[Kunlun]add collective ops for multi XPU cards training and add Kunlun multi XPU cards CI

[vslyu]

PR types

New features

PR changes

OPs

Describe

• add collective ops for multi XPU cards training
  add allreduce_sum/prod/max/min and reduce_sum/prod/max/min ops for XPU.

example:
fleetrun --xpus 0,1 example.py

import numpy as np
import paddle
from paddle.distributed import init_parallel_env

paddle.set_device('xpu:%d'%paddle.distributed.ParallelEnv().dev_id)
init_parallel_env()
if paddle.distributed.ParallelEnv().local_rank == 0:
    np_data = np.array([[4, 5, 6], [4, 5, 6]]).astype(np.float32)
else:
    np_data = np.array([[1, 2, 3], [1, 2, 3]]).astype(np.float32)
data = paddle.to_tensor(np_data)
paddle.distributed.reduce(data, 0)
out = data.numpy()
# [[5, 7, 9], [5, 7, 9]]

• add Kunlun multi XPU cards CI
  add dist xpu tests in python/paddle/fluid/tests/unittests/CMakeLists.txt.
  UT name must contain "_xpu" to be executed, redefined example as following:

# dist xpu tests: 
if (WITH_XPU_BKCL)
    py_test(test_collective_reduce_api_xpu SRCS "test_collective_reduce_api.py")
    py_test(test_collective_allreduce_api_xpu SRCS "test_collective_allreduce_api.py")
endif()

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wangxicoding]

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)

[vslyu]

done～

[wangxicoding]

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)

[vslyu]

done～

[QingshuChen]

LGTM