[slimtensor] Add from_etensor factory function for ETensor to SlimTensor conversion

backends/aoti/slim/factory/from_etensor.h

@@ -0,0 +1,108 @@
+/*
+ * 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/backends/aoti/slim/core/slim_tensor.h>
+#include <executorch/backends/aoti/slim/factory/empty.h>
+#include <executorch/backends/aoti/slim/util/array_ref_util.h>
+#include <executorch/runtime/core/error.h>
+#include <executorch/runtime/core/exec_aten/util/tensor_util.h>
+#include <executorch/runtime/core/portable_type/tensor.h>
+
+namespace executorch::backends::aoti::slim {
+
+/// Creates a SlimTensor from an ETensor (ExecuTorch portable tensor).
+///
+/// This factory function converts an ETensor to a SlimTensor, optionally
+/// copying the data to a target device. The source device can be specified
+/// to support ETensors residing on different devices (CPU, CUDA).
+///
+/// @param etensor The source ETensor.
+/// @param source_device The device where the ETensor data resides.
+/// @param target_device The target device for the output SlimTensor.
+/// @return A new SlimTensor with data copied to the target device.
+///
+/// @note ETensor uses int32_t (SizesType/StridesType) for sizes and strides,
+///       while SlimTensor uses int64_t. This function handles the conversion.
+///
+/// Example usage:
+/// @code
+///   // CPU ETensor to GPU SlimTensor
+///   auto* cpu_tensor = &(args[i]->toTensor());
+///   SlimTensor gpu_tensor = from_etensor(*cpu_tensor, CPU_DEVICE,
+///                                        DEFAULT_CUDA_DEVICE);
+///
+///   // GPU ETensor to GPU SlimTensor (same device copy)
+///   SlimTensor gpu_tensor2 = from_etensor(*gpu_etensor, DEFAULT_CUDA_DEVICE,
+///                                         DEFAULT_CUDA_DEVICE);
+///
+///   // GPU ETensor to CPU SlimTensor
+///   SlimTensor cpu_tensor = from_etensor(*gpu_etensor, DEFAULT_CUDA_DEVICE,
+///                                        CPU_DEVICE);
+/// @endcode
+inline SlimTensor from_etensor(
+    const executorch::runtime::etensor::Tensor& etensor,
+    const c10::Device& source_device = CPU_DEVICE,
+    const c10::Device& target_device = CPU_DEVICE) {
+  // Step 1: Extract metadata from ETensor
+  const auto ndim = static_cast<size_t>(etensor.dim());
+
+  // Convert sizes from exec_aten::SizesType (int32_t) to int64_t
+  std::vector<int64_t> sizes_vec(ndim);
+  for (size_t i = 0; i < ndim; ++i) {
+    sizes_vec[i] = static_cast<int64_t>(etensor.size(static_cast<ssize_t>(i)));
+  }
+
+  // Convert strides from exec_aten::StridesType (int32_t) to int64_t
+  std::vector<int64_t> strides_vec(ndim);
+  auto etensor_strides = etensor.strides();
+  for (size_t i = 0; i < ndim; ++i) {
+    strides_vec[i] = static_cast<int64_t>(etensor_strides[i]);
+  }
+
+  // Map ETensor ScalarType to SlimTensor ScalarType
+  c10::ScalarType dtype = static_cast<c10::ScalarType>(etensor.scalar_type());
+
+  // Step 2: Create SlimTensor on target device
+  SlimTensor result = empty_strided(
+      makeArrayRef(sizes_vec), makeArrayRef(strides_vec), dtype, target_device);
+
+  // Step 3: Copy data from ETensor (source device) to SlimTensor (target
+  // device) Supports CPU→CPU, CPU→CUDA, CUDA→CPU, or CUDA→CUDA copy
+  const void* src_data = etensor.const_data_ptr();
+  void* dst_data = result.data_ptr();
+  size_t nbytes = etensor.nbytes();
+
+  if (nbytes > 0) {
+    // const_cast is safe here because copy_ only reads from src_data
+    result.storage()->copy_(
+        dst_data, const_cast<void*>(src_data), nbytes, source_device);
+  }
+
+  return result;
+}
+
+/// Creates a SlimTensor from an ETensor pointer.
+///
+/// Convenience overload that accepts a pointer instead of a reference.
+///
+/// @param etensor Pointer to the source ETensor (must not be null).
+/// @param source_device The device where the ETensor data resides.
+/// @param target_device The target device for the output SlimTensor.
+/// @return A new SlimTensor with data copied to the target device.
+inline SlimTensor from_etensor(
+    const executorch::runtime::etensor::Tensor* etensor,
+    const c10::Device& source_device = CPU_DEVICE,
+    const c10::Device& target_device = CPU_DEVICE) {
+  ET_CHECK_MSG(
+      etensor != nullptr, "from_etensor: etensor pointer cannot be nullptr");
+  return from_etensor(*etensor, source_device, target_device);
+}
+
+} // namespace executorch::backends::aoti::slim

backends/aoti/slim/factory/targets.bzl

@@ -29,3 +29,16 @@ def define_common_targets():
             "//executorch/backends/aoti/slim/util:size_util",
         ],
     )
+
+    runtime.cxx_library(
+        name = "from_etensor",
+        headers = [
+            "from_etensor.h",
+        ],
+        visibility = ["@EXECUTORCH_CLIENTS"],
+        exported_deps = [
+            "//executorch/backends/aoti/slim/factory:empty",
+            "//executorch/backends/aoti/slim/util:array_ref_util",
+            "//executorch/runtime/core/portable_type:portable_type",
+        ],
+    )

backends/aoti/slim/factory/test/targets.bzl

@@ -44,3 +44,17 @@ def define_common_targets():
             ],
             **backend_kwargs
         )
+
+        runtime.cxx_test(
+            name = "test_from_etensor" + backend_suffix,
+            srcs = [
+                "test_from_etensor.cpp",
+            ],
+            deps = [
+                "//executorch/backends/aoti/slim/core:storage",
+                "//executorch/backends/aoti/slim/factory:empty",
+                "//executorch/backends/aoti/slim/factory:from_etensor",
+                "//executorch/runtime/core/exec_aten/testing_util:tensor_util",
+            ],
+            **backend_kwargs
+        )