Introduce torch::executor::TensorAccessor

Differential Revision: D66033489

Pull Request resolved: pytorch#6905

extension/tensor/targets.bzl

@@ -18,6 +18,7 @@ def define_common_targets():
             ],
             exported_headers = [
                 "tensor.h",
+                "tensor_accessor.h",
                 "tensor_ptr.h",
                 "tensor_ptr_maker.h",
             ],

extension/tensor/tensor_accessor.h

@@ -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.
+ */
+
+#pragma once
+
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+#include <executorch/runtime/core/result.h>
+
+namespace executorch {
+namespace extension {
+namespace internal {
+
+/**
+ * Base class template storing the underlying data with size and stride helpers.
+ * Inherited by TensorAccessor<> which requires specialization on rank.
+ */
+template <typename T, ssize_t N>
+class TensorAccessorBase {
+ public:
+  /// Returns the size of the underlying tensor at the given dimension.
+  executorch::aten::SizesType size(ssize_t i) const {
+    ET_CHECK_MSG(
+        i < dim_ && i >= 0,
+        "Dimension outside of [0, %zd], got %zd",
+        dim_ - 1,
+        i);
+    return sizes_[i];
+  }
+
+  /// Returns the stride of the underlying tensor at the given dimension.
+  executorch::aten::StridesType stride(ssize_t i) const {
+    ET_CHECK_MSG(
+        i < dim_ && i >= 0,
+        "Dimension outside of [0, %zd], got %zd",
+        dim_ - 1,
+        i);
+    return strides_[i];
+  }
+
+ protected:
+  TensorAccessorBase(
+      T* data,
+      const executorch::aten::SizesType* sizes,
+      const executorch::aten::StridesType* strides,
+      ssize_t dim)
+      : data_(data), sizes_(sizes), strides_(strides), dim_(dim) {}
+
+  T* data_;
+  const executorch::aten::SizesType* sizes_;
+  const executorch::aten::StridesType* strides_;
+  ssize_t dim_;
+};
+
+} // namespace internal
+
+/**
+ * TensorAccessor template with data type and rank as template parameters. No
+ * public constructors, can only be created using make_tensor_accessor from a
+ * given executorch::aten::Tensor. Use operator[] to index and obtain a lower
+ * rank accessor or the underlying scalar value.
+ */
+template <typename T, ssize_t N>
+class TensorAccessor : public internal::TensorAccessorBase<T, N> {
+ public:
+  /**
+   * Index into the the outer most dimension.
+   *
+   * @param i Index.
+   * @return If N > 1, a TensorAccessor with N-1 dimensions. If N == 1, a
+   * reference to the underlying scalar. Refer to the TensorAccessor<T, 1>
+   * specialization.
+   */
+  TensorAccessor<T, N - 1> operator[](ssize_t i) {
+    return TensorAccessor<T, N - 1>(
+        this->data_ + this->strides_[0] * i,
+        this->sizes_ + 1,
+        this->strides_ + 1,
+        N - 1);
+  }
+
+  /**
+   * Index into the the outer most dimension.
+   *
+   * @param i Index.
+   * @return If N > 1, a constant TensorAccessor with N-1 dimensions. If N == 1,
+   * a constant reference to the underlying scalar. Refer to the
+   * TensorAccessor<T, 1> specialization.
+   */
+  const TensorAccessor<T, N - 1> operator[](ssize_t i) const {
+    return TensorAccessor<T, N - 1>(
+        this->data_ + this->strides_[0] * i,
+        this->sizes_ + 1,
+        this->strides_ + 1,
+        N - 1);
+  }
+
+ private:
+  TensorAccessor(
+      T* data,
+      const executorch::aten::SizesType* sizes,
+      const executorch::aten::StridesType* strides,
+      ssize_t dim)
+      : internal::TensorAccessorBase<T, N>(data, sizes, strides, dim) {}
+
+  template <typename T2, ssize_t N2>
+  friend class TensorAccessor;
+
+  template <typename T2, ssize_t N2>
+  friend executorch::runtime::Result<TensorAccessor<T2, N2>>
+  make_tensor_accessor(const executorch::aten::Tensor& t);
+};
+
+/**
+ * TensorAccessor specialization for N == 1, where operator[] returns a
+ * reference to the underlying scalar.
+ */
+template <typename T>
+class TensorAccessor<T, 1> : public internal::TensorAccessorBase<T, 1> {
+ public:
+  /**
+   * Index into the the outer most dimension.
+   *
+   * @param i Index.
+   * @return Reference to the underlying scalar.
+   */
+  T& operator[](ssize_t i) {
+    return this->data_[this->strides_[0] * i];
+  }
+
+  /**
+   * Index into the the outer most dimension.
+   *
+   * @param i Index.
+   * @return Constant reference to the underlying scalar.
+   */
+  const T& operator[](ssize_t i) const {
+    return this->data_[this->strides_[0] * i];
+  }
+
+ private:
+  TensorAccessor(
+      T* data,
+      const executorch::aten::SizesType* sizes,
+      const executorch::aten::StridesType* strides,
+      ssize_t dim)
+      : internal::TensorAccessorBase<T, 1>(data, sizes, strides, dim) {}
+
+  template <typename T2, ssize_t N2>
+  friend class TensorAccessor;
+
+  template <typename T2, ssize_t N2>
+  friend executorch::runtime::Result<TensorAccessor<T2, N2>>
+  make_tensor_accessor(const executorch::aten::Tensor& t);
+};
+
+/**
+ * Creates a TensorAccessor<T, N> from the given tensor. The number of dimension
+ * N and the data type T's size must match those of the input tensor. For
+ * Executorch tensors, non-trivial dimension order is not supported.
+ *
+ * @param tensor Origin tensor. The TensorImpl inside must outlive the returned
+ * TensorAccessor.
+ * @return TensorAccessor of the input tensor.
+ * @retval Error::InvalidArgument Mismatch on data type or number of dimensions.
+ * @retval Error::NotSupported Input tensor has non-trivial dimension onrder.
+ */
+template <typename T, ssize_t N>
+executorch::runtime::Result<TensorAccessor<T, N>> make_tensor_accessor(
+    const executorch::aten::Tensor& tensor) {
+  static_assert(
+      N > 0,
+      "TensorAccessor is used for indexing tensors, for scalar use *_data_ptr<T>()");
+
+  if (N != tensor.dim()) {
+    ET_LOG(
+        Error, "Expecting %zd dimensions but tensor has %zd.", N, tensor.dim());
+    return executorch::runtime::Error::InvalidArgument;
+  }
+
+  if (sizeof(T) != tensor.element_size()) {
+    ET_LOG(
+        Error,
+        "Size of data type template argument (%zd) not equal to tensor element size (%zd)",
+        sizeof(T),
+        tensor.element_size());
+    return executorch::runtime::Error::InvalidArgument;
+  }
+
+#ifndef USE_ATEN_LIB
+  auto dim_order = tensor.dim_order();
+  for (ssize_t i = 0; i < dim_order.size(); i++) {
+    if (dim_order[i] != i) {
+      ET_LOG(Error, "Non-trival dim_order not supported.");
+      return executorch::runtime::Error::NotSupported;
+    }
+  }
+#endif
+
+  T* ptr = nullptr;
+  if constexpr (std::is_const_v<T>) {
+    ptr = tensor.const_data_ptr<T>();
+  } else {
+    ptr = tensor.mutable_data_ptr<T>();
+  }
+  return TensorAccessor<T, N>(
+      ptr, tensor.sizes().data(), tensor.strides().data(), N);
+}
+
+} // namespace extension
+} // namespace executorch

extension/tensor/test/targets.bzl

@@ -13,6 +13,7 @@ def define_common_targets():
         runtime.cxx_test(
             name = "test" + aten_suffix,
             srcs = [
+                "tensor_accessor_test.cpp",
                 "tensor_ptr_maker_test.cpp",
                 "tensor_ptr_test.cpp",
             ],

extension/tensor/test/tensor_accessor_test.cpp

@@ -0,0 +1,155 @@
+/*
+ * 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/tensor/tensor_accessor.h>
+
+#include <gtest/gtest.h>
+#include <vector>
+
+#include <executorch/extension/tensor/tensor_ptr.h>
+#include <executorch/runtime/platform/runtime.h>
+
+using executorch::extension::make_tensor_accessor;
+using executorch::extension::make_tensor_ptr;
+using executorch::extension::TensorAccessor;
+
+class TensorAccessorTest : public ::testing::Test {
+ protected:
+  static void SetUpTestSuite() {
+    executorch::runtime::runtime_init();
+  }
+};
+
+TEST_F(TensorAccessorTest, From1DTensor) {
+  constexpr int32_t kN = 16;
+  std::vector<uint8_t> data(kN, 0);
+  for (int32_t i = 0; i < kN; i++) {
+    data[i] = i;
+  }
+
+  auto tensor =
+      make_tensor_ptr({kN}, data.data(), executorch::aten::ScalarType::Byte);
+  auto tensor_accessor = make_tensor_accessor<uint8_t, 1>(*tensor.get());
+  EXPECT_TRUE(tensor_accessor.ok());
+  for (int32_t i = 0; i < kN; i++) {
+    EXPECT_EQ(tensor_accessor.get()[i], i);
+  }
+}
+
+int32_t
+value_at_pos_in_4d_int_tensor(int32_t n, int32_t c, int32_t h, int32_t w) {
+  // just encode the position into the value, assuming dimensions fit in 8 bits
+  return (n << 24) | (c << 16) | (h << 8) | w;
+}
+
+void check_4d_int_tensor_accessor(
+    TensorAccessor<int32_t, 4> accessor,
+    int32_t N,
+    int32_t C,
+    int32_t H,
+    int32_t W) {
+  for (int32_t n = 0; n < N; n++) {
+    for (int32_t c = 0; c < C; c++) {
+      for (int32_t h = 0; h < H; h++) {
+        for (int32_t w = 0; w < W; w++) {
+          EXPECT_EQ(
+              accessor[n][c][h][w], value_at_pos_in_4d_int_tensor(n, c, h, w));
+        }
+      }
+    }
+  }
+}
+
+TEST_F(TensorAccessorTest, From4DTensor) {
+  constexpr int32_t kN = 2;
+  constexpr int32_t kC = 8;
+  constexpr int32_t kH = 4;
+  constexpr int32_t kW = 6;
+  std::vector<int32_t> data(kN * kC * kH * kW, 0);
+  size_t idx = 0;
+  for (int32_t n = 0; n < kN; n++) {
+    for (int32_t c = 0; c < kC; c++) {
+      for (int32_t h = 0; h < kH; h++) {
+        for (int32_t w = 0; w < kW; w++) {
+          data[idx++] = value_at_pos_in_4d_int_tensor(n, c, h, w);
+        }
+      }
+    }
+  }
+
+  auto tensor = make_tensor_ptr(
+      {kN, kC, kH, kW}, data.data(), executorch::aten::ScalarType::Int);
+  auto accessor = make_tensor_accessor<int32_t, 4>(*tensor.get());
+  EXPECT_TRUE(accessor.ok());
+  check_4d_int_tensor_accessor(accessor.get(), kN, kC, kH, kW);
+}
+
+#ifdef USE_ATEN_LIB // Non-contiguous tensor is only allowed in ATen mode.
+TEST_F(TensorAccessorTest, FromNonContiguousTensor) {
+  constexpr int32_t kN = 2;
+  constexpr int32_t kC = 8;
+  constexpr int32_t kH = 4;
+  constexpr int32_t kW = 6;
+  constexpr int32_t kW_padded = 8;
+  std::vector<int32_t> data(kN * kC * kH * kW_padded, 0);
+  std::array<executorch::aten::SizesType, 4> sizes = {kN, kC, kH, kW};
+  std::array<executorch::aten::StridesType, 4> strides = {
+      kC * kH * kW_padded,
+      1, // channel last
+      kC * kW_padded, // width is padded
+      kC};
+
+  size_t idx = 0;
+  for (int32_t n = 0; n < kN; n++) {
+    for (int32_t h = 0; h < kH; h++) {
+      for (int32_t w = 0; w < kW_padded; w++) {
+        for (int32_t c = 0; c < kC; c++) {
+          data[idx++] = value_at_pos_in_4d_int_tensor(n, c, h, w);
+        }
+      }
+    }
+  }
+
+  auto tensor = at::from_blob(
+      data.data(), sizes, strides, at::TensorOptions().dtype(at::kInt));
+  auto accessor = make_tensor_accessor<int32_t, 4>(tensor);
+  EXPECT_TRUE(accessor.ok());
+  check_4d_int_tensor_accessor(accessor.get(), kN, kC, kH, kW);
+}
+#endif // ifdef USE_ATEN_LIB
+
+TEST_F(TensorAccessorTest, FailOnIncorrectDtypeOrRank) {
+  constexpr int32_t kN = 16;
+  std::vector<float> data(kN, 0);
+  auto tensor = make_tensor_ptr({kN}, data.data());
+
+  // Tensor has rank 1 but creating accessor with rank 2.
+  auto fail1 = make_tensor_accessor<float, 2>(*tensor.get());
+  EXPECT_FALSE(fail1.ok());
+
+  // Tensor has dtype float but creating accoessor with dtype uint8_t.
+  auto fail2 = make_tensor_accessor<uint8_t, 1>(*tensor.get());
+  EXPECT_FALSE(fail2.ok());
+}
+
+#ifndef USE_ATEN_LIB // Dim order is only defined for portable Tensor
+TEST_F(TensorAccessorTest, FailOnNonTrivialDimOrder) {
+  constexpr int32_t kN = 8;
+  constexpr int32_t kM = 16;
+  std::vector<float> data(kN * kM, 0);
+  auto tensor = make_tensor_ptr(
+      {kN, kM},
+      data.data(),
+      /*dim_order=*/{1, 0},
+      /*strides=*/{1, kN});
+
+  // Non trivial dim order is not supported.
+  auto fail = make_tensor_accessor<float, 2>(*tensor.get());
+  EXPECT_FALSE(fail.ok());
+}
+#endif // ifndef USE_ATEN_LIB