tracking_resource_adaptor.hpp

前往此文件的文档。

/*
 * Copyright (c) 2020-2025, NVIDIA CORPORATION.
 *
 * 根据 Apache 许可证 2.0 版（以下简称“许可证”）获得许可；
 * 除非符合许可证的规定，否则您不得使用此文件。
 * 您可以在以下网址获取许可证的副本：
 *
 * https://apache.ac.cn/licenses/LICENSE-2.0
 *
 * 除非适用法律要求或书面同意，根据许可证分发的软件
 * 按“现状”分发，不附带任何明示或暗示的担保或条件。
 * 有关管理权限和
 * 许可证下的限制，请参阅许可证。
 */

#pragma once

#include <rmm/detail/error.hpp>
#include <rmm/detail/export.hpp>
#include <rmm/detail/stack_trace.hpp>
#include <rmm/logger.hpp>
#include <rmm/mr/device/device_memory_resource.hpp>
#include <rmm/mr/device/per_device_resource.hpp>
#include <rmm/resource_ref.hpp>

#include <atomic>
#include <cstddef>
#include <map>
#include <mutex>
#include <shared_mutex>
#include <sstream>

namespace RMM_NAMESPACE {
namespace mr {
template <typename Upstream>
class tracking_resource_adaptor final : public device_memory_resource {
 public
 using read_lock_t =
 std::shared_lock<std::shared_mutex>;
 using write_lock_t =
 std::unique_lock<std::shared_mutex>;
 struct allocation_info {
 std::unique_ptr<rmm::detail::stack_trace> strace;
 std::size_t allocation_size;

 allocation_info() = delete;
 allocation_info(std::size_t size, bool capture_stack)
 : strace{[&]() {
 return capture_stack ? std::make_unique<rmm::detail::stack_trace>() : nullptr;
 }()},
 allocation_size{size} {};
 };

 tracking_resource_adaptor(device_async_resource_ref upstream, bool capture_stacks = false)
 : capture_stacks_{capture_stacks}, allocated_bytes_{0}, upstream_{upstream}
 {
 }

 tracking_resource_adaptor(Upstream* upstream, bool capture_stacks = false)
 : capture_stacks_{capture_stacks},
 allocated_bytes_{0},
 upstream_{to_device_async_resource_ref_checked(upstream)}
 {
 }

 tracking_resource_adaptor() = delete;
 ~tracking_resource_adaptor() override = default;
 tracking_resource_adaptor(tracking_resource_adaptor const&) = delete;
 tracking_resource_adaptor(tracking_resource_adaptor&&) noexcept =
 default;
 tracking_resource_adaptor& operator=(tracking_resource_adaptor const&) = delete;
 tracking_resource_adaptor& operator=(tracking_resource_adaptor&&) noexcept =
 default;

 [[nodiscard]] rmm::device_async_resource_ref get_upstream_resource() const noexcept
 {
 return upstream_;
 }

 std::map<void*, allocation_info> const& get_outstanding_allocations() const noexcept
 {
 return allocations_;
 }

 std::size_t get_allocated_bytes() const noexcept { return allocated_bytes_; }

 std::string get_outstanding_allocations_str() const
 {
 read_lock_t lock(mtx_);

 std::ostringstream oss;

 if (!allocations_.empty()) {
 for (auto const& alloc : allocations_) {
 oss << alloc.first << ": " << alloc.second.allocation_size << " B";
 if (alloc.second.strace != nullptr) {
 oss << " : callstack:" << std::endl << *alloc.second.strace;
 }
 oss << std::endl;
 }
 }

 return oss.str();
 }

 void log_outstanding_allocations() const
 {
#if RMM_LOG_ACTIVE_LEVEL <= RMM_LOG_LEVEL_DEBUG
 RMM_LOG_DEBUG("Outstanding Allocations: %s", get_outstanding_allocations_str());
#endif // RMM_LOG_ACTIVE_LEVEL <= RMM_LOG_LEVEL_DEBUG
 }

 private
 void* do_allocate(std::size_t bytes, cuda_stream_view stream) override
 {
 void* ptr = get_upstream_resource().allocate_async(bytes, stream);
 // 追踪它。
 {
 write_lock_t lock(mtx_);
 allocations_.emplace(ptr, allocation_info{bytes, capture_stacks_});
 }
 allocated_bytes_ += bytes;

 return ptr;
 }

 void do_deallocate(void* ptr, std::size_t bytes, cuda_stream_view stream) override
 {
 get_upstream_resource().deallocate_async(ptr, bytes, stream);
 {
 write_lock_t lock(mtx_);

 const auto found = allocations_.find(ptr);

 // 确保找到了分配，并且字节数匹配
 if (found == allocations_.end()) {
 // 不要抛出异常，而是记录错误。在析构函数（或任何 noexcept）中抛出异常将调用
 // std::terminate
 RMM_LOG_ERROR(
 "Deallocating a pointer that was not tracked. Ptr: %p [%zuB], Current Num. Allocations: "
 "%zu",
 ptr,
 bytes,
 this->allocations_.size());
 } else {
 allocations_.erase(found);

 auto allocated_bytes = found->second.allocation_size;

 if (allocated_bytes != bytes) {
 // 不要抛出异常，而是记录错误。在析构函数（或任何 noexcept）中抛出异常将调用
 // std::terminate
 RMM_LOG_ERROR(
 "Alloc bytes (%zu) and Dealloc bytes (%zu) do not match", allocated_bytes, bytes);

 bytes = allocated_bytes;
 }
 }
 }
 allocated_bytes_ -= bytes;
 }

 bool do_is_equal(device_memory_resource const& other) const noexcept override
 {
 if (this == &other) { return true; }
 auto cast = dynamic_cast<tracking_resource_adaptor<Upstream> const*>(&other);
 if (cast == nullptr) { return false; }
 return get_upstream_resource() == cast->get_upstream_resource();
 }

 bool capture_stacks_; // 是否捕获调用栈
 std::map<void*, allocation_info> allocations_; // 活跃分配的映射
 std::atomic<std::size_t> allocated_bytes_; // 当前分配的字节数
 std::shared_mutex mutable mtx_; // 用于线程安全访问 allocations_ 的互斥锁
 device_async_resource_ref upstream_; // 用于满足
 // 分配请求的上游资源
};
 // 组结束
} // 命名空间 mr
} // 命名空间 RMM_NAMESPACE