[MLU]add mlu kernel for allreduce (#39788)

paddle/fluid/operators/collective/c_allreduce_op.h

@@ -24,7 +24,8 @@ limitations under the License. */
 #include "paddle/fluid/platform/device/npu/npu_op_runner.h"
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL) ||          \
-    defined(PADDLE_WITH_ASCEND_CL) || defined(PADDLE_WITH_XPU_BKCL)
+    defined(PADDLE_WITH_ASCEND_CL) || defined(PADDLE_WITH_XPU_BKCL) || \
+    defined(PADDLE_WITH_CNCL)
 #include "paddle/fluid/platform/collective_helper.h"
 #endif
 
@@ -45,6 +46,10 @@ limitations under the License. */
 #include "paddle/fluid/platform/device/npu/hccl_helper.h"
 #endif
 
+#if defined(PADDLE_WITH_CNCL)
+#include "paddle/fluid/platform/device/mlu/cncl_helper.h"
+#endif
+
 #if defined(PADDLE_WITH_ASCEND_CL)
 DECLARE_bool(hccl_check_nan);
 #endif
@@ -398,6 +403,65 @@ class CAllReduceOpCUDAKernel : public framework::OpKernel<T> {
   }
 };
 
+template <ReduceType red_type, typename T>
+class CAllReduceOpMLUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+#if defined(PADDLE_WITH_CNCL)
+    auto in = ctx.Input<framework::Tensor>("X");
+    auto out = ctx.Output<framework::Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+    cnclDataType_t dtype =
+        platform::ToCNCLDataType(framework::TransToProtoVarType(in->type()));
+    int64_t numel = in->numel();
+    const void* sendbuff = in->data<T>();
+    out->Resize(in->dims());
+    void* recvbuff = out->mutable_data<T>(place);
+
+    int rid = ctx.Attr<int>("ring_id");
+    auto comm = platform::CNCLCommContext::Instance().Get(rid, place);
+
+    mluStream stream = nullptr;
+    if (ctx.Attr<bool>("use_calc_stream")) {
+      auto dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      stream = static_cast<platform::MLUDeviceContext*>(dev_ctx)->stream();
+    } else {
+      stream = comm->stream();
+    }
+
+    cnclReduceOp_t cncl_red_type = cnclSum;
+    switch (red_type) {
+      case kRedSum:
+        cncl_red_type = cnclSum;
+        break;
+
+      case kRedMax:
+        cncl_red_type = cnclMax;
+        break;
+
+      case kRedMin:
+        cncl_red_type = cnclMin;
+        break;
+
+      case kRedProd:
+        cncl_red_type = cnclProd;
+        break;
+
+      default:
+        PADDLE_THROW(platform::errors::InvalidArgument(
+            "Invalid reduce type: %d", red_type));
+    }
+
+    PADDLE_ENFORCE_MLU_SUCCESS(cnclAllReduce(
+        sendbuff, recvbuff, numel, dtype, cncl_red_type, comm->comm(), stream));
+#else
+    PADDLE_THROW(platform::errors::PreconditionNotMet(
+        "PaddlePaddle should compile with MLU."));
+#endif
+  }
+};
+
 class CAllReduceOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {

paddle/fluid/operators/collective/c_allreduce_sum_op_mlu.cc

+/* Copyright (c) 2022 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/fluid/operators/collective/c_allreduce_op.h"
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_MLU_KERNEL(c_allreduce_sum,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, float>,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, plat::float16>,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, int>,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, int16_t>,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, int8_t>,
+                       ops::CAllReduceOpMLUKernel<ops::kRedSum, uint8_t>)

paddle/fluid/operators/collective/c_broadcast_op_mlu.cc

@@ -30,7 +30,8 @@ class CBroadcastOPMLUKernel : public framework::OpKernel<T> {
     auto x = ctx.Input<framework::LoDTensor>("X");
     auto out = ctx.Output<framework::LoDTensor>("Out");
     int numel = x->numel();
-    cnclDataType_t dtype = platform::ToCNCLDataType(x->type());
+    cnclDataType_t dtype =
+        platform::ToCNCLDataType(framework::TransToProtoVarType(x->type()));
 
     int rid = ctx.Attr<int>("ring_id");
     auto place = ctx.GetPlace();

python/paddle/fluid/tests/unittests/mlu/CMakeLists.txt

@@ -6,4 +6,5 @@ if (WITH_MLU)
         py_test_modules(${TEST_OP} MODULES ${TEST_OP})
     endforeach(TEST_OP)
     set_tests_properties(test_collective_broadcast PROPERTIES TIMEOUT 120)
+    set_tests_properties(test_collective_allreduce PROPERTIES TIMEOUT 120)
 endif()

python/paddle/fluid/tests/unittests/mlu/collective_allreduce_op.py

+# Copyright (c) 2022 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.
+
+from __future__ import print_function
+
+import numpy as np
+import argparse
+import os
+import sys
+import signal
+import time
+import socket
+from contextlib import closing
+from six import string_types
+import math
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.profiler as profiler
+import paddle.fluid.unique_name as nameGen
+from paddle.fluid import core
+import unittest
+from multiprocessing import Process
+import paddle.fluid.layers as layers
+from functools import reduce
+from test_collective_base_mlu import TestCollectiveRunnerBase, runtime_main
+
+paddle.enable_static()
+
+
+class TestCollectiveAllreduce(TestCollectiveRunnerBase):
+    def __init__(self):
+        self.global_ring_id = 0
+
+    def get_model(self, main_prog, startup_program):
+        ring_id = 0
+        with fluid.program_guard(main_prog, startup_program):
+            tindata = layers.data(
+                name="tindata", shape=[10, 1000], dtype='float32')
+            toutdata = main_prog.current_block().create_var(
+                name="outofallreduce",
+                dtype='float32',
+                type=core.VarDesc.VarType.LOD_TENSOR,
+                persistable=False,
+                stop_gradient=False)
+            main_prog.global_block().append_op(
+                type="c_allreduce_sum",
+                inputs={'X': tindata},
+                attrs={'ring_id': ring_id},
+                outputs={'Out': toutdata})
+            main_prog.global_block().append_op(
+                type="c_sync_comm_stream",
+                inputs={'X': toutdata},
+                outputs={'Out': toutdata},
+                attrs={'ring_id': ring_id})
+            return toutdata
+
+
+if __name__ == "__main__":
+    runtime_main(TestCollectiveAllreduce, "allreduce", 0)

python/paddle/fluid/tests/unittests/mlu/test_collective_allreduce.py

+#   Copyright (c) 2022 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.
+
+from __future__ import print_function
+import sys
+import unittest
+import numpy as np
+import paddle
+
+from test_collective_base_mlu import TestDistBase
+
+paddle.enable_static()
+
+
+class TestCAllreduceOp(TestDistBase):
+    def _setup_config(self):
+        pass
+
+    def test_allreduce_fp32(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce",
+                              "float32")
+
+    def test_allreduce_fp16(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce",
+                              "float16")
+
+    def test_allreduce_int32(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce",
+                              "int32")
+
+    def test_allreduce_int16(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce",
+                              "int16")
+
+    def test_allreduce_int8(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce", "int8")
+
+    def test_allreduce_uint8(self):
+        self.check_with_place("collective_allreduce_op.py", "allreduce",
+                              "uint8")
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/mlu/test_collective_base_mlu.py

@@ -262,5 +262,13 @@ class TestDistBase(unittest.TestCase):
             need_result = input2
             self.assertTrue(np.allclose(tr0_out, need_result))
             self.assertTrue(np.allclose(tr1_out, need_result))
+        elif col_type == "allreduce":
+            need_result = input1 + input2
+            self.assertTrue(
+                np.allclose(
+                    tr0_out, need_result, rtol=1e-05, atol=1e-05))
+            self.assertTrue(
+                np.allclose(
+                    tr1_out, need_result, rtol=1e-05, atol=1e-05))
         else:
             pass