Apple HPC Tools

1
Apple HPC Tools

• Chris Mueller
• July 6, 2004

2
Altivec

• http//developer.apple.com/hardware/ve
• SIMD, 32x128-bit registers, 160 opcodes
• G5 has 2 units vector permute and vector ALU
  along with a streaming prefetch unit

3
Altivec Programming

• -faltivec
• C programming model
• Variables vector lttypegt varname
• Functions vec_ltopgt(arg, )
• Caveats
• Data must be 16 byte-aligned
• Its WYSIWYG, the compiler wont rearrange code
• However, the compiler will add loads/stores as
  needed within vector function calls

4
Example
void VecAdd(unsigned char data, long len,
unsigned char result) // Diagnol sum of
all the values in data long i 0 vector
unsigned char score, score1, score2, vperm,
newsum newsum vec_splat_u8(0) //
create a constant for(i 0 i lt len - 16
i) // Load each vector if((i
0x0000000f) 0) // aligned case
score vec_ld(0, (datai)) else
// unaligned case
score1 vec_ld(0, (datai)) score2
vec_ld(16, (datai)) vperm
vec_lvsl(0, (datai)) score
vec_perm(score1, score2, vperm)
newsum vec_add(score, newsum)
vec_st(newsum, 0, result) // aligned
store return
5
Performance

• Standard sum adds up all values in an array
• Vector sum performs a diagnol sum of all
  vectors in an array
• Register sum loads one value in to the register
  and adds it to itself repeatedly

6
Shark

• /Developer/Applications/Performance
  Tools/CHUD/Shark
• Apples main (high) performance measuerment tool
• Samples all running applications and reports
  hotspots and offers optmization suggestions
• Nifty source/asm viewer
• Compile with
• -g -DCOPY_PHASE_STRIPNO

7
amber/simg5/scrollpv

• Cycle accurate trace (amber), simulation (simg5)
  and visualization (scrollpv)
• Usage
• Warning amber turns off one CPU. Prematurely
  killing amber can leave you with one CPU!!!

amber -I -x 5000 ltprogramgt simg5
thread_001.tt6e 5000 100 1 simg5 -p 1 -b 1 -e
5000 scrollpv -pipe trace_001.pipe -config
trace_001.config