[GitHub] [tvm] tqchen commented on a change in pull request #6917: Add Relay option to link parameters into runtime Modules

[GitBox <gi...@apache.org>]

tqchen commented on a change in pull request #6917:
URL: https://github.com/apache/tvm/pull/6917#discussion_r529696620



##########
File path: src/target/source/codegen_params.cc
##########
@@ -0,0 +1,218 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you 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.
+ */
+
+/*!
+ * \file codegen_params.cc
+ */
+
+#include "codegen_params.h"
+
+#include <dlpack/dlpack.h>
+
+#include <cmath>
+#include <iomanip>
+#include <memory>
+#include <string>
+
+namespace tvm {
+namespace codegen {
+
+namespace {
+class DLManagedTensorDeleter {
+ public:
+  void operator()(DLManagedTensor* ptr) { ptr->deleter(ptr); }
+};
+}  // namespace
+
+static constexpr const int kMaxLineLength = 80;
+
+template <typename T, typename = std::enable_if<std::is_integral<T>::value>>
+void PrintArray(void* data, size_t num_elements, int elements_per_row, std::string indent_str,
+                std::ostream& os) {
+  for (size_t i = 0; i < num_elements; i++) {
+    int64_t elem = static_cast<T*>(data)[i];
+    if (std::is_signed<T>::value) {
+      uint64_t to_print;
+      if (elem < 0) {
+        os << "-";
+        to_print = -elem;
+      } else {
+        os << "+";
+        to_print = elem;
+      }
+      os << "0x" << std::setw(sizeof(T) * 8 / 4) << static_cast<std::uint64_t>(to_print);
+    } else {
+      os << "0x" << std::setw(sizeof(T) * 8 / 4) << static_cast<std::uint64_t>(elem);
+    }
+    if (i < num_elements - 1) {
+      os << ", ";
+    }
+    if (((i + 1) % elements_per_row) == 0) {
+      os << "\n" << indent_str;
+    }
+  }
+}
+
+template <typename T, typename = std::enable_if<std::is_floating_point<T>::value>>
+void PrintArray(void* data, size_t num_elements, int one_element_size_bytes, int elements_per_row,
+                std::string indent_str, std::ostream& os) {
+  std::stringstream ss;
+  if (std::is_signed<T>::value) {
+    ss.setf(std::ios::hex | std::ios::showbase | std::ios::fixed | std::ios::scientific,
+            std::ios::basefield | std::ios::showbase | std::ios::floatfield);
+  } else {
+    ss.setf(std::ios::hex | std::ios::fixed | std::ios::scientific,
+            std::ios::basefield | std::ios::showbase | std::ios::floatfield);
+  }
+  for (size_t i = 0; i < num_elements; i++) {
+    T elem = static_cast<T*>(data)[i];
+    if (std::isinf(elem)) {
+      // C99 standard.
+      os << (elem < 0 ? "-" : " ") << std::setw(one_element_size_bytes - 1) << "INFINITY";
+    } else if (std::isnan(elem)) {
+      // GNU extension, implemenatation-dependent.
+      os << std::setw(one_element_size_bytes) << "NAN";
+    } else {
+      ss << elem;
+      os << std::setw(one_element_size_bytes) << ss.str();
+      ss.str("");
+    }
+    if (i < num_elements - 1) {
+      os << ", ";
+    }
+    if (((i + 1) % elements_per_row) == 0) {
+      os << "\n" << indent_str;
+    }
+  }
+}
+
+void NDArrayDataToC(::tvm::runtime::NDArray arr, int indent_chars, std::ostream& os) {
+  auto arr_type = arr.DataType();
+  CHECK_EQ(arr_type.lanes(), 1) << "CodegenParams: only support generating 1-lane parameters; saw "
+                                << arr_type.lanes();
+
+  int one_element_size_bytes = (arr_type.bits() / 4) + (2 /* "0x" */) + (2 /* ", " */);
+  if (arr_type.code() == runtime::DataType::TypeCode::kInt) {
+    one_element_size_bytes += 1;  // sign character
+    if (arr_type.bits() > 32) {
+      one_element_size_bytes += 2;  // "LL"
+    }
+  } else if (arr_type.code() == runtime::DataType::TypeCode::kUInt) {
+    if (arr_type.bits() > 32) {
+      one_element_size_bytes += 3;  // "ULL"
+    }
+  } else if (arr_type.code() == runtime::DataType::TypeCode::kFloat) {
+    // Floats and doubles are printed as hex but casted.
+    one_element_size_bytes += 1 /* sign */ + 1 /* decimal point */ + 1 /* exponent sign */;
+    if (arr_type.bits() == 64) {
+      one_element_size_bytes += 2; /* 4 decimal digits in exponent, relative to bits / 4 */
+    } else if (arr_type.bits() == 32) {
+      one_element_size_bytes += 1; /* extra decimal digit in exponent, relative to bits / 4 */
+    }
+  }
+
+  int elements_per_row = 16;
+  while (elements_per_row > 1 &&
+         (elements_per_row * one_element_size_bytes) > (kMaxLineLength - indent_chars)) {
+    elements_per_row /= 2;
+  }
+
+  std::string indent_str(indent_chars, ' ');
+  os << indent_str;
+
+  auto shape = arr.Shape();
+  int num_elements = 1;
+  for (auto shape_elem : shape) {
+    num_elements *= shape_elem;
+  }
+
+  std::unique_ptr<DLManagedTensor, DLManagedTensorDeleter> tensor(arr.ToDLPack());

Review comment:
       We don't need ToDLPack, and can directly get DLTensor* from the arr.operator->

##########
File path: src/target/source/codegen_params.cc
##########
@@ -0,0 +1,214 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you 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.
+ */
+
+/*!
+ * \file codegen_params.cc
+ */
+
+#include "codegen_params.h"
+
+#include <dlpack/dlpack.h>
+
+#include <cmath>
+#include <iomanip>
+#include <memory>
+#include <string>
+
+namespace tvm {
+namespace codegen {
+
+namespace {
+class DLManagedTensorDeleter {
+ public:
+  void operator()(DLManagedTensor* ptr) { ptr->deleter(ptr); }
+};
+}  // namespace
+
+static constexpr const int kMaxLineLength = 80;
+
+template <typename T, typename = std::enable_if<std::is_integral<T>::value>>
+void PrintArray(void* data, size_t num_elements, int elements_per_row, std::string indent_str,
+                std::ostream& os) {
+  for (size_t i = 0; i < num_elements; i++) {
+    int64_t elem = static_cast<T*>(data)[i];
+    if (std::is_signed<T>::value) {
+      uint64_t to_print;
+      if (elem < 0) {
+        os << "-";
+        to_print = -elem;
+      } else {
+        os << "+";
+        to_print = elem;
+      }
+      os << "0x" << std::setw(sizeof(T) * 8 / 4) << static_cast<std::uint64_t>(to_print);
+    } else {
+      os << "0x" << std::setw(sizeof(T) * 8 / 4) << static_cast<std::uint64_t>(elem);
+    }
+    if (i < num_elements - 1) {
+      os << ", ";
+    }
+    if (((i + 1) % elements_per_row) == 0) {
+      os << "\n" << indent_str;
+    }
+  }
+}
+
+template <typename T, typename = std::enable_if<std::is_floating_point<T>::value>>
+void PrintArray(void* data, size_t num_elements, int one_element_size_bytes, int elements_per_row,
+                std::string indent_str, std::ostream& os) {
+  std::stringstream ss;
+  if (std::is_signed<T>::value) {
+    ss.setf(std::ios::hex | std::ios::showbase | std::ios::fixed | std::ios::scientific,
+            std::ios::basefield | std::ios::showbase | std::ios::floatfield);
+  } else {
+    ss.setf(std::ios::hex | std::ios::fixed | std::ios::scientific,
+            std::ios::basefield | std::ios::showbase | std::ios::floatfield);
+  }
+  for (size_t i = 0; i < num_elements; i++) {
+    T elem = static_cast<T*>(data)[i];
+    if (std::isinf(elem)) {
+      // C99 standard.
+      os << (elem < 0 ? "-" : " ") << std::setw(one_element_size_bytes - 1) << "INFINITY";
+    } else if (std::isnan(elem)) {
+      // GNU extension, implemenatation-dependent.
+      os << std::setw(one_element_size_bytes) << "NAN";
+    } else {
+      ss << elem;
+      os << std::setw(one_element_size_bytes) << ss.str();
+      ss.str("");
+    }
+    if (i < num_elements - 1) {
+      os << ", ";
+    }
+    if (((i + 1) % elements_per_row) == 0) {
+      os << "\n" << indent_str;
+    }
+  }
+}
+
+void NDArrayDataToC(::tvm::runtime::NDArray arr, int indent_chars, std::ostream& os) {
+  auto arr_type = arr.DataType();
+  CHECK_EQ(arr_type.lanes(), 1) << "CodegenParams: only support generating 1-lane parameters; saw "
+                                << arr_type.lanes();
+
+  int one_element_size_bytes = (arr_type.bits() / 4) + (2 /* "0x" */) + (2 /* ", " */);
+  if (arr_type.code() == runtime::DataType::TypeCode::kInt) {
+    one_element_size_bytes += 1;  // sign character
+    if (arr_type.bits() > 32) {
+      one_element_size_bytes += 2;  // "LL"
+    }
+  } else if (arr_type.code() == runtime::DataType::TypeCode::kUInt) {
+    if (arr_type.bits() > 32) {
+      one_element_size_bytes += 3;  // "ULL"
+    }
+  } else if (arr_type.code() == runtime::DataType::TypeCode::kFloat) {
+    // Floats and doubles are printed as hex but casted.
+    one_element_size_bytes += 1 /* sign */ + 1 /* decimal point */ + 1 /* exponent sign */ +
+                              1 /* extra decimal digit in exponent */;
+  }
+
+  int elements_per_row = 16;
+  while (elements_per_row > 1 &&
+         (elements_per_row * one_element_size_bytes) > (kMaxLineLength - indent_chars)) {
+    elements_per_row /= 2;
+  }
+
+  std::string indent_str(indent_chars, ' ');
+  os << indent_str;
+
+  auto shape = arr.Shape();
+  int num_elements = 1;
+  for (auto shape_elem : shape) {
+    num_elements *= shape_elem;
+  }
+
+  std::unique_ptr<DLManagedTensor, DLManagedTensorDeleter> tensor(arr.ToDLPack());
+  auto old_fmtflags = os.flags();
+  os.setf(std::ios::internal | std::ios::hex,
+          std::ios::adjustfield | std::ios::basefield | std::ios::showbase);
+  os.fill('0');
+  switch (arr_type.code()) {
+    case runtime::DataType::kInt:
+      CHECK(arr_type.bits() == 8 || arr_type.bits() == 16 || arr_type.bits() == 32 ||
+            arr_type.bits() == 64)
+          << "CodegenParams: only support generating 8-, 16-, 32-, or 64-bit integer params; saw "
+          << arr_type.bits() << "-bit array";
+      if (arr_type.bits() == 8) {
+        PrintArray<int8_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str, os);
+      } else if (arr_type.bits() == 16) {
+        PrintArray<int16_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str, os);
+      } else if (arr_type.bits() == 32) {
+        PrintArray<int32_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str, os);
+      } else if (arr_type.bits() == 64) {
+        PrintArray<int64_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str, os);
+      } else {
+        CHECK(false) << "should not get here";
+      }
+      break;
+
+    case runtime::DataType::TypeCode::kUInt:
+      CHECK(arr_type.bits() == 8 || arr_type.bits() == 16 || arr_type.bits() == 32 ||
+            arr_type.bits() == 64)
+          << "CodegenParams: only support generating 8-, 16-, 32-, or 64-bit integer params; saw "
+          << arr_type.bits() << "-bit array";
+
+      if (arr_type.bits() == 8) {
+        PrintArray<uint8_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str, os);
+      } else if (arr_type.bits() == 16) {
+        PrintArray<uint16_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str,
+                             os);
+      } else if (arr_type.bits() == 32) {
+        PrintArray<uint32_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str,
+                             os);
+      } else if (arr_type.bits() == 64) {
+        PrintArray<uint64_t>(tensor->dl_tensor.data, num_elements, elements_per_row, indent_str,
+                             os);
+      } else {
+        CHECK(false) << "should not get here";
+      }
+      break;
+
+    case runtime::DataType::TypeCode::kFloat: {
+      os.fill(' ');
+      os.setf(std::ios::left, std::ios::adjustfield);
+      if (arr_type.bits() == 32) {
+        PrintArray<float>(tensor->dl_tensor.data, num_elements, one_element_size_bytes,

Review comment:
       To put it in another way. We will need support for bfloat16 and fp16 now given that they are already part of the type system. Addtiionally, there might also be need to support customized data types. 
   
   The main reason to go for the uint repr is its bit accurate precision(we don't need to worry about printing out INF, nan, or loss of accuracy due to printing and parsing of floating pt -- although partly addressed by hex printing, still somewhat complicated) and simplicity in implementation.
   
   Additionally, for big endian machines and small endian machines we are already using the bits as the only information in our runtime https://github.com/apache/tvm/blob/main/include/tvm/runtime/ndarray.h#L429, which means our runtime won't simply work beyond such cases. Right now we also did not see an example that goes beyond these two cases.
   
   If we do think it is important to keep float and double printing as their own type. I think we could do that, but need to make sure the rest of the codepath uses uint so that it works for bfloat16, fp16 and other potential extension data types now.




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org