[SYSTEMDS-1038] Add support for aggregate unary operations on GPU

• Type: Sub-task
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Fix Version/s: SystemML 0.11
• Component/s: Compiler, Runtime
• Assignee: Unassigned
• Reporter: Niketan Pansare
• Created: 12/Oct/16 22:57
• Updated: 12/Oct/16 22:58
• Resolved: 12/Oct/16 22:58

Description

1. Update UnaryOp to enable GPU for these instructions.
2. Add AggregateUnaryGPUInstruction and invoke kernels corresponding to https://github.com/apache/incubator-systemml/blob/master/src/main/java/org/apache/sysml/runtime/matrix/data/MatrixBlock.java#L4451

Note nrow, ncol and length should always be a CP instruction. Also, we can break this into subtasks for sparse if necessary.

The opcodes related to this tasks are: uak+, uark+, uack+, uasqk+, uarsqk+, uacsqk+, uamean, uarmean, uacmean, uavar, uarvar, uacvar, uamax, uarmax, uarimax, uacmax, uamin, uarmin, uarimin, uacmin, ua+, uar+, uac+, ua*, uatrace, uaktrace and the operations of these are (with/without kahan) sum, trace, rowsum, colsum, sum of squares, row sum of squares, col sum of squares, cumsum, cumprod, cummin, cummax, min, max, min_index, max_index, mean, variance and product.

nakul02 This is a mega-task. Hopefully we can generalize our kernel implementation to make this easier.

Issue Links

• duplicates: SYSTEMDS-1039 Add support for aggregate unary operations on GPU