add all_to_all phi operator (#54797)

* add all_to_all phi operator, kernel, api

* add all_to_all ut

* tinyfix

paddle/phi/api/yaml/static_ops.yaml

@@ -26,6 +26,16 @@
     func : all_reduce
     param: [x, reduce_type]
 
+- op : all_to_all
+  args : (Tensor x, int ring_id = 0)
+  output : Tensor(out)
+  infer_meta :
+    func : AllToAllInferMeta
+    param: [x]
+  kernel :
+    func : all_to_all
+    param: [x]
+
 - op : amax
   args : (Tensor x, IntArray axis={0}, bool keepdim=false, bool reduce_all=false, int in_dtype=-1, int out_dtype=-1)
   output : Tensor(out)

paddle/phi/infermeta/unary.cc

@@ -136,6 +136,13 @@ void AllReduceInferMeta(const MetaTensor& x, MetaTensor* out) {
   out->set_dims(x.dims());
 }
 
+void AllToAllInferMeta(const MetaTensor& x, MetaTensor* out) {
+  auto dim = x.dims();
+  if (dim[0] < 0) dim[0] = -1;
+  out->set_dtype(x.dtype());
+  out->set_dims(dim);
+}
+
 void ArgMinMaxInferMeta(const MetaTensor& x,
                         const Scalar& axis,
                         bool keepdims,

paddle/phi/infermeta/unary.h

@@ -43,6 +43,8 @@ void AllGatherInferMeta(const MetaTensor& x, int nranks, MetaTensor* out);
 
 void AllReduceInferMeta(const MetaTensor& x, MetaTensor* out);
 
+void AllToAllInferMeta(const MetaTensor& x, MetaTensor* out);
+
 void ArgMinMaxInferMeta(const MetaTensor& x,
                         const Scalar& axis,
                         bool keepdims,

paddle/phi/kernels/all_to_all_kernel.h

@@ -0,0 +1,26 @@
+// Copyright (c) 2023 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
+//
+//     http://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.
+
+#pragma once
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void AllToAllKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    DenseTensor* out);
+
+}  // namespace phi

paddle/phi/kernels/cpu/all_to_all_kernel.cc

@@ -0,0 +1,43 @@
+// Copyright (c) 2023 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
+//
+//     http://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 "paddle/phi/kernels/all_to_all_kernel.h"
+
+#include "paddle/phi/backends/all_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void AllToAllKernel(const Context& dev_ctx UNUSED,
+                    const DenseTensor& x UNUSED,
+                    DenseTensor* out UNUSED) {
+  PADDLE_THROW(
+      errors::Unimplemented("Unimplemented cpu kernel for all_to_all."));
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(all_to_all,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::AllToAllKernel,
+                   float,
+                   double,
+                   int,
+                   bool,
+                   int8_t,
+                   uint8_t,
+                   int64_t,
+                   phi::dtype::float16) {}

paddle/phi/kernels/gpu/all_to_all_kernel.cu

@@ -0,0 +1,114 @@
+// Copyright (c) 2023 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
+//
+//     http://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 "paddle/phi/kernels/all_to_all_kernel.h"
+#include "glog/logging.h"
+
+#include "paddle/phi/backends/all_context.h"
+#include "paddle/phi/core/kernel_registry.h"
+
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+#include "paddle/phi/core/distributed/nccl_comm_context.h"
+#include "paddle/phi/core/distributed/utils.h"
+#endif
+
+namespace phi {
+
+template <typename T, typename Context>
+void AllToAllKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    DenseTensor* out) {
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+#if NCCL_VERSION_CODE >= 2703
+  auto x_dims = x.dims();
+  out->Resize(x_dims);
+  dev_ctx.template Alloc<T>(out);
+
+  auto comm_ctx =
+      static_cast<distributed::NCCLCommContext*>(dev_ctx.GetCommContext());
+  PADDLE_ENFORCE_NE(
+      comm_ctx,
+      nullptr,
+      errors::Unavailable("NCCLCommContext is nullptr, collective op should "
+                          "has ring_id attr."));
+  gpuStream_t stream = dev_ctx.stream();
+  PADDLE_ENFORCE_NOT_NULL(stream,
+                          errors::NotFound("Should initialize NCCL firstly."));
+
+  int nranks = comm_ctx->GetSize();
+  int send_numel = x.numel() / nranks;
+  size_t offset = 0;
+
+  PADDLE_ENFORCE_EQ(
+      x_dims[0] % nranks,
+      0,
+      errors::InvalidArgument(
+          "The first dimension size (%d) of the input tensor must be "
+          "divisible by the number of ranks (%d).",
+          x_dims[0],
+          nranks));
+
+  comm_ctx->GroupStart();
+
+  const auto* send_buf = x.data<T>();
+  auto* recv_buf = out->data<T>();
+  for (auto i = 0; i < nranks; ++i) {
+    auto send_buf = phi::distributed::GetPartialTensor(x, offset, send_numel);
+    comm_ctx->Send(send_buf, send_numel, i, stream);
+    auto recv_buf =
+        phi::distributed::GetPartialTensor(*out, offset, send_numel);
+    comm_ctx->Recv(&recv_buf, send_numel, i, stream);
+    offset += send_numel;
+  }
+  comm_ctx->GroupEnd();
+#else
+  PADDLE_THROW(
+      platform::errors::Unavailable("NCCL version >= 2.7.3 is needed."));
+#endif
+#else
+  PADDLE_THROW(
+      errors::PreconditionNotMet("PaddlePaddle should compile with GPU."));
+#endif
+}
+
+}  // namespace phi
+
+#if NCCL_VERSION_CODE >= 21000 && CUDA_VERSION >= 11000
+PD_REGISTER_KERNEL(all_to_all,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::AllToAllKernel,
+                   float,
+                   double,
+                   int,
+                   int8_t,
+                   uint8_t,
+                   int64_t,
+                   bool,
+                   phi::dtype::bfloat16,
+                   phi::dtype::float16) {}
+#else
+PD_REGISTER_KERNEL(all_to_all,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::AllToAllKernel,
+                   float,
+                   double,
+                   int,
+                   int8_t,
+                   uint8_t,
+                   int64_t,
+                   bool,
+                   phi::dtype::float16) {}
+#endif

test/collective/collective_alltoall_api.py

@@ -18,11 +18,86 @@
 )
 
 import paddle
-from paddle import fluid
+import paddle.distributed as dist
+from paddle import fluid, framework
+from paddle.fluid import data_feeder
 
 paddle.enable_static()
 
 
+def alltoall_new(
+    in_tensor_or_tensor_list,
+    out_tensor_or_tensor_list,
+    group=None,
+    sync_op=True,
+):
+    op_type = 'all_to_all'
+
+    ring_id = 0 if group is None else group.id
+    nranks = dist.get_world_size()
+    helper = framework.LayerHelper(op_type, **locals())
+
+    in_tensor = in_tensor_or_tensor_list
+    if isinstance(in_tensor_or_tensor_list, list):
+        if len(in_tensor_or_tensor_list) == 0:
+            raise RuntimeError("The input tensor_list should not be empty.")
+        # 0-D use stack/unstack while others use concat/split
+        if len(in_tensor_or_tensor_list[0].shape) == 0:
+            in_tensor = paddle.stack(in_tensor_or_tensor_list, axis=0)
+        else:
+            in_tensor = paddle.concat(in_tensor_or_tensor_list, axis=0)
+
+    out_tensor = out_tensor_or_tensor_list
+    if isinstance(out_tensor_or_tensor_list, list):
+        if len(out_tensor_or_tensor_list) != 0:
+            raise ValueError(
+                "The 'out_tensor_list' for all_to_all " "must be an empty list."
+            )
+        out_tensor = helper.create_variable_for_type_inference(
+            dtype=in_tensor.dtype
+        )
+
+    data_feeder.check_variable_and_dtype(
+        in_tensor,
+        'in_tensor',
+        [
+            'float16',
+            'float32',
+            'float64',
+            'int32',
+            'int64',
+            'int8',
+            'uint8',
+            'bool',
+            'uint16',
+        ],
+        'all_to_all',
+    )
+    helper.append_op(
+        type=op_type,
+        inputs={'x': [in_tensor]},
+        outputs={'out': [out_tensor]},
+        attrs={
+            'ring_id': ring_id,
+        },
+    )
+    # NOTE(liyurui): If the argument `out_tensor_or_tensor_list` is a tensor_list,
+    # we need to split the result. So we should wait the result of all_to_all
+    # before split if the communication is not on calc stream.
+    if isinstance(out_tensor_or_tensor_list, list):
+        if not sync_op:
+            dist.wait(out_tensor, use_calc_stream=False)
+        # 0-D use stack/unstack while others use concat/split
+        if len(in_tensor_or_tensor_list[0].shape) == 0:
+            out_tensor_or_tensor_list.extend(paddle.unstack(out_tensor, 0))
+        else:
+            out_tensor_or_tensor_list.extend(
+                paddle.split(out_tensor, nranks, 0)
+            )
+
+    return None
+
+
 class TestCollectiveAllToAllAPI(TestCollectiveAPIRunnerBase):
     def __init__(self):
         self.global_ring_id = 0
@@ -38,6 +113,19 @@ def get_model(self, main_prog, startup_program, rank):
             paddle.distributed.alltoall(tindata, tout_data)
             return tout_data
 
+    def get_model_new(
+        self, main_prog, startup_program, rank, dtype=None, reduce_type=None
+    ):
+        with fluid.program_guard(main_prog, startup_program):
+            tindata = paddle.static.data(
+                name="tindata", shape=[-1, 10, 1000], dtype=dtype
+            )
+            tindata.desc.set_need_check_feed(False)
+            tindata = paddle.split(tindata, 2, axis=0)
+            tout_data = []
+            alltoall_new(tindata, tout_data)
+            return tout_data
+
 
 if __name__ == "__main__":
     runtime_main(TestCollectiveAllToAllAPI, "alltoall")

test/collective/test_collective_alltoall_api.py

@@ -28,6 +28,21 @@ def _setup_config(self):
     def test_alltoall_nccl(self):
         self.check_with_place("collective_alltoall_api.py", "alltoall", "nccl")
 
+    def test_alltoall_nccl_with_comm_context(self):
+        dtypes_to_test = [
+            "float32",
+        ]
+        if self._nccl_version >= 21000:
+            dtypes_to_test.append("bfloat16")
+        for dtype in dtypes_to_test:
+            self.check_with_place(
+                "collective_alltoall_api.py",
+                "alltoall",
+                "nccl",
+                dtype=dtype,
+                need_envs={"USE_COMM_CONTEXT": "1"},
+            )
+
     def test_alltoall_nccl_dygraph(self):
         dtypes_to_test = [
             "float16",