Refactor norm_linear into struct functor

[NorthmanPKU]

Description of changes:
Refactor norm_linear kernel into struct functor style.

NormLinearKernelSpec: Maintain compile-time configuration and constants
ProcessAtomFunctor: Core computation logic for processing one single OUTPUT_ATOM_SIZE output tiles
NormLinearHandler: Top-level control flow and memory management. Later can have more fine-grained functions

Now the style to run norm_linear kernel is as:

using KernelSpec = NormLinearKernelSpec<T, BATCH_SIZE, OUTPUT_SIZE, REDUCTION_SIZE, O_STRIDE, K_PIPE_MAX>;
NormLinearHandler<KernelSpec> handler(input_ptr, norm_weight_ptr, weight_ptr, eps, output_ptr);
handler.run();

The refactored code has exactly the same register usage as the old one (123). Also, no efficiency difference has been observed:
ptxas info : Used 123 registers, used 1 barriers, 392 bytes cmem[0]

An imagination of what we could do in the future:

using KernelSpec = NormLinearKernelSpec<T, BATCH_SIZE, OUTPUT_SIZE, REDUCTION_SIZE, O_STRIDE, K_PIPE_MAX>;
NormLinearHandler<KernelSpec> handler(input_ptr, norm_weight_ptr, weight_ptr, eps, output_ptr);
handler.load_independent_data(); // inter-layer overlap
/* Some sync logic*/
handler.main_logic();

Related Issues:

Linked Issues:

• Issue #

Issues closed by this PR:

• Closes #

[jiazhihao]

Maybe this is because __CUDA_ARCH__ wasn't passed into this header file.

[NorthmanPKU]

Seems like __CUDA_ARCH__ can only be used in the implementation of GPU functions

[NorthmanPKU]

I think we could use cudaGetDeviceProperties in host side to get & set the maximum smem size. And use CUDA_ARCH in device side to do so.

[jiazhihao]

@NorthmanPKU Is this PR ready for review?

[NorthmanPKU]

@NorthmanPKU Is this PR ready for review?

Yes

[jiazhihao]

@NorthmanPKU Do we still want to merge this?