I guess that means not. Then I'll use virtualization or sell this Nvidia card and buy an AMD one. Using WebGL appears to be an alternative too, but I guess it is very slow.
Well Tensorflow can run on CPU too, I just have a spare GPU I haven't sold, so I thought I use that instead of stressing the CPU. It is a sequential pattern mining project, I want to compare normal algorithms like SPADE or PrefixSpan to machine learning.
Graphics/OpenCL Wiki (https://wiki.freebsd.org/Graphics/OpenCL) does not even mention nvidia..
Due to constant problems with my RADEON RX580 I swapped it with NVIDIA GTX1060.. using binary driver version 550.. all problems are gone.. and I was wondering how to use OpenCL/CUDA with that card.. any hints welcome
I am a bit surprised to see ZERO native support from NVIDIA to FreeBSD in CUDA / OpenCL area. AMD supports OpenCL / Clover. 
% pkg info -l x11/nvidia-driver | grep -n icd
124: /usr/local/share/vulkan/icd.d/nvidia_icd.json
% pkg info -l x11/linux-nvidia-libs | grep -n icd
2: /compat/linux/etc/OpenCL/vendors/nvidia.icd
3: /compat/linux/etc/vulkan/icd.d/nvidia_icd.json
%
Which minor version of 535? I've tried ones listed below before.Also 535 driver from nvidia website does not build on FreeBSD 14
NVIDIA_OVERRIDE_VERSION= 535.146.02
.if ${.CURDIR:M/usr/ports/x11/nvidia-driver} && defined (NVIDIA_OVERRIDE_VERSION)
DISTVERSION= ${NVIDIA_OVERRIDE_VERSION}
NO_CHECKSUM= YES
.endif
.if ${.CURDIR:M/usr/ports/x11/linux-nvidia-libs} && defined (NVIDIA_OVERRIDE_VERSION)
DISTVERSION= ${NVIDIA_OVERRIDE_VERSION}
NO_CHECKSUM= YES
.endifMaybe it would be because of the Debian's philosopy, dislike proprietary softwares.
% nv-sglrun clpeak
shim init
Platform: NVIDIA CUDA
Device: NVIDIA GeForce GTX 1660
Driver version : 550.127.05 (FreeBSD)
Compute units : 22
Clock frequency : 1830 MHz
Global memory bandwidth (GBPS)
float : 153.98
float2 : 160.37
float4 : 164.72
float8 : 160.68
float16 : 159.13
Single-precision compute (GFLOPS)
float : 5580.16
float2 : 5519.13
float4 : 5503.80
float8 : 5451.26
float16 : 5396.63
No half precision support! Skipped
Double-precision compute (GFLOPS)
double : 176.49
double2 : 176.27
double4 : 175.76
double8 : 175.04
double16 : 173.29
Integer compute (GIOPS)
int : 4860.36
int2 : 4770.71
int4 : 4774.76
int8 : 4799.89
int16 : 4777.37
Integer compute Fast 24bit (GIOPS)
int : 4628.80
int2 : 4761.41
int4 : 4770.80
int8 : 4744.14
int16 : 4692.37
Integer char (8bit) compute (GIOPS)
char : 4041.64
char2 : 3993.92
char4 : 4049.56
char8 : 4063.33
char16 : 3389.27
Integer short (16bit) compute (GIOPS)
short : 4011.38
short2 : 3832.12
short4 : 3959.47
short8 : 3474.62
short16 : 3339.47
Transfer bandwidth (GBPS)
enqueueWriteBuffer : 12.55
enqueueReadBuffer : 12.75
enqueueWriteBuffer non-blocking : 11.90
enqueueReadBuffer non-blocking : 12.14
enqueueMapBuffer(for read) : 3.88
memcpy from mapped ptr : 19.61
enqueueUnmap(after write) : 13.13
memcpy to mapped ptr : 19.86
Kernel launch latency : 6.49 us
% pkg which -p nv-sglrun
/usr/local/bin/nv-sglrun was installed by package libc6-shim-20240512
% pkg info | grep nvidia
linux-nvidia-libs-550.127.05 NVidia graphics libraries and programs (Linux version)
nvidia-driver-550.127.05.1402000 NVidia graphics card binary drivers for hardware OpenGL rendering
nvidia-settings-535.146.02_1 Display Control Panel for X NVidia driver
nvidia-xconfig-525.116.04 Tool to manipulate X configuration files for the NVidia driverClover is entirely unusable: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19385.
I have requested for analog RL9 x11/linux-nvidia-libs as C7 is fading away but its great to know this stuff works.. would be nice to have few words on how to make it working in Handbook If you are happy with "poor-man's" compute (generally misusing OpenGL for generic compute use-cases), then for desktop OpenGL (i.e not WebGL/OpenGL|ES) you can make use of compute shaders:
x11/nvidia-driver (550) and building x11/linux-nvidia-libs with DEFAULT_VERSIONS+=linux=rl9 set in /etc/make.conf CUDA and OpenCL is now working with Linux-RL9 layer (default port still uses C7). Programs that are supposed to run OpenCL/CUDA applications whould be wrapped by nv-sglrun that comes with libc6-shim-20240512 package. For instance nvidia-smi that shows GPU capabilities (bundled with the driver) will not reveal CUDA until it is wrapped with nv-sglrun nvidia-smi (something like padsp to wrap OSS audio over PulseAudio). I can see now that computations can be shortened from 6 weeks to 4 days on relatively old NVIDIA GTX1060 GPU !!BIG THANK YOU!! # nv-sglrun nvidia-smi
shim init
Tue Feb 4 11:16:02 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.127.05 Driver Version: 550.127.05 CUDA Version: 12.4 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce GTX 1060 6GB Off | 00000000:01:00.0 On | N/A |
| 29% 50C P0 26W / 120W | 1220MiB / 6144MiB | 4% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
+-----------------------------------------------------------------------------------------+# nv-sglrun clpeak
shim init
Platform: NVIDIA CUDA
Device: NVIDIA GeForce GTX 1060 6GB
Driver version : 550.127.05 (FreeBSD)
Compute units : 10
Clock frequency : 1708 MHz
Global memory bandwidth (GBPS)
float : 138.25
float2 : 142.37
float4 : 147.13
float8 : 147.19
float16 : 98.19
Single-precision compute (GFLOPS)
float : 4108.05
float2 : 4312.85
float4 : 4283.72
float8 : 4249.30
float16 : 4226.22
No half precision support! Skipped
Double-precision compute (GFLOPS)
double : 140.94
double2 : 140.18
double4 : 139.88
double8 : 138.68
double16 : 139.25
Integer compute (GIOPS)
int : 1441.66
int2 : 1421.13
int4 : 1431.67
int8 : 1316.01
int16 : 1299.19
Integer compute Fast 24bit (GIOPS)
int : 1428.16
int2 : 1394.27
int4 : 1415.95
int8 : 1411.34
int16 : 1388.43
Integer char (8bit) compute (GIOPS)
char : 3871.67
char2 : 4115.16
char4 : 4133.75
char8 : 4086.24
char16 : 4042.09
Integer short (16bit) compute (GIOPS)
short : 3798.35
short2 : 3925.07
short4 : 4028.50
short8 : 4101.96
short16 : 4036.17
Transfer bandwidth (GBPS)
enqueueWriteBuffer : 5.14
enqueueReadBuffer : 5.82
enqueueWriteBuffer non-blocking : 4.99
enqueueReadBuffer non-blocking : 5.46
enqueueMapBuffer(for read) : 5.86
memcpy from mapped ptr : 4.16
enqueueUnmap(after write) : 5.93
memcpy to mapped ptr : 4.18
Kernel launch latency : 7.88 usUSES= linux in ports Makefile make it depending on default Linuxulator (currently still c7) and can be overridden with rl9 by USES= linux:rl9 per-port basis. This is on the maintainer of the each port.DEFAUT_VERSIONS+= linux=rl9 in /etc/make.conf. This is on the admin of the each computer.
Clang , and I need to use the nv-sglrun command to run the binaries. nv-sglrun is a wrapper that enables application to access underlying hardware acceleration (something like padsp to pass audio from oss applications to pulseaudio). Turns out FreeBSD's Linuxlator is already so good it can do this sort of tricks And by my initial experimentation it is very important that nvidia drivers and libs have the same version, so having all this stuff in ports keeps things in sync It would be quite trivial if the pkg name r9-linux-nvidia-libs or linux-nvidia-libs-rl9 is OK, but not sure it's accepted by the group maintainers (x11@). I'm not a part of them.I think the best way to offer this x11/linux-nvidia-libs in packages is to just add FLAVOR for c7 and rl9 so it builds both for c7 and rl9 and its still a single port easy to maitnain?