[MLU] add compile ci scripts for MLU, test=mlu_ci (#39122)

paddle/fluid/distributed/index_dataset/CMakeLists.txt

 proto_library(index_dataset_proto SRCS index_dataset.proto)
 cc_library(index_wrapper SRCS index_wrapper.cc DEPS index_dataset_proto fs)
 if(WITH_MKLDNN)
-  cc_library(index_sampler SRCS index_sampler.cc DEPS xxhash index_wrapper mkldnn)
+  cc_library(index_sampler SRCS index_sampler.cc DEPS xxhash index_wrapper eigen3 mkldnn)
 else()
-  cc_library(index_sampler SRCS index_sampler.cc DEPS xxhash index_wrapper)
+  cc_library(index_sampler SRCS index_sampler.cc DEPS xxhash index_wrapper eigen3)
 endif()
 if(WITH_PYTHON)
   py_proto_compile(index_dataset_py_proto SRCS index_dataset.proto)

paddle/fluid/operators/concat_op_mlu.cc

@@ -45,7 +45,8 @@ class ConcatMLUKernel : public framework::OpKernel<T> {
         ins_dims[i] = ins[i]->dims();
       }
 
-      framework::DDim out_dims = ComputeAndCheckShape(true, ins_dims, axis);
+      framework::DDim out_dims =
+          pten::funcs::ComputeAndCheckShape(true, ins_dims, axis);
       out->Resize(out_dims);
     }
     const int axis_t = axis;

paddle/fluid/operators/fill_constant_op_mlu.cc

@@ -63,7 +63,8 @@ class FillConstantMLUKernel : public framework::OpKernel<T> {
       framework::Tensor mlu_tensor;
       auto tmp_place = value_tensor->place();
       if (platform::is_mlu_place(tmp_place)) {
-        TensorCopySync(*value_tensor, platform::CPUPlace(), &mlu_tensor);
+        framework::TensorCopySync(*value_tensor, platform::CPUPlace(),
+                                  &mlu_tensor);
         tensor_data = mlu_tensor.data<T>();
       }
       value = tensor_data[0];

paddle/fluid/operators/mlu/CMakeLists.txt

 
 IF(WITH_MLU)
-    cc_library(mlu_baseop SRCS mlu_baseop.cc DEPS neuware_lib framework_proto xxhash)
+    cc_library(mlu_baseop SRCS mlu_baseop.cc DEPS neuware_lib device_context)
     cc_test(activation_op_mlu_test SRCS activation_op_mlu_test.cc DEPS op_registry activation_op scope device_context executor)
 ENDIF()

paddle/fluid/operators/split_op_mlu.cc

@@ -58,7 +58,6 @@ class SplitMLUKernel : public framework::OpKernel<T> {
     std::vector<void*> vct_tensor;
     std::vector<MLUCnnlTensorDesc> output_descs;
     std::vector<cnnlTensorDescriptor_t> desc_vector;
-    auto place = ctx.GetPlace();
     for (size_t i = 0; i < outs.size(); i++) {
       outs[i]->mutable_data<T>(ctx.GetPlace());
       output_descs.emplace_back(MLUCnnlTensorDesc(

paddle/fluid/operators/sum_op_mlu.cc

@@ -33,7 +33,7 @@ class SumMLUKernel : public framework::OpKernel<T> {
       auto place = ctx.GetPlace();
       int ins_size = static_cast<int>(ins.size());
       if (ins_size == 1) {
-        TensorCopy(*ins[0], place, out);
+        framework::TensorCopy(*ins[0], place, out);
         return;
       }
 

paddle/fluid/platform/CMakeLists.txt

@@ -55,6 +55,12 @@ ENDIF()
 cc_library(place SRCS place.cc DEPS enforce boost pten_place)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
 
+IF(WITH_MKLDNN)
+    set(MKLDNN_CTX_DEPS mkldnn)
+ELSE()
+    set(MKLDNN_CTX_DEPS)
+ENDIF()
+
 add_subdirectory(device)
 add_subdirectory(dynload)
 add_subdirectory(stream)
@@ -85,12 +91,6 @@ IF(WITH_MLU)
     set(MLU_CTX_DEPS mlu_device_context)
 ENDIF()
 
-IF(WITH_MKLDNN)
-    set(MKLDNN_CTX_DEPS mkldnn)
-ELSE()
-    set(MKLDNN_CTX_DEPS)
-ENDIF()
-
 IF(WITH_ASCEND_CL OR WITH_MLU)
 cc_library(stream_callback_manager SRCS stream_callback_manager.cc DEPS simple_threadpool enforce)
 ENDIF()

paddle/fluid/platform/device/CMakeLists.txt

+# GPU
+IF(WITH_GPU OR WITH_ROCM)
+  add_subdirectory(gpu)
+ENDIF()
+
 # XPU
 IF(WITH_XPU)
   add_subdirectory(xpu)
 ENDIF()
-IF(WITH_GPU OR WITH_ROCM)
-  add_subdirectory(gpu)
-ENDIF()
 
 # NPU
 IF(WITH_ASCEND OR WITH_ASCEND_CL)

paddle/fluid/platform/device/mlu/CMakeLists.txt

-IF(WITH_MLU)
-    cc_test(mlu_enforce_test SRCS enforce_test.cc DEPS stringpiece)
 
-    cc_library(mlu_info SRCS mlu_info.cc DEPS enforce glog monitor neuware_lib)
+if(NOT WITH_MLU)
+    return()
+endif()
 
-    cc_library(mlu_stream SRCS mlu_stream.cc DEPS boost mlu_info stream_callback_manager)
-    
-    cc_library(mlu_device_context SRCS device_context.cc DEPS mlu_stream eigen3)
-    cc_test(mlu_device_context_test SRCS device_context_test.cc DEPS mlu_device_context)
-ENDIF()
+cc_test(mlu_enforce_test SRCS enforce_test.cc DEPS stringpiece)
+cc_library(mlu_info SRCS mlu_info.cc DEPS enforce glog monitor neuware_lib)
+cc_library(mlu_stream SRCS mlu_stream.cc DEPS boost mlu_info stream_callback_manager eigen3 ${MKLDNN_CTX_DEPS})
+cc_library(mlu_device_context SRCS device_context.cc DEPS mlu_stream)
+cc_test(mlu_device_context_test SRCS device_context_test.cc DEPS mlu_device_context)

paddle/fluid/platform/stream/CMakeLists.txt

-IF(WITH_MKLDNN)
-    set(MKLDNN_CTX_DEPS mkldnn)
-ELSE()
-    set(MKLDNN_CTX_DEPS)
-ENDIF()
-
 IF(WITH_GPU OR WITH_ROCM)
-cc_library(cuda_stream SRCS cuda_stream.cc DEPS enforce boost ${MKLDNN_CTX_DEPS})
+    cc_library(cuda_stream SRCS cuda_stream.cc DEPS enforce boost eigen3 ${MKLDNN_CTX_DEPS})
 ENDIF()

paddle/scripts/paddle_build.sh

@@ -227,6 +227,7 @@ function cmake_base() {
         -DWITH_GLOO=${gloo_flag}
         -DWITH_LITE=${WITH_LITE:-OFF}
         -DWITH_XPU=${WITH_XPU:-OFF}
+        -DWITH_MLU=${WITH_MLU:-OFF}
         -DLITE_GIT_TAG=release/v2.10
         -DWITH_UNITY_BUILD=${WITH_UNITY_BUILD:-OFF}
         -DWITH_XPU_BKCL=${WITH_XPU_BKCL:-OFF}
@@ -275,6 +276,7 @@ EOF
         -DWITH_GLOO=${gloo_flag} \
         -DLITE_GIT_TAG=release/v2.10 \
         -DWITH_XPU=${WITH_XPU:-OFF} \
+        -DWITH_MLU=${WITH_MLU:-OFF} \
         -DXPU_SDK_ROOT=${XPU_SDK_ROOT:-""} \
         -DWITH_LITE=${WITH_LITE:-OFF} \
         -DWITH_XPU_BKCL=${WITH_XPU_BKCL:-OFF} \

tools/dockerfile/Dockerfile.paddle-mlu-build

+FROM registry.baidubce.com/qili93/paddle-mlu:neuware-latest
+RUN apt-get install pigz -y
+RUN apt-get remove -y openjdk*
+CMD ["/bin/bash"]
+EXPOSE 22