[FEA] Support different index/stride types

[cliffburdick]

index_t is currently used for both index and stride computations, and users select which one they want at compile time. 64b stride/index is the default. In most cases users would never need a 64b index, and it's wasted instructions using them. Most CUDA libraries don't even support 64b indexes and only strides anyways, so it doesn't make sense to allow this.

This feature is to create a second type offset_t that is used for stride calculations only, and the user can select which they want to use.

[MashrafeeAryan]

Hi @cliffburdick, is this issue still open? Can I work on it please?

[cliffburdick]

Hi @MashrafeeAryan, yes, you can look at it if you'd like.

[MashrafeeAryan]

Hi @cliffburdick, I noticed that the current index_t for indexes and strides is 64-bit by default, and if users want to switch to 32-bit, they usually build MatX with something like:

cmake -DMATX_BUILD_32_BIT=ON

I want to add a similar CMake option for offset_t so users can do something like MATX_BUILD_32_BIT_OFFSETS=ON to switch offsets to 32-bit.

My questions are:

1. May I use this name for the build configuration: MATX_BUILD_32_BIT_OFFSETS? Is there another name you would prefer?
2. To switch index_t to 32-bit, I am keeping the original name: MATX_BUILD_32_BIT. Should this option control only index_t after the change?
3. Should we change the default index_t type to 32-bit, since users will probably not need more than 32-bit indexes?

[cliffburdick]

Hi @MashrafeeAryan, the current state is somewhat of a mix where index_t is used for both sizes and strides. It should most likely just be used for sizes. Switching 32b on effectively sets both strides and sizes to 32b.

1. Yes, I would leave DMATX_BUILD_32_BIT, but change it to set two subflags for SIZES and STRIDES (or offsets?) to 32, but they could also set them both separately if they want.
2. What this really should be controlling is here

   MatX/include/matx/core/tensor_desc.h

   Lines 503 to 509 in a7c4a51

   	#ifdef MATX_INDEX_32_BIT
   	template <int RANK>
   	using DefaultDescriptor = tensor_desc_cr_ds_32_32_t<RANK>;
   	#else
   	template <int RANK>
   	using DefaultDescriptor = tensor_desc_cr_ds_64_64_t<RANK>;
   	#endif

   . This defines the strides and sizes separately, but right now they're tied together.
3. That would likely be breaking behavior at this point. I would need to think about anyone this impacts first.