OpenCL Raw SHA-256 [Openwall Community Wiki]

OpenCL Raw SHA-256

John can crack raw SHA-256 on OpenCL enabled devices. To use it, type:
john –format=raw-sha256-opencl [other options]

All available GPU power is used while John is running, so the computer can become less responsive, especially if the GPU is used to control your monitor.
The hint is: if your computer seems to hang and you have only one GPU, your X Server is busy and you can't do anything. Just wait 5 or 10 minutes. If nothing happens, reboot your computer. I saw cases on Radeon HD 6770 that after some minutes waiting everything comes to normality.

Despite of that, you can do real crack on your GPU, even if you use it as your graphics controller. Although, you might experience some slowness.

The maximum password length is: 55 (56 - 1) bytes.

In order to change it, anyone can open the opencl_rawsha512.h, modify the statements seen below and recompile the project. Note that some memory mis-alignment could lead to wrong computation results.

#define RAW_PLAINTEXT_LENGTH    56

This software was tested using (see hardware and software details below):
- Board: C-01. On S-01 and S-02 software.
- Board: C-03. On S-01 and S-02 software.
- Board: C-02, C-04 and C-05 on Ubuntu X86_64.
1. Not sure how it will behave under different hardware/software.

Set Up

John can try to figure out what is the best configuration to use in the running hardware. In this set up mode, John will make some benchmark and take some measures. When finished, it will analyse the results and select the best configuration. Although it is not deterministic, the results (at least) give good hints about the best configuration.

There are 3 environment variables that could be set to configure John's behavior.

LWS: local work size. Define workgroup/warp size. It means how (many) items are going to be grouped together.
GWS: global work size. Define how many candidate keys are going to be sent to hardware at once. To hide latency, at least a few thousands should be picked.
STEP. Optional, default is 65536. Define the step size that should be used in order to find the best GWS. It is especially useful on poor or very good cards (where default value might be sub-optimal), when set up takes too much time to finish or while experimenting. This variable is only used while probing GWS.
Also, the user can set STEP=0. In this case, starting on 65536, the GWS values will be multiplied by 2. In this situation, the search is going to be faster, but, in today's hardware, the best value could be missed.
It is a good idea to use only power of two in LWS, GWS and STEP. However, as you can see below, the best benchmark numbers got on AMD are not following this rule.

To access the auto-configuration mode, the user should set LWS and/or GWS to zero (LWS=0 and/or GWS=0).

Usage:

LWS=0 GWS=65536 ./john -fo:raw-sha256-opencl -t
LWS=0 GWS=0 STEP=15386 ./john -fo:raw-sha256-opencl -t
LWS=64 GWS=0 STEP=1024 ./john -fo:raw-sha256-opencl -t
LWS=128 GWS=0 STEP=0 ./john -fo:raw-sha256-opencl -t

One should:

Pay special attention to how the work is going to be divided/organized. It means that some time should be spent to select the best work-group size and threads number (LWS and GWS sizes). A good LWS+GWS set can make the performance increases an order of magnitude.
A safe start point could be: LWS=64 and GWS=8192. Test it using the command below:

LWS=64 GWS=8192 ./john -fo:raw-sha256-opencl -t

My default first try on new hardware is:

LWS=0 GWS=0 STEP=65536 ./john -fo:raw-sha256-opencl -t

===== Configuration ===== ==== Tested Hardware ==== ID : processor, memory * C-01: AMD Phenom™ II X6 1075T Processor × 6, 4GB DDR3 1800MHz * C-02: AMD FX™-8120 Eight-Core Processor * C-03: AMD Radeon HD 6770 (Juniper) * C-04: AMD Radeon HD 7970 (Tahiti) * C-05: GeForce GTX 570 ==== Tested Software ==== * S-01: * SDK 2.6 OpenCL 1.1 AMD-APP (898.1) * Catalyst 12.2 * Driver 8.95-120214a-134397C-ATI * Ubuntu 11.10 x86_64 * S-02: * SDK 2.6 OpenCL 1.2 AMD-APP (923.1) * Catalyst 12.2 * Driver 8.95-120214a-134397C-ATI * Ubuntu 12.04 x86_64 ==== Benchmarks ==== ID : configuration. Operating system. * C-01. S-02. * Local work size (LWS) 768, global work size (GWS) 6136320 * Benchmarking: Raw SHA-256 (pwlen < 32) [OpenCL]… DONE * Raw: 9897K c/s real, 2158K c/s virtual * C-02. * Local work size (LWS) 512, global work size (GWS) 1507328 * Benchmarking: Raw SHA-256 (pwlen < 32) [OpenCL]… DONE * Raw: 10467K c/s real, 1684K c/s virtual * C-03. S-02. * Local worksize (LWS) 64, global worksize (GWS) 2097152 * Benchmarking: Raw SHA-256 [OpenCL (inefficient, development use mostly)]… DONE * Raw: 45147K c/s real, 59101K c/s virtual * C-04. * Local worksize (LWS) 64, global worksize (GWS) 4194304 * Benchmarking: Raw SHA-256 [OpenCL (inefficient, development use mostly)]… DONE * Raw: 54708K c/s real, 56679K c/s virtual * C-05. * Local worksize (LWS) 64, global worksize (GWS) 4423680 * Benchmarking: Raw SHA-256 [OpenCL (inefficient, development use mostly)]… DONE * Raw: 57202K c/s real, 56713K c/s virtual Reference (july 2012)
⇒ John Jumbo on CPU C-01 (1 core)
Benchmarking: Raw SHA-256 [32/64 OpenSSL]… DONE
Raw: 2796K c/s real, 2777K c/s virtual
⇒ John Jumbo with OMP on CPU C-01 (6 cores)
Benchmarking: Raw SHA-256 [32/64 OpenSSL]… (6xOMP) DONE
Raw: 9451K c/s real, 1591K c/s virtual
⇒ John Jumbo on CPU C-02 (1 core)
Benchmarking: Raw SHA-256 [32/64 OpenSSL]… DONE
Raw: 2291K c/s real, 2291K c/s virtual
⇒ John Jumbo with OMP on CPU C-02 (8 cores)
Benchmarking: Raw SHA-256 [32/64 OpenSSL]… (8xOMP) DONE
Raw: 8380K c/s real, 1048K c/s virtual