General Optimization Issues

1
General Optimization Issues

• M. Smith

2
To be tackled today

• Most optimized TigerSHARC instruction
• Integer and float
• Systematic optimization procedure
• SISD and SIMD modes
• Exercises

3
Most optimized SIMD Floating point(32-bit)TigerSH
ARC instruction

• xR30 CB Qj0 4 yR30 CB Qk0 4
  xyFR4 R5 R6 xyFR7 R8 R9, FR10 R8 -
  R9
• xR30 CB Qj0 4 / Fetches 4 values on J
  BUS into x compute registers XR3, XR2,
  XR1, XR0 Increments J register and
  adjusts for circular buffer
  operation /
• yR30 CB Qk0 4 / Fetches 4 values on J
  BUS into x compute registers XR3, XR2,
  XR1, XR0 Increments J register and
  adjusts for circular buffer
  operation /
• xyFR4 R5 R6 / Two multiplications XFR5
  XFR6 and YFR5 YFR6 /
• xyFR7 R8 R9, FR10 R8 - R9 / Two
  additions XFR8 XFR9 and YFR8 YFR9 AND Two
  subtractions XFR8 - XFR9 and YFR8 - YFR9 /
• / Same register must be used either side
  of and operators /

4
Most optimized SIMD Integer (short)(16-bit)TigerS
HARC instruction

• xR30 CB Qj0 4 yR30 CB Qk0 4
  R76 R54 R32 xySR98 R76R10,SR1110
  R76-R10
• xR30 CB Qj0 4 / Fetches 4 values on J
  BUS into x compute registers XR3, XR2,
  XR1, XR0 Increments J register and
  adjusts for circular buffer
  operation /
• yR30 CB Qk0 4 / Fetches 4 values on J
  BUS into x compute registers XR3, XR2,
  XR1, XR0 Increments J register and
  adjusts for circular buffer
  operation /
• xyR76 R54 R32 / Eight multiplications
  XR5.H XR3.H, and XR5.L XR3.L, XR4.H
  XR2.H, XR4.L XR3.L ditto YR /
• xySR98 R76 R10, R1110 R76 R10
  / Eight additions ???????
  AND Eight subtractions
  ????????????????? /

5
ExerciseWrite out the 16 operations performed

• xySR98 R76 R10, R1110 R76 R10
  / Eight additions ???????
  AND Eight subtractions
  ????????????????? /
• Now do a sideways add on xySR98 and get a value

6
Steps to optimize

• Get the algorithm to work in C
• Determine how much time is available
• If Timing already okay quit
• Determine maximum number of each type of
  operation (add, subtract, multiple, memory
  fetches)
• Divide the calculated maximum by the number of
  available resources for that type of operation
• The largest division result is the in theory
  number of cycles needed for the algorithm
• If that minimum time is more than 100 of the
  time available find a new algorithm
• If that minimum time is less than 40 of the time
  available perhaps you can optimize the code to
  meet the speed requirements

7
Code optimization 32 bit integersor 32-bit
floats
2 SIZE additions 2 SIZE Memory fetches If
done correctly Can do 2 additions AND 2 memory
fetches each cycle Therefore optimum isSIZE
cycles IFF can find all optimizations
8
Code optimization 32 bit integersor 32-bit
floats
2 SIZE additions 2 SIZE Memory fetches Left
fetched on J-bus And done in X-compute Right
fetched on K-bus And done in Y-compute
9
16-bit integers (short int) might be okay in some
circumstances
2 SIZE additions 2 SIZE Memory fetches If
done correctly Can do 8 short additions AND 32
short memory fetches each cycle Therefore
optimum isSIZE / 4 cycles IFF can find all
optimizations
10
FIR optimization
SIZE additions SIZE multiplications SIZE 2
memory fetches 2 additions, 2 multiplications
and 8 fetches per cycles Should be able to do it
in SIZE / 2 cycles
11
FIR optimization
SIZE additions SIZE multiplications SIZE 2
memory fetches Fetch 2 values along J-bus into
XA and YA compute Fetch 2 coefficients along
K-bus into XB and YB compute
12
Need a systematic approach to handling the
optimization of code

• Get the C code to work
• Rewrite code in simplest format one operation
  per line
• Recommend rewrite code using register names
• Unwrap the loop start with twice
• Rewrite the second part of the loop using
  different register names avoids setting up
  unexpected dependencies
• Overlap the first and second parts of loops
• Rearrange start-up and ending code

13
STAGE 1Get the C code to work
14
Need a systematic approach to handling the
optimization of code

• Get the C code to work
• Rewrite code in simplest format one operation
  per line
• Recommend rewrite code using register names
• Unwrap the loop start with twice
• Rewrite the second part of the loop using
  different register names avoids setting up
  unexpected dependencies
• Overlap the first and second parts of loops
• Rearrange start-up and ending code

15
Stage 2 Rewrite in simplest format
Note naming convention Single operation per
line Note other changes
16
Need a systematic approach to handling the
optimization of code

• Get the C code to work
• Rewrite code in simplest format one operation
  per line
• Recommend rewrite code using register names
• Unwrap the loop start with twice
• Rewrite the second part of the loop using
  different register names avoids setting up
  unexpected dependencies
• Overlap the first and second parts of loops
• Rearrange start-up and ending code

17
Step 3 -- Unwrap the loop
Again Note naming convention
18
Need a systematic approach to handling the
optimization of code

• Get the C code to work
• Rewrite code in simplest format one operation
  per line
• Recommend rewrite code using register names
• Unwrap the loop start with twice
• Rewrite the second part of the loop using
  different register names avoids setting up
  unexpected dependencies
• Overlap the first and second parts of loops
• Rearrange start-up and ending code

19
Step 4Overlap the first and second parts of
loops
Note The C code goes no faster, but using
this format for translating into parallel
assembly code will Step 1 -- 4 N Step 3 8
(N / 2) 2 Step 4 6 (N / 2) 2
20
Need a systematic approach to handling the
optimization of code

• Get the C code to work
• Rewrite code in simplest format one operation
  per line
• Recommend rewrite code using register names
• Unwrap the loop start with twice
• Rewrite the second part of the loop using
  different register names avoids setting up
  unexpected dependencies
• Overlap the first and second parts of loops
• Rearrange start-up and ending code

21
Step 5A - Rearrange start-up and ending code
Software Pipeline Move first read outside Need
to add extra read at the end of the
loop Timing 2 (N/2 1) 6 Need to adjust
loop start (Is it done correctly? Are we
one-out) CAUTION NEED TO FIX
22
Step 5B - Rearrange start-up and ending code
Can now parallel additional adds and memory
fetches Note loop still in error
23
Exercise -- Get the loop control correct
24
Exercise 1 -- Get the loop control correct
BUFFER_SIZE 1 BUFFER_SIZE 2 BUFFER_SIZE
4 BUFFER_SIZE 5 BUFFER_SIZE 8 BUFFER_SIZE
128
25
Exercise 2 -- Rewrite the code when it is known
that BUFFER_SIZE 127
BUFFER_SIZE 1 N 2 N 4 N 5 N 8 N 128
26
Code to this point is SISD parallel optimization

• SISD single instruction single data
• Using X_compute block and J memory bus
• Next stage SIMD single instruction multiple
  data
• Using X_compute block and J memory bus for left
• Using Y_compute block and K memory bus for right
• Will need similar but different code when you are
  doing FIR in Lab. 3

27
Exercise 3 -- BUFFER_SIZE 128Rewrite so that
X and Y ops done together
BUFFER_SIZE 1 N 2 N 4 N 5 N 8 N 128
28
Exercise 4 -- BUFFER_SIZE 128Rewrite so that
expect no data dependency stalls
BUFFER_SIZE 1 N 2 N 4 N 5 N 8 N 128
29
To be tackled today

• Most optimized TigerSHARC instruction
• Integer and float
• Systematic optimization procedure
• SISD and SIMD modes
• Exercises