!13262 Support string in GPU Print

From: @TFbunny
Reviewed-by: 
Signed-off-by:

mindspore/ccsrc/backend/kernel_compiler/gpu/debug/print_gpu_kernel.h

@@ -48,7 +48,7 @@ class PrintGpuKernel : public GpuKernel {
     }
     int *output_address = GetDeviceAddress<int>(outputs, 0);
     // host initialization
-    std::vector<std::unique_ptr<T[]> > input_host_data;
+    std::vector<std::unique_ptr<T[]>> input_host_data;
     for (size_t i = 0; i < input_size_.size(); i++) {
       std::unique_ptr<T[]> value = std::make_unique<T[]>(input_size_[i]);
       input_host_data.push_back(std::move(value));
@@ -60,19 +60,25 @@ class PrintGpuKernel : public GpuKernel {
       MS_LOG(EXCEPTION) << "GPU print does not support the input type.";
     }
     // print core function
-    for (size_t i = 0; i < input_host_data.size(); i++) {
-      std::string error_msg = "cudaMemcpy print loop failed at input_device_data[";
-      error_msg.append(std::to_string(i));
-      error_msg.append("].");
-      CHECK_CUDA_RET_WITH_EXCEPT(
-        kernel_node_,
-        cudaMemcpy(input_host_data[i].get(), input_device_data_[i], input_size_[i] * sizeof(T), cudaMemcpyDeviceToHost),
-        error_msg);
-      ShapeVector shape;
-      (void)std::transform(input_shape_[i].begin(), input_shape_[i].end(), std::back_inserter(shape),
-                           [](const size_t &value) { return static_cast<int64_t>(value); });
-      Tensor current_tensor(type_id, shape, input_host_data[i].get(), input_size_[i] * sizeof(T));
-      std::cout << current_tensor.ToString() << std::endl;
+    size_t string_idx = 0;
+    for (size_t i = 0; i < input_flag_.size(); i++) {
+      if (input_flag_[i] == -1) {
+        std::cout << string_value_[string_idx] << std::endl;
+        string_idx++;
+      } else {
+        size_t tensor_idx = LongToSize(input_flag_[i]);
+        std::string error_msg = "cudaMemcpyAsync print loop failed at input_device_data[";
+        error_msg.append(std::to_string(tensor_idx));
+        error_msg.append("].");
+        CHECK_CUDA_RET_WITH_EXCEPT(kernel_node_,
+                                   cudaMemcpyAsync(input_host_data[tensor_idx].get(), input_device_data_[tensor_idx],
+                                                   input_size_[tensor_idx] * sizeof(T), cudaMemcpyDeviceToHost,
+                                                   reinterpret_cast<cudaStream_t>(stream_ptr)),
+                                   error_msg);
+        CHECK_CUDA_RET_WITH_EXCEPT(kernel_node_, cudaDeviceSynchronize(), "cudaDeviceSyncFailed - Print");
+        auto current_string = GetTensorString(&input_shape_, tensor_idx, type_id, &input_host_data, &input_size_);
+        std::cout << current_string << std::endl;
+      }
     }
     int output = 1;
     CHECK_CUDA_RET_WITH_EXCEPT(kernel_node_,
@@ -84,7 +90,12 @@ class PrintGpuKernel : public GpuKernel {
 
   bool Init(const CNodePtr &kernel_node) override {
     kernel_node_ = kernel_node;
+    if (AnfAlgo::HasNodeAttr("string_pos", kernel_node)) {
+      string_value_ = GetAttr<std::vector<std::string>>(kernel_node, "string_value");
+      string_pos_ = GetAttr<std::vector<int64_t>>(kernel_node, "string_pos");
+    }
     size_t input_tensor_num = AnfAlgo::GetInputTensorNum(kernel_node);
+    input_flag_ = SetInputFlag(&string_pos_, input_tensor_num);
     input_device_data_ = std::make_unique<T *[]>(input_tensor_num);
     std::vector<size_t> value_shape;
     for (size_t i = 0; i < input_tensor_num; i++) {
@@ -103,6 +114,9 @@ class PrintGpuKernel : public GpuKernel {
   }
 
   void ResetResource() noexcept override {
+    string_value_.clear();
+    string_pos_.clear();
+    input_flag_.clear();
     input_device_data_ = nullptr;
     input_size_.clear();
     input_shape_.clear();
@@ -146,14 +160,52 @@ class PrintGpuKernel : public GpuKernel {
     return kTypeUnknown;
   }
 
+  std::vector<int64_t> SetInputFlag(std::vector<int64_t> *string_pos, size_t input_tensor_num) {
+    // -1 -> string position
+    // others -> input tensor position
+    std::vector<int64_t> res(string_pos->size() + input_tensor_num);
+    // without string inputs
+    int64_t value = 0;
+    if (res.size() == input_tensor_num) {
+      std::generate(res.begin(), res.end(), [&value]() { return value++; });
+      return res;
+    }
+    for (size_t i = 0; i < string_pos->size(); i++) {
+      if ((*string_pos)[i] < 0) {
+        MS_LOG(EXCEPTION) << "string_pos cannot be a negative value";
+      }
+      auto index = IntToSize((*string_pos)[i]);
+      res[index] = -1;
+    }
+    for (size_t i = 0; i < res.size(); i++) {
+      if (res[i] != -1) {
+        res[i] += value;
+        value++;
+      }
+    }
+    return res;
+  }
+
+  std::string GetTensorString(std::vector<std::vector<size_t>> *input_shape, size_t index, TypeId type_id,
+                              std::vector<std::unique_ptr<T[]>> *input_host_data, std::vector<size_t> *input_size) {
+    ShapeVector shape;
+    (void)std::transform((*input_shape)[index].begin(), (*input_shape)[index].end(), std::back_inserter(shape),
+                         [](const size_t &value) { return static_cast<int64_t>(value); });
+    Tensor current_tensor(type_id, shape, (*input_host_data)[index].get(), (*input_size)[index] * sizeof(T));
+    return current_tensor.ToStringNoLimit();
+  }
+
  private:
+  std::vector<std::string> string_value_;
+  std::vector<int64_t> string_pos_;
+  std::vector<int64_t> input_flag_;
   std::unique_ptr<T *[]> input_device_data_;
   std::vector<size_t> input_size_;
-  std::vector<std::vector<size_t> > input_shape_;
+  std::vector<std::vector<size_t>> input_shape_;
   std::vector<size_t> input_size_list_;
   std::vector<size_t> output_size_list_;
   std::vector<size_t> workspace_size_list_;
-};  // namespace kernel
+};
 }  // namespace kernel
 }  // namespace mindspore
 #endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_DEBUG_PRINT_GPU_KERNEL_H_

mindspore/ccsrc/backend/optimizer/gpu/print_reduce_fusion.cc

@@ -0,0 +1,154 @@
+/**
+ * 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/optimizer/gpu/print_reduce_fusion.h"
+
+#include <memory>
+#include <vector>
+#include <string>
+
+#include "backend/session/anf_runtime_algorithm.h"
+#include "ir/primitive.h"
+#include "utils/utils.h"
+#include "backend/optimizer/common/helper.h"
+
+namespace mindspore {
+namespace opt {
+kernel::KernelBuildInfoPtr GenerateKernelBuildInfo(CNodePtr node) {
+  std::vector<std::string> inputs_format;
+  std::vector<std::string> outputs_format;
+  std::vector<TypeId> inputs_type;
+  std::vector<TypeId> outputs_type;
+  kernel::KernelBuildInfo::KernelBuildInfoBuilder builder;
+
+  size_t input_num = AnfAlgo::GetInputTensorNum(node);
+  for (size_t input_index = 0; input_index < input_num; input_index++) {
+    inputs_format.push_back(kOpFormat_DEFAULT);
+    inputs_type.push_back(AnfAlgo::GetPrevNodeOutputInferDataType(node, input_index));
+  }
+  size_t output_num = AnfAlgo::GetOutputTensorNum(node);
+  for (size_t output_index = 0; output_index < output_num; output_index++) {
+    outputs_format.push_back(kOpFormat_DEFAULT);
+    outputs_type.push_back(AnfAlgo::GetOutputInferDataType(node, output_index));
+  }
+
+  builder.SetInputsFormat(inputs_format);
+  builder.SetOutputsFormat(outputs_format);
+  builder.SetInputsDeviceType(inputs_type);
+  builder.SetOutputsDeviceType(outputs_type);
+  return builder.Build();
+}
+
+bool GetOptList(const std::vector<AnfNodePtr> &node_list, std::vector<AnfNodePtr> *opt_list,
+                std::vector<std::vector<int64_t>> *string_pos_vec,
+                std::vector<std::vector<std::string>> *string_value_vec) {
+  for (auto &node : node_list) {
+    // {prim::kPrimPrint} only print with string will be reduced
+    std::vector<int64_t> string_pos;
+    std::vector<std::string> string_value;
+    if (IsPrimitiveCNode(node, prim::kPrimPrint)) {
+      size_t input_num = AnfAlgo::GetInputTensorNum(node);
+      for (size_t i = 0; i < input_num; i++) {
+        auto current_node = AnfAlgo::GetInputNode(utils::cast<CNodePtr>(node), i);
+        // not a string
+        if (current_node->cast<ValueNodePtr>() == nullptr) {
+          continue;
+        }
+        auto value_node = current_node->cast<ValueNodePtr>()->value();
+        if (value_node->type()->generic_type_id() == kObjectTypeString) {
+          auto current_string_value = GetValue<std::string>(value_node);
+          string_pos.push_back(i);
+          string_value.push_back(std::string(current_string_value));
+        } else {
+          MS_LOG(EXCEPTION) << "Current value node is not string or tensor";
+        }
+      }
+      if (string_pos.size() != 0) {
+        opt_list->push_back(node);
+        string_pos_vec->push_back(string_pos);
+        string_value_vec->push_back(string_value);
+      }
+    }
+  }
+  if (opt_list->size() == 0) {
+    return false;
+  }
+  return true;
+}
+
+bool PrintReduceFusion::Run(const FuncGraphPtr &graph) {
+  MS_EXCEPTION_IF_NULL(graph);
+  auto manager = graph->manager();
+  MS_EXCEPTION_IF_NULL(manager);
+  std::vector<AnfNodePtr> node_list = TopoSort(graph->get_return());
+  std::vector<AnfNodePtr> opt_list;
+  std::vector<std::vector<int64_t>> string_pos_vec;
+  std::vector<std::vector<std::string>> string_value_vec;
+  if (!GetOptList(node_list, &opt_list, &string_pos_vec, &string_value_vec)) {
+    return false;
+  }
+  for (size_t idx = 0; idx < opt_list.size(); idx++) {
+    auto node = opt_list[idx];
+    CNodePtr cnode = utils::cast<CNodePtr>(node);
+    size_t input_num = AnfAlgo::GetInputTensorNum(cnode);
+    auto prim = std::make_shared<Primitive>("Print");
+    std::vector<AnfNodePtr> inputs = {NewValueNode(prim)};
+    auto string_pos = string_pos_vec[idx];
+    std::vector<int64_t> input_flag(input_num);
+    for (size_t i = 0; i < string_pos.size(); i++) {
+      if (string_pos[i] < 0) {
+        MS_LOG(EXCEPTION) << "string_pos cannot be a negative value";
+      }
+      size_t index = LongToSize(string_pos[i]);
+      input_flag[index] = -1;
+    }
+    for (size_t i = 0; i < input_flag.size(); i++) {
+      if (input_flag[i] == -1) {
+        continue;
+      }
+      auto input_tensor = AnfAlgo::GetInputNode(cnode, i);
+      MS_EXCEPTION_IF_NULL(input_tensor);
+      inputs.push_back(input_tensor);
+    }
+    // add monad
+    auto monad_node = AnfAlgo::GetInputNode(cnode, input_flag.size());
+    MS_EXCEPTION_IF_NULL(monad_node);
+    inputs.push_back(monad_node);
+    auto string_value = string_value_vec[idx];
+    // create new cnode
+    auto print_fused = graph->NewCNode(inputs);
+    // hand over the attrs to new print
+    AnfAlgo::SetNodeAttr("string_pos", MakeValue<std::vector<int64_t>>(string_pos), print_fused);
+    AnfAlgo::SetNodeAttr("string_value", MakeValue<std::vector<std::string>>(string_value), print_fused);
+    // set output type and shape
+    std::vector<TypeId> types;
+    std::vector<std::vector<size_t>> shapes;
+    size_t output_num = AnfAlgo::GetOutputTensorNum(cnode);
+    for (size_t i = 0; i < output_num; i++) {
+      types.push_back(AnfAlgo::GetOutputInferDataType(cnode, i));
+      shapes.push_back(AnfAlgo::GetOutputInferShape(cnode, i));
+    }
+    AnfAlgo::SetOutputInferTypeAndShape(types, shapes, print_fused.get());
+    // add build info
+    auto build_info = GenerateKernelBuildInfo(print_fused);
+    AnfAlgo::SetSelectKernelBuildInfo(build_info, print_fused.get());
+    if (!manager->Replace(cnode, print_fused)) {
+      MS_LOG(EXCEPTION) << "manager replace node failed in print reduce fusion.";
+    }
+  }
+  return true;
+}
+}  // namespace opt
+}  // namespace mindspore

mindspore/ccsrc/backend/optimizer/gpu/print_reduce_fusion.h

@@ -0,0 +1,32 @@
+/**
+ * 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.
+ */
+#ifndef MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_PRINT_REDUCE_FUSION_H_
+#define MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_PRINT_REDUCE_FUSION_H_
+
+#include <string>
+#include "backend/optimizer/common/optimizer.h"
+
+namespace mindspore {
+namespace opt {
+class PrintReduceFusion : public Pass {
+ public:
+  explicit PrintReduceFusion(const std::string &name) : Pass("print_reduce") {}
+  ~PrintReduceFusion() override = default;
+  bool Run(const FuncGraphPtr &graph) override;
+};
+}  // namespace opt
+}  // namespace mindspore
+#endif  // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_GPU_PRINT_REDUCE_FUSION_H_

mindspore/ccsrc/backend/session/gpu_session.cc

@@ -37,6 +37,7 @@
 #include "backend/optimizer/gpu/insert_format_transform_op.h"
 #include "backend/optimizer/gpu/replace_momentum_cast_fusion.h"
 #include "backend/optimizer/gpu/replace_addn_fusion.h"
+#include "backend/optimizer/gpu/print_reduce_fusion.h"
 #include "backend/optimizer/gpu/remove_format_transform_pair.h"
 #include "backend/optimizer/gpu/remove_redundant_format_transform.h"
 #include "backend/optimizer/gpu/reduce_precision_fusion.h"
@@ -141,6 +142,7 @@ void GPUSession::Optimize(const std::shared_ptr<KernelGraph> &kernel_graph) {
   pm->AddPass(std::make_shared<opt::CombineMomentumFusion>("combine_momentum"));
   pm->AddPass(std::make_shared<opt::ReplaceMomentumCastFusion>());
   pm->AddPass(std::make_shared<opt::ReplaceAddNFusion>());
+  pm->AddPass(std::make_shared<opt::PrintReduceFusion>("print_reduce"));
   optimizer->AddPassManager(pm);
   (void)optimizer->Optimize(kernel_graph);
   kernel_graph->SetExecOrderByDefault();

mindspore/ops/operations/debug_ops.py

@@ -346,11 +346,10 @@ class Print(PrimitiveWithInfer):
         In pynative mode, please use python print function.
         In graph mode, the bool, int, float, tuple, and list would be converted into Tensor to print,
         str remains unchanged.
-        In GPU, all input elements should be the same type and string is not supported.
 
     Inputs:
         - **input_x** (Union[Tensor, bool, int, float, str, tuple, list]) - The graph node to attach to.
-          Supports multiple inputs which are separated by ','. GPU does not support string as an input.
+          Supports multiple inputs which are separated by ','.
 
     Supported Platforms:
         ``Ascend`` ``GPU``

tests/st/ops/gpu/test_print_op.py

@@ -71,6 +71,27 @@ def print_testcase(nptype):
     net_2(x, y)
     net_3(x)
 
+class PrintNetString(nn.Cell):
+    def __init__(self):
+        super(PrintNetString, self).__init__()
+        self.op = P.Print()
+
+    def construct(self, x, y):
+        self.op("The first Tensor is", x)
+        self.op("The second Tensor is", y)
+        self.op("This line only prints string", "Another line")
+        self.op("The first Tensor is", x, y, "is the second Tensor")
+        return x
+
+def print_testcase_string(nptype):
+    x = np.ones(18).astype(nptype)
+    y = np.arange(9).reshape(3, 3).astype(nptype)
+    x = Tensor(x)
+    y = Tensor(y)
+    # graph mode
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+    net = PrintNetString()
+    net(x, y)
 
 @pytest.mark.level0
 @pytest.mark.platform_x86_gpu_training
@@ -147,3 +168,10 @@ def test_print_float16():
 @pytest.mark.env_onecard
 def test_print_float32():
     print_testcase(np.float32)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_print_string():
+    print_testcase_string(np.float32)