microsoft / OpenCLOn12

The OpenCL-on-D3D12 mapping layer
MIT License
104 stars 13 forks source link

Geekbench 5 "Feature Matching" benchmark error.. #22

Open oscarbg opened 3 years ago

oscarbg commented 3 years ago

Hi, just testing CLon12 perf/correctness under Geekbench 5.. CLon12 package is Win Store version 1.2105.3.0..

all tests, except the shown in title, seem to run correctly.. run with: geekbench5 --compute OpenCL --compute-platform-id 2 --compute-device-id 0 err shown is:

Running Feature Matching
[0512/065245:ERROR:compute_workload.cpp(111)] workload 311 failed validation

other thing, is performance is an order of magnitude slower than native OpenCL at least on Nvidia.. https://browser.geekbench.com/v5/compute/2812026 compared to native OpenCL: https://browser.geekbench.com/v5/compute/1185552

log:

Administrator permissions are needed for Geekbench to report complete system
information. Use an administrator command prompt for a complete system
information report.

Geekbench 5.2.3 Tryout : https://www.geekbench.com/

Geekbench 5 is in tryout mode.

Geekbench 5 requires an active Internet connection when in tryout mode, and
automatically uploads test results to the Geekbench Browser. Other features
are unavailable in tryout mode.

Buy a Geekbench 5 license to enable offline use and remove the limitations of
tryout mode.

If you would like to purchase Geekbench you can do so online:

  https://store.primatelabs.com/v5

If you have already purchased Geekbench, enter your email address and license
key from your email receipt with the following command line:

  geekbench5 -r <email address> <license key>

  Running Gathering system information
System Information
  Operating System              Microsoft Windows 10 Pro Insider Preview (64-bit)
  Operating System              2004
  Model                         Gigabyte Technology Co., Ltd. Z170X-Gaming 7
  Motherboard                   Gigabyte Technology Co., Ltd. Z170X-Gaming 7
  BIOS                          American Megatrends Inc. F22j

Processor Information
  Name                          Intel Core i5-6600K
  Topology                      1 Processor, 4 Cores
  Identifier                    GenuineIntel Family 6 Model 94 Stepping 3
  Base Frequency                3.50 GHz
  Maximum Frequency             0.00 Hz
  L1 Instruction Cache          32.0 KB x 2
  L1 Data Cache                 32.0 KB x 2
  L2 Cache                      256 KB x 2
  L3 Cache                      6.00 MB

Memory Information
  Size                          32.0 GB

OpenCL Information
  Platform Vendor               Microsoft
  Platform Name                 OpenCLOn12
  Device Vendor                 Microsoft
  Device Name                   NVIDIA TITAN V
  Device Driver Version         1.0 .0
  Maximum Frequency             12.0 MHz
  Compute Units                 1
  Device Memory                 11.8 GB

OpenCL
  Running Sobel
  Running Canny
  Running Stereo Matching
  Running Histogram Equalization
  Running Gaussian Blur
  Running Depth of Field
  Running Face Detection
  Running Horizon Detection
  Running Feature Matching
[0512/065245:ERROR:compute_workload.cpp(111)] workload 311 failed validation
  Running Particle Physics
  Running SFFT

Uploading results to the Geekbench Browser. This could take a minute or two
depending on the speed of your internet connection.

Upload succeeded. Visit the following link and view your results online:

  https://browser.geekbench.com/v5/compute/2812026
jenatali commented 1 year ago

Looks like the feature matching test uses doubles... guess I need to consider implementing the fp64 extension.

jenatali commented 1 year ago

Ok adding -cl-single-precision-constant removed the use of doubles, now I'm reproducing the same failure reported. Let's see if I can find out what's going on...

jenatali commented 1 year ago

Without the ability to isolate the specific failing test I'm not sure I have a good way to debug this... It fails the same way with WARP as it does with hardware so it is consistent.

jenatali commented 1 year ago

I'm getting a report that this seems to be fixed with the latest package. Need to verify locally and then I'll close it.