Linux: Amdgpupro 16.60
#1
Hi guys,

Amdgpupro 16.60 has just been released. Have you tried it out yet? Does it fix the slow performance in rule-based attack that we encountered with 16.50?
#2
I've just tested it:
  • Positive: Installation went good
  • Positive: Performance in BF mode is still good. Here's a full table: https://docs.google.com/spreadsheets/d/1...1273965171
  • Negative: Drivers too aggressive clock management still overrides "manual" setting in power_dpm_force_performance_level at random events. Such jumping clocks during benchmark still occur and make benchmark meassurements inaccurate. This is one of the problems that makes us developers go crazy especially when you try to optimize kernels you never knew if the driver changed the clocks or if it was your code change that actually changed the performance. The fact that there's no way to turn off this "feature" makes this .... I don't find a word.. I mean why is it called manual then?
  • Negative: The drivers OpenCL JiT compiler still segfaults when compiling SCRYPT based kernels 8900 and 9300 and PeopleSoft Token mode 13500. This started with AMDGPU-Pro 16.50. The only way to override was to patch benchmark.c and to make sure the compiler never calls them
  • Negative: Still had to manually patch the installation script to not need to install 32 bit support
  • Negative: The problem with the huge slowdown on rule-based cracking still exist

AMDGPU-Pro 16.40 is still the only working driver and shows the best performance in benchmarks
#3
under the current archlinux (Linux lenovo-s440 4.10.3-1-ARCH #1 SMP PREEMPT Wed Mar 15 09:17:17 CET 2017 x86_64 GNU/Linux) and with the opencl part from the amdgpu driver. https://aur.archlinux.org/packages/opencl-amd/ at least the benchmark works fine.

Code:
Hashtype: scrypt
Speed.Dev.#1.....:    22297 H/s (111.95ms)
--
Hashtype: Cisco $9$
Speed.Dev.#1.....:      547 H/s (464.73ms)
--
Hashtype: PeopleSoft PS_TOKEN
Speed.Dev.#1.....:   171.9 MH/s (60.50ms)

Code:
OpenCL Info:

Platform ID #1
 Vendor  : Advanced Micro Devices, Inc.
 Name    : AMD Accelerated Parallel Processing
 Version : OpenCL 2.0 AMD-APP (2264.10)

 Device ID #1
   Type           : GPU
   Vendor ID      : 1
   Vendor         : Advanced Micro Devices, Inc.
   Name           : Hainan
   Version        : OpenCL 1.2 AMD-APP (2264.10)
   Processor(s)   : 5
   Clock          : 975
   Memory         : 1791/2043 MB allocatable
   OpenCL Version : OpenCL C 1.2
   Driver Version : 2264.10

 Device ID #2
   Type           : CPU
   Vendor ID      : 128
   Vendor         : GenuineIntel
   Name           : Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz
   Version        : OpenCL 1.2 AMD-APP (2264.10)
   Processor(s)   : 4
   Clock          : 2443
   Memory         : 7904/7904 MB allocatable
   OpenCL Version : OpenCL C 1.2
   Driver Version : 2264.10 (sse2,avx)

Platform ID #2
 Vendor  : Intel(R) Corporation
 Name    : Intel(R) OpenCL
 Version : OpenCL 1.2

 Device ID #3
   Type           : GPU
   Vendor ID      : 8
   Vendor         : Intel(R) Corporation
   Name           : Intel(R) HD Graphics
   Version        : OpenCL 1.2
   Processor(s)   : 20
   Clock          : 1000
   Memory         : 815/1630 MB allocatable
   OpenCL Version : OpenCL C 1.2
   Driver Version : r4.0.59481

 Device ID #4
   Type           : CPU
   Vendor ID      : 8
   Vendor         : Intel(R) Corporation
   Name           : Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz
   Version        : OpenCL 1.2 (Build 25)
   Processor(s)   : 4
   Clock          : 1700
   Memory         : 1976/7904 MB allocatable
   OpenCL Version : OpenCL C 1.2
   Driver Version : 1.2.0.25

Platform ID #3
 Vendor  : Intel(R) Corporation
 Name    : Intel(R) OpenCL
 Version : OpenCL 1.2 LINUX

 Device ID #5
   Type           : CPU
   Vendor ID      : 8
   Vendor         : Intel(R) Corporation
   Name           : Intel(R) Core(TM) i5-4210U CPU @ 1.70GHz
   Version        : OpenCL 1.2 (Build 25)
   Processor(s)   : 4
   Clock          : 1700
   Memory         : 1976/7904 MB allocatable
   OpenCL Version : OpenCL C 1.2
   Driver Version : 1.2.0.25
#4
Just checked new 17.10 driver, no changes to 16.60 - all bugs still exist. Currently 16.40 is still the best amgpu-pro driver.
#5
New 17.20 driver out there.

Has anybody had the chance to try it out already?
#6
I've just tested it (on Ubuntu 16.4):
  • Positive: Installation still works good.
  • Positive: The most important problem with the huge slowdown on rule-based cracking was resolved! Speed actually increased, see benchmark table in the link below.
  • Positive: Performance in BF mode is still good. Here's a full table: https://docs.google.com/spreadsheets/d/1...=661638232
  • Negative: Still have to manually remove "amdgpu-pro-lib32" from the amdgpu-pro-install program. This problem exist since ever AMDGPU-Pro exist.
  • Negative: Drivers too aggressive clock management still overrides "manual" setting in power_dpm_force_performance_level at random events. Such jumping clocks during benchmark still occur and make benchmark meassurements inaccurate. This is one of the problems that makes us developers go crazy especially when you try to optimize kernels you never knew if the driver changed the clocks or if it was your code change that actually changed the performance. The fact that there's no way to turn off this "feature" makes this .... I don't find a word.. I mean why is it called manual then? This problem exist since ever AMDGPU-Pro exist.
  • Negative: The drivers OpenCL JiT compiler still segfaults when compiling SCRYPT based kernels 8900 and 9300 and PeopleSoft Token mode 13500. This started with AMDGPU-Pro 16.50. The only way to override was to patch benchmark.c and to make sure the compiler never calls them

I'll mark it as good driver in github master
#7
(06-29-2017, 02:16 PM)atom Wrote: I've just tested it (on Ubuntu 16.4):
  • Positive: Installation still works good.
  • Positive: The most important problem with the huge slowdown on rule-based cracking was resolved! Speed actually increased, see benchmark table in the link below.
  • Positive: Performance in BF mode is still good. Here's a full table: https://docs.google.com/spreadsheets/d/1...=661638232
  • Negative: Still have to manually remove "amdgpu-pro-lib32" from the amdgpu-pro-install program. This problem exist since ever AMDGPU-Pro exist.
  • Negative: Drivers too aggressive clock management still overrides "manual" setting in power_dpm_force_performance_level at random events. Such jumping clocks during benchmark still occur and make benchmark meassurements inaccurate. This is one of the problems that makes us developers go crazy especially when you try to optimize kernels you never knew if the driver changed the clocks or if it was your code change that actually changed the performance. The fact that there's no way to turn off this "feature" makes this .... I don't find a word.. I mean why is it called manual then? This problem exist since ever AMDGPU-Pro exist.
  • Negative: The drivers OpenCL JiT compiler still segfaults when compiling SCRYPT based kernels 8900 and 9300 and PeopleSoft Token mode 13500. This started with AMDGPU-Pro 16.50. The only way to override was to patch benchmark.c and to make sure the compiler never calls them

I'll mark it as good driver in github master


Thanks !!! This is great news !! I was going crazy with the rule thing issue.
#8
I've just tried it (running 17.20 right now). I must be doing something wrong, because speed is still an issue with rule-based attacks:
Speed.Dev.#1.....:   236.2 MH/s (72.77ms)

Same wordlist, mask attack: 
Speed.Dev.#1.....:  4046.6 MH/s (78.35ms)


(* Device #1: Hawaii, 3469/3609 MB allocatable, 40MCU)
#9
I've tried on my RX480, it's possible they fixed it only for newer GPU generations. Also make sure to use latest beta for testing from https://hashcat.net/beta/ or github master version.