gem5
v20.1.0.0
|
The stack distance calculator is a passive object that merely observes the addresses pass to it. More...
#include <stack_dist_calc.hh>
Classes | |
struct | Node |
Node which takes form of Leaf, INode or Root. More... | |
Public Member Functions | |
StackDistCalc (bool verify_stack=false) | |
~StackDistCalc () | |
std::pair< uint64_t, bool > | calcStackDist (const Addr r_address, bool mark=false) |
Process the given address. More... | |
std::pair< uint64_t, bool > | calcStackDistAndUpdate (const Addr r_address, bool addNewNode=true) |
Process the given address: More... | |
Static Public Attributes | |
static constexpr uint64_t | Infinity = std::numeric_limits<uint64_t>::max() |
A convenient way of refering to infinity. More... | |
Private Types | |
typedef std::map< uint64_t, Node * > | IndexNodeMap |
typedef std::map< Addr, uint64_t > | AddressIndexMap |
typedef std::vector< IndexNodeMap > | TreeType |
Private Member Functions | |
uint64_t | getSum (Node *node, bool from_left, uint64_t sum_from_below, uint64_t stack_dist, uint64_t level) const |
Gets sum from the node upwards recursively till the root. More... | |
uint64_t | getSumsLeavesToRoot (Node *node) const |
Gets the sum from the leaf node specified. More... | |
uint64_t | updateSum (Node *node, bool from_left, uint64_t sum_from_below, uint64_t level, uint64_t stack_dist, bool discard_node) |
Updates the nodes upwards recursively till the root. More... | |
uint64_t | updateSumsLeavesToRoot (Node *node, bool is_new_leaf) |
Updates the leaf nodes and nodes above. More... | |
void | updateTree () |
updateTree is a tree balancing operation, which maintains the binary tree structure. More... | |
void | sanityCheckTree (const Node *node, uint64_t level=0) const |
This method is used for verification purposes It recursively traverses upwards from the given node till the root to check if the ultimate parent node (root-node) points to null. More... | |
uint64_t | getIndex () const |
Return the counter for address accesses (unique and non-unique). More... | |
uint64_t | getTreeDepth () const |
Query depth of the tree (tree[0] represents leaf layer while tree[treeDepth] represents the root layer, all layers in between contain intermediate nodes) More... | |
void | printStack (int n=5) const |
Print the last n items on the stack. More... | |
uint64_t | verifyStackDist (const Addr r_address, bool update_stack=false) |
This is an alternative implementation of the stack-distance in a naive way. More... | |
Private Attributes | |
uint64_t | index |
Internal counter for address accesses (unique and non-unique) This counter increments everytime the calcStackDist() method is called. More... | |
TreeType | tree |
AddressIndexMap | aiMap |
std::vector< uint64_t > | nextIndex |
std::vector< uint64_t > | stack |
const bool | verifyStack |
The stack distance calculator is a passive object that merely observes the addresses pass to it.
It calculates stack distances of incoming addresses based on the partial sum hierarchy tree algorithm described by Alamasi et al. http://doi.acm.org/10.1145/773039.773043.
A tree structure is maintained and updated at each transaction (unique or non-unique). The tree is implemented as an STL vector with layers of the form <map> Each layer in this tree is an ordered map <uint64_t, Node*>. Nodes are structs which take form of leaf, intermediate and root nodes. For example, in a tree with 3 layers, tree[0][5] gives a leaf node pointer for key=5 tree[1][1] gives an intermediate node pointer for key=1 tree[2][0] gives the root node in the tree.
At every transaction a hash-map (aiMap) is looked up to check if the address was already encountered before. Based on this lookup a transaction can be termed as unique or non-unique.
In addition to the normal stack distance calculation, a feature to mark an old node in the tree is added. This is useful if it is required to see the reuse pattern. For example, BackInvalidates from a lower level (e.g. membus to L2), can be marked (isMarked flag of Node set to True). Then later if this same address is accessed (by L1), the value of the isMarked flag would be True. This would give some insight on how the BackInvalidates policy of the lower level affect the read/write accesses in an application.
There are two functions provided to interface with the calculator:
At every non-unique transaction the tree is traversed from the leaf at the returned index to the root, the old node is deleted from the tree, and the sums (to the right are collected) and decremented. The collected sum represets the stack distance of the found node. If this node was marked then a bool flag set to True is returned with the stack_distance. During this operation a node is discarded at the leaf layer always. Moreover during the traversal upwards using the updateSum() method, if an intermediate node is found with no children connected to it, then that is discarded too.
The return value of this function is a pair representing the stack_distance and the value of the marked flag.
At every unique transaction the stack-distance is returned as a constant representing INFINITY.
At every non-unique transaction the tree is traversed from the leaf at the returned index to the root, and the sums (to the right) are collected. The collected sum represets the stack distance of the found node.
This function does NOT Modify the stack. (No node is added or deleted). It is just used to mark a node already created and get its stack distance.
The return value of this function is a pair representing the stack distance and the value of the marked flag.
The table below depicts the usage of the Algorithm using the functions: pair<uint64_t Stack_dist, bool isMarked> calcStackDistAndUpdate (Addr r_address, bool addNewNode) pair<uint64_t Stack_dist, bool isMarked> calcStackDist (Addr r_address, bool mark)
| Function | Arguments |Return Val |Use For| |calcStackDistAndUpdate|r_address, True|I/SD,False |A,GD,GM| |calcStackDistAndUpdate|r_address,False|SD,prevMark|D,GD,GM| |calcStackDist |r_address,False|SD,prevMark| GD,GM| |calcStackDist |r_address, True|SD,prevMark| GD,GM|
(*A: Allocate an address in stack, if old entry present then it is deleted, *U: Delete old-address from stack, no new entry is added *GD: Get-Stack distance of an address, *GM: Get value of Mark flag, indicates if that address has been touched before, *I: stack-distance = infinity, *SD: Stack Distance *r_address: address to be added, *prevMark: value of isMarked flag of the Node)
Invalidates refer to a type of packet that removes something from a cache, either autonoumously (due-to cache's own replacement policy), or snoops from other caches which invalidate something inside our cache.
Usage | Function to use |Typical Use | Add new entry |calcStackDistAndUpdate|Read/Write Allocate | Delete Old Entry |calcStackDistAndUpdate|Writebacks/Cleanevicts| Dist.of Old entry|calcStackDist |Cleanevicts/Invalidate|
Node Balancing: The tree structure is maintained by an updateTree() operation called when an intermediate node is required. The update operation is roughly categorized as a root update or intermediate layer update. When number of leaf nodes grow over a power of 2 then a new layer is added at the top of the tree and a new root node is initialized. The old node at the lower layer is connected to this. In an intermediate node update operation a new intermediate node is added to the required layer.
Debugging: Debugging can be enabled by setting the verifyStack flag true. Debugging is implemented using a dummy stack that behaves in a naive way, using STL vectors (i.e each unique address is pushed on the top of an STL vector stack, and SD is returned as Infinity. If a non unique address is encountered then the previous entry in the STL vector is removed, all the entities above it are pushed down, and the address is pushed at the top of the stack).
A printStack(int numOfEntitiesToPrint) is provided to print top n entities in both (tree and STL based dummy stack).
Definition at line 171 of file stack_dist_calc.hh.
|
private |
Definition at line 179 of file stack_dist_calc.hh.
|
private |
Definition at line 176 of file stack_dist_calc.hh.
|
private |
Definition at line 180 of file stack_dist_calc.hh.
StackDistCalc::StackDistCalc | ( | bool | verify_stack = false | ) |
Definition at line 45 of file stack_dist_calc.cc.
References nextIndex, StackDistCalc::Node::nodeIndex, and tree.
StackDistCalc::~StackDistCalc | ( | ) |
std::pair< uint64_t, bool > StackDistCalc::calcStackDist | ( | const Addr | r_address, |
bool | mark = false |
||
) |
Process the given address.
If Mark is true then set the mark flag of the leaf node. This function returns the stack distance of the incoming address and the previous status of the mark flag.
r_address | The current address to process |
mark | set the mark flag for the address. |
Definition at line 459 of file stack_dist_calc.cc.
References aiMap, getSumsLeavesToRoot(), Infinity, panic_if, printStack(), tree, verifyStack, and verifyStackDist().
std::pair< uint64_t, bool > StackDistCalc::calcStackDistAndUpdate | ( | const Addr | r_address, |
bool | addNewNode = true |
||
) |
Process the given address:
r_address | The current address to process |
addNewNode | If true, a new node is added to the tree |
Definition at line 357 of file stack_dist_calc.cc.
References aiMap, index, Infinity, StackDistCalc::Node::isLeftNode, StackDistCalc::Node::isMarked, nextIndex, StackDistCalc::Node::nodeIndex, panic_if, StackDistCalc::Node::parent, printStack(), sanityCheckTree(), tree, updateSumsLeavesToRoot(), updateTree(), verifyStack, and verifyStackDist().
Referenced by StackDistProbe::handleRequest().
|
inlineprivate |
Return the counter for address accesses (unique and non-unique).
This is further used to dump stats at regular intervals.
Definition at line 270 of file stack_dist_calc.hh.
References index.
|
private |
Gets sum from the node upwards recursively till the root.
This function is called first by getSumsLeavesToRoot, and then recursively calls itself.
node | pointer to the node which is updated |
from_left | variable which says that the request arrived from the left |
sum_from_below | Sum of left and right children below |
level | level in the tree the calling node is located |
stack_dist | stack distance of the node below |
Definition at line 211 of file stack_dist_calc.cc.
References StackDistCalc::Node::isLeftNode, X86ISA::level, StackDistCalc::Node::parent, StackDistCalc::Node::sumLeft, and StackDistCalc::Node::sumRight.
Referenced by getSumsLeavesToRoot().
|
private |
Gets the sum from the leaf node specified.
This function is called by calcStackDist.
node | pointer to the node which is updated |
Definition at line 234 of file stack_dist_calc.cc.
References getSum(), StackDistCalc::Node::isLeftNode, and StackDistCalc::Node::parent.
Referenced by calcStackDist().
|
inlineprivate |
Query depth of the tree (tree[0] represents leaf layer while tree[treeDepth] represents the root layer, all layers in between contain intermediate nodes)
Definition at line 279 of file stack_dist_calc.hh.
References tree.
Referenced by printStack(), sanityCheckTree(), and updateTree().
|
private |
Print the last n items on the stack.
This method prints top n entries in the tree based implementation as well as dummy stack.
n | Number of entries to print |
Definition at line 563 of file stack_dist_calc.cc.
References ArmISA::a, aiMap, X86ISA::count, DPRINTF, getTreeDepth(), ArmISA::n, StackDistCalc::Node::nodeIndex, stack, tree, and verifyStack.
Referenced by calcStackDist(), and calcStackDistAndUpdate().
|
private |
This method is used for verification purposes It recursively traverses upwards from the given node till the root to check if the ultimate parent node (root-node) points to null.
node | pointer to the node whose sanity is being checked |
level | the level at which this node is located in the tree |
Definition at line 512 of file stack_dist_calc.cc.
References getTreeDepth(), ArmISA::i, X86ISA::level, StackDistCalc::Node::nodeIndex, panic_if, and StackDistCalc::Node::parent.
Referenced by calcStackDistAndUpdate().
|
private |
Updates the nodes upwards recursively till the root.
This function is first called by updateSumsLeavesToRoot, and then it recursively calls itself.
node | pointer to the node which is updated |
from_left | variable which says that the request arrived from the left |
sum_from_below | Sum of left and right children below |
level | level in the tree the calling node is located |
stack_dist | stack distance of the node below |
discard_node | whether the calling node was discarded or not |
Definition at line 97 of file stack_dist_calc.cc.
References StackDistCalc::Node::discardLeft, StackDistCalc::Node::discardRight, StackDistCalc::Node::isLeftNode, X86ISA::level, StackDistCalc::Node::nodeIndex, panic_if, StackDistCalc::Node::parent, StackDistCalc::Node::sumLeft, StackDistCalc::Node::sumRight, tree, and verifyStack.
Referenced by updateSumsLeavesToRoot().
|
private |
Updates the leaf nodes and nodes above.
This function is called by the calcStackDistAndUpdate.
node | pointer to the node which is updated |
is_new_leaf | is true if this is a newly added node |
Definition at line 186 of file stack_dist_calc.cc.
References Infinity, StackDistCalc::Node::isLeftNode, X86ISA::level, StackDistCalc::Node::parent, StackDistCalc::Node::sumLeft, and updateSum().
Referenced by calcStackDistAndUpdate().
|
private |
updateTree is a tree balancing operation, which maintains the binary tree structure.
This method is called whenever index%2 == 0 (i.e. every alternate cycle) The two main operation are : OP1. Moving the root node one layer up if index counter crosses power of 2 OP2. Addition of intermediate nodes as and when required and linking them to their parents in the layer above.
Definition at line 248 of file stack_dist_calc.cc.
References StackDistCalc::Node::discardLeft, getTreeDepth(), ArmISA::i, index, StackDistCalc::Node::isLeftNode, isPowerOf2(), nextIndex, StackDistCalc::Node::nodeIndex, StackDistCalc::Node::parent, StackDistCalc::Node::sumLeft, and tree.
Referenced by calcStackDistAndUpdate().
|
private |
This is an alternative implementation of the stack-distance in a naive way.
It uses simple STL vector to represent the stack. It can be used in parallel for debugging purposes. It is 10x slower than the tree based implemenation.
r_address | The current address to process |
update_stack | Flag to indicate if stack should be updated |
Definition at line 532 of file stack_dist_calc.cc.
References ArmISA::a, Infinity, and stack.
Referenced by calcStackDist(), and calcStackDistAndUpdate().
|
private |
Definition at line 400 of file stack_dist_calc.hh.
Referenced by calcStackDist(), calcStackDistAndUpdate(), printStack(), and ~StackDistCalc().
|
private |
Internal counter for address accesses (unique and non-unique) This counter increments everytime the calcStackDist() method is called.
This counter is used as a key for the hash- map at the leaf layer. Practically at every call to the calculator this counter is incremented and a new leaf node is added in the tree at the leaf layer using this counter value as the key.
Definition at line 394 of file stack_dist_calc.hh.
Referenced by calcStackDistAndUpdate(), getIndex(), and updateTree().
|
staticconstexpr |
A convenient way of refering to infinity.
Definition at line 312 of file stack_dist_calc.hh.
Referenced by calcStackDist(), calcStackDistAndUpdate(), StackDistProbe::handleRequest(), updateSumsLeavesToRoot(), and verifyStackDist().
|
private |
Definition at line 404 of file stack_dist_calc.hh.
Referenced by calcStackDistAndUpdate(), StackDistCalc(), updateTree(), and ~StackDistCalc().
|
private |
Definition at line 407 of file stack_dist_calc.hh.
Referenced by printStack(), verifyStackDist(), and ~StackDistCalc().
|
private |
Definition at line 397 of file stack_dist_calc.hh.
Referenced by calcStackDist(), calcStackDistAndUpdate(), getTreeDepth(), printStack(), StackDistCalc(), updateSum(), updateTree(), and ~StackDistCalc().
|
private |
Definition at line 410 of file stack_dist_calc.hh.
Referenced by calcStackDist(), calcStackDistAndUpdate(), printStack(), and updateSum().