[executorch][flat_tensor] DataMap implementation

extension/flat_tensor/named_data_map/TARGETS

@@ -0,0 +1,6 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+load(":targets.bzl", "define_common_targets")
+
+oncall("executorch")
+
+define_common_targets()

extension/flat_tensor/named_data_map/data_map.cpp

@@ -0,0 +1,215 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/extension/flat_tensor/named_data_map/data_map.h>
+#include <executorch/extension/flat_tensor/serialize/flat_tensor_header.h>
+#include <executorch/extension/flat_tensor/serialize/schema_generated.h>
+#include <executorch/runtime/core/error.h>
+#include <executorch/runtime/core/exec_aten/util/tensor_util.h>
+#include <executorch/runtime/core/freeable_buffer.h>
+#include <executorch/runtime/core/result.h>
+#include <executorch/runtime/core/span.h>
+#include <executorch/runtime/platform/compiler.h>
+
+#include <tuple>
+#include <unordered_map>
+
+using executorch::runtime::Error;
+using executorch::runtime::FreeableBuffer;
+using executorch::runtime::Result;
+using executorch::runtime::Span;
+
+using executorch::aten::ScalarType;
+using executorch::runtime::DataLoader;
+using executorch::runtime::TensorLayout;
+
+namespace executorch {
+namespace extension {
+
+namespace {
+/**
+ * FlatTensor data must be aligned to this value to properly parse it. Must be a
+ * power of 2. Note that max_align_t is the alignment that malloc() and new
+ * guarantee.
+ */
+constexpr size_t kMinimumAlignment = alignof(std::max_align_t);
+
+bool IsAligned(const void* data) {
+  uintptr_t addr = reinterpret_cast<uintptr_t>(data);
+  return addr % kMinimumAlignment == 0;
+}
+} // namespace
+
+ET_NODISCARD Result<const TensorLayout> DataMap::get_metadata(
+    const char* key) const {
+  auto result = _name_to_tensor.find(key);
+  if (result == _name_to_tensor.end()) {
+    return Error::InvalidArgument;
+  }
+  // value is a tuple of (segment_index, offset, tensor_layout)
+  return std::get<2>(result->second);
+}
+
+ET_NODISCARD Result<FreeableBuffer> DataMap::get_data(const char* key) const {
+  auto result = _name_to_tensor.find(key);
+  if (result == _name_to_tensor.end()) {
+    return Error::InvalidArgument;
+  }
+  int offset = std::get<1>(result->second);
+  TensorLayout tensor = std::get<2>(result->second);
+
+  const uint8_t* data = static_cast<const uint8_t*>(_data_ro.data()) + offset;
+  return FreeableBuffer(data, tensor.nbytes(), nullptr);
+}
+
+ET_NODISCARD Result<size_t>
+DataMap::load_data_into(const char* key, void* buffer, size_t size) const {
+  return Error::NotImplemented;
+}
+
+ET_NODISCARD Result<size_t> DataMap::get_num_keys() const {
+  return _name_to_tensor.size();
+}
+
+ET_NODISCARD Result<const char*> DataMap::get_key(size_t index) const {
+  if (index < 0 || index >= _name_to_tensor.size()) {
+    return Error::InvalidArgument;
+  }
+
+  auto iter = _name_to_tensor.begin();
+  for (int i = 0; i < index; ++i) {
+    ++iter;
+  }
+  return iter->first.c_str();
+}
+
+/* static */ Result<DataMap> DataMap::load(DataLoader* loader) {
+  // Load data map.
+  size_t flatbuffer_offset = 0;
+  size_t flatbuffer_size = 0;
+  size_t segment_base_offset = 0;
+  size_t segment_data_size = 0;
+  {
+    // Check header.
+    Result<FreeableBuffer> header = loader->load(
+        /*offset=*/0,
+        FlatTensorHeader::kNumHeadBytes,
+        DataLoader::SegmentInfo(DataLoader::SegmentInfo::Type::External));
+    if (!header.ok()) {
+      return header.error();
+    }
+    Result<FlatTensorHeader> fh =
+        FlatTensorHeader::Parse(header->data(), header->size());
+    if (fh.ok()) {
+      // The header has the data map size.
+      flatbuffer_offset = fh->flatbuffer_offset;
+      flatbuffer_size = fh->flatbuffer_size;
+      segment_base_offset = fh->segment_base_offset;
+      segment_data_size = fh->segment_data_size;
+    } else if (fh.error() == Error::NotFound) {
+      // No header, throw error.
+      ET_LOG(Error, "No FlatTensorHeader found.");
+      return fh.error();
+    } else {
+      // corruption, throw error.
+      ET_LOG(Error, "Flat tensor header may be corrupt.");
+      return fh.error();
+    }
+  }
+
+  // Load flatbuffer data as a segment.
+  Result<FreeableBuffer> flat_tensor_data = loader->load(
+      /*offset=*/flatbuffer_offset,
+      flatbuffer_size,
+      DataLoader::SegmentInfo(DataLoader::SegmentInfo::Type::External));
+  if (!flat_tensor_data.ok()) {
+    return flat_tensor_data.error();
+  }
+
+  // Make sure magic matches.
+  if (!flat_tensor_flatbuffer::FlatTensorBufferHasIdentifier(
+          flat_tensor_data->data())) {
+    ET_LOG(
+        Error,
+        "FlatTensor identifier '%.4s' != expected '%.4s'",
+        flatbuffers::GetBufferIdentifier(flat_tensor_data->data()),
+        flat_tensor_flatbuffer::FlatTensorIdentifier());
+    return Error::InvalidExternalData;
+  }
+
+  // The flatbuffer data must start at an aligned address to ensure internal
+  // alignment of flatbuffer fields.
+  ET_CHECK_OR_RETURN_ERROR(
+      IsAligned(flat_tensor_data->data()),
+      InvalidArgument,
+      "FlatTensor data 0x%p must be aligned to %zu",
+      flat_tensor_data->data(),
+      kMinimumAlignment);
+
+  // Get pointer to root of flatbuffer table.
+  const flat_tensor_flatbuffer::FlatTensor* flat_tensor =
+      flat_tensor_flatbuffer::GetFlatTensor(flat_tensor_data->data());
+
+  // Get pointer to tensor metadata.
+  const auto* s_tensor_metadata = flat_tensor->tensors();
+  assert(s_tensor_metadata != nullptr);
+
+  std::unordered_map<std::string, std::tuple<int, int, TensorLayout>>
+      name_to_tensor = {};
+  for (int i = 0; i < s_tensor_metadata->size(); i++) {
+    // Create TensorLayouts.
+    ScalarType scalar_type =
+        static_cast<ScalarType>(s_tensor_metadata->Get(i)->scalar_type());
+    const int dim = s_tensor_metadata->Get(i)->sizes()->size();
+
+    const auto serialized_sizes = s_tensor_metadata->Get(i)->sizes()->data();
+    const auto serialized_dim_order =
+        s_tensor_metadata->Get(i)->dim_order()->data();
+    TensorLayout tensor_layout = TensorLayout(
+        scalar_type,
+        Span<const int32_t>(serialized_sizes, dim),
+        Span<const uint8_t>(serialized_dim_order, dim));
+
+    int segment_index = s_tensor_metadata->Get(i)->segment_index();
+    int offset = s_tensor_metadata->Get(i)->offset();
+    std::string key = s_tensor_metadata->Get(i)->fully_qualified_name()->str();
+
+    auto val = std::make_tuple(segment_index, offset, tensor_layout);
+    name_to_tensor.insert({key, std::move(val)});
+  }
+
+  // Load constant data.
+  const auto* s_data_segment = flat_tensor->segments();
+
+  // Only support one segment for now.
+  assert(s_data_segment->size() == 1);
+  // First segment offset should be 0.
+  int segment_offset = s_data_segment->Get(0)->offset();
+  assert(segment_offset == 0);
+  // First segment size should be <= the total segment data size.
+  int segment_size = s_data_segment->Get(0)->size();
+  assert(segment_size <= segment_data_size);
+
+  Result<FreeableBuffer> _data_ro = loader->load(
+      /*offset=*/segment_base_offset + segment_offset,
+      segment_size,
+      DataLoader::SegmentInfo(DataLoader::SegmentInfo::Type::External));
+  if (!_data_ro.ok()) {
+    return _data_ro.error();
+  }
+
+  return DataMap(
+      std::move(flat_tensor_data.get()),
+      std::move(name_to_tensor),
+      std::move(_data_ro.get()));
+}
+
+DataMap::~DataMap() {}
+
+} // namespace extension
+} // namespace executorch

extension/flat_tensor/named_data_map/data_map.h

@@ -0,0 +1,84 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#pragma once
+
+#include <executorch/runtime/core/data_loader.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/named_data_map.h>
+#include <executorch/runtime/core/result.h>
+#include <executorch/runtime/core/tensor_layout.h>
+#include <executorch/runtime/platform/compiler.h>
+
+#include <unordered_map>
+#include <utility>
+
+// Forward declare flatbuffer types. This is a public header and must not
+// include the generated flatbuffer header.
+namespace flat_tensor_flatbuffer {
+struct FlatTensor;
+} // namespace flat_tensor
+
+namespace executorch {
+namespace extension {
+
+class DataMap final : public executorch::runtime::NamedDataMap {
+ public:
+  static executorch::runtime::Result<DataMap> load(
+      executorch::runtime::DataLoader* loader);
+
+  ET_NODISCARD
+  executorch::runtime::Result<const executorch::runtime::TensorLayout>
+  get_metadata(const char* key) const override;
+  ET_NODISCARD
+  executorch::runtime::Result<executorch::runtime::FreeableBuffer> get_data(
+      const char* key) const override;
+  ET_NODISCARD executorch::runtime::Result<size_t>
+  load_data_into(const char* key, void* buffer, size_t size) const override;
+
+  ET_NODISCARD executorch::runtime::Result<size_t> get_num_keys()
+      const override;
+  ET_NODISCARD executorch::runtime::Result<const char*> get_key(
+      size_t index) const override;
+
+  DataMap(DataMap&&) noexcept = default;
+  ~DataMap() override;
+
+ private:
+  DataMap(
+      executorch::runtime::FreeableBuffer&& flat_tensor_data,
+      std::unordered_map<
+          std::string,
+          std::tuple<int, int, executorch::runtime::TensorLayout>>
+          name_to_tensor,
+      executorch::runtime::FreeableBuffer&& data_ro)
+      : _flat_tensor_data(std::move(flat_tensor_data)),
+        _name_to_tensor(std::move(name_to_tensor)),
+        _data_ro(std::move(data_ro)) {}
+
+  // Not copyable or assignable.
+  DataMap(const DataMap& rhs) = delete;
+  DataMap& operator=(DataMap&& rhs) noexcept = delete;
+  DataMap& operator=(const DataMap& rhs) = delete;
+
+  // FlatTensor flatbuffer data. Contains the data backing up
+  // TensorLayout information in the _name_to_tensor map; must outlive it.
+  executorch::runtime::FreeableBuffer _flat_tensor_data;
+
+  // Map of name to {segment index, offset, TensorLayout}.
+  std::unordered_map<
+      std::string,
+      std::tuple<int, int, executorch::runtime::TensorLayout>>
+      _name_to_tensor;
+
+  // Raw, read-only tensor data.
+  executorch::runtime::FreeableBuffer _data_ro;
+};
+
+} // namespace extension
+} // namespace executorch

extension/flat_tensor/named_data_map/targets.bzl

@@ -0,0 +1,23 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+def define_common_targets():
+    runtime.cxx_library(
+        name = "data_map",
+        srcs = [
+            "data_map.cpp",
+        ],
+        exported_headers = ["data_map.h"],
+        deps = [
+            "//executorch/extension/flat_tensor/serialize:schema",
+            "//executorch/extension/flat_tensor/serialize:serialize",
+            "//executorch/extension/flat_tensor/serialize:generated_headers",
+            "//executorch/extension/flat_tensor/serialize:flat_tensor_header",
+            "//executorch/runtime/core:core",
+            "//executorch/runtime/core:evalue",
+            "//executorch/runtime/core/exec_aten:lib",
+            "//executorch/runtime/core/exec_aten/util:tensor_util",
+        ],
+        visibility = [
+            "//executorch/...",
+        ],
+    )

extension/flat_tensor/test/TARGETS

@@ -6,7 +6,7 @@ load(":targets.bzl", "define_common_targets")
 
 oncall("executorch")
 
-define_common_targets()
+define_common_targets(is_fbcode=True)
 
 python_unittest(
     name = "serialize",