Searched refs:ElementInputB (Results 1 – 3 of 3) sorted by relevance
| /external/pytorch/aten/src/ATen/native/sparse/cuda/ |
| D | SparseSemiStructuredLinear.cu | 41 typename ElementInputB,
69 constexpr int AlignmentInputB = 128 / cutlass::sizeof_bits<ElementInputB>::value;
149 ElementInputB,
240 cutlass::TensorRef<ElementInputB, LayoutInputB>(
241 (ElementInputB*)tensor_b.data_ptr(), layout_b);
314 typename ElementInputB,
342 ElementInputB,
364 ElementInputB,
386 ElementInputB,
408 ElementInputB,
[all …]
|
| D | SparseSemiStructuredOps.cu | 38 typename ElementInputB,
64 constexpr int AlignmentInputB = 128 / cutlass::sizeof_bits<ElementInputB>::value;
141 ElementInputB,
205 cutlass::TensorRef<ElementInputB, LayoutInputB>(
206 (ElementInputB*)tensor_b.data_ptr(), layout_b);
302 typename ElementInputB,
330 ElementInputB,
355 ElementInputB,
380 ElementInputB,
405 ElementInputB,
[all …]
|
| /external/pytorch/aten/src/ATen/native/cuda/ |
| D | MixedDtypesLinear.cu | 38 template<typename ElementInputA, typename ElementInputB, typename EpilogueTag>
57 constexpr auto ElementsPerCacheLine = 128 * 8 / cutlass::sizeof_bits<ElementInputB>::value;
65 constexpr auto ElementsPerAccessB = 128 / cutlass::sizeof_bits<ElementInputB>::value;
89 ElementInputB,
121 {(ElementInputB*)weight.data_ptr(), ldb},
156 template<typename ElementInputA, typename ElementInputB>
165 ElementInputB,
176 ElementInputB,
181 ElementInputB,
186 ElementInputB,
|