SIMT challenge

[daydreamer]

I think I return to SIMT task I earlier solved with SIMD + LUT, and follow the student rules to solve it with SIMT
First a way to check with api or user enter number of cores on their computer
1: generate x million random numbers
Probably call rand in a loop
2: check if prime number
3: add those prime numbers together

Main start x number of threads,sending 1,2,3 to thread
On 16 cores cpu, 1,2,3 is choice depending on which is most computing expensive and least computing expensive
Each thread starts with Switch/case
Case 1
Invoke random
Case 2
Invoke primes
Case 3
Invoke sum

Option is return when finished task 1 and 2 and thread joins helping other threads that aren't finished yet

Other ideas is
Two threads random numbers are 1 thread starts from beginning of array and second thread starts from end of array, stops when they reach same adress

[NoCforMe]

I don't get it; what is the point of this whole exercise?

[daydreamer]

When even old computers have 4 or more cores, it's easier with several threads coding standard x86 gp reg code than exotic SIMD opcodes and more flexible able to run many different Algos /proc on each core than 128 bit SSE 4 x32 bit of same algo

I have 4 cores on my newest and want try learn use all 4

[Siekmanski]

I don't get it; what is the point of this whole exercise?

A challenge to excel in programming by using your logical and analytical part of your brain. 

[zedd151]

All I see here is a word salad. 

And only a 'concept' of an 'idea'.   

[daydreamer]

I have no scalar code prime testing code yet,sieve and divide, I only have experimental SSE : real4's using many divps using 2.0 to around 300.0 with a bug :
its not capable of detect all those lower 2.0-300.0 primes and limited to REAL4 range of ca 9 million

 

[sinsi]

Some older (but still recent) CPUs would reduce the clock if full AVX was used on multi cores because of the heat.
With Intel CPUs you also need to pick a CPU that's not hyper-threaded.

Multithreading isn't easy, that's why most apps don't use it.
Also, CPU multithreading is a whole lot different to using an OS's thread functions.

[daydreamer]

Sinsi, I don't want to end up in the future stuck in only 1 core coding, 1 of 192 cores = 0.5% cpu usage : Hall of shame 

Gonna Start code/debug now
Ideas for several different proc's that might be used on several different threads / cores
 

Code Select Expand

SIMT exercise
Xorps xmm7,xmm7
Movaps sums,xmm7

1:
Movaps xmm0,start
Mulps xmm0,seeds
Addps xmm0,extra
Movaps result, xmm0
2b :
Movaps xmm0,result
Movss xmm1,[ebx*4+prime list ]
Shuffle to all 4 floats
Subps xmm0,Xmm1
Comiss xmm0,zeros
Jne
Shift xmm0, 32 bits
Comiss xmm0,zeros
Shift xmm0, 32 bits
Comiss xmm0,zeros
Shift xmm0, 32 bits
Comiss xmm0,zeros
Add ebx,1 ; check next number
Add ecx,1
Cmp ecx,length prime list
Jne l2 b
2 c :
Prime list Db,0,0,2,3,5,7,11,13,...
Mov ecx,number of random numbers
Mov Al,random byte
Xlat prime list
Add edx,eax
Dec ecx
Jne l2 c

2:

L2:
Movaps xmm1, result
Divps xmm1,[prime list+ebx]
subps xmm1,%
Comiss xmm1,zeros (*4 comiss)

3:
If primes addps sums,primes
Haddps sum,sums

Proc Init
Invoke malloc ( if 64 bit mode,possible to malloc huge 4+ GB prime LUT)
Open file,"prime LUT.bin"
If eax!=true Invoke Create LUT
Create file, "prime LUT.bin"
Block write LUT
Close file
Else
Block read file
Close file