Set up R OpenCL with an AMD GPU on Windows

[Quirinms]

I already installed the R package OpenCL with a Nvidia GPU (GTX 1050 Ti) and the Nvidia toolkit on Windows. Now I am trying to set up R OpenCL for an AMD GPU (Radeon 7900 XTX) on Windows.

I installed the appropriate drivers for the AMD GPU from the official website and downloaded the AMD SDK and set up the OCL environment variable to "C:\Program Files (x86)\AMD APP SDK\3.0". Then I tried to install the OpenCL package with install.packages("OpenCL", INSTALL_opts = '--no-test-load') and install.packages("OpenCL", INSTALL_opts = c('--no-clean-on-error', '--no-test-load')). The compilation failed.

The following message "OCL64LIB not set, assuming C:/Program Files (x86)/AMD APP SDK/3.0/lib/x64/OpenCL.lib OCL32LIB not set, assuming C:/Program Files (x86)/AMD APP SDK/3.0/lib/x86/OpenCL.lib OCLINC not set, assuming -IC:/Program Files (x86)/AMD APP SDK/3.0/include" made me wonder, if the paths could be wrong and indeed, there are paths "C:\Program Files (x86)\AMD APP SDK\3.0\lib\x86" and "C:\Program Files (x86)\AMD APP SDK\3.0\lib\x86_64". I dont know, why there is a x86_x64 instead of an expected x64, but R will not find the right way for x64 with the previous assumption.

Setting the paths directly to the files didnt work. Renaming the AMD APP SDK directory does result in a seemingly successful building of the package, but OpenCL code, which could be used before (on the Nvidia GPU) was not working due to OpenCL errors. Also, only the CPU built-in GPU with the was recognized by OpenCL and not the AMD GPU.

[s-u]

@Quirinms Sorry, but what is you question? Your simply describing how you successfully installed OpenCL, which is great. The only issue I can see is that you tested it with some OpenCL code that is not supported by the AMD driver and/or your GPU. That has nothing to do with the OpenCL package (and you didn't post the code nor the errors) so I'm not sure what is the issue you're reporting.

[Quirinms]

The question is: how to set up OpenCL in R in a way, so that it recognizes the AMD GPU? Was something wrong in my set up? Was the "trick" I did not valid? Is it a driver issue that OpenCL in R does not recognize the AMD GPU?

[aaronpuchert]

Does the GPU show up in the output of clinfo? And can you compare that with the output of

lapply(oclPlatforms(), oclInfo)
lapply(oclDevices(), oclInfo)

in R?

[Quirinms]

The GPU does show up in the output of clinfo, but not in the oclDevices():

OpenCL::oclDevices() [[1]] OpenCL device 'Intel(R) UHD Graphics 770'

lapply(oclPlatforms(), oclInfo) [[1]] [[1]]$name [1] "Intel(R) OpenCL HD Graphics"

[[1]]$vendor [1] "Intel(R) Corporation"

[[1]]$version [1] "OpenCL 3.0 "

[[1]]$profile [1] "FULL_PROFILE"

[[1]]$exts [1] "cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_icd cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_intel_command_queue_families cl_intel_subgroups cl_intel_required_subgroup_size cl_intel_subgroups_short cl_khr_spir cl_intel_accelerator cl_intel_driver_diagnostics cl_khr_priority_hints cl_khr_throttle_hints cl_khr_create_command_queue cl_intel_subgroups_char cl_intel_subgroups_long cl_khr_il_program cl_intel_mem_force_host_memory cl_khr_subgroup_extended_types cl_khr_subgroup_non_uniform_vote cl_khr_subgroup_ballot cl_khr_subgroup_non_uniform_arithmetic cl_khr_subgroup_shuffle cl_khr_subgroup_shuffle_relative cl_khr_subgroup_clustered_reduce cl_intel_device_attribute_query cl_khr_suggested_local_work_size cl_intel_spirv_media_block_io cl_intel_spirv_subgroups cl_khr_spirv_no_integer_wrap_decoration cl_intel_unified_shared_memory_preview cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_intel_planar_yuv cl_intel_packed_yuv cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_image2d_from_buffer cl_khr_depth_images cl_khr_3d_image_writes cl_intel_media_block_io cl_khr_gl_sharing cl_khr_gl_depth_images cl_khr_gl_event cl_khr_gl_msaa_sharing cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_intel_d3d11_nv12_media_sharing cl_intel_sharing_format_query cl_khr_pci_bus_info cl_intel_subgroup_local_block_io cl_intel_simultaneous_sharing "

[[2]] [[2]]$name [1] "AMD Accelerated Parallel Processing"

[[2]]$vendor [1] "Advanced Micro Devices, Inc."

[[2]]$version [1] "OpenCL 2.1 AMD-APP (3504.0)"

[[2]]$profile [1] "FULL_PROFILE"

[[2]]$exts [1] "cl_khr_icd cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_amd_event_callback cl_amd_offline_devices "

lapply(oclDevices(), oclInfo) [[1]] [[1]]$name [1] "Intel(R) UHD Graphics 770"

[[1]]$vendor [1] "Intel(R) Corporation"

[[1]]$version [1] "OpenCL 3.0 NEO "

[[1]]$profile [1] "FULL_PROFILE"

[[1]]$exts [1] "cl_khr_byte_addressable_store" "cl_khr_fp16"
[3] "cl_khr_global_int32_base_atomics" "cl_khr_global_int32_extended_atomics"
[5] "cl_khr_icd" "cl_khr_local_int32_base_atomics"
[7] "cl_khr_local_int32_extended_atomics" "cl_intel_command_queue_families"
[9] "cl_intel_subgroups" "cl_intel_required_subgroup_size"
[11] "cl_intel_subgroups_short" "cl_khr_spir"
[13] "cl_intel_accelerator" "cl_intel_driver_diagnostics"
[15] "cl_khr_priority_hints" "cl_khr_throttle_hints"
[17] "cl_khr_create_command_queue" "cl_intel_subgroups_char"
[19] "cl_intel_subgroups_long" "cl_khr_il_program"
[21] "cl_intel_mem_force_host_memory" "cl_khr_subgroup_extended_types"
[23] "cl_khr_subgroup_non_uniform_vote" "cl_khr_subgroup_ballot"
[25] "cl_khr_subgroup_non_uniform_arithmetic" "cl_khr_subgroup_shuffle"
[27] "cl_khr_subgroup_shuffle_relative" "cl_khr_subgroup_clustered_reduce"
[29] "cl_intel_device_attribute_query" "cl_khr_suggested_local_work_size"
[31] "cl_intel_spirv_media_block_io" "cl_intel_spirv_subgroups"
[33] "cl_khr_spirv_no_integer_wrap_decoration" "cl_intel_unified_shared_memory_preview" [35] "cl_khr_mipmap_image" "cl_khr_mipmap_image_writes"
[37] "cl_intel_planar_yuv" "cl_intel_packed_yuv"
[39] "cl_khr_int64_base_atomics" "cl_khr_int64_extended_atomics"
[41] "cl_khr_image2d_from_buffer" "cl_khr_depth_images"
[43] "cl_khr_3d_image_writes" "cl_intel_media_block_io"
[45] "cl_khr_gl_sharing" "cl_khr_gl_depth_images"
[47] "cl_khr_gl_event" "cl_khr_gl_msaa_sharing"
[49] "cl_khr_d3d10_sharing" "cl_khr_d3d11_sharing"
[51] "cl_intel_d3d11_nv12_media_sharing" "cl_intel_sharing_format_query"
[53] "cl_khr_pci_bus_info" "cl_intel_subgroup_local_block_io"
[55] "cl_intel_simultaneous_sharing"

[[1]]$driver.ver [1] "30.0.101.1273"

[[1]]$max.frequency [1] 1500

[s-u]

You only asked for device in the Intel platform. If you want the AMD device, you have to use the AMD platform (see documentation).

[aaronpuchert]

Indeed, I should have asked you to run lapply(oclPlatforms(), oclDevices) or lapply(lapply(oclPlatforms(), oclDevices), oclInfo). So you'll need a device provided by oclDevices(oclPlatforms()[[2]]).

[Quirinms]

oclDevices(oclPlatforms()[[2]]) [[1]] OpenCL device 'gfx1100'

[aaronpuchert]

Can't find an official source, but it seems that GFX1100 stands for RDNA 3 Navi 31, marketed as "Radeon RX 7900 XT[X]". So that should be the card that you're looking for.