Add debugging helper to dump string, vector and thread id (#21224)

### Description

Add some macro to help print data to console for debugging purpose.

Example usage:
```
int input_id;
vector<int> some_vector;

DUMP_CPU_TENSOR_INIT()
DUMP_CPU_TENSOR("some vector", some_vector);
DUMP_STRING("input_id=", input_id);
```

- To enable dump thread id, set environment variable
`ORT_DUMP_THREAD_ID=0`.
- User can disable dumping by environment variable
`ORT_ENABLE_CPU_DUMP=0`.

### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->

onnxruntime/contrib_ops/cpu/utils/console_dumper.h

@@ -37,6 +37,8 @@ class IConsoleDumper {
   virtual void Print(const char* name, int index, bool end_line) const = 0;
   virtual void Print(const char* name, const std::string& value, bool end_line) const = 0;
 
+  virtual void Print(const std::string& value) const = 0;
+
  protected:
   bool is_enabled_;
 };

onnxruntime/contrib_ops/cpu/utils/debug_macros.h

@@ -1,4 +1,5 @@
 #pragma once
+#include "core/common/make_string.h"
 
 // #define DEBUG_GENERATION 1  // uncomment it for debugging generation (like beam search etc)
 
@@ -14,9 +15,11 @@
 #if DUMP_CPU_TENSOR_LEVEL > 0
 #define DUMP_CPU_TENSOR_INIT() onnxruntime::contrib::CpuTensorConsoleDumper cpu_dumper
 #define DUMP_CPU_TENSOR(...) cpu_dumper.Print(__VA_ARGS__)
+#define DUMP_STRING(...) cpu_dumper.Print(::onnxruntime::MakeString(__VA_ARGS__))
 #else
 #define DUMP_CPU_TENSOR_INIT()
 #define DUMP_CPU_TENSOR(...)
+#define DUMP_STRING(...)
 #endif
 
 #if DUMP_CPU_TENSOR_LEVEL > 1

onnxruntime/contrib_ops/cpu/utils/dump_tensor.cc

@@ -1,18 +1,38 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
-#include <iomanip>
 #include "contrib_ops/cpu/utils/dump_tensor.h"
+#include <iomanip>
+#include <mutex>
+#include <thread>
+#include <iostream>
 #include "core/framework/print_tensor_utils.h"
 #include "contrib_ops/cpu/utils/debug_macros.h"
+#include "core/platform/env_var_utils.h"
 
 namespace onnxruntime {
 namespace contrib {
 
 #if DUMP_CPU_TENSOR_LEVEL > 0
 
+// Environment variable to enable/disable dumping
+constexpr const char* kEnableCpuTensorDumper = "ORT_ENABLE_CPU_DUMP";
+
+// Environment variable to enable/disable dumping thread id
+constexpr const char* kDumpThreadId = "ORT_DUMP_THREAD_ID";
+
+// To avoid dumping at the same time from multiple threads
+static std::mutex s_mutex;
+
+static bool s_output_thread_id = false;
+
 template <typename T>
 void DumpCpuTensor(const char* name, const T* tensor, int dim0, int dim1) {
+  std::unique_lock<std::mutex> lock(s_mutex);
+
+  if (s_output_thread_id)
+    std::cout << "Thread ID:" << std::this_thread::get_id() << std::endl;
+
   if (nullptr != name) {
     std::cout << std::string(name) << std::endl;
   }
@@ -26,6 +46,11 @@ void DumpCpuTensor(const char* name, const T* tensor, int dim0, int dim1) {
 
 template <typename T>
 void DumpCpuTensor(const char* name, const T* tensor, int dim0, int dim1, int dim2) {
+  std::unique_lock<std::mutex> lock(s_mutex);
+
+  if (s_output_thread_id)
+    std::cout << "Thread ID:" << std::this_thread::get_id() << std::endl;
+
   if (nullptr != name) {
     std::cout << std::string(name) << std::endl;
   }
@@ -93,6 +118,21 @@ void DumpCpuTensor(const char* name, const Tensor& tensor) {
   DumpCpuTensor(nullptr, tensor, static_cast<int>(num_rows), static_cast<int>(row_size));
 }
 
+CpuTensorConsoleDumper::CpuTensorConsoleDumper() {
+  is_enabled_ = ParseEnvironmentVariableWithDefault<int>(kEnableCpuTensorDumper, 1) != 0;
+  s_output_thread_id = ParseEnvironmentVariableWithDefault<int>(kDumpThreadId, 0) != 0;
+}
+
+void CpuTensorConsoleDumper::Print(const std::string& value) const {
+  if (!is_enabled_)
+    return;
+
+  std::unique_lock<std::mutex> lock(s_mutex);
+  if (s_output_thread_id)
+    std::cout << "Thread ID:" << std::this_thread::get_id() << std::endl;
+  std::cout << value << std::endl;
+}
+
 void CpuTensorConsoleDumper::Print(const char* name, const float* tensor, int dim0, int dim1) const {
   if (!is_enabled_)
     return;
@@ -185,6 +225,8 @@ void CpuTensorConsoleDumper::Print(const char* name, const OrtValue& value) cons
 void CpuTensorConsoleDumper::Print(const char* name, int index, bool end_line) const {
   if (!is_enabled_)
     return;
+
+  std::unique_lock<std::mutex> lock(s_mutex);
   std::cout << std::string(name) << "[" << index << "]";
 
   if (end_line) {
@@ -196,6 +238,7 @@ void CpuTensorConsoleDumper::Print(const char* name, const std::string& value, b
   if (!is_enabled_)
     return;
 
+  std::unique_lock<std::mutex> lock(s_mutex);
   std::cout << std::string(name) << "=" << value;
 
   if (end_line) {
@@ -204,6 +247,12 @@ void CpuTensorConsoleDumper::Print(const char* name, const std::string& value, b
 }
 
 #else
+CpuTensorConsoleDumper::CpuTensorConsoleDumper() {
+}
+
+void CpuTensorConsoleDumper::Print(const std::string&) const {
+}
+
 void CpuTensorConsoleDumper::Print(const char*, const float*, int, int) const {
 }
 
@@ -254,7 +303,6 @@ void CpuTensorConsoleDumper::Print(const char*, int, bool) const {
 
 void CpuTensorConsoleDumper::Print(const char*, const std::string&, bool) const {
 }
-
 #endif
 
 }  // namespace contrib

onnxruntime/contrib_ops/cpu/utils/dump_tensor.h

@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 #pragma once
+#include <vector>
 #include <string>
 #include "core/framework/ort_value.h"
 #include "contrib_ops/cpu/utils/console_dumper.h"
@@ -11,7 +12,7 @@ namespace contrib {
 
 class CpuTensorConsoleDumper : public IConsoleDumper {
  public:
-  CpuTensorConsoleDumper() = default;
+  CpuTensorConsoleDumper();
   virtual ~CpuTensorConsoleDumper() {}
   void Print(const char* name, const float* tensor, int dim0, int dim1) const override;
   void Print(const char* name, const MLFloat16* tensor, int dim0, int dim1) const override;
@@ -33,6 +34,14 @@ class CpuTensorConsoleDumper : public IConsoleDumper {
   void Print(const char* name, const OrtValue& value) const override;
   void Print(const char* name, int index, bool end_line) const override;
   void Print(const char* name, const std::string& value, bool end_line) const override;
+
+  void Print(const std::string& value) const override;
+
+  // Output a vector with a threshold for max number of elements to output. Default threshold 0 means no limit.
+  template <typename T>
+  void Print(const char* name, const std::vector<T>& vec, size_t max_count = 0) const {
+    this->Print(name, vec.data(), 1, static_cast<int>(std::min(max_count, vec.size())));
+  }
 };
 
 }  // namespace contrib

onnxruntime/contrib_ops/cuda/utils/dump_cuda_tensor.cc

@@ -202,6 +202,10 @@ void DumpGpuTensor(const char* name, const Tensor& tensor) {
   DumpGpuTensor(nullptr, tensor, static_cast<int>(num_rows), static_cast<int>(row_size));
 }
 
+void CudaTensorConsoleDumper::Print(const std::string& value) const {
+  std::cout << value << std::endl;
+}
+
 void CudaTensorConsoleDumper::Print(const char* name, const size_t* tensor, int dim0, int dim1) const {
   if (is_enabled_)
     DumpGpuTensor<size_t>(name, tensor, dim0, dim1, true);
@@ -325,6 +329,10 @@ void CudaTensorConsoleDumper::Print(const char* name, const std::string& value,
 }
 
 #else
+
+void CudaTensorConsoleDumper::Print(const std::string&) const {
+}
+
 void CudaTensorConsoleDumper::Print(const char*, const size_t*, int, int) const {
 }
 

onnxruntime/contrib_ops/cuda/utils/dump_cuda_tensor.h

@@ -46,6 +46,8 @@ class CudaTensorConsoleDumper : public onnxruntime::contrib::IConsoleDumper {
   void Print(const char* name, const OrtValue& value) const override;
   void Print(const char* name, int index, bool end_line) const override;
   void Print(const char* name, const std::string& value, bool end_line) const override;
+
+  void Print(const std::string& value) const override;
 };
 
 }  // namespace cuda