1/33
Looks like no tags are added yet.
Thread level parallelism
multicore processor and multithreading take advantage of _ _ _
Within a program
exploit thread-level parallelism is
Across application
exploit process-level parallelism is
Duplicate core, add extra logic
multicore, relatively simple extensions
Add storage, Associated Pipeline
multithreading, relatively simple extensions
ILP
executes multiple independent instructions from a single instruction stream at the same time
Pipeline, Superscalar design
ILP using
Instruction Level Parallelism
meaning of ILP
DLP
applies the same operation to multiple data items simultaneously
SIMD, GPUs
DLP using
Single Instruction Multiple Data
meaning of SIMD
Graphics Processing Unit
meaning of GPU
Data Level Parallelism
meaning of DLP
Independent Work, Process Parallel
prerequisite of data-level parallelism
Data level Parallelism
vector processors explicitly exploit _ _ _
Large Vector Register
vector processors gather data from memory into _ _ _
Register Register Operation
essentially performing many _ _ _ with the same opcode
Energy Efficient Computation
for vectorizable data, vector processors (1)
Hide memory latencies
for vectorizable data, vector processors (2)
Vector Register
holds a sequence of elements
Vector Instruction
applies one operation to corresponding elements
Vector Length
number of active elements for this instruction
Vector Lanes
parallel/pipelined datapaths that process elements
Vector Load/Store
moves contiguous or strided data to/from vector registers
Vector Registers, Scalar Register
Vector functional units (FUs) are fed by
Vector Register
provide independent data to operate on
Scalar Register
provide additional data or memory addresses
Fully Pipelined
vector FUs are _ _
SIMD
execution brings a (constrained) form of vector processing to general-purpose CPUs
32 bit register
scalar execution would put each value in its own _ _ _
Vector Register
SIMD packs them into a _ _
256 bit vector of 16 x 16 bit data
vector register
GPU
is originally developed as specialized hardware for graphics computation
Usage Patterns
GPU hardware evolution closely tied to evolution in _ _