TiledTensor / TiledCUDA

TiledCUDA/tests/cpp/cell/test_s2f_copy.cu

Lines 351 to 356 in 1043cbf

    
           using TemporalExecShared = TileShape<2, 1>; 
        
           using WarpLayout = TileShape<2, 2>; 
        
           using ThreadLayout = TileShape<16, 2>;  // fixed when using ldmatrix. 
        
           using ElemDataTileShared = TileShape<2, 16>; 
        
           // the final copy plan for accessing shared memory 
        
           using Shared = SharedTile<Element, TemporalExecShared, WarpLayout,

In the code snippet provided, a layout such as row-major or column-major is required to fully describe the copy plan. Currently, all existing implementations default to row-major.

	using TemporalExecShared = TileShape<2, 1>;
	using WarpLayout = TileShape<2, 2>;
	using ThreadLayout = TileShape<16, 2>; // fixed when using ldmatrix.
	using ElemDataTileShared = TileShape<2, 16>;
	// the final copy plan for accessing shared memory
	using Shared = SharedTile<Element, TemporalExecShared, WarpLayout,

`TileShape` is insufficient to fully describe a copy plan.