Patent Grant
10657306
Computer Aided Design (CAD) and Electronic Design Automation (EDA) problems are conventionally solved with sophisticated but general purpose software tools. Deep learning techniques offer the potential to go beyond these generic tools and train software on company-specific data to learn from company practices and internal knowledge that is not publicly available. “Deep learning” refers to an artificial intelligence function that imitates the workings of biological neural networks in processing data and creating patterns for use in decision making. Deep learning is a subset of machine learning utilizing artificial neural networks capable of learning based on experience, usually, initially, from a training data set. Deep learning is often implemented using deep neural networks. “Deep neural network” refers to an artificial neural network with one or more hidden layers. In addition, after a deep neural network (DNN) has been trained, it may provide faster results than traditional software techniques.
A challenge in applying deep learning (DL) techniques to CAD and EDA problems is mapping the problems into a DL solution space. Representation of netlists is particularly difficult because netlists are a class of directed acyclic graph (DAG) in which each node of the graph is assigned with specific logic block characteristics such as unique combinatorial logic. “Combinatorial logic” refers to logic to convert one or more input signals into output signals. Non-limiting examples of combinatorial logic includes Boolean logic blocks such as AND gates, OR gates, NOR gates etc.; flops, latches, and inverters; and many other types of circuits of varying complexity. “Logic block” refers to combinatorial logic with defined inputs and outputs. “Logic block characteristics” refers to characteristics of a logic block such as power consumption, propagation delay, inputs, outputs, size, and many other possibilities depending on the nature of the logic block and it's use in a circuit. The logic block characteristics may vary according to the class of the logic block: RDL (resistor-diode logic), RTL (resistor-transistor logic), DTL (diode-transistor logic), TTL (transistor-transistor logic), and CMOS, for example. Sub-variants of broad logic block classes may also have logic block characteristics particular to their sub-variant type, e.g. basic CMOS logic vs. advanced CMOS types with optimizations for improving the speed or reducing the power consumption of basic CMOS-class logic blocks. “Directed acyclic graph” refers to a finite directed graph with no directed cycles. That is, it consists of finitely many nodes and edges with each edge directed from one node to another, such that there is no way to start at any node v and follow a consistently-directed sequence of edges that eventually loops back to v again. Equivalently, a directed acyclic graph is a directed graph that has a topological ordering, a sequence of the vertices such that every edge is directed from earlier to later in the sequence. “Netlist” refers to a description of the connectivity of an electronic circuit. In its simplest form, a netlist comprises of a list of the electronic components in a circuit and the circuit nodes (a junction point between two components) that each is connected to. The netlist data must be preprocessed before it can be ingested by a DNN, and the preprocessing step requires human expert considerations and design trade-offs. Some data and structure of the circuit is typically preserved for processing by the DNN; other data and structure may be lost.
To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.
Embodiments of a pre-processor for netlists to be analyzed by a deep neural network is disclosed. “Pre-processor” refers to logic executed to generate the inputs to subsequent stages of processing, such as classification. At a high level, the pre-processor (1) performs levelization on the netlist by identifying each node level as a distance from primary inputs to primary outputs, (2) computes controllability-to-0, controllability-to-1, and observability for each node, and (3) formulates the netlist as a collection of fan-in nodes and fan-out nodes. “Fan-out node” refers to a node receiving the output of a node in a netlist or circuit. “Fan-in node” refers to a node supplying inputs to a node in a netlist or circuit. “Controllability-to-1” refers to whether the logic value at a node in a circuit can be set to logical ‘1’. “Controllability-to-0” refers to whether the logic value at a node in a circuit can be set to logical ‘0’. “Primary output” refers to outputs from a last node level of a netlist or circuit. Primary outputs are not applied to another node in the netlist or circuit. “Primary input” refers to inputs to the first node level of a netlist or circuit. Primary inputs are not generated from nodes inside the circuit or netlist, but rather are applied from outside the circuit or netlist. “Node level” refers to the level of a node within a levelized netlist. The node level is determined by how many other nodes are connected between the node's input and the primary input.
The pre-processor and other techniques disclosed herein may be applied to generate a modified graph/netlist for a circuit that can then be applied to manufacture the circuit with improved testability, and hence improved reliability.
Controllability is the ability to propagate ‘0’ (controllability-to-0) and ‘1’ (controllability-to-1) to the test node from either primary inputs or scan input ports. This is useful to test the node's behavior to a ‘1’ or ‘0’ input. Observability is the ability to measure the state of a logic value at the test node by shifting it to the scan out ports, which is again useful for testing the node's behavior to inputs.
Netlist connectivity is encoded using the levelization of the circuit design. Combinatorial logic gate characteristics are applied to a SCOAP algorithm (the SCOAP) for individual combinatorial logic elements of the netlist. “SCOAP” refers to the Sandia Controllability Observability Analysis algorithm for computing testability metrics at a gate or other logic block, and which typically has linear complexity. The SCOAP algorithm (see FIG. 2) inherently includes the logic block characteristics when computing the controllability-to-0, controllability-to-1, and observability of combinatorial logic elements.
The netlist is levelized from primary inputs to primary outputs with the primary inputs set to 0. Then each output node is assigned a level using breadth-first-search (BFS) followed by depth-first-search (DFS). “Depth-first-search” refers to an algorithm for traversing or searching tree or graph data structures. The algorithm starts at the root node (selecting some arbitrary node as the root node in the case of a graph) and explores as far as possible along each branch before backtracking. “Breadth-first-search” refers to an algorithm for traversing or searching tree or graph data structures. It starts at the tree root (or some arbitrary node of a graph, sometimes referred to as a ‘search key’), and explores all of the neighbor nodes at the present depth prior to moving on to the nodes at the next depth level. “Output node” refers to a node in a circuit or netlist providing the primary outputs of the circuit or netlist. See the levelized netlist 100 embodiment of
Inputs to the deep neural network are (1) a collection of fan-in nodes and fan-out nodes with one or more of controllability-to-0, controllability-to-1, or observability, or (2) a randomly selected non-controllable/non-observable node (a so-called difficult to test node or DTN). “Difficult to test node” refers to a node in a circuit for which either observability or controllability are difficult or impossible. Such a node is resistant to testing by conventional means.
The following description may use certain acronyms and abbreviations as follows:
Embodiments of a pre-processor will now be described for applying a graph convolutional network (GCN) to process a graph representing a netlist. “Graph convolutional network” refers to type of neural network designed to work directly on graphs and leverage their structural information. Formally, a graph convolutional network (GCN) is a neural network that given a graph G=(V, E), takes as input
At each layer, node features are aggregated to form the next layer's features using a propagation rule. In this way, features become increasingly more abstract at each consecutive layer. “Node” refers to a representation of combinatorial logic in a graph, or a location in a circuit, for example between logic blocks in the circuit. Whether or not the term refers to a graph element or a point in a circuit will be evident from the context in which the term appears.
The netlist is a levelized netlist 100 as depicted in
The predictions about whether nodes are difficult to test nodes or not is applied to determine whether or not to insert a test point at the nodes in the analyzed circuit. Test points improve the observability and controllability at the node, and hence they improve test coverage, but the also take up space in the circuit and consume power. Test point insertion (TPI) is a broadly used approach in design for testability (DFT) to modify circuits to improve their testability.
Vector spaces are more amenable to analysis by deep neural networks than are graphs. Vector operations are often simpler and faster to process than the equivalent graph operations, especially in deep learning applications. Node embedding is the process of representing graphs and other network structures in vector space, while preserving their properties, including connectivity. The embeddings are input as features to a model, and parameters of the graph or nodes are learned based on training data. This obviates the need for complex classification algorithms applied directly on the graph. More specifically, the node embeddings are derived from the local neighborhood of the current node 406. “Local neighborhood” refers to the nodes of a graph within D hops of a selected node, where D is the local neighborhood depth.
The deep neural network then predicts whether nodes in the graph are difficult to test nodes or not, and test point insertion logic uses the predictions to insert test nodes.
Each node embedding is represented as a vector. The ratio of the DTNs to non-DTN nodes in the netlist-derived graph 400 is typically low, e.g., 3:97, therefore selecting randomly from the set of non-DTNs may create a 1:1 test set. In one embodiment the inputs to the deep neural network are a set of vectors for each current node 406, as follows:
Each node includes at least four attributes to convey the node level and SCOAP values. In some embodiments more logic block characteristics may be included for each node to improve training and inference. The particular logic block characteristics stored with each node may be implementation specific, and depending on the EDA platform, the logic block characteristics for each node may be different.
Neural network classifiers predict whether or not an input has the characteristics matching a particular classification. The output of such classifiers may be generated by an output layer implementing a SoftMax function. “SoftMax function” refers to a function of the form f(xi)=exp(xi)/sum(exp(x)) where the sum is taken over a set of x. SoftMax is used at different layers (often at the output layer) of artificial neural networks to predict classifications for inputs to those layers. The SoftMax function calculates the probabilities distribution of the event xi over ‘n’ different events. In general sense, this function calculates the probabilities of each target class over all possible target classes. The calculated probabilities are helpful for predicting that the target class is represented in the inputs. The main advantage of using SoftMax is the output probabilities range. The range will 0 to 1, and the sum of all the probabilities will be equal to one. If the SoftMax function used for multi-classification model it returns the probabilities of each class and the target class will have the high probability. The formula computes the exponential (e-power) of the given input value and the sum of exponential values of all the values in the inputs. Then the ratio of the exponential of the input value and the sum of exponential values is the output of the SoftMax function. Multi-class neural network classifiers output more than one classification prediction for an input. In one embodiment, the deep neural network outputs a set of two or three one-hot vector predictions: “One-hot vector” refers to a vector in which exactly one value is set and the other values are not set. An example of a one-hot vector is [0, 0, 1, 0].
In one embodiment the deep neural network comprises one input layer, three hidden layers, and one output layer. Each layer uses a linear artificial neuron input function F(X)=W×X+B where W is the weight of the input connection and B is the bias on the input.
Each hidden layer may utilize a normal distribution with initialization of random weights and zero bias along with zero mean and 0.1 standard deviation. The deep neural network may also utilize a Rectified Linear Unit (ReLU) activation function, or a hyperbolic tangent or other sigmoid type activation. “ReLU” refers to a rectifier function, an activation function defined as the positive part of its input. It is also known as a ramp function and is analogous to half-wave rectification in electrical signal theory. ReLu is a popular activation function in deep neural networks. The output layer is similar to a hidden layer followed by a SoftMax function layer utilizing cross entropy as a loss function. “Loss function” refers to also referred to as the cost function or error function (not to be confused with the Gauss error function), is a function that maps values of one or more variables onto a real number intuitively representing some “cost” associated with those values. An Adam Optimizer backpropagation algorithm may also be used to adjust network parameters such as the weights W. “Backpropagation” refers to an algorithm used in artificial neural networks to calculate a gradient that is needed in the calculation of the weights to be used in the network. It is commonly used to train deep neural networks, a term referring to neural networks with more than one hidden layer. For backpropagation, the loss function calculates the difference between the network output and its expected output, after a case propagates through the network. “Adam Optimizer” refers to an optimization algorithm that can used instead of the classical stochastic gradient descent procedure to update network weights iterative based in training data. Stochastic gradient descent maintains a single learning rate (termed alpha) for all weight updates and the learning rate does not change during training. A learning rate is maintained for each network weight (parameter) and separately adapted as learning unfolds. Adam combines advantages of two other extensions of stochastic gradient descent: Adaptive Gradient Algorithm (AdaGrad) that maintains a per-parameter learning rate that improves performance on problems with sparse gradients (e.g. natural language and computer vision problems), and Root Mean Square Propagation (RMSProp) that also maintains per-parameter learning rates that are adapted based on the average of recent magnitudes of the gradients for the weight (e.g. how quickly it is changing). This means the algorithm does well on online and non-stationary problems (e.g. noisy). Adam realizes the benefits of both AdaGrad and RMSProp. Instead of adapting the parameter learning rates based on the average first moment (the mean) as in RMSProp, Adam also makes use of the average of the second moments of the gradients (the uncentered variance). Specifically, the algorithm calculates an exponential moving average of the gradient and the squared gradient, and the parameters beta1 and beta2 control the decay rates of these moving averages. The initial value of the moving averages and beta1 and beta2 values close to 1.0 (recommended) result in a bias of moment estimates towards zero. This bias is overcome by first calculating the biased estimates before then calculating bias-corrected estimates.
In one embodiment the fan-in and fan out of each node is limited, e.g., to 500, resulting in an input vector of at most 500×4+500×4+4=4004 inputs, and this input is reduced to the two output classification predictions (DTN or Non-DTN) by the output layer.
In one embodiment training data for the deep neural network is organized into a Pickle Dump after shuffling and converting the data into batches (e.g., 25 batches) for training. “Pickle Dump” refers to an algorithm used in Python to convert structured data objects into serialized byte streams. Different batches are used for validation and for testing. Each of the epochs is trained through all the training batches. “Epoch” refers to a number of times the neural network is exposed to an entire data set. Each time the neural network processes all the batches in a data set, such as a training, an epoch has completed. An iteration describes the number of times a batch of data passed through the neural network. That means the batch completes a forward pass and backward pass (e.g., backpropagation occurs). Thus every time a batch of data is processed through the neural network, an iteration is completed.
For example consider a dataset of 10 samples with a batch size of 2. The neural network may then be executed for 3 epochs. Therefore, in each epoch, 5 batches are processed (10/2=5). Each batch gets processed through the algorithm, therefore there are 5 iterations per epoch. Because there are 3 epochs, there are a total of 15 iterations (5*3=15) for training. The validation batch is then applied to validate the model represented by the trained deep neural network.
Test point selection and insertion in a graph may be performed iteratively. In each iteration, the label of each node in the graph is predicted using a trained GCN model (e.g., graph convolutional network 1000), followed by selection and insertion of test points, if any. The process completes when some conditions are satisfied such as when the total number of inserted test points is achieved, or when the number of positive (DTN) predictions is less than a desired threshold.
To classify a node in the graph, the neural network generates its node embeddings based on the node's own attributes and structural information of the graph from the node's local neighborhood. These embeddings are applied to a classifier that inputs the node embeddings and predicts a label for the node. To achieve this, three kinds of layers are utilized in the graph convolutional network: aggregators, encoder layers, and classifier layers. Aggregators and encoders are used to generate node embeddings by exploiting node attributes and the neighborhood information. The classifier predicts the label for each node in the graph based on its embedding.
The node embedding is generated using aggregation and encoding layers of the deep neural network. Each of them performs a specific operation on the node. An aggregator layer gathers the feature information from the node's neighbors using an aggregation function Agg(⋅). This is similar to how a filter works in convolutional neural networks that operate on images, for example. An encoder layer propagates information between different layers using a weight matrix. The embedding computation process, i.e., aggregation and encoding, is performed iteratively. Fully-connected layers are used for the classifier, which inputs the node embedding and outputs a predicted label for the node.
Given a graph G(V, E) and node attributes {x(v):∀v∈V}, the node embeddings {e(v):∀v∈V} may be generated per the algorithm in
where wpr and wsu are weights for predecessors and successors, respectively, and remain unchanged within each iteration of the outer loop. Next, a non-linear transformation is performed to encode the aggregated representation using a weight matrix Wd∈RKd-1×Kd and an activation function σ(⋅) (line 5). Kd is the dimension of the embedding after the d-th iteration where K0 is 4, the initial attribute dimension. A concrete example is the node embedding computation 900 depicted in
Making the derived classification model scalable to large graphs is important especially because fast inference is desired. The algorithm 700 depicted in
Here A∈R6×6, Gd∈R6×Kd-1, and the v-th row is the representation for node v after aggregation in the d-th iteration. The entire inner loop in the algorithm 700 (line 3-line 6) may be simply formulated as
Ed=σ(Gd·Wd)=σ((A·Ed-1)·Wd). Equation 3
In this form, the algorithm is constructed as a series of matrix multiplications which are efficiently computed without duplicated computation. One issue with this modified approach is that the radix of adjacency matrix A is N×N, which is extremely large in many cases and cannot be stored all at once in a typical processor memory. However the fact that the matrix A is a sparse matrix may then be exploited. The matrix A can be represented in a compressed sparse coordinate format (COO), as a list of (value, row_index, column_index) tuples. A matrix compressed in this fashion can be stored in a typical processor memory all at once to enable the matrix multiplication. For instance, in the COO representation of A in Equation 3 each column is a tuple including the value and indices of a non-zero element in the matrix, as follows:
value: [1, w1, w1, w1, w2, 1, w1, w2, 1, w2, w2, 1, w2, 1, w1, 1]
r_index: [1, 2, 3, 4, 1, 2, 5, 1, 3, 6, 1, 4, 2, 5, 3, 6],
c_index: [1, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 5, 5, 6, 6]
An additional benefit is that the COO format is efficient for incremental matrix construction which facilitates graph modifications.
Each aggregator and each encoder may be implemented as a layer of the graph convolutional network 1000. Each performs a specific operation on a node. An aggregator gathers the feature information from the node's neighbors using an aggregation algorithm. An encoder propagates values between layers of the graph convolutional network 1000 using a weight matrix. The embedding computation process, i.e., aggregation and encoding, is performed iteratively. In
In graph 1106, the arrowed lines indicate additional flow and aggregation of features toward the node 1108. In the second aggregation step, the node 1108 is black indicating the continued aggregation of information from the local neighborhood nodes 1110. Compared with the first aggregation step in graph 1104, graph 1106 aggregates all the features from the 2-hop local neighborhood of the node 1108.
After each aggregation step is completed, an encoder is utilized to project the feature vector of each node from m-dimensional space to a higher n-dimensional space by multiplying by a weight matrix Wm×n. This may be done when the original features are in a low-dimensional space, or following each aggregation step.
Referencing
Inserting observation points modifies the netlist, thus the graph should be modified, including the graph structure and node attributes. Inserting one observation point to a target node v results in (1) adding a new node p to the graph, and (2) adding an edge from the target node v to new node p. The adjacency matrix A and initial embedding matrix E0 must therefore be updated. One challenge in this iterative flow is how to update the graph efficiently. The matrix A may be incrementally updated by adding a column and a row, and setting corresponding entries as wpr or wsu, which can be done efficiently when using the COO format by appending three tuples (wpr,p,v), (wsu,v,p) and (1,p,p). E0 is updated by appending the attributes of new node p, which are set to [0,1,1,0]. Then only the attributes of the nodes in the fan-in cone of the new node p need be updated using the SCOAP algorithm. Because the graph convolutional network model is inductive, the updated A and E0 may be directly applied to the classifier for prediction.
DTNs typically only comprise a small fraction of the input graph, a condition referred to as “data imbalance”. To deal with data imbalance data pruning may be performed to reduce the number of negative (non-DTN) samples. Some embodiments may apply three pruning conditions to limit the graph to a set of likely non-DTN nodes: the flops in the netlist, the cells driving flops, and the cells with the zero observability.
Referencing
A multi-step classification may be applied utilizing a series of cascaded neural network stages. At each stage, negative samples are filtered out without eliminating many positive samples. Each network stage is trained separately. In the multi-stage classification 1300 depiction in
A multi-stage graph convolutional network may be utilized to address the undesirable performance effects associated with such data imbalance. In each stage, one graph convolutional network is trained while filtering out some of the negative classified nodes 1302 with high confidence. The filtered set is passed to the next stage. This is achieved by imposing a large weight on the positive classified nodes 1304 such that the backpropagated penalty for misclassifying them is large. In this way, most positive classified nodes 1304 remain on the right side of the decision boundary 1306 until negative classified nodes 1302 are substantially reduced. After a few stages, the data set becomes relatively balanced and a network can make the final predictions.
A conventional data-parallelism scheme cannot directly be applied for training the deep neural network with multiple graphs because the inputs of graph and sparse adjacency matrix cannot be split into equal chunks as is done in most data-parallelism schemes. Instead of splitting the input tensors into equal chunks, the inputs are divided among multiple GPUs. The graphs may have different sizes. The feedforward computation for each graph is completed on a corresponding GPU in parallel, and the overall output is aggregated and applied to one GPU for back-propagation to train the deep neural network. An example of this is illustrated in the parallel GPU processing 1500 of
The algorithms and techniques disclosed herein may be executed by computing devices utilizing one or more graphic processing unit and/or general purpose data processor (e.g., a ‘central processing unit or CPU). Exemplary machine architectures will now be described that may be used to carry out the techniques disclosed herein.
Parallel Processing Unit
One or more parallel processing unit 1600 modules may be configured to accelerate thousands of High Performance Computing (HPC), data center, and machine learning applications. The parallel processing unit 1600 may be configured to accelerate numerous deep learning systems and applications including autonomous vehicle platforms, deep learning, high-accuracy speech, image, and text recognition systems, intelligent video analytics, molecular simulations, drug discovery, disease diagnosis, weather forecasting, big data analytics, astronomy, molecular dynamics simulation, financial modeling, robotics, factory automation, real-time language translation, online search optimizations, and personalized user recommendations, and the like.
As shown in
The NVLink 1608 interconnect enables systems to scale and include one or more parallel processing unit 1600 modules combined with one or more CPUs, supports cache coherence between the parallel processing unit 1600 modules and CPUs, and CPU mastering. Data and/or commands may be transmitted by the NVLink 1608 through the hub 1616 to/from other units of the parallel processing unit 1600 such as one or more copy engines, a video encoder, a video decoder, a power management unit, etc. (not explicitly shown). The NVLink 1608 is described in more detail in conjunction with
The I/O unit 1606 is configured to transmit and receive communications (e.g., commands, data, etc.) from a host processor (not shown) over the interconnect 1602. The I/O unit 1606 may communicate with the host processor directly via the interconnect 1602 or through one or more intermediate devices such as a memory bridge. In an embodiment, the I/O unit 1606 may communicate with one or more other processors, such as one or more parallel processing unit 1600 modules via the interconnect 1602. In an embodiment, the I/O unit 1606 implements a Peripheral Component Interconnect Express (PCIe) interface for communications over a PCIe bus and the interconnect 1602 is a PCIe bus. In alternative embodiments, the I/O unit 1606 may implement other types of well-known interfaces for communicating with external devices.
The I/O unit 1606 decodes packets received via the interconnect 1602. In an embodiment, the packets represent commands configured to cause the parallel processing unit 1600 to perform various operations. The I/O unit 1606 transmits the decoded commands to various other units of the parallel processing unit 1600 as the commands may specify. For example, some commands may be transmitted to the front end unit 1610. Other commands may be transmitted to the hub 1616 or other units of the parallel processing unit 1600 such as one or more copy engines, a video encoder, a video decoder, a power management unit, etc. (not explicitly shown). In other words, the I/O unit 1606 is configured to route communications between and among the various logical units of the parallel processing unit 1600.
In an embodiment, a program executed by the host processor encodes a command stream in a buffer that provides workloads to the parallel processing unit 1600 for processing. A workload may comprise several instructions and data to be processed by those instructions. The buffer is a region in a memory that is accessible (e.g., read/write) by both the host processor and the parallel processing unit 1600. For example, the I/O unit 1606 may be configured to access the buffer in a system memory connected to the interconnect 1602 via memory requests transmitted over the interconnect 1602. In an embodiment, the host processor writes the command stream to the buffer and then transmits a pointer to the start of the command stream to the parallel processing unit 1600. The front end unit 1610 receives pointers to one or more command streams. The front end unit 1610 manages the one or more streams, reading commands from the streams and forwarding commands to the various units of the parallel processing unit 1600.
The front end unit 1610 is coupled to a scheduler unit 1612 that configures the various GPC 1700 modules to process tasks defined by the one or more streams. The scheduler unit 1612 is configured to track state information related to the various tasks managed by the scheduler unit 1612. The state may indicate which GPC 1700 a task is assigned to, whether the task is active or inactive, a priority level associated with the task, and so forth. The scheduler unit 1612 manages the execution of a plurality of tasks on the one or more GPC 1700 modules.
The scheduler unit 1612 is coupled to a work distribution unit 1614 that is configured to dispatch tasks for execution on the GPC 1700 modules. The work distribution unit 1614 may track a number of scheduled tasks received from the scheduler unit 1612. In an embodiment, the work distribution unit 1614 manages a pending task pool and an active task pool for each of the GPC 1700 modules. The pending task pool may comprise a number of slots (e.g., 32 slots) that contain tasks assigned to be processed by a particular GPC 1700. The active task pool may comprise a number of slots (e.g., 4 slots) for tasks that are actively being processed by the GPC 1700 modules. As a GPC 1700 finishes the execution of a task, that task is evicted from the active task pool for the GPC 1700 and one of the other tasks from the pending task pool is selected and scheduled for execution on the GPC 1700. If an active task has been idle on the GPC 1700, such as while waiting for a data dependency to be resolved, then the active task may be evicted from the GPC 1700 and returned to the pending task pool while another task in the pending task pool is selected and scheduled for execution on the GPC 1700.
The work distribution unit 1614 communicates with the one or more GPC 1700 modules via crossbar 1618. The crossbar 1618 is an interconnect network that couples many of the units of the parallel processing unit 1600 to other units of the parallel processing unit 1600. For example, the crossbar 1618 may be configured to couple the work distribution unit 1614 to a particular GPC 1700. Although not shown explicitly, one or more other units of the parallel processing unit 1600 may also be connected to the crossbar 1618 via the hub 1616.
The tasks are managed by the scheduler unit 1612 and dispatched to a GPC 1700 by the work distribution unit 1614. The GPC 1700 is configured to process the task and generate results. The results may be consumed by other tasks within the GPC 1700, routed to a different GPC 1700 via the crossbar 1618, or stored in the memory 1604. The results can be written to the memory 1604 via the memory partition unit 1800 modules, which implement a memory interface for reading and writing data to/from the memory 1604. The results can be transmitted to another parallel processing unit 1600 or CPU via the NVLink 1608. In an embodiment, the parallel processing unit 1600 includes a number U of memory partition unit 1800 modules that is equal to the number of separate and distinct memory 1604 devices coupled to the parallel processing unit 1600. A memory partition unit 1800 will be described in more detail below in conjunction with
In an embodiment, a host processor executes a driver kernel that implements an application programming interface (API) that enables one or more applications executing on the host processor to schedule operations for execution on the parallel processing unit 1600. In an embodiment, multiple compute applications are simultaneously executed by the parallel processing unit 1600 and the parallel processing unit 1600 provides isolation, quality of service (QoS), and independent address spaces for the multiple compute applications. An application may generate instructions (e.g., API calls) that cause the driver kernel to generate one or more tasks for execution by the parallel processing unit 1600. The driver kernel outputs tasks to one or more streams being processed by the parallel processing unit 1600. Each task may comprise one or more groups of related threads, referred to herein as a warp. In an embodiment, a warp comprises 32 related threads that may be executed in parallel. Cooperating threads may refer to a plurality of threads including instructions to perform the task and that may exchange data through shared memory. Threads and cooperating threads are described in more detail in conjunction with
In an embodiment, the operation of the GPC 1700 is controlled by the pipeline manager 1702. The pipeline manager 1702 manages the configuration of the one or more DPC 1706 modules for processing tasks allocated to the GPC 1700. In an embodiment, the pipeline manager 1702 may configure at least one of the one or more DPC 1706 modules to implement at least a portion of a graphics rendering pipeline. For example, a DPC 1706 may be configured to execute a vertex shader program on the programmable SM 1900. The pipeline manager 1702 may also be configured to route packets received from the work distribution unit 1614 to the appropriate logical units within the GPC 1700. For example, some packets may be routed to fixed function hardware units in the PROP 1704 and/or raster engine 1708 while other packets may be routed to the DPC 1706 modules for processing by the primitive engine 1712 or the SM 1900. In an embodiment, the pipeline manager 1702 may configure at least one of the one or more DPC 1706 modules to implement a neural network model and/or a computing pipeline.
The PROP 1704 is configured to route data generated by the raster engine 1708 and the DPC 1706 modules to a Raster Operations (ROP) unit, described in more detail in conjunction with
The raster engine 1708 includes a number of fixed function hardware units configured to perform various raster operations. In an embodiment, the raster engine 1708 includes a setup engine, a coarse raster engine, a culling engine, a clipping engine, a fine raster engine, and a tile coalescing engine. The setup engine receives transformed vertices and generates plane equations associated with the geometric primitive defined by the vertices. The plane equations are transmitted to the coarse raster engine to generate coverage information (e.g., an x, y coverage mask for a tile) for the primitive. The output of the coarse raster engine is transmitted to the culling engine where fragments associated with the primitive that fail a z-test are culled, and transmitted to a clipping engine where fragments lying outside a viewing frustum are clipped. Those fragments that survive clipping and culling may be passed to the fine raster engine to generate attributes for the pixel fragments based on the plane equations generated by the setup engine. The output of the raster engine 1708 comprises fragments to be processed, for example, by a fragment shader implemented within a DPC 1706.
Each DPC 1706 included in the GPC 1700 includes an MPC 1710, a primitive engine 1712, and one or more SM 1900 modules. The MPC 1710 controls the operation of the DPC 1706, routing packets received from the pipeline manager 1702 to the appropriate units in the DPC 1706. For example, packets associated with a vertex may be routed to the primitive engine 1712, which is configured to fetch vertex attributes associated with the vertex from the memory 1604. In contrast, packets associated with a shader program may be transmitted to the SM 1900.
The SM 1900 comprises a programmable streaming processor that is configured to process tasks represented by a number of threads. Each SM 1900 is multi-threaded and configured to execute a plurality of threads (e.g., 32 threads) from a particular group of threads concurrently. In an embodiment, the SM 1900 implements a Single-Instruction, Multiple-Data (SIMD) architecture where each thread in a group of threads (e.g., a warp) is configured to process a different set of data based on the same set of instructions. All threads in the group of threads execute the same instructions. In another embodiment, the SM 1900 implements a Single-Instruction, Multiple Thread (SIMT) architecture where each thread in a group of threads is configured to process a different set of data based on the same set of instructions, but where individual threads in the group of threads are allowed to diverge during execution. In an embodiment, a program counter, call stack, and execution state is maintained for each warp, enabling concurrency between warps and serial execution within warps when threads within the warp diverge. In another embodiment, a program counter, call stack, and execution state is maintained for each individual thread, enabling equal concurrency between all threads, within and between warps. When execution state is maintained for each individual thread, threads executing the same instructions may be converged and executed in parallel for maximum efficiency. The SM 1900 will be described in more detail below in conjunction with
The MMU 1716 provides an interface between the GPC 1700 and the memory partition unit 1800. The MMU 1716 may provide translation of virtual addresses into physical addresses, memory protection, and arbitration of memory requests. In an embodiment, the MMU 1716 provides one or more translation lookaside buffers (TLBs) for performing translation of virtual addresses into physical addresses in the memory 1604.
In an embodiment, the memory interface 1806 implements an HBM2 memory interface and Y equals half U. In an embodiment, the HBM2 memory stacks are located on the same physical package as the parallel processing unit 1600, providing substantial power and area savings compared with conventional GDDR5 SDRAM systems. In an embodiment, each HBM2 stack includes four memory dies and Y equals 4, with HBM2 stack including two 128-bit channels per die for a total of 8 channels and a data bus width of 1024 bits.
In an embodiment, the memory 1604 supports Single-Error Correcting Double-Error Detecting (SECDED) Error Correction Code (ECC) to protect data. ECC provides higher reliability for compute applications that are sensitive to data corruption. Reliability is especially important in large-scale cluster computing environments where parallel processing unit 1600 modules process very large datasets and/or run applications for extended periods.
In an embodiment, the parallel processing unit 1600 implements a multi-level memory hierarchy. In an embodiment, the memory partition unit 1800 supports a unified memory to provide a single unified virtual address space for CPU and parallel processing unit 1600 memory, enabling data sharing between virtual memory systems. In an embodiment the frequency of accesses by a parallel processing unit 1600 to memory located on other processors is traced to ensure that memory pages are moved to the physical memory of the parallel processing unit 1600 that is accessing the pages more frequently. In an embodiment, the NVLink 1608 supports address translation services allowing the parallel processing unit 1600 to directly access a CPU's page tables and providing full access to CPU memory by the parallel processing unit 1600.
In an embodiment, copy engines transfer data between multiple parallel processing unit 1600 modules or between parallel processing unit 1600 modules and CPUs. The copy engines can generate page faults for addresses that are not mapped into the page tables. The memory partition unit 1800 can then service the page faults, mapping the addresses into the page table, after which the copy engine can perform the transfer. In a conventional system, memory is pinned (e.g., non-pageable) for multiple copy engine operations between multiple processors, substantially reducing the available memory. With hardware page faulting, addresses can be passed to the copy engines without worrying if the memory pages are resident, and the copy process is transparent.
Data from the memory 1604 or other system memory may be fetched by the memory partition unit 1800 and stored in the L2 cache 1804, which is located on-chip and is shared between the various GPC 1700 modules. As shown, each memory partition unit 1800 includes a portion of the L2 cache 1804 associated with a corresponding memory 1604 device. Lower level caches may then be implemented in various units within the GPC 1700 modules. For example, each of the SM 1900 modules may implement an L1 cache. The L1 cache is private memory that is dedicated to a particular SM 1900. Data from the L2 cache 1804 may be fetched and stored in each of the L1 caches for processing in the functional units of the SM 1900 modules. The L2 cache 1804 is coupled to the memory interface 1806 and the crossbar 1618.
The ROP 1802 performs graphics raster operations related to pixel color, such as color compression, pixel blending, and the like. The ROP 1802 also implements depth testing in conjunction with the raster engine 1708, receiving a depth for a sample location associated with a pixel fragment from the culling engine of the raster engine 1708. The depth is tested against a corresponding depth in a depth buffer for a sample location associated with the fragment. If the fragment passes the depth test for the sample location, then the ROP 1802 updates the depth buffer and transmits a result of the depth test to the raster engine 1708. It will be appreciated that the number of partition memory partition unit 1800 modules may be different than the number of GPC 1700 modules and, therefore, each ROP 1802 may be coupled to each of the GPC 1700 modules. The ROP 1802 tracks packets received from the different GPC 1700 modules and determines which GPC 1700 that a result generated by the ROP 1802 is routed to through the crossbar 1618. Although the ROP 1802 is included within the memory partition unit 1800 in
As described above, the work distribution unit 1614 dispatches tasks for execution on the GPC 1700 modules of the parallel processing unit 1600. The tasks are allocated to a particular DPC 1706 within a GPC 1700 and, if the task is associated with a shader program, the task may be allocated to an SM 1900. The scheduler unit 1612 receives the tasks from the work distribution unit 1614 and manages instruction scheduling for one or more thread blocks assigned to the SM 1900. The scheduler unit 1904 schedules thread blocks for execution as warps of parallel threads, where each thread block is allocated at least one warp. In an embodiment, each warp executes 32 threads. The scheduler unit 1904 may manage a plurality of different thread blocks, allocating the warps to the different thread blocks and then dispatching instructions from the plurality of different cooperative groups to the various functional units (e.g., core 1910 modules, SFU 1912 modules, and LSU 1914 modules) during each clock cycle.
Cooperative Groups is a programming model for organizing groups of communicating threads that allows developers to express the granularity at which threads are communicating, enabling the expression of richer, more efficient parallel decompositions. Cooperative launch APIs support synchronization amongst thread blocks for the execution of parallel algorithms. Conventional programming models provide a single, simple construct for synchronizing cooperating threads: a barrier across all threads of a thread block (e.g., the syncthreads( ) function). However, programmers would often like to define groups of threads at smaller than thread block granularities and synchronize within the defined groups to enable greater performance, design flexibility, and software reuse in the form of collective group-wide function interfaces.
Cooperative Groups enables programmers to define groups of threads explicitly at sub-block (e.g., as small as a single thread) and multi-block granularities, and to perform collective operations such as synchronization on the threads in a cooperative group. The programming model supports clean composition across software boundaries, so that libraries and utility functions can synchronize safely within their local context without having to make assumptions about convergence. Cooperative Groups primitives enable new patterns of cooperative parallelism, including producer-consumer parallelism, opportunistic parallelism, and global synchronization across an entire grid of thread blocks.
A dispatch 1906 unit is configured within the scheduler unit 1904 to transmit instructions to one or more of the functional units. In one embodiment, the scheduler unit 1904 includes two dispatch 1906 units that enable two different instructions from the same warp to be dispatched during each clock cycle. In alternative embodiments, each scheduler unit 1904 may include a single dispatch 1906 unit or additional dispatch 1906 units.
Each SM 1900 includes a register file 1908 that provides a set of registers for the functional units of the SM 1900. In an embodiment, the register file 1908 is divided between each of the functional units such that each functional unit is allocated a dedicated portion of the register file 1908. In another embodiment, the register file 1908 is divided between the different warps being executed by the SM 1900. The register file 1908 provides temporary storage for operands connected to the data paths of the functional units.
Each SM 1900 comprises L processing core 1910 modules. In an embodiment, the SM 1900 includes a large number (e.g., 128, etc.) of distinct processing core 1910 modules. Each core 1910 may include a fully-pipelined, single-precision, double-precision, and/or mixed precision processing unit that includes a floating point arithmetic logic unit and an integer arithmetic logic unit. In an embodiment, the floating point arithmetic logic units implement the IEEE 754-2008 standard for floating point arithmetic. In an embodiment, the core 1910 modules include 64 single-precision (32-bit) floating point cores, 64 integer cores, 32 double-precision (64-bit) floating point cores, and 8 tensor cores.
Tensor cores configured to perform matrix operations, and, in an embodiment, one or more tensor cores are included in the core 1910 modules. In particular, the tensor cores are configured to perform deep learning matrix arithmetic, such as convolution operations for neural network training and inferencing. In an embodiment, each tensor core operates on a 4×4 matrix and performs a matrix multiply and accumulate operation D=A′B+C, where A, B, C, and D are 4×4 matrices.
In an embodiment, the matrix multiply inputs A and B are 16-bit floating point matrices, while the accumulation matrices C and D may be 16-bit floating point or 32-bit floating point matrices. Tensor Cores operate on 16-bit floating point input data with 32-bit floating point accumulation. The 16-bit floating point multiply requires 64 operations and results in a full precision product that is then accumulated using 32-bit floating point addition with the other intermediate products for a 4×4×4 matrix multiply. In practice, Tensor Cores are used to perform much larger two-dimensional or higher dimensional matrix operations, built up from these smaller elements. An API, such as CUDA 9 C++ API, exposes specialized matrix load, matrix multiply and accumulate, and matrix store operations to efficiently use Tensor Cores from a CUDA-C++ program. At the CUDA level, the warp-level interface assumes 16×16 size matrices spanning all 32 threads of the warp.
Each SM 1900 also comprises M SFU 1912 modules that perform special functions (e.g., attribute evaluation, reciprocal square root, and the like). In an embodiment, the SFU 1912 modules may include a tree traversal unit configured to traverse a hierarchical tree data structure. In an embodiment, the SFU 1912 modules may include texture unit configured to perform texture map filtering operations. In an embodiment, the texture units are configured to load texture maps (e.g., a 2D array of texels) from the memory 1604 and sample the texture maps to produce sampled texture values for use in shader programs executed by the SM 1900. In an embodiment, the texture maps are stored in the shared memory/L1 cache 1918. The texture units implement texture operations such as filtering operations using mip-maps (e.g., texture maps of varying levels of detail). In an embodiment, each SM 1900 includes two texture units.
Each SM 1900 also comprises N LSU 1914 modules that implement load and store operations between the shared memory/L1 cache 1918 and the register file 1908. Each SM 1900 includes an interconnect network 1916 that connects each of the functional units to the register file 1908 and the LSU 1914 to the register file 1908 and shared memory/L1 cache 1918. In an embodiment, the interconnect network 1916 is a crossbar that can be configured to connect any of the functional units to any of the registers in the register file 1908 and connect the LSU 1914 modules to the register file 1908 and memory locations in shared memory/L1 cache 1918.
The shared memory/L1 cache 1918 is an array of on-chip memory that allows for data storage and communication between the SM 1900 and the primitive engine 1712 and between threads in the SM 1900. In an embodiment, the shared memory/L1 cache 1918 comprises 128 KB of storage capacity and is in the path from the SM 1900 to the memory partition unit 1800. The shared memory/L1 cache 1918 can be used to cache reads and writes. One or more of the shared memory/L1 cache 1918, L2 cache 1804, and memory 1604 are backing stores.
Combining data cache and shared memory functionality into a single memory block provides the best overall performance for both types of memory accesses. The capacity is usable as a cache by programs that do not use shared memory. For example, if shared memory is configured to use half of the capacity, texture and load/store operations can use the remaining capacity. Integration within the shared memory/L1 cache 1918 enables the shared memory/L1 cache 1918 to function as a high-throughput conduit for streaming data while simultaneously providing high-bandwidth and low-latency access to frequently reused data.
When configured for general purpose parallel computation, a simpler configuration can be used compared with graphics processing. Specifically, the fixed function graphics processing units shown in
The parallel processing unit 1600 may be included in a desktop computer, a laptop computer, a tablet computer, servers, supercomputers, a smart-phone (e.g., a wireless, hand-held device), personal digital assistant (PDA), a digital camera, a vehicle, a head mounted display, a hand-held electronic device, and the like. In an embodiment, the parallel processing unit 1600 is embodied on a single semiconductor substrate. In another embodiment, the parallel processing unit 1600 is included in a system-on-a-chip (SoC) along with one or more other devices such as additional parallel processing unit 1600 modules, the memory 1604, a reduced instruction set computer (RISC) CPU, a memory management unit (MMU), a digital-to-analog converter (DAC), and the like.
In an embodiment, the parallel processing unit 1600 may be included on a graphics card that includes one or more memory devices 304. The graphics card may be configured to interface with a PCIe slot on a motherboard of a desktop computer. In yet another embodiment, the parallel processing unit 1600 may be an integrated graphics processing unit (iGPU) or parallel processor included in the chipset of the motherboard.
Exemplary Computing System
Systems with multiple GPUs and CPUs are used in a variety of industries as developers expose and leverage more parallelism in applications such as artificial intelligence computing. High-performance GPU-accelerated systems with tens to many thousands of compute nodes are deployed in data centers, research facilities, and supercomputers to solve ever larger problems. As the number of processing devices within the high-performance systems increases, the communication and data transfer mechanisms need to scale to support the increased bandwidth.
In another embodiment (not shown), the NVLink 1608 provides one or more high-speed communication links between each of the parallel processing unit 1600 modules and the CPU 2006 and the switch 2002 interfaces between the interconnect 1602 and each of the parallel processing unit 1600 modules. The parallel processing unit 1600 modules, memory 1604 modules, and interconnect 1602 may be situated on a single semiconductor platform to form a parallel processing module 2004. In yet another embodiment (not shown), the interconnect 1602 provides one or more communication links between each of the parallel processing unit 1600 modules and the CPU 2006 and the switch 2002 interfaces between each of the parallel processing unit 1600 modules using the NVLink 1608 to provide one or more high-speed communication links between the parallel processing unit 1600 modules. In another embodiment (not shown), the NVLink 1608 provides one or more high-speed communication links between the parallel processing unit 1600 modules and the CPU 2006 through the switch 2002. In yet another embodiment (not shown), the interconnect 1602 provides one or more communication links between each of the parallel processing unit 1600 modules directly. One or more of the NVLink 1608 high-speed communication links may be implemented as a physical NVLink interconnect or either an on-chip or on-die interconnect using the same protocol as the NVLink 1608.
In the context of the present description, a single semiconductor platform may refer to a sole unitary semiconductor-based integrated circuit fabricated on a die or chip. It should be noted that the term single semiconductor platform may also refer to multi-chip modules with increased connectivity which simulate on-chip operation and make substantial improvements over utilizing a conventional bus implementation. Of course, the various circuits or devices may also be situated separately or in various combinations of semiconductor platforms per the desires of the user. Alternately, the parallel processing module 2004 may be implemented as a circuit board substrate and each of the parallel processing unit 1600 modules and/or memory 1604 modules may be packaged devices. In an embodiment, the CPU 2006, switch 2002, and the parallel processing module 2004 are situated on a single semiconductor platform.
In an embodiment, the signaling rate of each NVLink 1608 is 20 to 25 Gigabits/second and each parallel processing unit 1600 includes six NVLink 1608 interfaces (as shown in
In an embodiment, the NVLink 1608 allows direct load/store/atomic access from the CPU 2006 to each parallel processing unit 1600 module's memory 1604. In an embodiment, the NVLink 1608 supports coherency operations, allowing data read from the memory 1604 modules to be stored in the cache hierarchy of the CPU 2006, reducing cache access latency for the CPU 2006. In an embodiment, the NVLink 1608 includes support for Address Translation Services (ATS), allowing the parallel processing unit 1600 to directly access page tables within the CPU 2006. One or more of the NVLink 1608 may also be configured to operate in a low-power mode.
The exemplary processing system 2100 also includes input devices 2108, the parallel processing module 2004, and display devices 2106, e.g. a conventional CRT (cathode ray tube), LCD (liquid crystal display), LED (light emitting diode), plasma display or the like. User input may be received from the input devices 2108, e.g., keyboard, mouse, touchpad, microphone, and the like. Each of the foregoing modules and/or devices may even be situated on a single semiconductor platform to form the exemplary processing system 2100. Alternately, the various modules may also be situated separately or in various combinations of semiconductor platforms per the desires of the user.
Further, the exemplary processing system 2100 may be coupled to a network (e.g., a telecommunications network, local area network (LAN), wireless network, wide area network (WAN) such as the Internet, peer-to-peer network, cable network, or the like) through a network interface 2102 for communication purposes.
The exemplary processing system 2100 may also include a secondary storage (not shown). The secondary storage includes, for example, a hard disk drive and/or a removable storage drive, representing a floppy disk drive, a magnetic tape drive, a compact disk drive, digital versatile disk (DVD) drive, recording device, universal serial bus (USB) flash memory. The removable storage drive reads from and/or writes to a removable storage unit in a well-known manner.
Computer programs, or computer control logic algorithms, may be stored in the main memory 2104 and/or the secondary storage. Such computer programs, when executed, enable the exemplary processing system 2100 to perform various functions. The main memory 2104, the storage, and/or any other storage are possible examples of computer-readable media.
The architecture and/or functionality of the various previous figures may be implemented in the context of a general computer system, a circuit board system, a game console system dedicated for entertainment purposes, an application-specific system, and/or any other desired system. For example, the exemplary processing system 2100 may take the form of a desktop computer, a laptop computer, a tablet computer, servers, supercomputers, a smart-phone (e.g., a wireless, hand-held device), personal digital assistant (PDA), a digital camera, a vehicle, a head mounted display, a hand-held electronic device, a mobile phone device, a television, workstation, game consoles, embedded system, and/or any other type of logic.
While various embodiments have been described above, it should be understood that they have been presented by way of example only, and not limitation. Thus, the breadth and scope of a preferred embodiment should not be limited by any of the above-described exemplary embodiments, but should be defined only in accordance with the following claims and their equivalents.
Graphics Processing Pipeline
An application writes model data for a scene (e.g., a collection of vertices and attributes) to a memory such as a system memory or memory 1604. The model data defines each of the objects that may be visible on a display. The application then makes an API call to the driver kernel that requests the model data to be rendered and displayed. The driver kernel reads the model data and writes commands to the one or more streams to perform operations to process the model data. The commands may reference different shader programs to be implemented on the SM 1900 modules of the parallel processing unit 1600 including one or more of a vertex shader, hull shader, domain shader, geometry shader, and a pixel shader. For example, one or more of the SM 1900 modules may be configured to execute a vertex shader program that processes a number of vertices defined by the model data. In an embodiment, the different SM 1900 modules may be configured to execute different shader programs concurrently. For example, a first subset of SM 1900 modules may be configured to execute a vertex shader program while a second subset of SM 1900 modules may be configured to execute a pixel shader program. The first subset of SM 1900 modules processes vertex data to produce processed vertex data and writes the processed vertex data to the L2 cache 1804 and/or the memory 1604. After the processed vertex data is rasterized (e.g., transformed from three-dimensional data into two-dimensional data in screen space) to produce fragment data, the second subset of SM 1900 modules executes a pixel shader to produce processed fragment data, which is then blended with other processed fragment data and written to the frame buffer in memory 1604. The vertex shader program and pixel shader program may execute concurrently, processing different data from the same scene in a pipelined fashion until all of the model data for the scene has been rendered to the frame buffer. Then, the contents of the frame buffer are transmitted to a display controller for display on a display device.
The graphics processing pipeline 2200 is an abstract flow diagram of the processing steps implemented to generate 2D computer-generated images from 3D geometry data. As is well-known, pipeline architectures may perform long latency operations more efficiently by splitting up the operation into a plurality of stages, where the output of each stage is coupled to the input of the next successive stage. Thus, the graphics processing pipeline 2200 receives input data 601 that is transmitted from one stage to the next stage of the graphics processing pipeline 2200 to generate output data 2204. In an embodiment, the graphics processing pipeline 2200 may represent a graphics processing pipeline defined by the OpenGL® API. As an option, the graphics processing pipeline 2200 may be implemented in the context of the functionality and architecture of the previous Figures and/or any subsequent Figure(s).
As shown in
The data assembly 2206 stage receives the input data 2202 that specifies vertex data for high-order surfaces, primitives, or the like. The data assembly 2206 stage collects the vertex data in a temporary storage or queue, such as by receiving a command from the host processor that includes a pointer to a buffer in memory and reading the vertex data from the buffer. The vertex data is then transmitted to the vertex shading 2208 stage for processing.
The vertex shading 2208 stage processes vertex data by performing a set of operations (e.g., a vertex shader or a program) once for each of the vertices. Vertices may be, e.g., specified as a 4-coordinate vector (e.g., <x, y, z, w>) associated with one or more vertex attributes (e.g., color, texture coordinates, surface normal, etc.). The vertex shading 2208 stage may manipulate individual vertex attributes such as position, color, texture coordinates, and the like. In other words, the vertex shading 2208 stage performs operations on the vertex coordinates or other vertex attributes associated with a vertex. Such operations commonly including lighting operations (e.g., modifying color attributes for a vertex) and transformation operations (e.g., modifying the coordinate space for a vertex). For example, vertices may be specified using coordinates in an object-coordinate space, which are transformed by multiplying the coordinates by a matrix that translates the coordinates from the object-coordinate space into a world space or a normalized-device-coordinate (NCD) space. The vertex shading 2208 stage generates transformed vertex data that is transmitted to the primitive assembly 2210 stage.
The primitive assembly 2210 stage collects vertices output by the vertex shading 2208 stage and groups the vertices into geometric primitives for processing by the geometry shading 2212 stage. For example, the primitive assembly 2210 stage may be configured to group every three consecutive vertices as a geometric primitive (e.g., a triangle) for transmission to the geometry shading 2212 stage. In some embodiments, specific vertices may be reused for consecutive geometric primitives (e.g., two consecutive triangles in a triangle strip may share two vertices). The primitive assembly 2210 stage transmits geometric primitives (e.g., a collection of associated vertices) to the geometry shading 2212 stage.
The geometry shading 2212 stage processes geometric primitives by performing a set of operations (e.g., a geometry shader or program) on the geometric primitives. Tessellation operations may generate one or more geometric primitives from each geometric primitive. In other words, the geometry shading 2212 stage may subdivide each geometric primitive into a finer mesh of two or more geometric primitives for processing by the rest of the graphics processing pipeline 2200. The geometry shading 2212 stage transmits geometric primitives to the viewport SCC 2214 stage.
In an embodiment, the graphics processing pipeline 2200 may operate within a streaming multiprocessor and the vertex shading 2208 stage, the primitive assembly 2210 stage, the geometry shading 2212 stage, the fragment shading 2218 stage, and/or hardware/software associated therewith, may sequentially perform processing operations. Once the sequential processing operations are complete, in an embodiment, the viewport SCC 2214 stage may utilize the data. In an embodiment, primitive data processed by one or more of the stages in the graphics processing pipeline 2200 may be written to a cache (e.g. L1 cache, a vertex cache, etc.). In this case, in an embodiment, the viewport SCC 2214 stage may access the data in the cache. In an embodiment, the viewport SCC 2214 stage and the rasterization 2216 stage are implemented as fixed function circuitry.
The viewport SCC 2214 stage performs viewport scaling, culling, and clipping of the geometric primitives. Each surface being rendered to is associated with an abstract camera position. The camera position represents a location of a viewer looking at the scene and defines a viewing frustum that encloses the objects of the scene. The viewing frustum may include a viewing plane, a rear plane, and four clipping planes. Any geometric primitive entirely outside of the viewing frustum may be culled (e.g., discarded) because the geometric primitive will not contribute to the final rendered scene. Any geometric primitive that is partially inside the viewing frustum and partially outside the viewing frustum may be clipped (e.g., transformed into a new geometric primitive that is enclosed within the viewing frustum. Furthermore, geometric primitives may each be scaled based on a depth of the viewing frustum. All potentially visible geometric primitives are then transmitted to the rasterization 2216 stage.
The rasterization 2216 stage converts the 3D geometric primitives into 2D fragments (e.g. capable of being utilized for display, etc.). The rasterization 2216 stage may be configured to utilize the vertices of the geometric primitives to setup a set of plane equations from which various attributes can be interpolated. The rasterization 2216 stage may also compute a coverage mask for a plurality of pixels that indicates whether one or more sample locations for the pixel intercept the geometric primitive. In an embodiment, z-testing may also be performed to determine if the geometric primitive is occluded by other geometric primitives that have already been rasterized. The rasterization 2216 stage generates fragment data (e.g., interpolated vertex attributes associated with a particular sample location for each covered pixel) that are transmitted to the fragment shading 2218 stage.
The fragment shading 2218 stage processes fragment data by performing a set of operations (e.g., a fragment shader or a program) on each of the fragments. The fragment shading 2218 stage may generate pixel data (e.g., color values) for the fragment such as by performing lighting operations or sampling texture maps using interpolated texture coordinates for the fragment. The fragment shading 2218 stage generates pixel data that is transmitted to the raster operations 2220 stage.
The raster operations 2220 stage may perform various operations on the pixel data such as performing alpha tests, stencil tests, and blending the pixel data with other pixel data corresponding to other fragments associated with the pixel. When the raster operations 2220 stage has finished processing the pixel data (e.g., the output data 2204), the pixel data may be written to a render target such as a frame buffer, a color buffer, or the like.
It will be appreciated that one or more additional stages may be included in the graphics processing pipeline 2200 in addition to or in lieu of one or more of the stages described above. Various implementations of the abstract graphics processing pipeline may implement different stages. Furthermore, one or more of the stages described above may be excluded from the graphics processing pipeline in some embodiments (such as the geometry shading 2212 stage). Other types of graphics processing pipelines are contemplated as being within the scope of the present disclosure. Furthermore, any of the stages of the graphics processing pipeline 2200 may be implemented by one or more dedicated hardware units within a graphics processor such as parallel processing unit 1600. Other stages of the graphics processing pipeline 2200 may be implemented by programmable hardware units such as the SM 1900 of the parallel processing unit 1600.
The graphics processing pipeline 2200 may be implemented via an application executed by a host processor, such as a CPU. In an embodiment, a device driver may implement an application programming interface (API) that defines various functions that can be utilized by an application in order to generate graphical data for display. The device driver is a software program that includes a plurality of instructions that control the operation of the parallel processing unit 1600. The API provides an abstraction for a programmer that lets a programmer utilize specialized graphics hardware, such as the parallel processing unit 1600, to generate the graphical data without requiring the programmer to utilize the specific instruction set for the parallel processing unit 1600. The application may include an API call that is routed to the device driver for the parallel processing unit 1600. The device driver interprets the API call and performs various operations to respond to the API call. In some instances, the device driver may perform operations by executing instructions on the CPU. In other instances, the device driver may perform operations, at least in part, by launching operations on the parallel processing unit 1600 utilizing an input/output interface between the CPU and the parallel processing unit 1600. In an embodiment, the device driver is configured to implement the graphics processing pipeline 2200 utilizing the hardware of the parallel processing unit 1600.
Various programs may be executed within the parallel processing unit 1600 in order to implement the various stages of the graphics processing pipeline 2200. For example, the device driver may launch a kernel on the parallel processing unit 1600 to perform the vertex shading 2208 stage on one SM 1900 (or multiple SM 1900 modules). The device driver (or the initial kernel executed by the parallel processing unit 1600) may also launch other kernels on the parallel processing unit 1600 to perform other stages of the graphics processing pipeline 2200, such as the geometry shading 2212 stage and the fragment shading 2218 stage. In addition, some of the stages of the graphics processing pipeline 2200 may be implemented on fixed unit hardware such as a rasterizer or a data assembler implemented within the parallel processing unit 1600. It will be appreciated that results from one kernel may be processed by one or more intervening fixed function hardware units before being processed by a subsequent kernel on an SM 1900.
Various logic functional operations described herein may be implemented in logic that is referred to using a noun or noun phrase reflecting said operation or function. For example, an association operation may be carried out by an “associator” or “correlator”. Likewise, switching may be carried out by a “switch”, selection by a “selector”, and so on.
“Logic” is used herein to machine memory circuits, non transitory machine readable media, and/or circuitry which by way of its material and/or material-energy configuration comprises control and/or procedural signals, and/or settings and values (such as resistance, impedance, capacitance, inductance, current/voltage ratings, etc.), that may be applied to influence the operation of a device. Magnetic media, electronic circuits, electrical and optical memory (both volatile and nonvolatile), and firmware are examples of logic. Logic specifically excludes pure signals or software per se (however does not exclude machine memories comprising software and thereby forming configurations of matter).
Within this disclosure, different entities (which may variously be referred to as “units,” “circuits,” other components, etc.) may be described or claimed as “configured” to perform one or more tasks or operations. This formulation—[entity] configured to [perform one or more tasks]—is used herein to refer to structure (i.e., something physical, such as an electronic circuit). More specifically, this formulation is used to indicate that this structure is arranged to perform the one or more tasks during operation. A structure can be said to be “configured to” perform some task even if the structure is not currently being operated. A “credit distribution circuit configured to distribute credits to a plurality of processor cores” is intended to cover, for example, an integrated circuit that has circuitry that performs this function during operation, even if the integrated circuit in question is not currently being used (e.g., a power supply is not connected to it). Thus, an entity described or recited as “configured to” perform some task refers to something physical, such as a device, circuit, memory storing program instructions executable to implement the task, etc. This phrase is not used herein to refer to something intangible.
The term “configured to” is not intended to mean “configurable to.” An unprogrammed FPGA, for example, would not be considered to be “configured to” perform some specific function, although it may be “configurable to” perform that function after programming.
Reciting in the appended claims that a structure is “configured to” perform one or more tasks is expressly intended not to invoke 35 U.S.C. § 112(f) for that claim element. Accordingly, claims in this application that do not otherwise include the “means for” [performing a function] construct should not be interpreted under 35 U.S.C § 112(f).
As used herein, the term “based on” is used to describe one or more factors that affect a determination. This term does not foreclose the possibility that additional factors may affect the determination. That is, a determination may be solely based on specified factors or based on the specified factors as well as other, unspecified factors. Consider the phrase “determine A based on B.” This phrase specifies that B is a factor that is used to determine A or that affects the determination of A. This phrase does not foreclose that the determination of A may also be based on some other factor, such as C. This phrase is also intended to cover an embodiment in which A is determined based solely on B. As used herein, the phrase “based on” is synonymous with the phrase “based at least in part on.”
As used herein, the phrase “in response to” describes one or more factors that trigger an effect. This phrase does not foreclose the possibility that additional factors may affect or otherwise trigger the effect. That is, an effect may be solely in response to those factors, or may be in response to the specified factors as well as other, unspecified factors. Consider the phrase “perform A in response to B.” This phrase specifies that B is a factor that triggers the performance of A. This phrase does not foreclose that performing A may also be in response to some other factor, such as C. This phrase is also intended to cover an embodiment in which A is performed solely in response to B.
As used herein, the terms “first,” “second,” etc. are used as labels for nouns that they precede, and do not imply any type of ordering (e.g., spatial, temporal, logical, etc.), unless stated otherwise. For example, in a register file having eight registers, the terms “first register” and “second register” can be used to refer to any two of the eight registers, and not, for example, just logical registers 0 and 1.
When used in the claims, the term “or” is used as an inclusive or and not as an exclusive or. For example, the phrase “at least one of x, y, or z” means any one of x, y, and z, as well as any combination thereof.
Having thus described illustrative embodiments in detail, it will be apparent that modifications and variations are possible without departing from the scope of the invention as claimed.
This application claims priority and benefit of U.S. application Ser. No. 62/758,337, titled “Testability Analysis with Graph Convolutional Networks”, filed on Nov. 9, 2018, the contents of which are incorporated herein by reference in their entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 10460359 | Huang | Oct 2019 | B1 |
| 20100161680 | Atre | Jun 2010 | A1 |
| 20150350324 | Hu | Dec 2015 | A1 |
| 20170221238 | Limberger | Aug 2017 | A1 |
| 20180248904 | Villella | Aug 2018 | A1 |
| 20190213067 | Vadivelu | Jul 2019 | A1 |
| Entry |
|---|
| Palash Goyal and Emilio Ferrara et al, Graph Embedding Techniques, Applications, and Performance: A Survey, ARXIV:1705.02801V4 [CS.SI] Dec. 22, 2017. |
| Rex Ying et al, Graph Convolutional Neural Networks for Web-Scale Recommender Systems, ARXIV:1806.01973V1 [CS.IR] Jun. 6, 2018. |
| Smriti Bhagat et al, Node Classification in Social Networks, ARXIV:1101.3291V1 [CS.SI] Jan. 17, 2011. |
| Number | Date | Country | |
|---|---|---|---|
| 62758337 | Nov 2018 | US |
