Windows: GPU architecture detection for artifact selection is untested on real hardware

[luraess]

On Linux the ROCm artifact is selected from gfx_target_version in the KFD topology. Windows has no KFD, so rocm_arch() falls back to matching GPU marketing names reported by EnumDisplayDevicesW against a fixed table (ROCm_Runtime/.pkg/platform_augmentation.jl:43):

const device_name_archs = [
    ["8050s", "8060s", "device 1586"] => "gfx1151",
    ["880m", "890m"]                  => "gfx1150",
    ["r9700", "9060", "9070"]         => "gfx120X",
    ["7700", "7800", "7900", "v710"]  => "gfx110X",
    ["6800", "6700", "6600", "6500"]  => "gfx103X",
]

The adapter string must also contain radeon or amd, otherwise the entry is skipped.

This path is not yet tested.

MWE

On Windows, with any AMD GPU:

using AMDGPU
AMDGPU.versioninfo()

RR = AMDGPU.ROCm_Runtime
@show RR.windows_video_device_names()
@show RR.rocm_arch()
@show RR.is_available()
@show AMDGPU.functional()

Please report your GPU model and the output.

Expected: windows_video_device_names() lists your adapter, and rocm_arch() maps it to the matching gfx family (for example an RX 6700 XT should give gfx103X). An empty rocm_arch() or is_available() == false means the adapter string did not match the table and no artifact was resolved.

As a workaround in that case, the architecture can be set explicitly:

AMDGPU.set_rocm_version!(arch="gfx1031")  # substitute your own target

Reports of the exact adapter string are useful even when detection works.

[ejmeitz]

I'm on Windows 11 with a 6700 XT. It seems to have run correctly. AMD.versioninfo does hang for a lot longer than I would expect, maybe 20 seconds. After compilation maybe 1-2 seconds. Not sure why windows_video_device_names shows that I have 7 GPUs. I only have one.

julia> AMDGPU.versioninfo()
AMDGPU versioninfo
ROCm provider: downloaded artifacts
┌───────────┬──────────────────┬────────────┬────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ Available │ Name             │ Version    │ Path                                                                                                                                           │
├───────────┼──────────────────┼────────────┼────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┤
│     +     │ LLD              │ -          │ C:\\Users\\ejmei\\.julia\\artifacts\\578b36cac4eb31b77226d65946ba921d70e2b52d\\bin\\lld.exe                                                    │
│     +     │ Device Libraries │ -          │ C:\\Users\\ejmei\\.julia\\artifacts\\5ad5ecb46e3c334821f54c1feecc6c152b7b6a45\\amdgcn/bitcode                                                  │
│     +     │ HIP              │ 7.14.60850 │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\amdhip64_7.dll │
│     +     │ rocBLAS          │ 5.5.0      │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\rocblas.dll    │
│     +     │ rocSOLVER        │ 3.35.0     │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\rocsolver.dll  │
│     +     │ rocSPARSE        │ 4.7.0      │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\rocsparse.dll  │
│     +     │ rocRAND          │ 4.5.0      │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\rocrand.dll    │
│     +     │ rocFFT           │ 1.0.38     │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\rocfft.dll     │
│     -     │ hipTENSOR        │ -          │ -                                                                                                                                              │
│     +     │ MIOpen           │ 3.5.2      │ C:\\Users\\ejmei\\.julia\\artifacts\\2174846a8ed3e2bd9818b1123a3c538710397211\\therock-dist-windows-gfx103X-all-7.14.0rc3\\bin\\MIOpen.dll     │
└───────────┴──────────────────┴────────────┴────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘

AMDGPU devices
┌────┬───────────────────────┬──────────┬───────────┬────────────┬───────────────┐
│ Id │                  Name │ GCN arch │ Wavefront │     Memory │ Shared Memory │
├────┼───────────────────────┼──────────┼───────────┼────────────┼───────────────┤
│  1 │ AMD Radeon RX 6700 XT │  gfx1031 │        32 │ 11.984 GiB │    64.000 KiB │
└────┴───────────────────────┴──────────┴───────────┴────────────┴───────────────┘


julia>

julia> RR = AMDGPU.ROCm_Runtime
ROCm_Runtime

julia> @show RR.windows_video_device_names()
RR.windows_video_device_names() = ["AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT", "AMD Radeon RX 6700 XT"]
7-element Vector{String}:
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"
 "AMD Radeon RX 6700 XT"

julia> @show RR.rocm_arch()
RR.rocm_arch() = ["gfx103X"]
1-element Vector{String}:
 "gfx103X"

julia> @show RR.is_available()
RR.is_available() = true
true

julia> @show AMDGPU.functional()
AMDGPU.functional() = true
true

[simeonschaub]

Thanks for testing, this is great news, it looks like everything is working! Did you test whether you can actually compile and run kernels too? Ideally run the whole testsuite, but only if you have the time, this has already been super helpful!

AMD.versioninfo does hang for a lot longer than I would expect, maybe 20 seconds.

@luraess I am pretty sure this is due to #1001, I wonder whether it's really worth adding this extra complexity if the installation is just broken anyways? At the very least, we probably shouldn't be loading all of AMDGPU in the subprocess and instead call Libdl directly

Not sure why windows_video_device_names shows that I have 7 GPUs. I only have one.

It seems like this corresponds to the number of display adapter sources (the potential monitor connections the driver supports) not physical GPUs. We deduplicate these for the artifact selection anyways, so it shouldn't matter in the end

[luraess]

this is due to #1001, I wonder whether it's really worth adding this extra complexity if the installation is just broken anyways? At the very least, we probably shouldn't be loading all of AMDGPU in the subprocess and instead call Libdl directly

Good point. Will look into that.

@ejmeitz thanks a lot for the report

[ejmeitz]

I'll run the tests and report the results soon-ish. Will ping on here when that's done.