|
| 1 | +/* |
| 2 | + * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | + * All rights reserved. |
| 4 | + * |
| 5 | + * This source code is licensed under the BSD-style license found in the |
| 6 | + * LICENSE file in the root directory of this source tree. |
| 7 | + */ |
| 8 | + |
| 9 | +#pragma once |
| 10 | + |
| 11 | +#include <executorch/backends/aoti/slim/core/slim_tensor.h> |
| 12 | +#include <executorch/backends/aoti/slim/factory/empty.h> |
| 13 | +#include <executorch/backends/aoti/slim/util/array_ref_util.h> |
| 14 | +#include <executorch/runtime/core/error.h> |
| 15 | +#include <executorch/runtime/core/exec_aten/util/tensor_util.h> |
| 16 | +#include <executorch/runtime/core/portable_type/tensor.h> |
| 17 | + |
| 18 | +namespace executorch::backends::aoti::slim { |
| 19 | + |
| 20 | +/// Creates a SlimTensor from an ETensor (ExecuTorch portable tensor). |
| 21 | +/// |
| 22 | +/// This factory function converts an ETensor to a SlimTensor, optionally |
| 23 | +/// copying the data to a target device. The source device can be specified |
| 24 | +/// to support ETensors residing on different devices (CPU, CUDA). |
| 25 | +/// |
| 26 | +/// @param etensor The source ETensor. |
| 27 | +/// @param source_device The device where the ETensor data resides. |
| 28 | +/// @param target_device The target device for the output SlimTensor. |
| 29 | +/// @return A new SlimTensor with data copied to the target device. |
| 30 | +/// |
| 31 | +/// @note ETensor uses int32_t (SizesType/StridesType) for sizes and strides, |
| 32 | +/// while SlimTensor uses int64_t. This function handles the conversion. |
| 33 | +/// |
| 34 | +/// Example usage: |
| 35 | +/// @code |
| 36 | +/// // CPU ETensor to GPU SlimTensor |
| 37 | +/// auto* cpu_tensor = &(args[i]->toTensor()); |
| 38 | +/// SlimTensor gpu_tensor = from_etensor(*cpu_tensor, CPU_DEVICE, |
| 39 | +/// DEFAULT_CUDA_DEVICE); |
| 40 | +/// |
| 41 | +/// // GPU ETensor to GPU SlimTensor (same device copy) |
| 42 | +/// SlimTensor gpu_tensor2 = from_etensor(*gpu_etensor, DEFAULT_CUDA_DEVICE, |
| 43 | +/// DEFAULT_CUDA_DEVICE); |
| 44 | +/// |
| 45 | +/// // GPU ETensor to CPU SlimTensor |
| 46 | +/// SlimTensor cpu_tensor = from_etensor(*gpu_etensor, DEFAULT_CUDA_DEVICE, |
| 47 | +/// CPU_DEVICE); |
| 48 | +/// @endcode |
| 49 | +inline SlimTensor from_etensor( |
| 50 | + const executorch::runtime::etensor::Tensor& etensor, |
| 51 | + const c10::Device& source_device = CPU_DEVICE, |
| 52 | + const c10::Device& target_device = CPU_DEVICE) { |
| 53 | + // Step 1: Extract metadata from ETensor |
| 54 | + const auto ndim = static_cast<size_t>(etensor.dim()); |
| 55 | + |
| 56 | + // Convert sizes from exec_aten::SizesType (int32_t) to int64_t |
| 57 | + std::vector<int64_t> sizes_vec(ndim); |
| 58 | + for (size_t i = 0; i < ndim; ++i) { |
| 59 | + sizes_vec[i] = static_cast<int64_t>(etensor.size(static_cast<ssize_t>(i))); |
| 60 | + } |
| 61 | + |
| 62 | + // Convert strides from exec_aten::StridesType (int32_t) to int64_t |
| 63 | + std::vector<int64_t> strides_vec(ndim); |
| 64 | + auto etensor_strides = etensor.strides(); |
| 65 | + for (size_t i = 0; i < ndim; ++i) { |
| 66 | + strides_vec[i] = static_cast<int64_t>(etensor_strides[i]); |
| 67 | + } |
| 68 | + |
| 69 | + // Map ETensor ScalarType to SlimTensor ScalarType |
| 70 | + c10::ScalarType dtype = static_cast<c10::ScalarType>(etensor.scalar_type()); |
| 71 | + |
| 72 | + // Step 2: Create SlimTensor on target device |
| 73 | + SlimTensor result = empty_strided( |
| 74 | + makeArrayRef(sizes_vec), makeArrayRef(strides_vec), dtype, target_device); |
| 75 | + |
| 76 | + // Step 3: Copy data from ETensor (source device) to SlimTensor (target |
| 77 | + // device) Supports CPU→CPU, CPU→CUDA, CUDA→CPU, or CUDA→CUDA copy |
| 78 | + const void* src_data = etensor.const_data_ptr(); |
| 79 | + void* dst_data = result.data_ptr(); |
| 80 | + size_t nbytes = etensor.nbytes(); |
| 81 | + |
| 82 | + if (nbytes > 0) { |
| 83 | + // const_cast is safe here because copy_ only reads from src_data |
| 84 | + result.storage()->copy_( |
| 85 | + dst_data, const_cast<void*>(src_data), nbytes, source_device); |
| 86 | + } |
| 87 | + |
| 88 | + return result; |
| 89 | +} |
| 90 | + |
| 91 | +/// Creates a SlimTensor from an ETensor pointer. |
| 92 | +/// |
| 93 | +/// Convenience overload that accepts a pointer instead of a reference. |
| 94 | +/// |
| 95 | +/// @param etensor Pointer to the source ETensor (must not be null). |
| 96 | +/// @param source_device The device where the ETensor data resides. |
| 97 | +/// @param target_device The target device for the output SlimTensor. |
| 98 | +/// @return A new SlimTensor with data copied to the target device. |
| 99 | +inline SlimTensor from_etensor( |
| 100 | + const executorch::runtime::etensor::Tensor* etensor, |
| 101 | + const c10::Device& source_device = CPU_DEVICE, |
| 102 | + const c10::Device& target_device = CPU_DEVICE) { |
| 103 | + ET_CHECK_MSG( |
| 104 | + etensor != nullptr, "from_etensor: etensor pointer cannot be nullptr"); |
| 105 | + return from_etensor(*etensor, source_device, target_device); |
| 106 | +} |
| 107 | + |
| 108 | +} // namespace executorch::backends::aoti::slim |
0 commit comments