cory-master/BabelStream
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Inline AMDGPU's hard_wait

Browse Source 
Show the selected implementation and not a constant "threaded"
This commit is contained in:
Tom Lin 2021-06-30 18:09:54 +01:00
parent d799535c96
commit 4e6c56729b
2 changed files with 15 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
JuliaStream.jl/src/AMDGPUStream.jl
View File

@ -56,11 +56,6 @@ function make_stream(
return data return data
end end
function hard_wait(kernel)
# soft wait causes HSA_REFCOUNT overflow issues
AMDGPU.wait(kernel, soft = false)
end
function init_arrays!(data::ROCData{T}, init::Tuple{T,T,T}) where {T} function init_arrays!(data::ROCData{T}, init::Tuple{T,T,T}) where {T}
AMDGPU.fill!(data.a, init[1]) AMDGPU.fill!(data.a, init[1])
AMDGPU.fill!(data.b, init[2]) AMDGPU.fill!(data.b, init[2])
@ -73,7 +68,10 @@ function copy!(data::ROCData{T}) where {T}
@inbounds c[i] = a[i] @inbounds c[i] = a[i]
return return
end end
hard_wait(@roc groupsize = TBSize gridsize = gridsize(data) kernel(data.a, data.c)) AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = gridsize(data) kernel(data.a, data.c)
)
end end
function mul!(data::ROCData{T}) where {T} function mul!(data::ROCData{T}) where {T}
@ -82,7 +80,8 @@ function mul!(data::ROCData{T}) where {T}
@inbounds b[i] = scalar * c[i] @inbounds b[i] = scalar * c[i]
return return
end end
hard_wait( AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = gridsize(data) kernel(data.b, data.c, data.scalar) @roc groupsize = TBSize gridsize = gridsize(data) kernel(data.b, data.c, data.scalar)
) )
end end
@ -93,7 +92,8 @@ function add!(data::ROCData{T}) where {T}
@inbounds c[i] = a[i] + b[i] @inbounds c[i] = a[i] + b[i]
return return
end end
hard_wait( AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = gridsize(data) kernel(data.a, data.b, data.c) @roc groupsize = TBSize gridsize = gridsize(data) kernel(data.a, data.b, data.c)
) )
end end
@ -104,7 +104,8 @@ function triad!(data::ROCData{T}) where {T}
@inbounds a[i] = b[i] + (scalar * c[i]) @inbounds a[i] = b[i] + (scalar * c[i])
return return
end end
hard_wait( AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = gridsize(data) kernel( @roc groupsize = TBSize gridsize = gridsize(data) kernel(
data.a, data.a,
data.b, data.b,
@ -120,7 +121,8 @@ function nstream!(data::ROCData{T}) where {T}
@inbounds a[i] += b[i] + scalar * c[i] @inbounds a[i] += b[i] + scalar * c[i]
return return
end end
hard_wait( AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = gridsize(data) kernel( @roc groupsize = TBSize gridsize = gridsize(data) kernel(
data.a, data.a,
data.b, data.b,
@ -160,7 +162,8 @@ function dot(data::ROCData{T}) where {T}
return return
end end
partial_sum = ROCArray{T}(undef, DotBlocks) partial_sum = ROCArray{T}(undef, DotBlocks)
hard_wait( AMDGPU.wait(
soft = false, # soft wait causes HSA_REFCOUNT overflow issues
@roc groupsize = TBSize gridsize = TBSize * DotBlocks kernel( @roc groupsize = TBSize gridsize = TBSize * DotBlocks kernel(
data.a, data.a,
data.b, data.b,

3
JuliaStream.jl/src/Stream.jl
View File

@ -99,7 +99,6 @@ end
@with_kw mutable struct Config @with_kw mutable struct Config
list::Bool = false list::Bool = false
impl::String = "threaded"
device::Int = 1 device::Int = 1
numtimes::Int = 100 numtimes::Int = 100
arraysize::Int = 33554432 arraysize::Int = 33554432
@ -201,7 +200,7 @@ function main()
if !config.csv if !config.csv
println("""BabelStream println("""BabelStream
Version: $Version Version: $Version
Implementation: Julia; $(config.impl)""") Implementation: Julia; $(PROGRAM_FILE)""")
println("Running kernels $(config.numtimes) times") println("Running kernels $(config.numtimes) times")
if benchmark == Triad if benchmark == Triad
println("Number of elements: $(config.arraysize)") println("Number of elements: $(config.arraysize)")