PaddlePaddle
diff --git a/‎paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc
Copy file name to clipboard
-181Lines changed: 0 additions & 181 deletions b/‎paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc
Copy file name to clipboard
-181Lines changed: 0 additions & 181 deletions
diff --git a/‎paddle/fluid/operators/fused/fusion_seqpool_concat_op.h
Copy file name to clipboard
-38Lines changed: 0 additions & 38 deletions b/‎paddle/fluid/operators/fused/fusion_seqpool_concat_op.h
Copy file name to clipboard
-38Lines changed: 0 additions & 38 deletions
diff --git a/‎paddle/phi/infermeta/fusion.cc
Copy file name to clipboard
+43Lines changed: 43 additions & 0 deletions b/‎paddle/phi/infermeta/fusion.cc
Copy file name to clipboard
+43Lines changed: 43 additions & 0 deletions
diff --git a/‎paddle/phi/infermeta/fusion.h
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions b/‎paddle/phi/infermeta/fusion.h
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/fusion_seqpool_concat_kernel.cc
Copy file name to clipboard
+100Lines changed: 100 additions & 0 deletions b/‎paddle/phi/kernels/cpu/fusion_seqpool_concat_kernel.cc
Copy file name to clipboard
+100Lines changed: 100 additions & 0 deletions
diff --git a/‎paddle/phi/ops/yaml/fused_ops.yaml
Copy file name to clipboard
+10Lines changed: 10 additions & 0 deletions b/‎paddle/phi/ops/yaml/fused_ops.yaml
Copy file name to clipboard
+10Lines changed: 10 additions & 0 deletions
@@ -5069,4 +5069,47 @@ void FusedSeqpoolCvmGradInferMeta(
   }
 }
 
+void FusionSeqpoolConcatInferMeta(const std::vector<const MetaTensor*>& x,
+                                  const std::string& pooltype,
+                                  int axis,
+                                  MetaTensor* out,
+                                  MetaConfig config) {
+  PADDLE_ENFORCE_GE(x.size(),
+                    1UL,
+                    phi::errors::InvalidArgument(
+                        "Inputs(X) of FusionSeqPoolConcatOp should be greater "
+                        "than 1, but received value is %d.",
+                        x.size()));
+  PADDLE_ENFORCE_EQ(
+      axis,
+      1,
+      phi::errors::InvalidArgument("FusionSeqPoolConcatOp only supports concat "
+                                   "axis=1 yet, but received axis value is %d",
+                                   axis));
+  std::vector<DDim> ins_dims;
+  ins_dims.reserve(x.size());
+  std::transform(x.begin(),
+                 x.end(),
+                 std::back_inserter(ins_dims),
+                 [](const MetaTensor* var) { return var->dims(); });
+  const size_t n = ins_dims.size();
+  PADDLE_ENFORCE_GT(n,
+                    0UL,
+                    phi::errors::InvalidArgument(
+                        "Input tensors count should be greater than 0, "
+                        "but received value is %d.",
+                        n));
+
+  // The output height should be confirmed in Compute,
+  // since input lod is not accessible here.
+  PADDLE_ENFORCE_EQ(ins_dims[0].size(),
+                    2,
+                    phi::errors::InvalidArgument(
+                        "The dims size of first input should be equal to 2, "
+                        "but received value is %d.",
+                        ins_dims[0].size()));
+  out->set_dims({-1, ins_dims[0][axis] * static_cast<int>(n)});
+  out->set_dtype(x[0]->dtype());
+}
+
 }  // namespace phi
@@ -1127,4 +1127,10 @@ void FusedSeqpoolCvmGradInferMeta(
     MetaTensor* cvm_grad,
     MetaConfig config = MetaConfig());
 
+void FusionSeqpoolConcatInferMeta(const std::vector<const MetaTensor*>& x,
+                                  const std::string& pooltype,
+                                  int axis,
+                                  MetaTensor* out,
+                                  MetaConfig config = MetaConfig());
+
 }  // namespace phi
@@ -0,0 +1,100 @@
+// Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <string>
+#include <vector>
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/core/tensor_utils.h"
+#include "paddle/phi/kernels/funcs/jit/kernels.h"
+
+namespace phi {
+template <typename T, typename Context>
+void FusionSeqPoolConcatKernel(const Context& dev_ctx,
+                               const std::vector<const DenseTensor*>& x,
+                               const std::string& pooltype,
+                               int axis,
+                               DenseTensor* out) {
+  auto ins = x;
+
+  auto x0_lod = ins[0]->lod();
+  const auto& x0_dims = ins[0]->dims();
+  const auto& y_dims = out->dims();
+  size_t bs = x0_lod[0].size() - 1;
+  out->Resize({static_cast<int64_t>(bs), y_dims[1]});
+  phi::LoD y_lod(1);
+  y_lod[0].resize(bs + 1);
+  for (size_t i = 0; i <= bs; ++i) {
+    y_lod[0][i] = i;
+  }
+  out->set_lod(y_lod);
+  T* y_data = dev_ctx.template Alloc<T>(out);
+
+  int w = static_cast<int>(ins[0]->numel() / x0_dims[0]);
+  PADDLE_ENFORCE_EQ(y_dims[1] % w,
+                    0,
+                    phi::errors::InvalidArgument(
+                        "The output of dims[1] should be dividable of w, but "
+                        "dims[1] is %d, w is %d.",
+                        y_dims[1],
+                        w));
+  phi::jit::seq_pool_attr_t attr(w, phi::jit::SeqPoolType::kSum);
+  if (pooltype == "AVERAGE") {
+    attr.type = phi::jit::SeqPoolType::kAvg;
+  } else if (pooltype == "SQRT") {
+    attr.type = phi::jit::SeqPoolType::kSqrt;
+  }
+  auto seqpool =
+      phi::jit::KernelFuncs<phi::jit::SeqPoolTuple<T>, phi::CPUPlace>::Cache()
+          .At(attr);
+  size_t n = ins.size();
+  size_t dst_step_size = n * w;
+  for (size_t i = 0; i < n; ++i) {
+    const auto& x_dims = ins[i]->dims();
+    auto x_lod = ins[i]->lod()[0];
+    const T* src = ins[i]->data<T>();
+    T* dst = y_data + i * w;
+    PADDLE_ENFORCE_EQ(
+        static_cast<int>(ins[i]->numel() / x_dims[0]),
+        w,
+        phi::errors::InvalidArgument(
+            "Width of all inputs should be equal, but the width of the %d-th "
+            "input %d is not equal to the previous %d",
+            i,
+            static_cast<int>(ins[i]->numel() / x_dims[0]),
+            w));
+    PADDLE_ENFORCE_EQ(
+        x_lod.size(),
+        bs + 1,
+        phi::errors::InvalidArgument(
+            "Batchsize of all inputs should be equal, but the value of the "
+            "%d-th %d is not equal to the previous %d.",
+            i,
+            x_lod.size(),
+            bs + 1));
+    for (size_t j = 0; j < bs; ++j) {
+      attr.h = static_cast<int>(x_lod[j + 1] - x_lod[j]);
+      seqpool(src, dst, &attr);
+      dst += dst_step_size;
+      src += attr.h * attr.w;
+    }
+  }
+}
+
+}  // namespace phi
+PD_REGISTER_KERNEL(fusion_seqpool_concat,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::FusionSeqPoolConcatKernel,
+                   float,
+                   double) {}
@@ -521,6 +521,16 @@
   optional : fc_bias
   intermediate : fc_out
 
+- op : fusion_seqpool_concat
+  args: (Tensor[] x, str pooltype = "SUM", int axis = 1)
+  output: Tensor (out)
+  infer_meta:
+    func: FusionSeqpoolConcatInferMeta
+  kernel:
+    func: fusion_seqpool_concat
+    data_type: x
+  support_dygraph_mode : true
+
 - op : fusion_seqpool_cvm_concat
   args: (Tensor[] x, Tensor cvm, str pooltype = "SUM", bool use_cvm = true, int axis = 1)
   output: Tensor (out)