[NPU] Support npu op `expand` (#31405)

* [npu] support npu kernel  for `expand`

paddle/fluid/operators/CMakeLists.txt

@@ -156,12 +156,11 @@ cc_library(tensor_formatter SRCS tensor_formatter.cc DEPS ${OP_HEADER_DEPS})
 if (WITH_PYTHON)
   cc_library(py_func_op SRCS py_func_op.cc DEPS op_registry python pybind)
 endif()
-if (WITH_ASCEND_CL)
-    cc_test(lookup_table_v2_op_npu_test SRCS lookup_table_v2_op_npu_test.cc DEPS op_registry lookup_table_v2_op scope device_context enforce executor compare_op)
-endif()
 
 if (WITH_ASCEND_CL)
   cc_test(range_op_npu_test SRCS range_op_npu_test.cc DEPS op_registry range_op scope device_context enforce executor)
+  cc_test(lookup_table_v2_op_npu_test SRCS lookup_table_v2_op_npu_test.cc DEPS op_registry lookup_table_v2_op scope device_context enforce executor compare_op)
+  cc_test(expand_op_npu_test SRCS expand_op_npu_test.cc DEPS op_registry expand_op scope device_context enforce executor compare_op)
 endif()
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

paddle/fluid/operators/expand_op.h

@@ -56,6 +56,12 @@ inline std::vector<int> get_expand_times(
       TensorCopySync(*expand_tensor, platform::CPUPlace(), &cpu_expand_tensor);
       expand_data = cpu_expand_tensor.data<int>();
     }
+#ifdef PADDLE_WITH_ASCEND_CL
+    if (platform::is_npu_place(expand_tensor->place())) {
+      TensorCopySync(*expand_tensor, platform::CPUPlace(), &cpu_expand_tensor);
+      expand_data = cpu_expand_tensor.data<int>();
+    }
+#endif
 #ifdef PADDLE_WITH_XPU
     if (platform::is_xpu_place(expand_tensor->place())) {
       TensorCopySync(*expand_tensor, platform::CPUPlace(), &cpu_expand_tensor);

paddle/fluid/operators/expand_op_npu.cc

+/* Copyright (c) 2021 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. */
+
+#ifdef PADDLE_WITH_ASCEND_CL
+#include <iostream>
+#include <memory>
+#include <string>
+
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/expand_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class ExpandNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto rank = context.Input<Tensor>("X")->dims().size();
+    PADDLE_ENFORCE_GE(
+        rank, 1,
+        platform::errors::InvalidArgument(
+            "The number of dimensions of the input 'x' for Op(expand) "
+            "must be greater than or equal to 1, but the value received is %d.",
+            rank));
+    PADDLE_ENFORCE_LE(
+        rank, MAX_RANK_SUPPORTED,
+        platform::errors::InvalidArgument(
+            "The number of dimensions of the input 'x' for Op(expand) "
+            "must be less than or equal to %d, but the value received is %d.",
+            MAX_RANK_SUPPORTED, rank));
+    switch (rank) { REP_EXPAND_TEMPLATE(MAX_RANK_SUPPORTED) }
+  }
+
+ protected:
+  template <int Rank>
+  void Expand(const framework::ExecutionContext& context) const {
+    auto* in0 = context.Input<framework::LoDTensor>("X");
+    auto in_dims = in0->dims();
+    auto expand_times = get_expand_times(context);
+    PADDLE_ENFORCE_EQ(
+        static_cast<size_t>(in_dims.size()), expand_times.size(),
+        platform::errors::InvalidArgument(
+            "The number of elements (%d) of 'expand_times' for "
+            "Op(expand) must be equal to the number "
+            "of dimensions (%d) of the input.",
+            expand_times.size(), static_cast<size_t>(in_dims.size())));
+    auto* out0 = context.Output<framework::LoDTensor>("Out");
+    framework::DDim out_dims(in_dims);
+    for (size_t i = 0; i < expand_times.size(); ++i) {
+      out_dims[i] *= expand_times[i];
+    }
+    out0->Resize(out_dims);
+    out0->mutable_data<T>(context.device_context().GetPlace());
+    auto runner = NpuOpRunner("TileD", {*in0}, {*out0}, {{"multiples", expand_times}});
+    auto stream =
+        context.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_NPU_KERNEL(
+    expand, ops::ExpandNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::ExpandNPUKernel<paddle::platform::NPUDeviceContext,
+                         paddle::platform::float16>);
+
+#endif

paddle/fluid/operators/expand_op_npu_test.cc

+/* Copyright (c) 2021 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. */
+
+#ifndef _WIN32
+#include <unistd.h>
+#endif
+
+#include <iostream>
+#include <string>
+#include <thread>  // NOLINT
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/operators/dropout_op.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace f = paddle::framework;
+namespace p = paddle::platform;
+namespace m = paddle::operators::math;
+
+USE_OP(expand);
+USE_OP_DEVICE_KERNEL(expand, NPU);
+
+template <typename T>
+void Compare(f::Scope* scope, const p::DeviceContext& ctx) {
+  // init
+  auto in = scope->Var("X");
+  auto expand_times = scope->Var("ExpandTimes");
+  auto out = scope->Var("Out");
+  auto in_t = in->GetMutable<f::LoDTensor>();
+  auto out_t = out->GetMutable<f::LoDTensor>();
+  auto expand_times_t = expand_times->GetMutable<f::LoDTensor>();
+
+  auto place = ctx.GetPlace();
+  TensorFromVector(std::vector<T>(3 * 1 * 7, 1), ctx, in_t);
+  TensorFromVector(std::vector<int>({1, 10, 1}), ctx, expand_times_t);
+
+  in_t->Resize(f::make_ddim({3, 1, 7}));
+  expand_times_t->Resize(f::make_ddim({3}));
+  out_t->Resize(f::make_ddim({3, 10, 7}));
+  out_t->mutable_data<T>(place);
+
+  f::AttributeMap attrs = {{}};
+  auto op = f::OpRegistry::CreateOp(
+      "expand", {{"X", {"X"}}, {"ExpandTimes", {"ExpandTimes"}}},
+      {{"Out", {"Out"}}}, attrs);
+  op->Run(*scope, place);
+  ctx.Wait();
+
+  auto out_dim = out_t->dims();
+  EXPECT_EQ(out_dim.at(0), 3);
+  EXPECT_EQ(out_dim.at(1), 10);
+  EXPECT_EQ(out_dim.at(2), 7);
+}
+
+TEST(expand, NPU_fp32) {
+  f::Scope scope;
+  p::NPUDeviceContext ctx(p::NPUPlace(0));
+  Compare<float>(&scope, ctx);
+}

python/paddle/fluid/tests/unittests/npu/test_expand_op_npu.py

+#  Copyright (c) 2021 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 unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestExpand(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "expand"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.randn(3,1,7).astype(self.dtype)
+        out = np.tile(x, [1,10,1])
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {'expand_times': [1,10,1]}
+        self.outputs = {'Out': out}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    # TODO(ascendrc): Add grad test
+    # def test_check_grad(self):
+    #     if self.dtype == np.float16:
+    #         return
+    #     self.check_grad(['X'], 'Out')
+    #
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestExpandV2(TestExpand):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "expand"
+        self.place = paddle.NPUPlace(0)
+
+        self.init_dtype()
+        np.random.seed(SEED)
+        x = np.random.randn(3,1,7).astype(self.dtype)
+        out = np.tile(x, [1,10,1])
+        expand_times = np.array([1,10,1]).astype(np.int32)
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x), 'ExpandTimes':  OpTest.np_dtype_to_fluid_dtype(expand_times)}
+        self.attrs = {}
+        self.outputs = {'Out': out}
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestExpandFp16(TestExpand):
+     no_need_check_grad = True
+     def init_dtype(self):
+         self.dtype = np.float16
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestExpandNet(unittest.TestCase):
+    def _test(self, run_npu=True):
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = SEED
+        startup_prog.random_seed = SEED
+        np.random.seed(SEED)
+
+        a_np = np.random.random(size=(32, 1)).astype('float32')
+        label_np = np.random.randint(2, size=(32, 1)).astype('int64')
+
+        with paddle.static.program_guard(main_prog, startup_prog):
+            a = paddle.static.data(name="a", shape=[32, 1], dtype='float32')
+            label = paddle.static.data(
+                name="label", shape=[32, 1], dtype='int64')
+
+            res = paddle.fluid.layers.expand(a, [1,32])
+            loss = res.sum()
+            sgd = fluid.optimizer.SGD(learning_rate=0.01)
+            sgd.minimize(loss)
+
+        if run_npu:
+            place = paddle.NPUPlace(0)
+        else:
+            place = paddle.CPUPlace()
+
+        exe = paddle.static.Executor(place)
+        exe.run(startup_prog)
+
+        for epoch in range(100):
+
+            loss_res = exe.run(
+                main_prog,
+                feed={"a": a_np,
+                      "label": label_np},
+                fetch_list=[loss])
+            if epoch % 10 == 0:
+                print("Epoch {} | Loss: {}".format(epoch, loss))
+
+        return loss_res
+
+    def test_npu(self):
+        cpu_loss = self._test(False)
+        npu_loss = self._test(True)
+
+        self.assertTrue(np.allclose(npu_loss, cpu_loss))
+
+
+if __name__ == '__main__':
+    unittest.main()
+