
#define MIGRAPHX_DISABLE_AVAILABLE_GPU_MEMORY_CHECK 0
#endif
#endif

[ivarusic-amd]

This macro is intended for internal use only, to quick test specific cases. It will not be activated in production on Windows. The build option allows us to decide whether to include or exclude the memory check code during compilation, we don't want it always ON.

[pfultz2]

OK if its internal then definitely dont add cmake option, just do:

#ifdef MIGRAPHX_DISABLE_AVAILABLE_GPU_MEMORY_CHECK
#define MIGRAPHX_DISABLE_AVAILABLE_GPU_MEMORY_CHECK 0
#endif

You can set the option in cmake by passing it to the CMAKE_CXX_FLAGS.

[apwojcik]

@pfultz2 @causten We received an explanation from the HIP team that using hipMemGetInfo in this context is incorrect. The primary intent of the function is to track memory usage trends, rather than determining whether a subsequent hipMalloc call will succeed, especially on APUs. They asked us not to abort applications based solely on a hipMemGetInfo result. Recently, various frameworks have been changing that behaviour, too.
We want to check only if hipMalloc succeeded or failed, and permanently ignore the result of hipMemGetInfo.