Skip to content

CHANGELOG.md

Latest commit

 

History

History
591 lines (325 loc) · 14.4 KB

CHANGELOG.md

File metadata and controls

591 lines (325 loc) · 14.4 KB
Raw

List of major changes and improvements

[v2.0 -- v2.1] -- 2010-12-31

Added

  • Workload models.

    • Add ResNet-50.

  • Software engineering.

    • Port code to Python 3; drop Python 2 support.

    • Add pylintrc.

Changed

  • Software models.

    • Share the same partition between input layer and external layers. For layered LSTMs, the number of external layers could be quite a few (e.g., 8). The complete combination of all partition choices of these external layers is too high to explore. So we assume all input and external layers share the same partition scheme.

  • Software engineering.

    • Allow both relative and absolute overheads in approx_dividable.

Fixed

  • In LoopBlockingScheme, put weight-pinning code block after buffer sharing.

[v1.6 -- v2.0] -- 2018-03-01

Added

  • Hardware models.

    • Access forwarding.

    • Buffer sharing scheme.

      • Use BufShrScheme class to represent and calculate NoC transfers.

  • Software models.

    • Add SchedulingConstraint class to specify loop blocking and partitioning constraints.

      • Add lazily updated rules to allow refine constraint with previous scheduling results at runtime.
      • Add subclass SchedulingConstraintLayerPipeline for layer pipelining constraints.

    • Add InterLayerPipeline.

      • Layers are organized into PipelineSegment, which are simultaneously mapped on to the resource both spatially and temporally.
      • Each layer in the segment has a 3-tuple scheduling index including segment index, spatial index, and temporal index.
      • Each layer in the segment has its resource allocation and scheduling constraint.
      • Use PipelineSegmentTiming to capture the timing relation of layers in the segment.
      • Specify maximum allowed execution time overhead due to layer pipelining in Option.
      • Specify maximum pipelining degree for layer pipelining in Option.

    • Add layer pipelining optimizations.

      • Ofmap forwarding: alternate layer loop ordering.
      • Ifmap forwarding: sharing the same inputs from memory to multiple regions.
      • Support model weight pinning when no resource time-multiplexing.
      • Allow disabling optimizations for layer pipelining to fall back to basic pipelining techniques.

Changed

  • Hardware models.

    • Allow data source/destination regions in Resource to be non-DATA type.

    • Allow NodeRegion to be folded along the w dimension in a zig-zag manner.

  • Software models.

    • LoopBlockingScheme supports access forwarding and buffer sharing.

    • LoopBlockingScheme supports remote node buffers as data regions (non-data type data regions).

    • partition unit number of hops calculation supports access forwarding and buffer sharing.

    • DataLayout supports closest-first forwarding data transfer for access forwarding and buffer sharing.

    • Refactor NNDataflow and NNDataflowScheme to incorporate inter-layer pipelining.

[v1.5 -- v1.6] -- 2018-01-31

Added

  • Workload models.

    • Add EltwiseLayer.

    • Allow only concatenation of layers; summation of layers turns into an additional EltwiseLayer.

    • Add external layers to networks, which is external directly input data.

    • Add various LSTMs.

    • Add data_loops attribute with type DataDimLoops to each type of layer.

  • Hardware models:

    • Add DRAM region in Resource.

    • Consider array bus width and its impact on data multicast latency.

    • Consider DRAM access time due to bandwidth limit.

  • Software models.

    • Add choices for optimization goal: E(nergy), D(elay), or ED.

  • Software engineering.

    • Record search time.

    • Add utility IntRange for integer ranges.

    • Add HashableDict class.

Changed

  • Hardware models.

    • 2D memory type is changed to constant four node on the chip corners.

    • NodeRegion adds dist attribute for inter-node distance.

    • NodeRegion renames DATA enum to DRAM.

    • Limit to single source/destination data regions in Resource.

  • Software models.

    • Cost uses static/idle unit cost for all nodes instead of one node.

    • Scheduling breaks loop/part cost into op/access/noc/static cost.

    • Scheduling breaks cost tie using time, using a compare key function of SchedulingResult.

    • Add external occupancy to MapStrategy and merge into NestedLoopDesc; use it for partitioning occupancy.

    • FmapRange.beg_end() returns an IntRange instance if with a single attribute argument, or a list of IntRange otherwise.

    • Move partitioning scheme sub-FmapRange method, which is used to get partitioned fmap ranges, to PartitionScheme.

    • Move partitioning scheme projection, which is used to generate ofmap layout, to PartitionScheme.

    • DataLayout refactored: use PartitionScheme to replace FmapRangeMap.

    • partition module refactored: use new DataLayout class and new PartitionScheme methods.

    • SchedulingResult uses a combined OrderedDict to replace dict_loop and dict_part.

    • In partitioning schemes, each partitioning must fully utilize one dimension before starting the other, except for fmap partitioning.

  • Software engineering.

    • Change NNDataflowScheme node-time product interface to explicitly be static cost.

    • Improve method names:

      • Remove DataDimLoops.data_cnt.
      • Change NodeRegion.node_iter to NodeRegion.iter_node.
      • Change Network method names to distinguish layer and layer name.

Fixed

  • Output dict PartitionScheme format fix.

  • idivc with inf arguments.

  • Integer division // vs. /.

  • ITCN access calculation for unit pass in MapStrategy.

  • FmapRange comparison.

  • Unit nhops calculation for filter data uses DRAM region.

  • Unit nhops calculation considers nodes with both non-empty ifmaps and ofmaps.

  • Replication size when underutilizing PE arrays.

[v1.4 -- v1.5] -- 2017-08-18

Added

  • Workload models.

    • Network method to return next layers.

    • Network uses None in previous/next layers for the input/output layers.

    • Network methods to return the first/last layers.

    • Add batch size argument to layer fmap size methods.

    • Add default filter size to FCLayer.

    • Add DataDimLoops class to denote loops that are dimensions of a data category.

    • Add neural neworks: MLP-L/M/S from PRIME ISCA 2016.

  • Software models.

    • Add statistic properties to SchedulingResult.

    • Add NNDataflowScheme class for overall NN dataflow.

  • Software engineering.

    • Add utilities to LoopBlockingScheme class.

    • Add negative operation to PhyDim2.

    • Add default arguments to Option.

  • Test.

    • Add unit tests.

Changed

  • Workload models.

    • Relax __len__ of Network to work before setting input layer.

    • Allow different height and width for filters in ConvLayer.

  • Hardware models:

    • Upgrade node dimensions to node region in Resource. The origins of node region and memory regions are all absolute.

    • Add type attribute to NodeRegion to differentiate processing and data node regions in Resource.

    • Change default cost of the NoC hop traversal.

  • Software models:

    • Add loop index generator to LoopBlockingScheme class.

    • PE array mapping for LocalRegionLayer reduces regfile size.

    • Loop blocking scheme result stats change from one node to all nodes.

    • Move partition occupation into LoopBlockingScheme constructor.

    • Move LoopBlockingScheme verification to tests.

    • Improve the workload partitioning for loop blocking exhaustive search.

    • Merge loopcnt attribute of NestedLoopDesc to a tuple.

    • Change LoopBlockingScheme interface for blocking factors and loop orders.

    • Loop blocking exhaustive search introduces regularized schemes and suboptimal schemes, to enable more skips. Also restrict the skips to CONV layer.

    • Refactor loop blocking bypass solvers, and restrict it to CONV layer.

    • Use row-stationary mapping to LocalRegionLayer, and merge with that of ConvLayer.

    • Generalize LoopBlockingScheme access model for arbitrary data loops.

    • Skip equivalence when generating PartitionScheme.

    • Check ifmap layout against layer parameters in Scheduling.

    • Add number of nodes to scheduling result.

    • Add type attribute to DataLayout to denote the type of the reside region.

    • Add guarantee to generate PartitionScheme.

  • Software engineering.

    • Lazily evaluate loop blocking stats.

    • Use rich comparison instead of __cmp__.

    • Convert RuntimeError exceptions to assertions.

    • Define __repr__ for class stringify, and remove StringifyClass.

    • Move map strategy class into NNDataflow constructor.

    • Reorganize package structure.

    • Use lower-case name for all modules.

    • Add local version number to output.

Fixed

  • Output data fetch count.

  • Error types and message typos.

  • FmapRange comparison: overlapping ranges cannot compare.

  • Multiple bugs fixed in Util.

  • Multiple bugs fixed in PartitionScheme.

  • Use GBUF unit access for DRAM when bypassing GBUF.

  • Partitioned ifmap range for LocalRegionLayer.

  • Clarify ITCN accesses to be number of individual transfers to each REGF.

  • Partition unit number of hops calculation ignores zero-sized data ranges.

[v1.3 -- v1.4] -- 2017-05-18

Added

  • Software models.

    • Partition schemes.
      • Input partitioning: partition different input fmaps (channels).

  • Explorers and solvers:

    • Loop blocking exhaustive search skips more equivalent schemes.
      • Adjacent same loops in different hierarchy levels.

  • Software engineering

    • Verbose mode.

Changed

  • Software models:

    • Loop blocking.
      • Avoid initial zero-value fetch for output data.

  • Software engineering.

    • Use a single global argument parser.

    • Introduce ContentHashClass.

Fixed

  • FmapRange comparison.

  • Map strategy bug when filters are folded.

[v1.2 -- v1.3] -- 2017-05-16

Added

  • Workload models:

    • Support loops: ifmap channel loop, ofmap channel loop, batch loop.

  • Software models:

    • Loop index generator for different loop blocking schemes.

    • Debug mode:

      • Verification of the loop blocking access model.

  • Explorers and solvers:

    • Loop blocking exhaustive search skips equivalent schemes.

Changed

  • Software models:

    • Loop blocking data buffer and reuse models.
      • Loop orders now also consider the order of batch loop.
      • Change the model for trivial loops (with blocking factor 1).

[v1.1 -- v1.2] -- 2017-05-10

Added

  • Explorers and solvers.

    • Performance improvements.
      • Add loop blocking scheme cache in Scheduling.
      • Use a single Scheduling instance for all same layers.

  • Software engineering

    • Class instance used as dict key.

      • Add value-based equality and hash to Layer.
      • Add value-based equality and hash to PartitionScheme.

    • Add version number to output json dump.

Changed

  • Explorers and solvers:

    • Better formatting of verification results against Eyeriss.

  • Software engineering

    • Replace numpy for better performance.

    • Move multiprocessing into loop blocking exploration for better performance scaling.

[v1.0 -- v1.1] -- 2017-05-04

Added

  • Workload models.

    • Network: a DAG of layers, rather than a linear pipeline.

    • New layer types: pooling layer (local region layer).

    • Enforce layer chaining has matched data size.

    • New neural network: GoogLeNet.

  • Hardware models.

    • NodeRegion.
      • Used to denote memory regions, i.e., relative positions and sizes of memories to the computation node NoC.
      • Support 2D memories, which are on the edges of the chip.

  • Software models.

    • FmapPosition and FmapRange: a position and a range in batched fmaps.

      • FmapRangeMap: efficient map structure of FmapPosition type.

    • DataLayout: describes the layer i/ofmap data layout.

      • Use a FmapRangeMap to map each data element to the stored node.

    • Partition schemes.

      • Batch partitioning: partition input data within a batch.

  • Explorers and solvers.

    • SchedulingResultDict: store layer scheduling results of a network.

    • More checks to enforce the schedules have the correct number of operations as the given workloads.

Changed

  • Workload models.

    • Update all network structures to include pooling layers.

  • Software models.

    • Allow different partitioning factor along height and width of a fmap, i.e., allow different height and width sizes of the partitioned fmap.

  • Explorers and solvers.

    • NNDataflow: new top-level class.

  • Software engineering:

    • Significant code refactoring to improve modularity.

      • More classes, e.g., MapStrategy, LoopBlockingScheme, Scheduling.

    • Code style lint.

    • Update option names to be more uniform.

    • Standardize class stringify.

Deprecated

  • Option --hybrid_partition2d, now is --hybrid_partition.

Fixed

  • Use of map() function in PhyDim2.

  • Name of namedtuple sublasses.

  • Structure configuration of ResNet152.

[init -- v1.0] -- 2017-01-21

Added

  • Workload models:

    • Two layer types: convolutional layer and fully-connected layer.

    • Supported neural neworks: AlexNet, VGG, VGG19, ZFNet, ResNet152.

    • Supported data categories: ifmaps, ofmaps, and weights.

  • Hardware models:

    • 2D Network-on-Chip (NoC) on the PE array (node) level.

    • 2D PE array on the PE level.

    • Memory hierarchy:

      • regf: register file in a PE.
      • itcn: interconnect between PEs in an array.
      • gbuf: global buffer of an array.
      • dram: main memory.

    • Cost (energy) of computation operations (MAC, etc.), memory hierarchy accesses, NoC hop traversals, and static leakage.

  • Software models:

    • Eyeriss Row-Stationary mapping to PE array (Chen et al., ISCA 2016).

    • Loop blocking schemes over ifmap channel, ofmap channel, and batch loops.

      • Loop reordering: exchange loop order.
      • Loop blocking: split a loop into multiple ones.

    • Partition schemes to split the workload of a layer to different nodes.

      • Fmap partitioning: partition height/width of a fmap.
      • Output partitioning: partition different fmaps (channels).

  • Explorers and solvers:

    • Per-layer schedule exploration.

    • Exhaustive search loop blocking schemes and partitioning schemes.

    • Analytically solve bypass loop ordering (Gao et al., ASPLOS 2017).

    • Naive partitioning scheme (Kim et al., ISCA 2016).

  • Software engineering

    • Support multi-process parallel processing.