Merge pull request #405 from JuliaHealth/multigpu

Extend GPU support to Metal, ROCm, and oneAPI backends

KomaMRICore/Project.toml

@@ -5,19 +5,35 @@ version = "0.8.3"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
-CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
+KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 KomaMRIBase = "d0bc0b20-b151-4d03-b2a4-6ca51751cb9c"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 ProgressMeter = "92933f4c-e287-5a05-a399-4b506db050ca"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 ThreadsX = "ac1d9e8a-700a-412c-b207-f0111f4b6c0d"
 
+[weakdeps]
+AMDGPU = "21141c5a-9bdb-4563-92ae-f87d6854732e"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+Metal = "dde4c033-4e86-420c-a63e-0dd931031962"
+oneAPI = "8f75cd03-7ff8-4ecb-9b8f-daf728133b1b"
+
+[extensions]
+KomaAMDGPUExt = "AMDGPU"
+KomaCUDAExt = "CUDA"
+KomaMetalExt = "Metal"
+KomaoneAPIExt = "oneAPI"
+
 [compat]
 Adapt = "3, 4"
+AMDGPU = "0.9"
 CUDA = "3, 4, 5"
 Functors = "0.4"
+KernelAbstractions = "0.9"
 KomaMRIBase = "0.8"
+Metal = "1"
+oneAPI = "1"
 Pkg = "1.4"
 ProgressMeter = "1"
 Reexport = "1"

KomaMRICore/ext/KomaAMDGPUExt.jl

@@ -0,0 +1,24 @@
+module KomaAMDGPUExt
+
+using AMDGPU
+import KomaMRICore
+
+KomaMRICore.name(::ROCBackend) = "AMDGPU"
+KomaMRICore.isfunctional(::ROCBackend) = AMDGPU.functional()
+KomaMRICore.set_device!(::ROCBackend, dev_idx::Integer) = AMDGPU.device_id!(dev_idx)
+KomaMRICore.set_device!(::ROCBackend, dev::AMDGPU.HIPDevice) = AMDGPU.device!(dev)
+KomaMRICore.device_name(::ROCBackend) = AMDGPU.device().name
+
+function KomaMRICore._print_devices(::ROCBackend)
+    devices = [
+        Symbol("($(i-1)$(i == 1 ? "*" : " "))") => d.name for
+        (i, d) in enumerate(AMDGPU.devices())
+    ]
+    @info "$(length(AMDGPU.devices())) AMD capable device(s)." devices...
+end
+
+function __init__()
+    push!(KomaMRICore.LOADED_BACKENDS[], ROCBackend())
+end
+
+end

KomaMRICore/ext/KomaCUDAExt.jl

@@ -0,0 +1,23 @@
+module KomaCUDAExt
+
+using CUDA
+import KomaMRICore
+
+KomaMRICore.name(::CUDABackend) = "CUDA"
+KomaMRICore.isfunctional(::CUDABackend) = CUDA.functional()
+KomaMRICore.set_device!(::CUDABackend, val) = CUDA.device!(val)
+KomaMRICore.device_name(::CUDABackend) = CUDA.name(CUDA.device())
+
+function KomaMRICore._print_devices(::CUDABackend)
+    devices = [
+        Symbol("($(i-1)$(i == 1 ? "*" : " "))") => CUDA.name(d) for
+        (i, d) in enumerate(CUDA.devices())
+    ]
+    @info "$(length(CUDA.devices())) CUDA capable device(s)." devices...
+end
+
+function __init__()
+    push!(KomaMRICore.LOADED_BACKENDS[], CUDABackend())
+end
+
+end

KomaMRICore/ext/KomaMetalExt.jl

@@ -0,0 +1,28 @@
+module KomaMetalExt
+
+using Metal
+import KomaMRICore
+
+KomaMRICore.name(::MetalBackend) = "Metal"
+KomaMRICore.isfunctional(::MetalBackend) = Metal.functional()
+KomaMRICore.set_device!(::MetalBackend, device_index::Integer) = device_index == 1 || @warn "Metal does not support multiple gpu devices. Ignoring the device setting."
+KomaMRICore.set_device!(::MetalBackend, dev::Metal.MTLDevice) = Metal.device!(dev)
+KomaMRICore.device_name(::MetalBackend) = String(Metal.current_device().name)
+
+function KomaMRICore._print_devices(::MetalBackend)
+    @info "Metal device type: $(KomaMRICore.device_name(MetalBackend()))"
+end
+
+#Temporary workaround for https://github.com/JuliaGPU/Metal.jl/issues/348
+#Once run_spin_excitation! and run_spin_precession! are kernel-based, this code
+#can be removed
+Base.cumsum(x::MtlVector) = convert(MtlVector, cumsum(KomaMRICore.cpu(x)))
+Base.cumsum(x::MtlArray{T}; dims) where T = convert(MtlArray{T}, cumsum(KomaMRICore.cpu(x), dims=dims))
+Base.findall(x::MtlVector{Bool}) = convert(MtlVector, findall(KomaMRICore.cpu(x)))
+
+function __init__()
+    push!(KomaMRICore.LOADED_BACKENDS[], MetalBackend())
+    @warn "Due to https://github.com/JuliaGPU/Metal.jl/issues/348, some functions may need to run on the CPU. Performance may be impacted as a result."
+end
+
+end

KomaMRICore/ext/KomaoneAPIExt.jl

@@ -0,0 +1,23 @@
+module KomaoneAPIExt
+
+using oneAPI
+import KomaMRICore
+
+KomaMRICore.name(::oneAPIBackend) = "oneAPI"
+KomaMRICore.isfunctional(::oneAPIBackend) = oneAPI.functional()
+KomaMRICore.set_device!(::oneAPIBackend, val) = oneAPI.device!(val)
+KomaMRICore.device_name(::oneAPIBackend) = oneAPI.properties(oneAPI.device()).name
+
+function KomaMRICore._print_devices(::oneAPIBackend)
+    devices = [
+        Symbol("($(i-1)$(i == 1 ? "*" : " "))") => oneAPI.properties(d).name for
+        (i, d) in enumerate(oneAPI.devices())
+    ]
+    @info "$(length(oneAPI.devices())) oneAPI capable device(s)." devices...
+end
+
+function __init__()
+    push!(KomaMRICore.LOADED_BACKENDS[], oneAPIBackend())
+end
+
+end

KomaMRICore/src/KomaMRICore.jl

@@ -2,12 +2,11 @@ module KomaMRICore
 
 # General
 import Base.*, Base.abs
+import KernelAbstractions as KA
 using Reexport
 using ThreadsX
 # Printing
 using ProgressMeter
-# Simulation
-using CUDA
 
 # KomaMRIBase
 @reexport using KomaMRIBase
@@ -18,6 +17,7 @@ include("rawdata/ISMRMRD.jl")
 include("datatypes/Spinor.jl")
 include("other/DiffusionModel.jl")
 # Simulator
+include("simulation/Functors.jl")
 include("simulation/GPUFunctions.jl")
 include("simulation/SimulatorCore.jl")
 

KomaMRICore/src/simulation/Functors.jl

@@ -0,0 +1,94 @@
+import Adapt: adapt, adapt_storage
+import Functors: @functor, functor, fmap, isleaf
+
+#Aux. funcitons to check if the variable we want to convert to CuArray is numeric
+_isleaf(x) = isleaf(x)
+_isleaf(::AbstractArray{<:Number}) = true
+_isleaf(::AbstractArray{T}) where T = isbitstype(T)
+_isleaf(::AbstractRange) = true
+
+"""
+	gpu(x)
+
+Tries to move `x` to the GPU backend specified in the 'backend' parameter. 
+
+This works for functions, and any struct marked with `@functor`.
+
+Use [`cpu`](@ref) to copy back to ordinary `Array`s.
+
+See also [`f32`](@ref) and [`f64`](@ref) to change element type only.
+
+# Examples
+```julia
+x = gpu(x, CUDABackend())
+```
+"""
+function gpu(x, backend::KA.GPU)
+    return fmap(x -> adapt(backend, x), x; exclude=_isleaf)
+end
+
+# To CPU
+"""
+	cpu(x)
+
+Tries to move object to CPU. This works for functions, and any struct marked with `@functor`.
+
+See also [`gpu`](@ref).
+
+# Examples
+```julia
+x = x |> cpu
+```
+"""
+cpu(x) = fmap(x -> adapt(KA.CPU(), x), x, exclude=_isleaf)
+
+#Precision
+paramtype(T::Type{<:Real}, m) = fmap(x -> adapt(T, x), m)
+adapt_storage(T::Type{<:Real}, xs::Real) = convert(T, xs)
+adapt_storage(T::Type{<:Real}, xs::AbstractArray{<:Real}) = convert.(T, xs)
+adapt_storage(T::Type{<:Real}, xs::AbstractArray{<:Complex}) = convert.(Complex{T}, xs)
+adapt_storage(T::Type{<:Real}, xs::AbstractArray{<:Bool}) = xs
+adapt_storage(T::Type{<:Real}, xs::SimpleMotion) = SimpleMotion(paramtype(T, xs.types))
+adapt_storage(T::Type{<:Real}, xs::NoMotion) = NoMotion{T}()
+function adapt_storage(T::Type{<:Real}, xs::ArbitraryMotion) 
+    fields = []
+    for field in fieldnames(ArbitraryMotion)
+        push!(fields, paramtype(T, getfield(xs, field)))
+    end
+    return ArbitraryMotion(fields...)
+end
+
+"""
+    f32(m)
+
+Converts the `eltype` of model's parameters to `Float32`
+Recurses into structs marked with `@functor`.
+
+See also [`f64`](@ref).
+"""
+f32(m) = paramtype(Float32, m)
+
+"""
+    f64(m)
+
+Converts the `eltype` of model's parameters to `Float64` (which is Koma's default)..
+Recurses into structs marked with `@functor`.
+
+See also [`f32`](@ref).
+"""
+f64(m) = paramtype(Float64, m)
+
+#The functor macro makes it easier to call a function in all the parameters
+@functor Phantom
+
+@functor Translation
+@functor Rotation
+@functor HeartBeat
+@functor PeriodicTranslation
+@functor PeriodicRotation
+@functor PeriodicHeartBeat
+
+@functor Spinor
+@functor DiscreteSequence
+
+export gpu, cpu, f32, f64