[INTEL_HPU] Add support of paddle.tile (PaddlePaddle#1691)

JianyuLi01 · web-flow · commit 71ead932dbdd · 2025-05-19T09:26:33.000+08:00
diff --git a/backends/intel_hpu/kernels/tile_kernel.cc b/backends/intel_hpu/kernels/tile_kernel.cc
@@ -0,0 +1,122 @@
+// 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 "habanalabs/perf_lib_layer_params.h"
+#include "kernels/funcs.h"
+#include "kernels/hpu_operator.h"
+#include "paddle/extension.h"
+#include "utils/utils.h"
+
+namespace custom_kernel {
+
+class Tile : public HpuOperator {
+ public:
+  Tile() : HpuOperator("repeat_pt_fwd_", false) {}
+
+  void AddNode(ConvertTensors& ct, ns_RepeatPt::Params& params) {
+    auto inputs = ct.GetTensors();
+    auto outputs = ct.GetTensors(false);
+
+    auto x_tensor = createTensor(inputs[0].dims.size(),
+                                 inputs[0].type,
+                                 inputs[0].dims,
+                                 true,
+                                 inputs[0].name);
+
+    auto out_tensor = createTensor(outputs[0].dims.size(),
+                                   outputs[0].type,
+                                   outputs[0].dims,
+                                   true,
+                                   outputs[0].name);
+
+    synTensor ins[] = {x_tensor};
+    synTensor outs[] = {out_tensor};
+    std::string guid = guid_ + SynDataTypeToStr(inputs[0].type);
+    synStatus status = synNodeCreate(graphHandle_,
+                                     ins,
+                                     outs,
+                                     1,
+                                     1,
+                                     &params,
+                                     sizeof(params),
+                                     guid.c_str(),
+                                     guid.c_str(),
+                                     nullptr,
+                                     nullptr);
+    PD_CHECK(
+        status == synSuccess, "[RUNTIME] synNodeCreate tile failed = ", status);
+  }
+};
+
+template <typename T, typename Context>
+void TileKernel(const Context& dev_ctx,
+                const phi::DenseTensor& x,
+                const phi::IntArray& repeat_times,
+                phi::DenseTensor* out) {
+  VLOG(6) << "call HPU TileKernel";
+
+  dev_ctx.template Alloc<T>(out);
+
+  ns_RepeatPt::Params params;
+  int max_repeat_times = sizeof(params.repeat) / sizeof(params.repeat[0]);
+  int repeat_times_size = repeat_times.size();
+  PADDLE_ENFORCE_GE(
+      max_repeat_times,
+      repeat_times_size,
+      phi::errors::ResourceExhausted("unsupported repeat_times size."));
+
+  for (int i = 0; i < repeat_times_size; i++) {
+    params.repeat[i] = repeat_times[i];
+  }
+  params.size = repeat_times_size;
+
+  ConvertTensors ct;
+  ct.Add(x);
+  ct.Add(out, false);
+  std::vector<DIMS> inputs_dims = ct.GetDims();
+
+  std::string guid_prefix = "TileKernel";
+  OpCacheOperator op_info;
+  op_info.prepareOpInfo<T, ns_RepeatPt::Params>(
+      guid_prefix, inputs_dims, &params);
+  auto recipe = op_info.GetRecipe();
+
+  if (recipe == nullptr) {
+    Tile op;
+
+    op.AddNode(ct, params);
+    op.Compile();
+    op_info.setOp(op);
+
+    recipe = op_info.GetRecipe();
+  }
+
+  std::map<std::string, uint64_t> tensors = ct.GetDeviceAddr();
+  RecipeRunner runner(recipe);
+  runner.Run(reinterpret_cast<C_Stream>(dev_ctx.stream()), tensors);
+}
+
+}  // namespace custom_kernel
+
+PD_REGISTER_PLUGIN_KERNEL(tile,
+                          intel_hpu,
+                          ALL_LAYOUT,
+                          custom_kernel::TileKernel,
+                          bool,
+                          float,
+                          double,
+                          int,
+                          int64_t,
+                          phi::dtype::float16,
+                          phi::dtype::bfloat16) {}
diff --git a/backends/intel_hpu/tests/unittests/test_tile.py b/backends/intel_hpu/tests/unittests/test_tile.py
@@ -0,0 +1,269 @@
+#   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.
+
+import unittest
+import numpy as np
+
+from tests.op_test import OpTest
+import paddle
+
+import os
+
+paddle.enable_static()
+np.random.seed(10)
+
+intel_hpus_module_id = os.environ.get("FLAGS_selected_intel_hpus", 0)
+
+
+# repeat_times is a list (without tensor)
+class TestTileOpRank1(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.python_api = paddle.scatter
+        self.init_data()
+
+        self.inputs = {"X": np.random.random(self.ori_shape).astype("float32")}
+        self.attrs = {"repeat_times": self.repeat_times}
+        output = np.tile(self.inputs["X"], self.repeat_times)
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.repeat_times = [2]
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+# with dimension expanding
+class TestTileOpRank2Expanding(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = [120]
+        self.repeat_times = [2, 2]
+
+
+class TestTileOpRank2(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = [12, 14]
+        self.repeat_times = [2, 3]
+
+
+class TestTileOpRank3_Corner(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 10, 5)
+        self.repeat_times = (1, 1, 1)
+
+
+class TestTileOpRank3_Corner2(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 10, 5)
+        self.repeat_times = (2, 2)
+
+
+class TestTileOpRank3(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 4, 15)
+        self.repeat_times = (2, 1, 4)
+
+
+class TestTileOpRank4(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = (2, 4, 5, 7)
+        self.repeat_times = (3, 2, 1, 2)
+
+
+class TestTileOpRank_ZeroDim1(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = []
+        self.repeat_times = []
+
+
+class TestTileOpRank_ZeroDim2(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = []
+        self.repeat_times = [2]
+
+
+class TestTileOpRank_ZeroDim3(TestTileOpRank1):
+    def init_data(self):
+        self.ori_shape = []
+        self.repeat_times = [2, 3]
+
+
+# repeat_times is a list (with tensor)
+class TestTileOpRank1_tensor_attr(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.init_data()
+        repeat_times_tensor = []
+        for index, ele in enumerate(self.repeat_times):
+            repeat_times_tensor.append(
+                ("x" + str(index), np.ones((1)).astype("int32") * ele)
+            )
+
+        self.inputs = {
+            "X": np.random.random(self.ori_shape).astype("float32"),
+            "repeat_times_tensor": repeat_times_tensor,
+        }
+        self.attrs = {"repeat_times": self.infer_repeat_times}
+        output = np.tile(self.inputs["X"], self.repeat_times)
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.repeat_times = [2]
+        self.infer_repeat_times = [-1]
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestTileOpRank2_Corner_tensor_attr(TestTileOpRank1_tensor_attr):
+    def init_data(self):
+        self.ori_shape = [12, 14]
+        self.repeat_times = [1, 1]
+        self.infer_repeat_times = [1, -1]
+
+
+class TestTileOpRank2_attr_tensor(TestTileOpRank1_tensor_attr):
+    def init_data(self):
+        self.ori_shape = [12, 14]
+        self.repeat_times = [2, 3]
+        self.infer_repeat_times = [-1, 3]
+
+
+# repeat_times is a tensor
+class TestTileOpRank1_tensor(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.init_data()
+
+        self.inputs = {
+            "X": np.random.random(self.ori_shape).astype("float32"),
+            "RepeatTimes": np.array(self.repeat_times).astype("int32"),
+        }
+        self.attrs = {}
+        output = np.tile(self.inputs["X"], self.repeat_times)
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def init_data(self):
+        self.ori_shape = [100]
+        self.repeat_times = [2]
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestTileOpRank2_tensor(TestTileOpRank1_tensor):
+    def init_data(self):
+        self.ori_shape = [12, 14]
+        self.repeat_times = [2, 3]
+
+
+# input x is Integer
+class TestTileOpInteger(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.inputs = {"X": np.random.randint(10, size=(4, 4, 5)).astype("int32")}
+        self.attrs = {"repeat_times": [2, 1, 4]}
+        output = np.tile(self.inputs["X"], (2, 1, 4))
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+# input x is Integer
+class TestTileOpInt64_t(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.inputs = {"X": np.random.randint(10, size=(2, 4, 5)).astype("int64")}
+        self.attrs = {"repeat_times": [2, 1, 4]}
+        output = np.tile(self.inputs["X"], (2, 1, 4))
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+# input x is Bool
+class TestTileOpBool(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.inputs = {"X": np.random.randint(1, size=(2, 4, 5)).astype("bool")}
+        self.attrs = {"repeat_times": [2, 1, 4]}
+        output = np.tile(self.inputs["X"], (2, 1, 4))
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+# input x is FP16
+class TestTileOpFloat16(OpTest):
+    def setUp(self):
+        self.set_intel_hpu()
+        self.place = paddle.CustomPlace("intel_hpu", int(intel_hpus_module_id))
+        self.op_type = "tile"
+        self.inputs = {"X": np.random.randint(10, size=(2, 10, 5)).astype("float16")}
+        self.attrs = {"repeat_times": [2, 1, 4]}
+        output = np.tile(self.inputs["X"], (2, 1, 4))
+        self.outputs = {"Out": output}
+
+    def set_intel_hpu(self):
+        self.__class__.use_custom_device = True
+        self.__class__.no_need_check_grad = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+if __name__ == "__main__":
+    unittest.main()
