[QST]Why Does CUTLASS Use 3-4-3 Swizzle?

[ziyuhuang123]

May I kindly ask why the swizzle configuration in CUTLASS is specifically set to 3, 4, and 3? I would greatly appreciate any insights or explanations regarding the rationale behind this design choice. Thank you so much in advance!

// K-major GMMA layouts in units of bits
using Layout_K_INTER_Atom_Bits  = ComposedLayout<Swizzle<0,4,3>, smem_ptr_flag, Layout<Shape<_8, _128>,Stride< _128,_1>>>;
using Layout_K_SW32_Atom_Bits   = ComposedLayout<Swizzle<1,4,3>, smem_ptr_flag, Layout<Shape<_8, _256>,Stride< _256,_1>>>;
using Layout_K_SW64_Atom_Bits   = ComposedLayout<Swizzle<2,4,3>, smem_ptr_flag, Layout<Shape<_8, _512>,Stride< _512,_1>>>;
using Layout_K_SW128_Atom_Bits  = ComposedLayout<Swizzle<3,4,3>, smem_ptr_flag, Layout<Shape<_8,_1024>,Stride<_1024,_1>>>;

[leimao]

I have my own interpretations.

[ziyuhuang123]

I have my own interpretations.

Thank you so much! Your post has been incredibly helpful! I actually studied L2 persistent before and remember coming across your post at that time. I'm part of a very active NV WeChat discussion group. I was wondering if you might be interested in joining us to explore CUDA technologies together? My WeChat ID is hermit_purple1.