fix(kernel): 为 Gather 支持负的 indices 值

Signed-off-by: YdrMaster <[email protected]>

Author: YdrMaster

src/04kernel/cuda/src/gather.cu

@@ -20,9 +20,10 @@ namespace refactor::kernel::cuda {
              tid += step) {
             auto i = tid / batch,
                  j = tid % batch;
-            auto index = __ldg(indices + i % midSizeO);
+            auto k = __ldg(indices + i % midSizeO);
+            auto quot = k >= 0 ? i / midSizeO : i / midSizeO + 1;
             optimizedMemcpy(unit * tid + output,
-                            unit * (batch * (i / midSizeO * midSizeI + index) + j) + data,
+                            unit * (batch * (quot * midSizeI + k) + j) + data,
                             unit);
         }
     }

src/04kernel/src/kernels/gather/cpu_kernel.cc

@@ -33,8 +33,9 @@ namespace refactor::kernel {
                 int64_t k = info.idxType == DataType::I64
                                 ? reinterpret_cast<int64_t const *>(inputs[1])[d.rem]
                                 : reinterpret_cast<int32_t const *>(inputs[1])[d.rem];
+                auto quot = k >= 0 ? d.quot : d.quot + 1;
                 std::memcpy(info.postfix * i + output,
-                            info.postfix * (d.quot * info.midSizeI + k) + data,
+                            info.postfix * (quot * info.midSizeI + k) + data,
                             info.postfix);
             });
         };