| 64 | |
| 65 | template<typename T> |
| 66 | void sortBatched(Param pVal, int dim, bool isAscending) { |
| 67 | af::dim4 inDims; |
| 68 | for (int i = 0; i < 4; i++) inDims[i] = pVal.info.dims[i]; |
| 69 | |
| 70 | // Sort dimension |
| 71 | // tileDims * seqDims = inDims |
| 72 | af::dim4 tileDims(1); |
| 73 | af::dim4 seqDims = inDims; |
| 74 | tileDims[dim] = inDims[dim]; |
| 75 | seqDims[dim] = 1; |
| 76 | |
| 77 | // Create/call iota |
| 78 | // Array<uint> pKey = createEmptyArray<uint>(inDims); |
| 79 | Array<uint> pKey = iota<uint>(seqDims, tileDims); |
| 80 | |
| 81 | pKey.setDataDims(inDims.elements()); |
| 82 | |
| 83 | // Flat |
| 84 | pVal.info.dims[0] = inDims.elements(); |
| 85 | pVal.info.strides[0] = 1; |
| 86 | for (int i = 1; i < 4; i++) { |
| 87 | pVal.info.dims[i] = 1; |
| 88 | pVal.info.strides[i] = pVal.info.strides[i - 1] * pVal.info.dims[i - 1]; |
| 89 | } |
| 90 | |
| 91 | // Sort indices |
| 92 | // sort_by_key<T, uint, isAscending>(*resVal, *resKey, val, key, 0); |
| 93 | // kernel::sort0_by_key<T, uint, isAscending>(pVal, pKey); |
| 94 | compute::command_queue c_queue(getQueue()()); |
| 95 | |
| 96 | compute::buffer pKey_buf((*pKey.get())()); |
| 97 | compute::buffer pVal_buf((*pVal.data)()); |
| 98 | |
| 99 | compute::buffer_iterator<type_t<T>> val0 = |
| 100 | compute::make_buffer_iterator<type_t<T>>(pVal_buf, 0); |
| 101 | compute::buffer_iterator<type_t<T>> valN = |
| 102 | compute::make_buffer_iterator<type_t<T>>(pVal_buf, +pVal.info.dims[0]); |
| 103 | compute::buffer_iterator<uint> key0 = |
| 104 | compute::make_buffer_iterator<uint>(pKey_buf, 0); |
| 105 | compute::buffer_iterator<uint> keyN = |
| 106 | compute::make_buffer_iterator<uint>(pKey_buf, pKey.dims()[0]); |
| 107 | if (isAscending) { |
| 108 | compute::sort_by_key(val0, valN, key0, c_queue); |
| 109 | } else { |
| 110 | compute::sort_by_key(val0, valN, key0, compute::greater<type_t<T>>(), |
| 111 | c_queue); |
| 112 | } |
| 113 | |
| 114 | // Needs to be ascending (true) in order to maintain the indices properly |
| 115 | // kernel::sort0_by_key<uint, T, true>(pKey, pVal); |
| 116 | compute::sort_by_key(key0, keyN, val0, c_queue); |
| 117 | |
| 118 | CL_DEBUG_FINISH(getQueue()); |
| 119 | } |
| 120 | |
| 121 | template<typename T> |
| 122 | void sort0(Param val, bool isAscending) { |
nothing calls this directly
no test coverage detected