Refactor stridingMemcpyKernel to be unroll friendly

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.

[Artem-B]

While loop over incrementing pointer is harder for compiler to analyze. Switching to directly indexed pointer accesses gives compiler more flexibility over unrolling the loop.

Also, work around clang-related issue which results in slow code when aggregates are swapped via a temporary variable. Copying uint4 element-wise avoids the problem, and works fine with NVCC.

[Artem-B]

Comparison between clang/nvcc old/new kernel: https://godbolt.org/z/onojWhr7a
Old kernel with clang shows the issue (local pipe_* variables are not eliminated), but it could be worked around by using __int128 type for the transfers via passing -Di128=__int128 to the clang compilation in the compiler explorer link above.