Processor Architecture Stream Arithmetic Logic Unit Programming - - PowerPoint PPT Presentation

Stream Programming

for high-performance computing

CS334 F09

• Prof. McGuire

Williams College

1

Processor Architecture

ALU Memory

Arithmetic Logic Unit 7 8 2

2

ALU Memory

Arithmetic Logic Unit 7

ALU

2 4 8 5 ?

ALU ALU ALU

Multi-Processor

3

Memory

Arithmetic Logic Unit 7 8

ALU ALU ALU ALU ALU

5

Stream Programming

2 4 8 9 3 5 6

4

Stream Primitives

• Map
• Reduce
• Scatter
• Gather

(map * ‘(1 2) (3 4)) ! ‘(3 8)

“kernel” function source destination

5

Map

• Apply a function to all

elements

• Matrix addition
• Audio adjustment

dst[i] = f(src[i])

(map * ‘(1 2) (3 4)) ! ‘(3 8)

6

Reduce (Filter)

• Drop elements that

fail a predicate

• Strip outliers
• Parse whitespace

(filter even? ‘(1 2 3 4)) ! ‘(2 4)

7

Reduce (Fold)

• Combine elements
• Summation
• Average
• Maximum

dst = f(src[0], src[1], ..., src[n]) (foldl + 0 ‘(1 2 3 4)) ! 10

8

Scatter (“Push”)

• Iterate over source,

compute destination index

• Radix, Insertion, Quick sort
• Hash table insertion
• Permutation
• Rasterization

dst[ f(i) ] = src[i]

9

Gather (“Pull”)

• Iterate over destination,

compute source index

• Merge sort
• Sparse matrix operations
• Ray tracing
• Database query

dst[i] = src[ f(i) ]

10

The Next Big Thing?

• OpenMP - Parallel C and Fortran extensions
• Chapel - Cray; streaming Fortran-ish
• X10 - IBM; streaming Java
• Fortress - Sun; Guy Steele’s Mathematica-Scheme
• HPC, ZPL, Cilk, Titanium, UPC, Cuda, CTM, Sh, ...

11