【Hackathon 6th Fundable Projects 3 No.97】fluid operator dgc

paddle/fluid/operators/CMakeLists.txt

@@ -82,7 +82,7 @@ endif()
 
 set(OP_HEADER_DEPS ${OP_HEADER_DEPS} phi common phi_utils static_prim_api get_expected_kernel_func)
 
-register_operators(EXCLUDES py_func_op dgc_op generated_op1 generated_op2 generated_op3 generated_op4 load_combine_op lstm_op run_program_op quantize_linear_op
+register_operators(EXCLUDES py_func_op generated_op1 generated_op2 generated_op3 generated_op4 load_combine_op lstm_op run_program_op quantize_linear_op
         recurrent_op save_combine_op sync_batch_norm_op activation_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS} processgroup_comm_utils)
 
 op_library(generated_op UNITY SRCS generated_op1.cc generated_op2.cc generated_op3.cc generated_op4.cc DEPS ${OP_HEADER_DEPS})
@@ -119,7 +119,6 @@ endif()
 set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
 
 if (WITH_DGC)
-    op_library(dgc_op DEPS dgc)
     set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dgc)
 endif()
 

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

@@ -476,14 +476,6 @@
     func: dequantize_log
     data_type: x
 
-- op : dgc
-  args : (Tensor u, Tensor v, Tensor grad, Tensor param, Tensor current_step, Tensor nranks, float[] sparsity, float m=0.9, bool use_nesterov=true, float rampup_begin_step=0.0, float rampup_step=0.0, float regular_coeff=0.0, int regular_type=0)
-  output : Tensor(u_out), Tensor(v_out), Tensor(encode_grad), Tensor(grad_out), Tensor(k), Tensor(gather_buff)
-  kernel :
-    func : dgc
-    param : [u, v, grad, param, current_step, nranks, m, use_nesterov, sparsity, rampup_begin_step, rampup_step, regular_coeff, regular_type]
-  optional: param
-
 - op : dgc_momentum
   args : (Tensor param, Tensor grad, Tensor velocity, Tensor learning_rate, Tensor master_param, Tensor current_step_tensor, Tensor nranks_tensor, float mu, bool use_nesterov=false, str regularization_method="", float regularization_coeff=0.0f, bool multi_precision=false, float rescale_grad=1.0f, float rampup_begin_step=-1.0f)
   output : Tensor(param_out), Tensor(velocity_out), Tensor(master_param_out), Tensor(grad_out)

paddle/phi/api/yaml/op_compat.yaml

@@ -3921,6 +3921,12 @@
   outputs:
     {u_out: U_out, v_out: V_out, encode_grad: EncodeGrad, grad_out: Grad_out, gather_buff: GatherBuff}
 
+- op: dgc
+  inputs:
+    {u : U, v : V, grad : Grad, param : Param}
+  outputs:
+    {u_out : U_out, v_out : V_out, encode_grad : EncodeGrad, grad_out : Grad_out, gather_buff : GatherBuff}
+
 - op: distribute_fpn_proposals
   inputs :
     {fpn_rois: FpnRois, rois_num: RoisNum}

paddle/phi/api/yaml/ops.yaml

@@ -881,6 +881,19 @@
     data_type: detect_res
   optional: has_state, pos_count, true_pos, false_pos
 
+- op : dgc
+  args : (Tensor u, Tensor v, Tensor grad, Tensor param, Tensor current_step, Tensor nranks, float m=0.9, bool use_nesterov=true, float[] sparsity={}, float rampup_begin_step=0.0, float rampup_step=0.0, float regular_coeff=0.0, int regular_type=0)
+  output : Tensor(u_out), Tensor(v_out), Tensor(encode_grad), Tensor(grad_out), Tensor(k), Tensor(gather_buff)
+  infer_meta:
+    func: DgcInferMeta
+    param : [u, v, grad, param, current_step, nranks]
+  kernel :
+    func : dgc
+    param : [u, v, grad, param, current_step, nranks, m, use_nesterov, sparsity, rampup_begin_step, rampup_step, regular_coeff, regular_type]
+  optional: param
+  data_transform :
+    skip_transform : current_step, nranks
+
 - op : diag
   args : (Tensor x, int offset = 0, float padding_value = 0.0)
   output : Tensor

paddle/phi/infermeta/multiary.cc

@@ -1661,6 +1661,19 @@ void DetectionMapInferMeta(const MetaTensor& detect_res,
   m_ap->set_dims(common::make_ddim({1}));
 }
 
+void DgcInferMeta(const MetaTensor& u,
+                  const MetaTensor& v,
+                  const MetaTensor& grad,
+                  const MetaTensor& param,
+                  const MetaTensor& current_step_tensor,
+                  const MetaTensor& nranks_tensor,
+                  MetaTensor* u_out,
+                  MetaTensor* v_out,
+                  MetaTensor* encode_grad_out,
+                  MetaTensor* grad_out,
+                  MetaTensor* k_out,
+                  MetaTensor* gather_buff) {}
+
 void DGCMomentumInferMeta(const MetaTensor& param,
                           const MetaTensor& grad,
                           const MetaTensor& velocity,

paddle/phi/infermeta/multiary.h

@@ -331,6 +331,19 @@ void DetectionMapInferMeta(const MetaTensor& detect_res,
                            MetaTensor* m_ap,
                            MetaConfig config = MetaConfig());
 
+void DgcInferMeta(const MetaTensor& u,
+                  const MetaTensor& v,
+                  const MetaTensor& grad,
+                  const MetaTensor& param,
+                  const MetaTensor& current_step_tensor,
+                  const MetaTensor& nranks_tensor,
+                  MetaTensor* u_out,
+                  MetaTensor* v_out,
+                  MetaTensor* encode_grad_out,
+                  MetaTensor* grad_out,
+                  MetaTensor* k_out,
+                  MetaTensor* gather_buff);
+
 void DGCMomentumInferMeta(const MetaTensor& param,
                           const MetaTensor& grad,
                           const MetaTensor& velocity,