已合并
feat(coo_ops): adapt 5 COO operators for Ascend NPU via CSR composition #22
lhp_lhp创建于 8月27日
feat(coo_ops): adapt 5 COO operators for Ascend NPU via CSR composition #22
已合并
Pull Request已成功合入, 合并人@knightgd
（感谢 lhp_lhp 的贡献）src/array/ascend/coo_ops.cc
已过期
@@ -0,0 +238,20 @@
238+template <typename IdType>
239+std::pair<bool, bool> COOIsSorted_Ascend(COOMatrix coo) {
240+ // Use metadata flags. If both flags say sorted, trust them.
241+ if (coo.row_sorted && coo.col_sorted) {
242+ return {true, true};
243+ }
244+ // When the metadata does not guarantee both sorted (row_sorted false or
245+ // col_sorted false), we cannot trust the flags alone: the data may still be
246+ // genuinely sorted but simply not flagged (e.g. is_sorted_srcdst always
247+ // constructs a COOMatrix with row_sorted=false). Fall back to a CPU scan
248+ // (D2H of the small index arrays) to report the true sorted state.
249+ //
250+ // COOIsSorted is called infrequently and returns only a bool pair, so the
251+ // CPU fallback is the pragmatic choice (mirrors the row_sorted-only branch).
252+ DGLContext cpu_ctx{kDGLCPU, 0};
253+ aclrtSynchronizeDevice();
254+ COOMatrix coo_cpu{
255+ coo.num_rows, coo.num_cols,
256+ coo.row.CopyTo(cpu_ctx), coo.col.CopyTo(cpu_ctx),
257+ coo.data.CopyTo(cpu_ctx), coo.row_sorted, coo.col_sorted};
258+ return COOIsSorted<kDGLCPU, IdType>(coo_cpu);
系统消息
9月4日 评论：
9月4日 评论：
changed this line on c981ac7a view diff detail
src/array/ascend/coo_ops.cc
已过期
@@ -0,0 +275,19 @@
275+template <typename IdType>
276+bool COOIsNonZero_Ascend(COOMatrix coo, int64_t row, int64_t col) {
277+ CSRMatrix csr = COOToCSR(coo);
278+ // Construct the single-element query arrays on CPU (host-accessible) and copy
279+ // to the NPU. Directly dereferencing the raw device pointer returned by
280+ // NDArray::data (e.g. `static_cast<IdType*>(row_arr->data)[0] = row`) is host
281+ // pointer aliasing of device memory -- undefined behaviour under the ACL
282+ // unified-memory model. Building on CPU + CopyTo mirrors the established
283+ // pattern in this file (EnsureCOODataForCSR) and coo2csr.cc.
284+ DGLContext cpu_ctx{kDGLCPU, 0};
285+ NDArray row_arr = NDArray::Empty({1}, coo.row->dtype, cpu_ctx);
286+ NDArray col_arr = NDArray::Empty({1}, coo.col->dtype, cpu_ctx);
287+ static_cast<IdType*>(row_arr->data)[0] = static_cast<IdType>(row);
288+ static_cast<IdType*>(col_arr->data)[0] = static_cast<IdType>(col);
289+ row_arr = row_arr.CopyTo(coo.row->ctx);
290+ col_arr = col_arr.CopyTo(coo.col->ctx);
291+ NDArray data = CSRGetData<kDGLAscend, IdType>(csr, row_arr, col_arr);
292+ IdType val = static_cast<IdType*>(data.CopyTo(cpu_ctx)->data)[0];
293+ return val != -1;
294+}
系统消息
9月4日 评论：
9月4日 评论：
changed this line on c981ac7a view diff detail
27 天前 修改标题为 “feat(coo_ops): adapt 5 COO operators for Ascend NPU via CSR composition”，原标题为“feat(coo_ops): adapt 7 COO operators for Ascend NPU via CSR composition”
Description
为 Ascend 910B3 (dav-2201) 适配 5 个 COO 算子的 NPU 原生路径，替换原有的 CPU 回退（D2H→CPU→H2D）和 LOG(FATAL) 崩溃。
适配策略
通过组合已有 Ascend 原生 CSR 算子实现：
COO → COOToCSR → CSR op → CSRToCOO（与csr_transpose.cc同范式），无需编写新的 AscendC kernel。适配算子
支持的数据类型
COO 算子为整数索引操作（无浮点特征），只涉及 IdType，不涉及 DType：
Checklist
Changes
Test Results
Excluded (Follow-up)