HiFi2 Lab

T ensilica HiFi 2 Audio Engine Lab

Task 1. Installing the HiFi Configuration-

1.Start up Xtensa Xplorer DE (Diamond Edition) or Xtensa Xplorer CE (Configurable

Edition)

2.If you are using Xplorer CE (Configurable Edition, not Diamond Edition), install the

DC_330HiFi Diamond configuration:

a.From the File menu, select New..., then Xtensa configuration.

b.Select Install a new build from a downloaded bundle as the new

configuration, then click Next.

c.Browse to and select DC_330HiFi_win32.tgz, click Add Build, and then

click Finish.

Task 2. Importing the dotprod source code project

1.From the File menu, select Import..., Xtensa Xplorer Workspace , and then HiFi2-

lab.xws.

2.Open dotprod.c and view golden_dotprod().

Note that we are doing regular 32x32 to 64-bit multiplications

Question 1.Why are data0[i] and data1[i] cast to (long long) before being multiplied?

Note that we are using INTEGER arithmetic, not fractional arithmetic

Task 3. Optimizing dot product algorithm for HiFi 2 Audio Engine

For parts 1a and 1b, we'll load and multiply only one pair of elements each iteration. For parts 2a and 2b, we'll load and multiply two pairs of elements each iteration.

1a) Do one multiplication per iteration, with wrong data type (ae_p24s) 1.Go to audio_engine_dotprod1a()

For 1a and 1b, we'll load one element at a time, which is why we chose to use ae_p24s pointers:

ae_p24s *p0, *p1;

Advantages of loading one element at a time:

Don't need to guarantee that data0 and data1 are 8-byte aligned

Don't need to guarantee that n is even

Disadvantage

Can only do one multiplication per iteration

Note that d0 and d1 are declared as ae_p24s. This seems natural, given that p0 and p1 are pointers to that type, but it will lead to problems (as we shall see)

https://www.360docs.net/doc/3714949246.html,plete the TODO tasks:

Set p0 and p1 to point to input data arrays

Load the next element from each array into d0 and d1

Multiply d0 and d1 together and add to sum

Store sum in result

3.Build the dotprod project for DC_330HiFi/Debug and make sure there are no errors

or warnings

4.Run, and make sure that part 1a passes (it's okay if parts 1b, 2a, 2b fail)

1b) Do one multiplication per iteration, with correct data type

(ae_p24x2s)

1.Go to audio_engine_dotprod1b()

Note that d0 and d1 are declared correctly this time:

ae_p24x2s d0, d1;

https://www.360docs.net/doc/3714949246.html,plete the TODO tasks:

Set p0 and p1 to point to input data arrays

Load the next element from each array into d0 and d1

Multiply d0 and d1 together and add to sum

Store sum in result

Question 2.Which element-selection suffix should you use: .HH / .HL / .LH / .LL ? 3.Build the dotprod project for DC_330HiFi/Debug and make sure there are no errors

or warnings.

4.Run, and make sure that parts 1a and 1b pass (it's okay if parts 2a, 2b fail).

Profiling the Code

1.Once all the code is running and the first two results PASS, set your build target as

Release, and profile the code!

Question 3.How many total cycles (including children!) for the following functions: golden_dotprod():

audio_engine_dotprod1a():

audio_engine_dotprod1b():

Question 4.Why does (1a) take so many more cycles than (1b)?

2a) Do two multiplications per iteration, using two single MACs

1. Go to audio_engine_dotprod2a(). For 2a and 2b, we'll load two elements at a time, which is why we chose to use ae_p24x2s pointers:

ae_p24x2s *p0, *p1;

Advantages of loading two elements at a time:

Can two multiplications per pair of loads

Disadvantages:

Need to guarantee that data0 and data1 are 8-byte aligned

Need to guarantee that n is even (or handle last product separately)

https://www.360docs.net/doc/3714949246.html,plete the TODO tasks:

Set p0 and p1 to point to input data arrays (you can assume that data0 and data1 are 8-byte aligned)

Load the next element pair from each array into d0 and d1

Multiply d0.H and d1.H together and add to sum

Multiply d0.L and d1.L together and add to sum

Store sum in result

3.Build the dotprod project for DC_330HiFi/Debug and make sure there are no errors

or warnings.

4.Run, and make sure that parts 1a, 1b, 2a pass (it's okay if part 2b fails)

2b) Do two multiplications per iteration, using dual MAC

1. Go to audio_engine_dotprod2b()

2. Complete the TODO tasks:

Set p0 and p1 to point to input data arrays (you can assume that data0 and data1 are 8-byte aligned)

Load the next element pair from each array into d0 and d1

Multiply the elements of d0 and the elements of d1 with a dual MAC

Store sum in result

3. Build the dotprod project for DC_330HiFi/Debug.

4. Make sure there are no errors or warnings.

5. Run, and make sure that all parts (1a, 1b, 2a, 2b) pass.

Profiling the code:

1. Once all the code is running and results all PASS, set your build target as Release, and

profile the code!

Question 5.How many total cycles (including children!) for the following functions: golden_dotprod():

audio_engine_dotprod1a():

audio_engine_dotprod1b():

audio_engine_dotprod2a():

audio_engine_dotprod2b():

Question 6.Why does (2a) take about the same number of cycles as (2b)?

Answers / Solutions

Question 1.Why are data0[i] and data1[i] cast to (long long) before being multiplied Answer:Otherwise, the result of the multiplication would be truncated to 32 bits.

Question 2.Which element-selection suffix should you use: .HH / .HL / .LH / .LL ? Answer:It doesn't matter -- any of them will work, because loading from an ae_p24s pointer fills both halves (H and L) of the p24x2s register with the same

value,

Question 3.How many total cycles (including children!) for the following functions: Answer:golden_dotprod(): 50426 = 9486 + 40940

audio_engine_dotprod1a(): 3092

audio_engine_dotprod1b(): 1042

Question 4.Why does (1a) take so many more cycles than (1b)?

Answer:The compiler has no automatic way to convert from p24s to p24x2s

variables in registers, so it must do the conversion via memory (extra store

and load)

Question 5.How many total cycles (including children!) for the following functions: golden_dotprod(): 50426 = 9486 + 40940

audio_engine_dotprod1a(): 3092

audio_engine_dotprod1b(): 1042

audio_engine_dotprod2a(): 542

audio_engine_dotprod2b(): 534

(These cycle counts were obtained with the RB-2006.0 version of XtensaTools.)

Question 6.Why does (2a) take about the same number of cycles as (2b)? Answer:Because we need two loads (two pairs) for every (dual) multiply

minimum two instructions in inner loop (one load per instruction) with

two loads and two MULs, still only two instructions (after

unrolling/software pipelining)