* Copyright (c) 2026 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
#warning \
"impl/tensor_api/tensor/tensor_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
#endif
* \file tensor_impl.h
* \brief
*/
#ifndef IMPL_TENSOR_API_TENSOR_TENSOR_IMPL_H
#define IMPL_TENSOR_API_TENSOR_TENSOR_IMPL_H
#include "impl/tensor_api/utils/utils_impl.h"
#include "impl/tensor_api/utils/npu_debug_assert.h"
#include "impl/tensor_api/utils/npu_debug_check.h"
#include "impl/tensor_api/tensor/layout_impl.h"
#include "impl/tensor_api/tensor/engine_impl.h"
#include "impl/tensor_api/tensor/pointer_pattern.h"
namespace asc {
namespace te {
template <typename LocationType, typename EngineT, typename LayoutT>
struct make_tensor_result;
template <typename EngineT, typename LayoutT>
__aicore__ inline base_tensor<EngineT, LayoutT>::base_tensor()
{}
template <typename EngineT, typename LayoutT>
__aicore__ inline base_tensor<EngineT, LayoutT>::base_tensor(const EngineT& engine, const LayoutT& layout)
: rep(layout, engine)
{}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::tensor() const
{
return *this;
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::engine() const
{
return Std::get<1>(rep);
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::engine()
{
return Std::get<1>(rep);
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::layout() const
{
return Std::get<0>(rep);
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::data() const
{
return engine().begin();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::data()
{
return engine().begin();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::shape() const
{
return layout().shape();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::stride() const
{
return layout().stride();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr auto base_tensor<EngineT, LayoutT>::size() const
{
return layout().size();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr auto base_tensor<EngineT, LayoutT>::capacity() const
{
return layout().capacity();
}
template <typename EngineT, typename LayoutT>
template <typename Coord>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator[](const Coord& coord)
{
TENSOR_API_DEBUG_CHECK(debug_check_coord, layout(), coord, "tensor operator[]");
return data()[layout()(coord)];
}
template <typename EngineT, typename LayoutT>
template <typename Coord>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator[](const Coord& coord) const
{
TENSOR_API_DEBUG_CHECK(debug_check_coord, layout(), coord, "tensor operator[]");
return data()[layout()(coord)];
}
template <typename EngineT, typename LayoutT>
template <typename Coord>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator()(const Coord& coord)
{
TENSOR_API_DEBUG_CHECK(debug_check_coord, layout(), coord, "tensor operator()");
auto slice_engine = engine() + layout()(coord);
auto coord_layout = make_coord_layout(coord, layout());
return make_sub_tensor(slice_engine, coord_layout);
}
template <typename EngineT, typename LayoutT>
template <typename Coord>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator()(const Coord& coord) const
{
TENSOR_API_DEBUG_CHECK(debug_check_coord, layout(), coord, "tensor operator()");
auto slice_engine = engine() + layout()(coord);
auto coord_layout = make_coord_layout(coord, layout());
return make_sub_tensor(slice_engine, coord_layout);
}
template <typename EngineT, typename LayoutT>
template <typename Coord0, typename Coord1, typename... Coords>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator()(
const Coord0& c0, const Coord1& c1, const Coords&... cs)
{
return operator()(make_coord(c0, c1, cs...));
}
template <typename EngineT, typename LayoutT>
template <typename Coord0, typename Coord1, typename... Coords>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::operator()(
const Coord0& c0, const Coord1& c1, const Coords&... cs) const
{
return operator()(make_coord(c0, c1, cs...));
}
template <typename EngineT, typename LayoutT>
template <typename Coord, typename Info>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::slice(const Coord& coord, const Info& info)
{
TENSOR_API_DEBUG_CHECK(debug_check_slice_args, layout(), coord, info);
auto slice_engine = engine() + layout()(coord);
auto coord_layout = make_slice_layout(coord, layout(), info);
return make_sub_tensor(slice_engine, coord_layout);
}
template <typename EngineT, typename LayoutT>
template <typename Coord, typename Info>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::slice(
const Coord& coord, const Info& info) const
{
TENSOR_API_DEBUG_CHECK(debug_check_slice_args, layout(), coord, info);
auto slice_engine = engine() + layout()(coord);
auto coord_layout = make_slice_layout(coord, layout(), info);
return make_sub_tensor(slice_engine, coord_layout);
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Tensor() const
{
return tensor();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Engine() const
{
return engine();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Engine()
{
return engine();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Layout() const
{
return layout();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Data() const
{
return data();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Data()
{
return data();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Shape() const
{
return shape();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Stride() const
{
return stride();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr auto base_tensor<EngineT, LayoutT>::Size() const
{
return size();
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr auto base_tensor<EngineT, LayoutT>::Capacity() const
{
return capacity();
}
template <typename EngineT, typename LayoutT>
template <typename Coord, typename Info>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Slice(const Coord& coord, const Info& info)
{
return slice(coord, info);
}
template <typename EngineT, typename LayoutT>
template <typename Coord, typename Info>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::Slice(
const Coord& coord, const Info& info) const
{
return slice(coord, info);
}
template <typename EngineT, typename LayoutT>
template <typename SliceEngine, typename SliceLayout>
__aicore__ inline constexpr decltype(auto) base_tensor<EngineT, LayoutT>::make_sub_tensor(
const SliceEngine& slice_engine, const SliceLayout& slice_layout)
{
using location = get_mem_location<SliceEngine>;
using result_tensor = typename make_tensor_result<location, SliceEngine, SliceLayout>::type;
return result_tensor{slice_engine, slice_layout};
}
template <typename EngineT, typename LayoutT>
__aicore__ inline global_tensor<EngineT, LayoutT>::global_tensor() = default;
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr cache_mode global_tensor<EngineT, LayoutT>::get_cache_mode() const
{
return static_cast<cache_mode>(this->engine().get_cache_mode());
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr void global_tensor<EngineT, LayoutT>::set_l2_cache_hint(cache_mode mode)
{
this->engine().set_cache_mode(mode);
}
template <typename EngineT, typename LayoutT>
__aicore__ inline constexpr void global_tensor<EngineT, LayoutT>::SetL2CacheHint(AscendC::Te::CacheMode mode)
{
set_l2_cache_hint(normalize_cache_mode(mode));
}
template <typename LocationType, typename EngineT, typename LayoutT>
struct make_tensor_result {
using type = typename Std::conditional<
Std::is_same_v<LocationType, location::gm>, global_tensor<EngineT, LayoutT>,
local_tensor<EngineT, LayoutT>>::type;
};
template <typename Tensor>
struct is_attr_tensor : Std::false_type {};
template <typename EngineT, typename LayoutT>
struct is_attr_tensor<global_tensor<EngineT, LayoutT>> : Std::true_type {};
template <typename EngineT, typename LayoutT>
struct is_attr_tensor<local_tensor<EngineT, LayoutT>> : Std::true_type {};
template <typename Tensor>
constexpr bool is_attr_tensor_v = is_attr_tensor<Std::remove_cvref_t<Tensor>>::value;
template <typename Iterator>
struct make_tensor_builder {
template <typename MemoryIterator, typename... Args>
__aicore__ inline constexpr auto operator()(const MemoryIterator& iterator, const Args&... args) const
{
using engine = view_engine<MemoryIterator>;
if constexpr (sizeof...(Args) == 1 && (is_layout_v<Args> && ...)) {
using layout = typename Std::tuple_element<0, Std::tuple<Args...>>::type;
using location = get_mem_location<engine>;
using result_tensor = typename make_tensor_result<location, engine, layout>::type;
return result_tensor{engine{iterator}, args...};
} else if constexpr (sizeof...(Args) == 2 && (Std::is_tuple_v<Std::remove_cvref_t<Args>> && ...)) {
using layout = decltype(make_layout(args...));
using location = get_mem_location<engine>;
using result_tensor = typename make_tensor_result<location, engine, layout>::type;
return result_tensor{engine{iterator}, make_layout(args...)};
} else {
static_assert(
sizeof...(Args) != sizeof...(Args),
"make_tensor expects a hardware memory pointer and data structure like Layout or Shape and Stride");
}
}
};
template <typename Iterator, typename... Args>
__aicore__ inline constexpr auto make_tensor(const Iterator& iter, const Args&... args)
{
static_assert(
is_hardware_mem_ptr_v<Iterator>, "make_tensor expects the first argument to be a memory pointer or iterator");
TENSOR_API_DEBUG_CHECK(debug_check_make_tensor_shape, args...);
return make_tensor_builder<Iterator>{}(iter, args...);
}
template <typename Tensor>
__aicore__ inline constexpr auto make_single_batch_sub_tensor(const Tensor& t, uint32_t batch_idx)
{
using layout_type = typename Tensor::layout_type;
static_assert(
layout_type::depth == three_dim_data || layout_type::depth == five_dim_data,
"make_single_batch_sub_tensor only supports batched layouts of depth 3 or 5.");
auto layout = t.layout();
auto batch_stride = get<0>(layout.stride());
auto sub_engine = t.engine() + batch_stride * batch_idx;
auto sub_layout = remove_batch_dim(layout);
using new_engine = Std::remove_cvref_t<decltype(sub_engine)>;
using new_layout = Std::remove_cvref_t<decltype(sub_layout)>;
using location = get_mem_location<new_engine>;
using new_tensor = typename make_tensor_result<location, new_engine, new_layout>::type;
return new_tensor{sub_engine, sub_layout};
}
template <
size_t... squeeze_dims, typename Tensor,
typename = Std::enable_if_t<(is_layout_v<Tensor> || is_attr_tensor_v<Tensor>) && (sizeof...(squeeze_dims) > 0)>>
__aicore__ inline constexpr auto squeeze(const Tensor& x)
{
if constexpr (is_attr_tensor_v<Tensor>) {
auto new_layout = squeeze_layout<squeeze_dims...>(x.layout());
using new_layout_type = Std::remove_cvref_t<decltype(new_layout)>;
using new_engine = typename Tensor::engine_type;
using location = get_mem_location<new_engine>;
using new_tensor = typename make_tensor_result<location, new_engine, new_layout_type>::type;
return new_tensor{x.engine(), new_layout};
} else {
static_assert(is_layout_v<Tensor>, "squeeze input must be a Layout or Tensor");
return squeeze_layout<squeeze_dims...>(x);
}
}
template <
typename Pattern, typename Tensor,
typename = Std::enable_if_t<
(is_layout_v<Tensor> || is_attr_tensor_v<Tensor>) && Std::is_tuple_v<Std::remove_cvref_t<Pattern>>>>
__aicore__ inline constexpr auto squeeze(const Tensor& x, const Pattern& pattern)
{
if constexpr (is_attr_tensor_v<Tensor>) {
auto new_layout = squeeze_layout(x.layout(), pattern);
using new_layout_type = Std::remove_cvref_t<decltype(new_layout)>;
using new_engine = typename Tensor::engine_type;
using location = get_mem_location<new_engine>;
using new_tensor = typename make_tensor_result<location, new_engine, new_layout_type>::type;
return new_tensor{x.engine(), new_layout};
} else {
static_assert(is_layout_v<Tensor>, "squeeze input must be a Layout or Tensor");
return squeeze_layout(x, pattern);
}
}
}
}
#endif
#if defined(UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC)
#undef ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
#undef UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
#endif