Reconfigurable Computing

1
Reconfigurable Computing (EN2911X,
Fall07) Lecture 09 RC Principles Software (2/4)
Prof. Sherief Reda Division of Engineering, Brown
University http//ic.engin.brown.edu
2
Behavioral code optimizing

• Tree-height reduction applies to arithmetic
  expression trees and strives to achieve the
  expression split into two-operand expressions to
  exploit parallelism
• The idea is to attempt to balance the expression
  tree as much as possible
• If we have n operations, what is the best height
  that can be achieved?
• Example x a b c d

b
c

b
c
a
d
a



d


x
x
3
Tree-height reduction
Exploiting the distributive property at the
expense of adding an operation

• x a(bcd e)

4
Constant and variable propagation

• Constant propagation consists of detecting
  constants operands and pre-computing the value of
  the operation with that operand. The result might
  a constant which can be propagated to other
  operations as input
• Example
• a 0 b a 1 c 2 b
• Replaced by ? a 0 b 1 c 2

• Variable propagation consists of detecting the
  copies of the variable and using the right-hand
  side in the following references in place of the
  left-hand side
• Example
• a x b a 1 c 2 a
• Replaced by ? a x b x 1 c 2 x

5
CSE and DCA

• Common Sub-expression Elimination (CSE) avoids
  unnecessary computations.
• Example
• a x y b a 1 c x y
• Can be replaced by ? a x y b a 1 c
  a
• Dead code elimination (DCA). Dead code consists
  of all operations that cannot be reached, or
  whose results is never referenced elsewhere.
• Example
• a x b x 1 c 2 x
• The first assignment can be removed if it is
  never subsequently referenced

6
Operator strength reduction code motion

• Operator strength reduction means reducing the
  cost of implementing an operator by using a
  weaker one (that uses less hardware / simpler
  and faster)
• Example
• a x2 b 3 x
• Replaced by ? a x x t x ltlt 1 b x t

• Code motion often applies to loop invariants,
  i.e., quantities that are computed inside an
  iterative construct but whose values fo not
  change from iteration to iteration.
• Example
• for (i 1 i lt a b)
• Replaced by ? t a b for ( i 1 i lt t)

7
Control-flow-based transformations

• Control-flow transformations are typically
  utilized to create more opportunities for
  data-flow transformations to be exercised
• Model expansion consists in flattening locally
  the model call hierarchy. Therefore the called
  model disappears, being swallowed by the calling
  one.
• A possible benefit is that the scope of
  application of some optimization techniques is
  enlarged yielding potentially a better circuit
• Example
• x a b y ab z func(x, y)
• where func(p, q) t q-ppq return t
• ? By expanding func, we get
• x a b y a b z a bab
• ? CSE x ab y ab z a-by

8
Conditional expansion

• A conditional construct can be always transformed
  in a parallel construct with a test in the end.
• Conditional expansion can increase the
  performance of the circuit when the conditional
  clause depends on some late-arriving signal.
• However, it can preclude the possibility of
  hardware sharing
• If (C) then xA else xB? compute A and B in
  parallel, x C ?AB

9
Loop expansion

• In loop expansion, or unrolling, a loop is
  replaced by as many instances of the body as the
  number of operations. The benefit is in expanding
  the scope of other transformations
• Example

x 0 for (i 1 i lt 12 i) x x
ai
10
Putting concepts into work Hardware acceleration
using custom instructions

• We studied the concepts HW/SW partitioning and
  code optimizations for high-level synthesis
• We will apply these concepts with the help of the
  Nios-II soft core processor
• Difference between Soft and Hard processors
• A hard processor is one that is implemented as a
  dedicated, predefined (hardwired) block
• As opposed to physically embedding a processor
  into the FPGA fabric, it is possible to configure
  a group of logic blocks to act as a soft
  processor
• What are the advantages and disadvantages of each?

11
The Nios II soft processor

• 32 bit soft processor from Altera
• 82 instructions
• Up to 256 custom instructions
• Optional multiply and divide depending on the
  flavor
• Comes in three flavors (number for Cyclone II
  implementations)
• Economy emphasizes minimum size 700 L.E and 17
  DMIPS.
• Standard performance/size balance 1400 L.E and
  54 DMIPS
• Fast best performance 1800 L. E and 92 DMIPS

12
Creating Nios based systems using SOPC and
program it using IDE
SOPC builder
Nios II IDE
13
Accelerating application within the Nios II
environment
custom instructions
Avalon component
Accelerator
peripherals
Avalon bus
Nios II processor
Memory (SRAM, or onchip)
14
Using customs instructions to accelerate
applications
15
HW assignment Accelerate C code to accelerate
palindrome detection

• A palindrome is a sequence of units a string
  that has the property of reading the same in
  either direction
• Examples
• Racecar
• Dennis sinned
• 425524
• HW is to write a C routine to detect whether a
  number is a palindrome or not then use it to
  write a C program to count the number of number
  palindromes between 0 and 1 billion. The count
  can be computed statically but the HW ask you to
  write a C program for the Nios II processor to
  compute the count using the routine you coded

16
HW assignment Accelerate C code to accelerate
palindrome detection

• After you write your program, report the runtime
  and then accelerate the program using custom
  instructions designed using the reconfigurable
  logic
• You are required to report the runtime before and
  after the acceleration. It will be also good to
  try your program on a general purpose workstation
  and report the runtime.
• You have to report the count of palindromes you
  found together with the runtimes. Here are my
  runtimes.
• Optional 2.4 GHz Xeon workstation 355 seconds
• Required Nios II (just software) 27400 seconds
• Required Nios II (software custom
  instructions) 105 seconds
• Grades 15/20 if you get all parts working
  correctly. 16/20 if your runtime is between
  500-1000 seconds, 17/20 if your runtime is
  between 200-500 and 18/20 if runtime is 100-200
  and 20/20 if runtime is lt 100