[pull] master from mindspore-ai:master

mindspore/_extends/graph_kernel/model/graph_split.py

@@ -584,6 +584,39 @@ def _matmul_depth(dom):
                     fused.append(a)
             return fused, False
 
+        def _transdata_pattern_support(dom, a):
+            transdata_op = dom.dom_op()
+
+            # Currently, if transdata has the pad, it is not used to fuse
+            def _has_pad():
+                res = False
+                input_shape = transdata_op.inputs[0].shape
+                output_shape = transdata_op.output.shape
+                cube_size = 16
+                for dim in input_shape[-2:]:
+                    if dim % cube_size != 0:
+                        res = True
+                for dim in output_shape[-2:]:
+                    if dim % cube_size != 0:
+                        res = True
+                return res
+            has_pad = _has_pad()
+            if has_pad:
+                return False
+
+            if a.dom_op().prim == "MatMul" and len(dom.ops) == 1:
+                return True
+            return False
+
+        def _transdata(dom):
+            if dom.dom_op().prim != "TransData":
+                return None
+            fused = []
+            for a, _ in dom.in_relations.items():
+                if _transdata_pattern_support(dom, a) and a.check_acyclic(dom):
+                    fused.append(a)
+            return fused, True
+
         changed = True
         while changed:
             changed = self.fuse(_reshape)
@@ -594,6 +627,8 @@ def _matmul_depth(dom):
             changed = self.fuse(_broadcast_depth) or changed
             changed = self.fuse(_broadcast_width) or changed
             changed = self.fuse(_matmul_depth) or changed
+        self.fuse(_transdata)
+
 
 def split(graph, target, flags):
     """Split graph"""

mindspore/_extends/graph_kernel/model/model.py

@@ -186,6 +186,7 @@ def unknown_relation(self, op, input_idx):
         'Tile': Prim(BROADCAST),
         'BroadcastTo': Prim(BROADCAST),
         'MatMul': Prim(OPAQUE),
+        'TransData': Prim(OPAQUE),
     }
 
     default_primtive = Prim(UNKNOWN)

mindspore/ccsrc/backend/kernel_compiler/cpu/scatter_arithmetic_cpu_kernel.cc

@@ -0,0 +1,200 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * 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 "backend/kernel_compiler/cpu/scatter_arithmetic_cpu_kernel.h"
+#include <map>
+#include <limits>
+#include "runtime/device/cpu/cpu_device_address.h"
+
+namespace mindspore {
+namespace kernel {
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::InitKernel(const CNodePtr &kernel_node) {
+  CheckParam(kernel_node);
+  kernel_name_ = AnfAlgo::GetCNodeName(kernel_node);
+  auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
+  input_size_ = 1;
+  inner_size_ = 1;
+  for (size_t i = 1; i < input_shape.size(); i++) {
+    inner_size_ *= input_shape[i];
+  }
+  input_size_ = input_shape[0] * inner_size_;
+  auto indices_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 1);
+  indices_size_ = 1;
+  for (size_t i = 0; i < indices_shape.size(); i++) {
+    indices_size_ *= indices_shape[i];
+  }
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::CheckParam(const CNodePtr &kernel_node) const {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  size_t input_num = AnfAlgo::GetInputTensorNum(kernel_node);
+  if (input_num != 3) {
+    MS_LOG(EXCEPTION) << "Input number is " << input_num << ", but ScatterAdd needs 3 inputs.";
+  }
+  size_t output_num = AnfAlgo::GetOutputTensorNum(kernel_node);
+  if (output_num != 1) {
+    MS_LOG(EXCEPTION) << "Output number is " << output_num << ", but ScatterAdd has 1 output.";
+  }
+}
+
+template <typename T>
+bool ScatterArithmeticCPUKernel<T>::Launch(const std::vector<kernel::AddressPtr> &inputs,
+                                           const std::vector<kernel::AddressPtr> &workspace,
+                                           const std::vector<kernel::AddressPtr> &outputs) {
+  static const std::map<std::string, std::function<void(ScatterArithmeticCPUKernel *, T *, const int *, const T *)>>
+    kScatterArithmeticBinOpFuncMap{{"ScatterAdd", &ScatterArithmeticCPUKernel<T>::ScatterAdd},
+                                   {"ScatterSub", &ScatterArithmeticCPUKernel<T>::ScatterSub},
+                                   {"ScatterMul", &ScatterArithmeticCPUKernel<T>::ScatterMul},
+                                   {"ScatterDiv", &ScatterArithmeticCPUKernel<T>::ScatterDiv},
+                                   {"ScatterMax", &ScatterArithmeticCPUKernel<T>::ScatterMax},
+                                   {"ScatterMin", &ScatterArithmeticCPUKernel<T>::ScatterMin},
+                                   {"ScatterUpdate", &ScatterArithmeticCPUKernel<T>::ScatterUpdate}};
+  if (kScatterArithmeticBinOpFuncMap.find(kernel_name_) != kScatterArithmeticBinOpFuncMap.end()) {
+    T *input = reinterpret_cast<T *>(inputs[0]->addr);
+    int *indices = reinterpret_cast<int *>(inputs[1]->addr);
+    T *updates = reinterpret_cast<T *>(inputs[2]->addr);
+    T *output = reinterpret_cast<T *>(outputs[0]->addr);
+    kScatterArithmeticBinOpFuncMap.at(kernel_name_)(this, input, indices, updates);
+    auto bufferSize = outputs[0]->size;
+    auto ret = memcpy_s(output, bufferSize, input, input_size_ * sizeof(T));
+    if (ret != EOK) {
+      MS_LOG(EXCEPTION) << "Memory copy failed!";
+    }
+  } else {
+    MS_LOG(EXCEPTION) << "Not support operator:" << kernel_name_;
+  }
+  return true;
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterAdd(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] += updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterSub(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] -= updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterMul(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] *= updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterDiv(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      for (size_t j = 0; j < inner_size_; j++) {
+        auto dividend = input[indices[i] * inner_size_ + j];
+        auto divisor = updates[i * inner_size_ + j];
+        if (divisor == 0) {
+          if (dividend == 0) {
+            input[indices[i] * inner_size_ + j] = std::numeric_limits<T>::quiet_NaN();
+            continue;
+          }
+          if (std::numeric_limits<T>::has_infinity) {
+            input[indices[i] * inner_size_ + j] =
+              dividend > 0 ? std::numeric_limits<T>::infinity() : -std::numeric_limits<T>::infinity();
+          } else {
+            input[indices[i] * inner_size_ + j] =
+              dividend > 0 ? std::numeric_limits<T>::max() : std::numeric_limits<T>::min();
+          }
+          continue;
+        }
+        input[indices[i] * inner_size_ + j] = dividend / divisor;
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterMax(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] = input[base_index_input + j] > updates[base_index_updates + j]
+                                        ? input[base_index_input + j]
+                                        : updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterMin(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] = input[base_index_input + j] < updates[base_index_updates + j]
+                                        ? input[base_index_input + j]
+                                        : updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+
+template <typename T>
+void ScatterArithmeticCPUKernel<T>::ScatterUpdate(T *input, const int *indices, const T *updates) {
+  auto task = [this, input, indices, updates](size_t start, size_t end) {
+    for (size_t i = start; i < end; i++) {
+      auto base_index_updates = i * inner_size_;
+      auto base_index_input = indices[i] * inner_size_;
+      for (size_t j = 0; j < inner_size_; j++) {
+        input[base_index_input + j] = updates[base_index_updates + j];
+      }
+    }
+  };
+  CPUKernelUtils::ParallelFor(task, indices_size_);
+}
+}  // namespace kernel
+}  // namespace mindspore