Wrapper for Blocksparse CuTensor code#3057
Merged
kshyatt merged 38 commits intoJuliaGPU:masterfrom Apr 22, 2026
Merged
Conversation
… to make it a union type of CuTensorBS and AbstractArray?
Contributor
|
Your PR requires formatting changes to meet the project's style guidelines. Click here to view the suggested changes.diff --git a/lib/cutensor/src/blocksparse/interfaces.jl b/lib/cutensor/src/blocksparse/interfaces.jl
index c6eef0e5b..0a479ddf8 100644
--- a/lib/cutensor/src/blocksparse/interfaces.jl
+++ b/lib/cutensor/src/blocksparse/interfaces.jl
@@ -1,4 +1,4 @@
-## For now call contract in ITensor and rely on UnallocatedArrays to make
+## For now call contract in ITensor and rely on UnallocatedArrays to make
## C in a dry-run of the contraction.
# function Base.:(*)(A::CuTensorBS, B::CuTensorBs)
# tC = promote_type(eltype(A), eltype(B))
@@ -18,11 +18,13 @@
using LinearAlgebra
function LinearAlgebra.mul!(C::CuTensorBS, A::CuTensorBS, B::CuTensorBS, α::Number, β::Number)
- contract!(α,
- A, A.inds, CUTENSOR_OP_IDENTITY,
- B, B.inds, CUTENSOR_OP_IDENTITY,
- β,
- C, C.inds, CUTENSOR_OP_IDENTITY,
- CUTENSOR_OP_IDENTITY; jit=CUTENSOR_JIT_MODE_DEFAULT)
- return C
-end
\ No newline at end of file
+ contract!(
+ α,
+ A, A.inds, CUTENSOR_OP_IDENTITY,
+ B, B.inds, CUTENSOR_OP_IDENTITY,
+ β,
+ C, C.inds, CUTENSOR_OP_IDENTITY,
+ CUTENSOR_OP_IDENTITY; jit = CUTENSOR_JIT_MODE_DEFAULT
+ )
+ return C
+end
diff --git a/lib/cutensor/src/blocksparse/operations.jl b/lib/cutensor/src/blocksparse/operations.jl
index 19542e5de..0f98c92ef 100644
--- a/lib/cutensor/src/blocksparse/operations.jl
+++ b/lib/cutensor/src/blocksparse/operations.jl
@@ -9,23 +9,26 @@ function contract!(
@nospecialize(beta::Number),
@nospecialize(C), Cinds::ModeType, opC::cutensorOperator_t,
opOut::cutensorOperator_t;
- jit::cutensorJitMode_t=JIT_MODE_NONE,
- workspace::cutensorWorksizePreference_t=WORKSPACE_DEFAULT,
- algo::cutensorAlgo_t=ALGO_DEFAULT,
- compute_type::Union{DataType, cutensorComputeDescriptorEnum, Nothing}=nothing,
- plan::Union{CuTensorPlan, Nothing}=nothing)
+ jit::cutensorJitMode_t = JIT_MODE_NONE,
+ workspace::cutensorWorksizePreference_t = WORKSPACE_DEFAULT,
+ algo::cutensorAlgo_t = ALGO_DEFAULT,
+ compute_type::Union{DataType, cutensorComputeDescriptorEnum, Nothing} = nothing,
+ plan::Union{CuTensorPlan, Nothing} = nothing
+ )
actual_plan = if plan === nothing
- plan_contraction(A, Ainds, opA, B, Binds, opB, C, Cinds, opC, opOut;
- jit, workspace, algo, compute_type)
+ plan_contraction(
+ A, Ainds, opA, B, Binds, opB, C, Cinds, opC, opOut;
+ jit, workspace, algo, compute_type
+ )
else
plan
end
contractBS!(actual_plan, alpha, nonzero_blocks(A), nonzero_blocks(B), beta, nonzero_blocks(C))
-
+
if plan === nothing
- CUDA.unsafe_free!(actual_plan)
+ CUDA.unsafe_free!(actual_plan)
end
return C
@@ -33,12 +36,14 @@ end
## This function assumes A, B, and C are Arrays of pointers to CuArrays.
## Please overwrite the `nonzero_blocks` function for your datatype to access this function from contract!
-function contractBS!(plan::CuTensorPlan,
- @nospecialize(alpha::Number),
- @nospecialize(A::AbstractArray),
- @nospecialize(B::AbstractArray),
- @nospecialize(beta::Number),
- @nospecialize(C::AbstractArray))
+function contractBS!(
+ plan::CuTensorPlan,
+ @nospecialize(alpha::Number),
+ @nospecialize(A::AbstractArray),
+ @nospecialize(B::AbstractArray),
+ @nospecialize(beta::Number),
+ @nospecialize(C::AbstractArray)
+ )
scalar_type = plan.scalar_type
# Extract GPU pointers from each CuArray block
@@ -46,11 +51,13 @@ function contractBS!(plan::CuTensorPlan,
A_ptrs = CuPtr{Cvoid}[pointer(block) for block in A]
B_ptrs = CuPtr{Cvoid}[pointer(block) for block in B]
C_ptrs = CuPtr{Cvoid}[pointer(block) for block in C]
-
- cutensorBlockSparseContract(handle(), plan,
- Ref{scalar_type}(alpha), A_ptrs, B_ptrs,
- Ref{scalar_type}(beta), C_ptrs, C_ptrs,
- plan.workspace, sizeof(plan.workspace), stream())
+
+ cutensorBlockSparseContract(
+ handle(), plan,
+ Ref{scalar_type}(alpha), A_ptrs, B_ptrs,
+ Ref{scalar_type}(beta), C_ptrs, C_ptrs,
+ plan.workspace, sizeof(plan.workspace), stream()
+ )
synchronize(stream())
return C
end
@@ -60,21 +67,22 @@ function plan_contraction(
@nospecialize(B), Binds::ModeType, opB::cutensorOperator_t,
@nospecialize(C), Cinds::ModeType, opC::cutensorOperator_t,
opOut::cutensorOperator_t;
- jit::cutensorJitMode_t=JIT_MODE_NONE,
- workspace::cutensorWorksizePreference_t=WORKSPACE_DEFAULT,
- algo::cutensorAlgo_t=ALGO_DEFAULT,
- compute_type::Union{DataType, cutensorComputeDescriptorEnum, Nothing}=nothing)
+ jit::cutensorJitMode_t = JIT_MODE_NONE,
+ workspace::cutensorWorksizePreference_t = WORKSPACE_DEFAULT,
+ algo::cutensorAlgo_t = ALGO_DEFAULT,
+ compute_type::Union{DataType, cutensorComputeDescriptorEnum, Nothing} = nothing
+ )
!is_unary(opA) && throw(ArgumentError("opA must be a unary op!"))
!is_unary(opB) && throw(ArgumentError("opB must be a unary op!"))
!is_unary(opC) && throw(ArgumentError("opC must be a unary op!"))
!is_unary(opOut) && throw(ArgumentError("opOut must be a unary op!"))
-
+
descA = CuTensorBSDescriptor(A)
descB = CuTensorBSDescriptor(B)
descC = CuTensorBSDescriptor(C)
# for now, D must be identical to C (and thus, descD must be identical to descC)
-
+
modeA = collect(Cint, Ainds)
modeB = collect(Cint, Binds)
modeC = collect(Cint, Cinds)
@@ -87,17 +95,19 @@ function plan_contraction(
desc = Ref{cutensorOperationDescriptor_t}()
- cutensorCreateBlockSparseContraction(handle(),
- desc,
- descA, modeA, opA,
- descB, modeB, opB,
- descC, modeC, opC,
- descC, modeC, actual_compute_type)
+ cutensorCreateBlockSparseContraction(
+ handle(),
+ desc,
+ descA, modeA, opA,
+ descB, modeB, opB,
+ descC, modeC, opC,
+ descC, modeC, actual_compute_type
+ )
plan_pref = Ref{cutensorPlanPreference_t}()
cutensorCreatePlanPreference(handle(), plan_pref, algo, jit)
- plan = CuTensorPlan(desc[], plan_pref[]; workspacePref=workspace)
+ plan = CuTensorPlan(desc[], plan_pref[]; workspacePref = workspace)
# cutensorDestroyOperationDescriptor(desc[])
cutensorDestroyPlanPreference(plan_pref[])
return plan
diff --git a/lib/cutensor/src/blocksparse/types.jl b/lib/cutensor/src/blocksparse/types.jl
index 292dc4d00..41cbebdbd 100644
--- a/lib/cutensor/src/blocksparse/types.jl
+++ b/lib/cutensor/src/blocksparse/types.jl
@@ -12,20 +12,26 @@ mutable struct CuTensorBS{T, N}
## This expects a Vector{Tuple(Int)} right now
nonzero_block_coords
- function CuTensorBS{T, N}(nonzero_data::Vector{<:CuArray},
- blocks_per_mode::Vector{Int}, block_extents, nonzero_block_coords, inds::Vector) where {T<:Number, N}
+ function CuTensorBS{T, N}(
+ nonzero_data::Vector{<:CuArray},
+ blocks_per_mode::Vector{Int}, block_extents, nonzero_block_coords, inds::Vector
+ ) where {T <: Number, N}
CuArrayT = eltype(nonzero_data)
@assert eltype(CuArrayT) == T
# @assert ndims(CuArrayT) == N
@assert length(block_extents) == N
- new(nonzero_data, inds, blocks_per_mode, block_extents, nonzero_block_coords)
+ return new(nonzero_data, inds, blocks_per_mode, block_extents, nonzero_block_coords)
end
end
-function CuTensorBS(nonzero_data::Vector{<:CuArray{T}},
- blocks_per_mode, block_extents, nonzero_block_coords, inds::Vector) where {T<:Number}
- CuTensorBS{T,length(block_extents)}(nonzero_data,
- blocks_per_mode, block_extents, nonzero_block_coords, inds)
+function CuTensorBS(
+ nonzero_data::Vector{<:CuArray{T}},
+ blocks_per_mode, block_extents, nonzero_block_coords, inds::Vector
+ ) where {T <: Number}
+ return CuTensorBS{T, length(block_extents)}(
+ nonzero_data,
+ blocks_per_mode, block_extents, nonzero_block_coords, inds
+ )
end
# array interface
function Base.size(T::CuTensorBS)
@@ -39,8 +45,8 @@ Base.strides(T::CuTensorBS) = vcat([[st...] for st in strides.(T.nonzero_data)].
Base.eltype(T::CuTensorBS) = eltype(eltype(T.nonzero_data))
function block_extents(T::CuTensorBS)
- extents = Vector{Int64}()
-
+ extents = Vector{Int64}()
+
for ex in T.block_extents
extents = vcat(extents, ex...)
end
@@ -66,18 +72,21 @@ mutable struct CuTensorBSDescriptor
handle::cutensorBlockSparseTensorDescriptor_t
# inner constructor handles creation and finalizer of the descriptor
function CuTensorBSDescriptor(
- numModes,
- numNonZeroBlocks,
- numSectionsPerMode,
- extent,
- nonZeroCoordinates,
- stride,
- eltype)
+ numModes,
+ numNonZeroBlocks,
+ numSectionsPerMode,
+ extent,
+ nonZeroCoordinates,
+ stride,
+ eltype
+ )
desc = Ref{cuTENSOR.cutensorBlockSparseTensorDescriptor_t}()
- cutensorCreateBlockSparseTensorDescriptor(handle(), desc,
- numModes, numNonZeroBlocks, numSectionsPerMode, extent, nonZeroCoordinates,
- stride, eltype)
+ cutensorCreateBlockSparseTensorDescriptor(
+ handle(), desc,
+ numModes, numNonZeroBlocks, numSectionsPerMode, extent, nonZeroCoordinates,
+ stride, eltype
+ )
obj = new(desc[])
finalizer(unsafe_destroy!, obj)
@@ -86,12 +95,13 @@ mutable struct CuTensorBSDescriptor
end
function CuTensorBSDescriptor(
- numModes,
- numNonZeroBlocks,
- numSectionsPerMode,
- extent,
- nonZeroCoordinates,
- eltype)
+ numModes,
+ numNonZeroBlocks,
+ numSectionsPerMode,
+ extent,
+ nonZeroCoordinates,
+ eltype
+ )
return CuTensorBSDescriptor(numModes, numNonZeroBlocks, numSectionsPerMode, extent, nonZeroCoordinates, C_NULL, eltype)
end
@@ -101,7 +111,7 @@ Base.show(io::IO, desc::CuTensorBSDescriptor) = @printf(io, "CuTensorBSDescripto
Base.unsafe_convert(::Type{cutensorBlockSparseTensorDescriptor_t}, obj::CuTensorBSDescriptor) = obj.handle
function unsafe_destroy!(obj::CuTensorBSDescriptor)
- cutensorDestroyBlockSparseTensorDescriptor(obj)
+ return cutensorDestroyBlockSparseTensorDescriptor(obj)
end
## Descriptor function for CuTensorBS type. Please overwrite for custom objects
@@ -110,11 +120,13 @@ function CuTensorBSDescriptor(A::CuTensorBS)
numNonZeroBlocks = Int64(length(A.nonzero_block_coords))
numSectionsPerMode = collect(Int32, A.blocks_per_mode)
extent = block_extents(A)
- nonZeroCoordinates = Int32.(vcat([[x...] for x in A.nonzero_block_coords]...) .- 1)
+ nonZeroCoordinates = Int32.(vcat([[x...] for x in A.nonzero_block_coords]...) .- 1)
st = strides(A)
- dataType = eltype(A)#convert(cuTENSOR.cutensorDataType_t, eltype(A))
+ dataType = eltype(A) #convert(cuTENSOR.cutensorDataType_t, eltype(A))
## Right now assume stride is NULL. I am not sure if stride works, need to discuss with cuTENSOR team.
- CuTensorBSDescriptor(numModes, numNonZeroBlocks,
- numSectionsPerMode, extent, nonZeroCoordinates, dataType)
+ return CuTensorBSDescriptor(
+ numModes, numNonZeroBlocks,
+ numSectionsPerMode, extent, nonZeroCoordinates, dataType
+ )
end
diff --git a/lib/cutensor/src/libcutensor.jl b/lib/cutensor/src/libcutensor.jl
index b33560b72..4e7ba168d 100644
--- a/lib/cutensor/src/libcutensor.jl
+++ b/lib/cutensor/src/libcutensor.jl
@@ -545,12 +545,12 @@ end
@gcsafe_ccall libcutensor.cutensorBlockSparseContract(handle::cutensorHandle_t,
plan::cutensorPlan_t,
alpha::Ptr{Cvoid},
- A::Ptr{CuPtr{Cvoid}},
- B::Ptr{CuPtr{Cvoid}},
+ A::Ptr{CuPtr{Cvoid}},
+ B::Ptr{CuPtr{Cvoid}},
beta::Ptr{Cvoid},
- C::Ptr{CuPtr{Cvoid}},
- D::Ptr{CuPtr{Cvoid}},
- workspace::CuPtr{Cvoid},
+ C::Ptr{CuPtr{Cvoid}},
+ D::Ptr{CuPtr{Cvoid}},
+ workspace::CuPtr{Cvoid},
workspaceSize::UInt64,
stream::cudaStream_t)::cutensorStatus_t
end
diff --git a/lib/cutensor/test/contractions.jl b/lib/cutensor/test/contractions.jl
index 636600a74..baf56949a 100644
--- a/lib/cutensor/test/contractions.jl
+++ b/lib/cutensor/test/contractions.jl
@@ -188,62 +188,73 @@ end
end
end
-eltypes_compact = [
- (Float32, Float32, Float32, Float32),
- (ComplexF32, ComplexF32, ComplexF32, Float32),
- (Float64, Float64, Float64, Float64),
- (ComplexF64, ComplexF64, ComplexF64, Float64)
-]
-@testset "Blocksparse Contraction" begin
- ## There are many unsupported types because this is a new functionality
- ## So I will test with Float32 and ComplexF32 only
- @testset for (eltyA, eltyB, eltyC, eltyCompute) in eltypes_compact
- ## i = [20,20,25]
- ## k = [10,10,15]
- ## l = [30,30,35]
- ## A = Tensor(k,i,l)
- ## Nonzero blocks are
- ## [1,1,1], [1,1,3], [1,3,1], [1,3,3], [3,1,1], [3,1,3], [3,3,1], [3,3,3]
- A = Vector{CuArray{eltyA, 3}}()
- for k in [10,15]
- for i in [20,25]
- for l in [30,35]
- push!(A, CuArray(ones(eltyA, k,i,l)))
+ eltypes_compact = [
+ (Float32, Float32, Float32, Float32),
+ (ComplexF32, ComplexF32, ComplexF32, Float32),
+ (Float64, Float64, Float64, Float64),
+ (ComplexF64, ComplexF64, ComplexF64, Float64),
+ ]
+ @testset "Blocksparse Contraction" begin
+ ## There are many unsupported types because this is a new functionality
+ ## So I will test with Float32 and ComplexF32 only
+ @testset for (eltyA, eltyB, eltyC, eltyCompute) in eltypes_compact
+ ## i = [20,20,25]
+ ## k = [10,10,15]
+ ## l = [30,30,35]
+ ## A = Tensor(k,i,l)
+ ## Nonzero blocks are
+ ## [1,1,1], [1,1,3], [1,3,1], [1,3,3], [3,1,1], [3,1,3], [3,3,1], [3,3,3]
+ A = Vector{CuArray{eltyA, 3}}()
+ for k in [10, 15]
+ for i in [20, 25]
+ for l in [30, 35]
+ push!(A, CuArray(ones(eltyA, k, i, l)))
+ end
end
end
- end
- ## B = Tensor(k,l)
- ## Nonzero blocks are
- ## [1,1], [2,3]
- B = Array{CuArray{eltyB, 2}}(
- [CuArray(randn(eltyB, 10, 30)),
- CuArray(randn(eltyB, 10, 35))])
-
- ## C = Tensor(i)
- ## Nonzero blocks are
- ## [1,], [3,]
- C = Vector{CuArray{eltyC, 1}}(
- [CuArray(zeros(eltyC, 20)),
- CuArray(zeros(eltyC, 25))]
- )
-
- cuTenA = cuTENSOR.CuTensorBS(A, [3,3,3],
- [(10,10,15), (20,20,25), (30,30,35)],
- [(1,1,1), (1,1,3), (1,3,1), (1,3,3), (3,1,1), (3,1,3), (3,3,1), (3,3,3)],
- [1,3,2])
- cuTenB = cuTENSOR.CuTensorBS(B, [3,3],
- [(10,10,15), (30,30,35)],
- [(1,1),(2,3)], [1,2], )
- cuTenC = cuTENSOR.CuTensorBS(C, [3],
- [(20,20,25)],[(1,),(3,)], [3])
-
- mul!(cuTenC, cuTenA, cuTenB, 1, 0)
- ## C[1] = A[1,1,1] * B[1,1]
- @test C[1] ≈ reshape(permutedims(A[1], (2,1,3)), (20, 10 * 30)) * reshape(B[1], (10 * 30))
- ## C[3] = A[1,3,1] * B[1,1]
- @test C[2] ≈ reshape(permutedims(A[3], (2,1,3)), (25, 10 * 30)) * reshape(B[1], (10 * 30))
+ ## B = Tensor(k,l)
+ ## Nonzero blocks are
+ ## [1,1], [2,3]
+ B = Array{CuArray{eltyB, 2}}(
+ [
+ CuArray(randn(eltyB, 10, 30)),
+ CuArray(randn(eltyB, 10, 35)),
+ ]
+ )
+
+ ## C = Tensor(i)
+ ## Nonzero blocks are
+ ## [1,], [3,]
+ C = Vector{CuArray{eltyC, 1}}(
+ [
+ CuArray(zeros(eltyC, 20)),
+ CuArray(zeros(eltyC, 25)),
+ ]
+ )
+
+ cuTenA = cuTENSOR.CuTensorBS(
+ A, [3, 3, 3],
+ [(10, 10, 15), (20, 20, 25), (30, 30, 35)],
+ [(1, 1, 1), (1, 1, 3), (1, 3, 1), (1, 3, 3), (3, 1, 1), (3, 1, 3), (3, 3, 1), (3, 3, 3)],
+ [1, 3, 2]
+ )
+ cuTenB = cuTENSOR.CuTensorBS(
+ B, [3, 3],
+ [(10, 10, 15), (30, 30, 35)],
+ [(1, 1), (2, 3)], [1, 2],
+ )
+ cuTenC = cuTENSOR.CuTensorBS(
+ C, [3],
+ [(20, 20, 25)], [(1,), (3,)], [3]
+ )
+
+ mul!(cuTenC, cuTenA, cuTenB, 1, 0)
+ ## C[1] = A[1,1,1] * B[1,1]
+ @test C[1] ≈ reshape(permutedims(A[1], (2, 1, 3)), (20, 10 * 30)) * reshape(B[1], (10 * 30))
+ ## C[3] = A[1,3,1] * B[1,1]
+ @test C[2] ≈ reshape(permutedims(A[3], (2, 1, 3)), (25, 10 * 30)) * reshape(B[1], (10 * 30))
+ end
end
-end
end |
Contributor
Author
|
There were some issues in the Clang.jl's conversion of the cuTENSOR.h file into Julia wrapper functions. Specifically I had a runtime issue when trying to convert arrays of cuarray into |
…mp5VT/CUDA.jl into kmp5/feature/wrap_blocksparse_cutensor
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## master #3057 +/- ##
==========================================
- Coverage 16.57% 16.41% -0.16%
==========================================
Files 120 123 +3
Lines 9586 9678 +92
==========================================
Hits 1589 1589
- Misses 7997 8089 +92 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
CUDA.jl Benchmarks
Details
| Benchmark suite | Current: f6bd83a | Previous: 22a3b2c | Ratio |
|---|---|---|---|
array/accumulate/Float32/1d |
100654 ns |
101179 ns |
0.99 |
array/accumulate/Float32/dims=1 |
76488.5 ns |
76332 ns |
1.00 |
array/accumulate/Float32/dims=1L |
1585634.5 ns |
1585653 ns |
1.00 |
array/accumulate/Float32/dims=2 |
143807.5 ns |
143537.5 ns |
1.00 |
array/accumulate/Float32/dims=2L |
657780 ns |
657611 ns |
1.00 |
array/accumulate/Int64/1d |
118415 ns |
118480 ns |
1.00 |
array/accumulate/Int64/dims=1 |
80053 ns |
79696 ns |
1.00 |
array/accumulate/Int64/dims=1L |
1694561 ns |
1706346 ns |
0.99 |
array/accumulate/Int64/dims=2 |
155872 ns |
156785 ns |
0.99 |
array/accumulate/Int64/dims=2L |
962013 ns |
961500 ns |
1.00 |
array/broadcast |
20151 ns |
20238 ns |
1.00 |
array/construct |
1230.25 ns |
1254.35 ns |
0.98 |
array/copy |
18017 ns |
18020 ns |
1.00 |
array/copyto!/cpu_to_gpu |
214317 ns |
212483.5 ns |
1.01 |
array/copyto!/gpu_to_cpu |
283467 ns |
281275 ns |
1.01 |
array/copyto!/gpu_to_gpu |
10772 ns |
10671 ns |
1.01 |
array/iteration/findall/bool |
135454 ns |
134806 ns |
1.00 |
array/iteration/findall/int |
150788 ns |
149640 ns |
1.01 |
array/iteration/findfirst/bool |
81312 ns |
81104 ns |
1.00 |
array/iteration/findfirst/int |
83536.5 ns |
83006 ns |
1.01 |
array/iteration/findmin/1d |
87117 ns |
83919 ns |
1.04 |
array/iteration/findmin/2d |
117060 ns |
116144 ns |
1.01 |
array/iteration/logical |
199068.5 ns |
197753 ns |
1.01 |
array/iteration/scalar |
66508 ns |
67066 ns |
0.99 |
array/permutedims/2d |
52271 ns |
52114 ns |
1.00 |
array/permutedims/3d |
52323 ns |
52672 ns |
0.99 |
array/permutedims/4d |
51254 ns |
51236 ns |
1.00 |
array/random/rand/Float32 |
13399 ns |
13391 ns |
1.00 |
array/random/rand/Int64 |
24921 ns |
25096 ns |
0.99 |
array/random/rand!/Float32 |
8705.666666666666 ns |
9978 ns |
0.87 |
array/random/rand!/Int64 |
21748 ns |
21790 ns |
1.00 |
array/random/randn/Float32 |
36984.5 ns |
43502 ns |
0.85 |
array/random/randn!/Float32 |
30804 ns |
30851 ns |
1.00 |
array/reductions/mapreduce/Float32/1d |
34371 ns |
34466 ns |
1.00 |
array/reductions/mapreduce/Float32/dims=1 |
39968.5 ns |
48755 ns |
0.82 |
array/reductions/mapreduce/Float32/dims=1L |
51502 ns |
51181 ns |
1.01 |
array/reductions/mapreduce/Float32/dims=2 |
56453 ns |
56334 ns |
1.00 |
array/reductions/mapreduce/Float32/dims=2L |
69257 ns |
69218 ns |
1.00 |
array/reductions/mapreduce/Int64/1d |
42013.5 ns |
41729 ns |
1.01 |
array/reductions/mapreduce/Int64/dims=1 |
41741 ns |
43165.5 ns |
0.97 |
array/reductions/mapreduce/Int64/dims=1L |
86937 ns |
87029.5 ns |
1.00 |
array/reductions/mapreduce/Int64/dims=2 |
59147 ns |
59231 ns |
1.00 |
array/reductions/mapreduce/Int64/dims=2L |
84663 ns |
84493 ns |
1.00 |
array/reductions/reduce/Float32/1d |
34553 ns |
34454 ns |
1.00 |
array/reductions/reduce/Float32/dims=1 |
39546.5 ns |
48965 ns |
0.81 |
array/reductions/reduce/Float32/dims=1L |
51569.5 ns |
51071 ns |
1.01 |
array/reductions/reduce/Float32/dims=2 |
56571.5 ns |
56187 ns |
1.01 |
array/reductions/reduce/Float32/dims=2L |
69904 ns |
69359.5 ns |
1.01 |
array/reductions/reduce/Int64/1d |
42565 ns |
41960 ns |
1.01 |
array/reductions/reduce/Int64/dims=1 |
42029 ns |
43285 ns |
0.97 |
array/reductions/reduce/Int64/dims=1L |
86852 ns |
86905 ns |
1.00 |
array/reductions/reduce/Int64/dims=2 |
59436 ns |
58992 ns |
1.01 |
array/reductions/reduce/Int64/dims=2L |
84430 ns |
84104 ns |
1.00 |
array/reverse/1d |
17564 ns |
17615 ns |
1.00 |
array/reverse/1dL |
68187 ns |
68235 ns |
1.00 |
array/reverse/1dL_inplace |
65614 ns |
65656 ns |
1.00 |
array/reverse/1d_inplace |
8381.333333333334 ns |
10268.666666666666 ns |
0.82 |
array/reverse/2d |
20639 ns |
20767 ns |
0.99 |
array/reverse/2dL |
72749 ns |
72880 ns |
1.00 |
array/reverse/2dL_inplace |
65690 ns |
65670 ns |
1.00 |
array/reverse/2d_inplace |
9751 ns |
9888 ns |
0.99 |
array/sorting/1d |
2735249 ns |
2736190 ns |
1.00 |
array/sorting/2d |
1069013 ns |
1068430 ns |
1.00 |
array/sorting/by |
3304195 ns |
3304193 ns |
1.00 |
cuda/synchronization/context/auto |
1125 ns |
1126.4 ns |
1.00 |
cuda/synchronization/context/blocking |
895 ns |
931.6315789473684 ns |
0.96 |
cuda/synchronization/context/nonblocking |
8038.700000000001 ns |
7346.299999999999 ns |
1.09 |
cuda/synchronization/stream/auto |
982.125 ns |
1042.6923076923076 ns |
0.94 |
cuda/synchronization/stream/blocking |
818.2666666666667 ns |
801.5346534653465 ns |
1.02 |
cuda/synchronization/stream/nonblocking |
7069.6 ns |
7426.700000000001 ns |
0.95 |
integration/byval/reference |
143589 ns |
143714 ns |
1.00 |
integration/byval/slices=1 |
145463 ns |
145836 ns |
1.00 |
integration/byval/slices=2 |
284181 ns |
284663 ns |
1.00 |
integration/byval/slices=3 |
422818.5 ns |
423042 ns |
1.00 |
integration/cudadevrt |
102326.5 ns |
102219 ns |
1.00 |
integration/volumerhs |
23418364 ns |
23416498.5 ns |
1.00 |
kernel/indexing |
13103 ns |
13008 ns |
1.01 |
kernel/indexing_checked |
13836 ns |
13732 ns |
1.01 |
kernel/launch |
2072.6 ns |
2108.4444444444443 ns |
0.98 |
kernel/occupancy |
697.6981132075472 ns |
665.65 ns |
1.05 |
kernel/rand |
14208 ns |
17312 ns |
0.82 |
latency/import |
3812714028.5 ns |
3818709770 ns |
1.00 |
latency/precompile |
4583803742.5 ns |
4584771622 ns |
1.00 |
latency/ttfp |
4394446986.5 ns |
4399290991.5 ns |
1.00 |
This comment was automatically generated by workflow using github-action-benchmark.
Member
|
Thanks very much for putting this together, I'm happy to help with the header issues if needed! |
…but the C++ code is still in flux)
Contributor
Author
|
@kshyatt I removed the extra code, made the functions that linked to the library relatively agnostic (i.e. you are not forced to use CuTensorBS but can buy in if you'd like) and added a unit test. If you could help with the Clang.jl issue, that would be amazing! |
Member
|
I'll try to take a look today! |
Member
|
Did you use the scripts in |
…mp5VT/CUDA.jl into kmp5/feature/wrap_blocksparse_cutensor
Contributor
Author
Yes I did use the scripts but this produced the ERROR: MethodError: no method matching unsafe_convert(::Type{Ptr{Nothing}}, ::CuPtr{Nothing})
The function `unsafe_convert` exists, but no method is defined for this combination of argument types.
Closest candidates are:
unsafe_convert(::Type{Ptr{Nothing}}, ::LibGit2.GitBlame)
@ LibGit2 ~/.julia/juliaup/julia-1.12.1+0.x64.linux.gnu/share/julia/stdlib/v1.12/LibGit2/src/types.jl:1096
unsafe_convert(::Type{Ptr{Nothing}}, ::LibGit2.GitRevWalker)
@ LibGit2 ~/.julia/juliaup/julia-1.12.1+0.x64.linux.gnu/share/julia/stdlib/v1.12/LibGit2/src/types.jl:1096
unsafe_convert(::Type{Ptr{Nothing}}, ::LibGit2.GitDiffStats)
@ LibGit2 ~/.julia/juliaup/julia-1.12.1+0.x64.linux.gnu/share/julia/stdlib/v1.12/LibGit2/src/types.jl:1096
...
Stacktrace:
[1] Ref{Ptr{Nothing}}(a::Vector{CuPtr{Nothing}})
@ Base ./refpointer.jl:166
[2] cconvert
@ ./refpointer.jl:178 [inlined]
[3] macro expansion
@ ~/.julia/dev/CUDA.jl/lib/cutensor/src/libcutensor.jl:545 [inlined]
[4] (::cuTENSOR.var"#cutensorBlockSparseContract##0#cutensorBlockSparseContract##1"{…})()
@ cuTENSOR ~/.julia/packages/GPUToolbox/JLBB1/src/ccalls.jl:34
[5] retry_reclaim
@ ~/.julia/packages/CUDA/Il00B/src/memory.jl:434 [inlined]
[6] check
@ ~/.julia/dev/CUDA.jl/lib/cutensor/src/libcutensor.jl:22 [inlined]
[7] cutensorBlockSparseContract
@ ~/.julia/packages/GPUToolbox/JLBB1/src/ccalls.jl:33 [inlined]
[8]
@ cuTENSOR ~/.julia/dev/CUDA.jl/lib/cutensor/src/blocksparse/operations.jl:50
[9] contract!(alpha::Number, A::Any, Ainds::Vector{…}, opA::cuTENSOR.cutensorOperator_t, B::Any, Binds::Vector{…}, opB::cuTENSOR.cutensorOperator_t, beta::Number, C::Any, Cinds::Vector{…}, opC::cuTENSOR.cutensorOperator_t, opOut::cuTENSOR.cutensorOperator_t; jit::cuTENSOR.cutensorJitMode_t, workspace::cuTENSOR.cutensorWorksizePreference_t, algo::cuTENSOR.cutensorAlgo_t, compute_type::Nothing, plan::Nothing)
@ cuTENSOR ~/.julia/dev/CUDA.jl/lib/cutensor/src/blocksparse/operations.jl:25
[10] mul!(C::CuTensorBS{Float64, 1}, A::CuTensorBS{Float64, 3}, B::CuTensorBS{Float64, 2}, α::Float64, β::Float64)
@ cuTENSOR ~/.julia/dev/CUDA.jl/lib/cutensor/src/blocksparse/interfaces.jl:21However, I found that If I modify the code to be |
Member
|
Probably you missed some of the weird esoterica in |
kshyatt
force-pushed
the
kmp5/feature/wrap_blocksparse_cutensor
branch
from
17806da to
cc4b826
Compare
lkdvos
reviewed
Mar 26, 2026
Contributor
lkdvos
left a comment
lkdvos
left a comment
There was a problem hiding this comment.
Left some remaining comments, but for me I think most of the parts that I would use are there, since I don't really see myself going through the CuTensorBS construction (we also never used the CuTensor in TensorOperations so that is completely fine)
Remove left over code. Will need to make something like this to define mul! in the future Co-authored-by: Lukas Devos <ldevos98@gmail.com>
…mp5VT/CUDA.jl into kmp5/feature/wrap_blocksparse_cutensor
…to a contigous memory block)
Contributor
Author
|
@kshyatt |
kshyatt
reviewed
Apr 22, 2026
kshyatt
reviewed
Apr 22, 2026
kshyatt
reviewed
Apr 22, 2026
kshyatt
approved these changes
Apr 22, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hi,
This is a wrapper type and functions to access the newly introduced blocksparse cutensor backend. Right now the code is expert level, i.e. users need to write a type that converts their object to CuTensorBS types or can achieve the low-level operations required by cutensor kernels. I am still writing a test but the code is fully operational.
Thanks,
Karl