guanxinq commented on a change in pull request #17569: Adding sparse support to
MXTensor for custom operators
URL: https://github.com/apache/incubator-mxnet/pull/17569#discussion_r387829780
##########
File path: include/mxnet/lib_api.h
##########
@@ -1193,19 +1312,57 @@ extern "C" {
const int64_t** outshapes, int* outdims, void**
outdata, int* outtypes,
size_t* outIDs, const char** outdev_type, int*
outdev_id, int num_out,
xpu_malloc_t cpu_malloc, void* cpu_alloc,
- xpu_malloc_t gpu_malloc, void* gpu_alloc, void*
stream) {
+ xpu_malloc_t gpu_malloc, void* gpu_alloc, void*
stream,
+ void** in_indices, void** in_indptr,
+ int64_t* in_indices_shapes, int64_t*
in_indptr_shapes,
+ std::vector<std::vector<float>>& tmp_data,
+ std::vector<std::vector<int64_t>>& col_idx,
+ std::vector<std::vector<int64_t>>& row_ptr) {
// create a vector of tensors for inputs
std::vector<MXTensor> inputs(num_in);
+ // create a vector for sparse inputs
+ std::vector<MXInSparse> in_sparse(num_in);
+
for (int i = 0; i < num_in; i++) {
- inputs[i].setTensor(indata[i], (MXDType)intypes[i], inshapes[i],
indims[i],
- inIDs[i], {indev_type[i], indev_id[i]});
+ // Dense representation.
+ if(!in_indices_shapes) {
+ inputs[i].setTensor(indata[i], (MXDType)intypes[i], inshapes[i],
indims[i],
+ inIDs[i], {indev_type[i], indev_id[i]},
kDefaultStorage);
+ }
+ // Sparse representation.
+ else {
+ MXStorageType type;
+ if(!in_indptr_shapes) {
+ type = kRowSparseStorage;
+ in_sparse[i].set(indata[i], inshapes[i], indims[i], in_indices[i],
in_indices_shapes[i]);
Review comment:
As what we discussed, we will assume all the operators are in the same
format. If we really need to support mixed format for an operator, we need to
add more test and update the code. Otherwise, we can add a check and throw
error here.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
