【PIR Dist Op Reg No.2】 reg distributed_push_sparse

paddle/fluid/pir/dialect/op_generator/ops_api_gen.py

@@ -124,6 +124,7 @@
     'c_reducescatter',
     'c_softmax_with_cross_entropy',
     'decayed_adagrad',
+    'distributed_push_sparse',
     'distributed_lookup_table',
     'dpsgd',
     'embedding_grad_sparse',

paddle/fluid/pir/dialect/operator/ir/ops.yaml

@@ -396,6 +396,15 @@
     func : distributed_lookup_table
     data_type : dtype
 
+- op : distributed_push_sparse
+  args : (Tensor[] ids, Tensor[] shows, Tensor[] clicks, int table_id = 0, int size = 8, bool is_distributed = false, str push_sparse_version = "push_sparse", int64_t padding_idx = -1, DataType dtype=DataType::FLOAT32, bool is_test = false, bool use_cvm_op = false)
+  output : Tensor[](output){ids.size()}
+  infer_meta :
+    func : DistributedPushSparseInferMeta
+  kernel :
+    func: distributed_push_sparse
+    data_type : dtype
+
 - op : divide
   args : (Tensor x, Tensor y)
   output : Tensor(out)

paddle/fluid/pir/dialect/operator/utils/utils.cc

@@ -41,6 +41,7 @@ const std::unordered_set<std::string> LegacyOpList = {
     CBroadcast_Op::name(),
     CSyncCalcStream_Op::name(),
     CSyncCommStream_Op::name(),
+    DistributedPushSparseOp::name(),
     FtrlOp::name(),
     FusedElemwiseAddActivationOp::name(),
     FusedElemwiseAddActivationGradOp::name(),

paddle/phi/api/yaml/op_compat.yaml

@@ -819,8 +819,10 @@
     out : Out
 
 - op : distributed_push_sparse
-  extra :
-    attrs : ['int[] slots = {}']
+  inputs :
+    {ids : Ids, shows : Shows, clicks: Clicks}
+  outputs :
+    out : Outputs
 
 - op : divide (elementwise_div)
   backward : divide_grad (elementwise_div_grad)

paddle/phi/infermeta/ternary.cc

@@ -255,6 +255,32 @@ void BoxCoderInferMeta(const MetaTensor& prior_box,
   output_box->set_dtype(target_box.dtype());
 }
 
+void DistributedPushSparseInferMeta(
+    const std::vector<const MetaTensor*>& ids,
+    const std::vector<const MetaTensor*>& shows,
+    const std::vector<const MetaTensor*>& clicks,
+    int table_id,
+    int size,
+    bool is_distributed,
+    const std::string& push_sparse_version,
+    int64_t padding_idx,
+    DataType dtype,
+    bool is_test,
+    bool use_cvm_op,
+    std::vector<MetaTensor*> output) {
+  auto ids_dims = ids[0]->dims();
+  PADDLE_ENFORCE_EQ(ids_dims.size(),
+                    2,
+                    phi::errors::InvalidArgument(
+                        "The dimension of the 'Ids' tensor must be 2."));
+  for (auto& out : output) {
+    if (out == nullptr) {
+      continue;
+    }
+    out->set_dtype(ids[0]->dtype());
+  }
+}
+
 void DpsgdInferMeta(const MetaTensor& param,
                     const MetaTensor& grad,
                     const MetaTensor& learning_rate,

paddle/phi/infermeta/ternary.h

@@ -63,6 +63,20 @@ void BoxCoderInferMeta(const MetaTensor& prior_box,
                        MetaTensor* output_box,
                        MetaConfig config = MetaConfig());
 
+void DistributedPushSparseInferMeta(
+    const std::vector<const MetaTensor*>& ids,
+    const std::vector<const MetaTensor*>& shows,
+    const std::vector<const MetaTensor*>& clicks,
+    int table_id,
+    int size,
+    bool is_distributed,
+    const std::string& push_sparse_version,
+    int64_t padding_idx,
+    DataType dtype,
+    bool is_test,
+    bool use_cvm_op,
+    std::vector<MetaTensor*> output);
+
 void DpsgdInferMeta(const MetaTensor& param,
                     const MetaTensor& grad,
                     const MetaTensor& learning_rate,