[SYSTEMDS-2065] Code generation for operator fusion v3

• Type: Epic
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Epic Name: Code generation for operator fusion v3
• Assignee: Unassigned
• Reporter: Matthias Boehm
• Created: 12/Jan/18 00:32
• Updated: 03/Aug/18 20:23

Description

This epic aims to make the existing code generation feature - introduced in SYSTEMML-1284 - production ready.

Apart from features for robustness, performance, and broader operation support this also includes the extension to generating native CPU and GPU operations, including deep-learning-specific operations.

Issue Links

• is related to: SYSTEMDS-2373 Codegen integration nary min/max
• relates to: SYSTEMDS-1284 Code generation for operator fusion

Sub-Tasks

• Candidate exploration w/ unknowns: In Progress
• Add support for convolution operations (forward/backward): Closed
• Add support for logical and bitwise logical operations: Closed
• Refactor codegen algorithm tests, incl new tests: Closed
• Add support for nary cbind operations: Closed
• Add support for ternary ifelse operations: Closed
• Add support for maxpooling operations (forward): Closed
• Add support for relu operations (forward/backward): Resolved
• Library of native vector operations: Open
• Native C/C++ template code generation: Open
• Native LLVM template code generation: Open
• Compiler utils for native code compilation/loading: Open
• Performance input/output transfers over JNI: Open
• Extended CSE for codegen plans: Open
• Extended rewrite framework for codegen plans: Closed
• Codegen compiler extension for matrix multiply chains: Open
• Add support specific datagen operations (seq, matrix): Closed
• Improve codegen optimizer (pruning effectiveness): Closed
• Add missing algorithm tests to codegen testsuite: Closed
• Configurable codegen cost model weights: Open
• Runtime adaptation of code model weights: Open
• Basic GPU codegen framework and integration: Open
• GPU code generation for dense inputs: Open
• GPU code generation for sparse inputs: Open
• GPU code generation for compressed inputs: Open
• Extended codegen optimizer (CPU/GPU placement): Open
• Add support for bias_add and bias_mult: Closed
• Add vector support for ternary codegen ops: Closed
• Support for sparse vector intermediates in row templates: Open
• Performance cell operations over compressed matrices: Closed
• Incorporate sparsity in lower bound costs: Closed
• Extend codegen algorithm tests for heuristics FA and FNR: Open
• Add support for nary rbind operations: Closed
• Update janino version from 3.0.0 to 3.0.8: Resolved
• Add support for prod/colProds/rowProds: Open
• Determine min number of temporary row vector intermediates: Closed
• Support rowMeans in codegen row templates: Closed