SHA1 optimization proposal, usable ?

Hi, I am relatively new to hashcat forgive for any foolishness


I was looking through the code to find ways to speedup recovering passwords for WPA2.


I found some, but is it worth it?


The second one is that I see a difference in speedups between NVIDIA/OpenCL (org compared to optimized both with OpenCL) and NVIDIA/CUDA (org compared to optimized both with CUDA).


I cannot explain it. I expected roughly the same speedups.



Result (showing only the relevant speedups from the short benchmark):




WPA-EAPOL-PBKDF2 (2500):


* without CUDA Toolkit installed: 4.14%


* with CUDA Toolkit installed: 2.11%




DPAPI masterkey file v1(15300):


* without CUDA Toolkit installed: 5.01%


* with CUDA Toolkit installed: 1.09%






- Do these results justify a pull request?


- Why are the achieved speedups for CUDA lower compared to OpenCL ?





More details below for the interested readers ..

---

Speedup stats  for other hashmodes in the short benchmark:




Without CUDA Toolkit installed:

• Worst -0.28%
  
• Best 0.77%
  
• Avg 0.10%
  
• Stdev 0.25%
  

With CUDA Toolkit installed:

• Worst -0.58%
  
• Best  0.49%
  
• Avg  -0.16%
  
• Stdev  0.24%
  

Optimization:



SHA1_transform(), SHA1_transform_vector() in inc_hash_sha1.cl




Based on: Exploiting an HMAC-SHA-1 optimization to speedup PBKDF2




Trade-off:


more memory (80 x u32_t instead of 16 x u32_t), less execution time (number of zero-optimizable xors increases by a)not destructively updating w0_t .. wf_t and b) by maximizing the righthand side usage of w0_t..wf_t)




Implemented:


basically equation 14 of the article, leaving still a general function which can be better optimized by the compilers with respect to zero based operations ( a XOR 0 = a, a + 0 = a). The remainder of the article presents some more execution optimizations and a major memory optimization, but these already assume zero based operation optimizations, making the resulting function less general. I still need to give that part of the article a second look.



Measurement method:

• i7-7700HQ, GTX-1050 (2GB) , low (laptop) specs I know, but my better one was busy.
  
• Based on v5.1.0-1484-gbfd95d42, Ubuntu 18.04, see below for CUDA/OpenCL versions
  
• Using only GPU acceleration.
  
• Benchmark command:
  $ for i in {0..10} ; do  hashcat -b --machine-readable ; done > ../result.txt
  
• From the 11 results discard the first one (warmup), average the remaining 10 per mode/run.
  
• Do this with:
  - original code
  - optimized code
  
• Do the above on a system  (is this usefull?):
  - without CUDA Toolkit installed
  - with CUDA Toolkit installed
  
• Compute speedup as: percentage(1-(average from org/average from optimized))
  

New bee (bzzz) questions:

The speedups with CUDA Toolkit do not match those without the toolkit.

I assumed that the same .cl code is used for both with- and without CUDA Toolkit, is this correct ?

Is this explainable ? What am I missing here?







--  CUDA / OpenCL info ----







hashcat --backendinfo (with CUDA Toolkit installed):



hashcat (v5.1.0-1484-gbfd95d42) starting...







CUDA Info:



==========







CUDA.Version.: 10.2







Backend Device ID #1 (Alias: #2)



  Name...........: GeForce GTX 1050



  Processor(s)...: 5



  Clock..........: 1493



  Memory.Total...: 2000 MB



  Memory.Free....: 1943 MB







OpenCL Info:



============







OpenCL Platform ID #1



  Vendor..: NVIDIA Corporation



  Name....: NVIDIA CUDA



  Version.: OpenCL 1.2 CUDA 10.2.95







  Backend Device ID #2 (Alias: #1)



    Type...........: GPU



    Vendor.ID......: 32



    Vendor.........: NVIDIA Corporation



    Name...........: GeForce GTX 1050



    Version........: OpenCL 1.2 CUDA



    Processor(s)...: 5



    Clock..........: 1493



    Memory.Total...: 2000 MB (limited to 500 MB allocatable in one block)



    Memory.Free....: 1920 MB



    OpenCL.Version.: OpenCL C 1.2



    Driver.Version.: 440.33.01