The OpenCL 2.0 CTS Can Now Run On Gallium3D Clover - But Doesn't Pass The Tests

[phoronix]

Phoronix: The OpenCL 2.0 CTS Can Now Run On Gallium3D Clover - But Doesn't Pass The Tests

Red Hat's Karol Herbst who has spent years now working on Nouveau SPIR-V support and other GPU open-source compute efforts around Mesa has provided a trivial implementation of clCreateCommandQueueWithProperties() that is now enough to begin running the OpenCL 2.0 conformance test suite on the Gallium3D "Clover" state tracker...

http://www.phoronix.com/scan.php?page=news_item&px=OpenCL-2.0-CTS-On-Clover

[FireBurn]

Does this statetracker still work for radeonsi?

[TemplarGR]

Does this statetracker still work for radeonsi?

It is really incomplete.

[jrch2k8]

Does this statetracker still work for radeonsi?

Yes.

Device Name AMD Radeon (TM) RX 470 Graphics (POLARIS10, DRM 3.35.0, 5.4.15-zen1-1-zen, LLVM 11.0.0)
Device Vendor AMD
Device Vendor ID 0x1002
Device Version OpenCL 1.1 Mesa 20.1.0-devel (git-35961b10da)
Driver Version 20.1.0-devel

but

Device OpenCL C Version OpenCL C 1.1

but when it works is faster than amdgpu-pro opencl implementation or at least on luxrender benchmarks

[Veerappan]

Yes.

Device Name AMD Radeon (TM) RX 470 Graphics (POLARIS10, DRM 3.35.0, 5.4.15-zen1-1-zen, LLVM 11.0.0)
Device Vendor AMD
Device Vendor ID 0x1002
Device Version OpenCL 1.1 Mesa 20.1.0-devel (git-35961b10da)
Driver Version 20.1.0-devel

but

Device OpenCL C Version OpenCL C 1.1

but when it works is faster than amdgpu-pro opencl implementation or at least on luxrender benchmarks

Note: The biggest reason that it's reporting CL 1.1 instead of 1.2 is probably that the libclc project doesn't support the 'printf()' function for debugging kernel execution, yet. You can override the CL/CLC versions exposed via environment variables in clover, which also enables several CL 1.2 features (e.g. volatile, extern, static keywords). Maybe there's other things, but I've always avoided pushing to report 1.2 capabilities in clover because we might as well finish up 1.1 (and get a passing CTS run on at least one GPU) before we start telling people we support an incomplete 1.2 implementation...

[nuetzel]

Note: The biggest reason that it's reporting CL 1.1 instead of 1.2 is probably that the libclc project doesn't support the 'printf()' function for debugging kernel execution, yet. You can override the CL/CLC versions exposed via environment variables in clover, which also enables several CL 1.2 features (e.g. volatile, extern, static keywords). Maybe there's other things, but I've always avoided pushing to report 1.2 capabilities in clover because we might as well finish up 1.1 (and get a passing CTS run on at least one GPU) before we start telling people we support an incomplete 1.2 implementation...

But compilation of

current LLVM git (e.g. 10.0-rcX) with SPIRV-LLVM-Translator (included)

and then

Mesa Clover with '-Dopencl-spirv=true'

is still broken.

See mesa-devel and my posts to Karol.

[nuetzel]

Yes.

Device Name AMD Radeon (TM) RX 470 Graphics (POLARIS10, DRM 3.35.0, 5.4.15-zen1-1-zen, LLVM 11.0.0)
Device Vendor AMD
Device Vendor ID 0x1002
Device Version OpenCL 1.1 Mesa 20.1.0-devel (git-35961b10da)
Driver Version 20.1.0-devel

but

Device OpenCL C Version OpenCL C 1.1

but when it works is faster than amdgpu-pro opencl implementation or at least on luxrender benchmarks

Can second that. For all _3_ luxrender 'versions'.
v2.0/v3.0/v3.1

Device Name Radeon RX 580 Series (POLARIS10, DRM 3.36.0, 5.5.2-2.g994cf1f-default, LLVM 10.0.0)
Device Vendor AMD
Device Vendor ID 0x1002
Device Version OpenCL 1.1 Mesa 20.1.0-devel (git-2799676218)
Driver Version 20.1.0-devel

[uid0]

You can override the CL/CLC versions exposed via environment variables in clover, which also enables several CL 1.2 features (e.g. volatile, extern, static keywords).

How?

Could you please list these variables? I was not able to find any documentation for Clover or libclc regarding this; did a quick search through the code, but no luck.

[Veerappan]

How?

Could you please list these variables? I was not able to find any documentation for Clover or libclc regarding this; did a quick search through the code, but no luck.

Something along these lines (ripped from a shell script I use when working on stuff):

CLOVER_PLATFORM_VERSION_OVERRIDE=1.2 CLOVER_DEVICE_VERSION_OVERRIDE=1.2 CLOVER_DEVICE_CLC_VERSION_OVERRIDE=1.2 ./program_you_want_to_run

[Veerappan]

But compilation of

current LLVM git (e.g. 10.0-rcX) with SPIRV-LLVM-Translator (included)

and then

Mesa Clover with '-Dopencl-spirv=true'

is still broken.

See mesa-devel and my posts to Karol.

Yes, but I was talking about CL 1.2, not 2.0. CL 1.2 on at least the radeon cards is not dependent on NIR/SPIR-V support.