Errors on small array inputs

[GiggleLiu]

The BLAS.gemmEx! function errors on array size <128. I want to experiment the functionality on small array size like 16x16.

Error message

ERROR: LoadError: ArgumentError: Grid dimensions should be non-null
Stacktrace:
 [1] launch(::CuFunction, ::CuDeviceArray{Float16,2,1}, ::CuDeviceArray{Float16,2,1}, ::CuDeviceArray{Float32,2,1}, ::CuDeviceArray{Float32,2,1}, ::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#1#3"{Int64,Int64}}, ::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#2#4"{Int64}}; blocks::Tuple{Int64,Int64}, threads::Int64, cooperative::Bool, shmem::Int64, stream::CuStream) at /home/leo/.julia/dev/CUDA/lib/cudadrv/execution.jl:57
 [2] #599 at /home/leo/.julia/dev/CUDA/lib/cudadrv/execution.jl:138 [inlined]
 [3] macro expansion at /home/leo/.julia/dev/CUDA/lib/cudadrv/execution.jl:97 [inlined]
 [4] convert_arguments at /home/leo/.julia/dev/CUDA/lib/cudadrv/execution.jl:79 [inlined]
 [5] #cudacall#598 at /home/leo/.julia/dev/CUDA/lib/cudadrv/execution.jl:137 [inlined]
 [6] #cudacall#790 at /home/leo/.julia/dev/CUDA/src/compiler/execution.jl:219 [inlined]
 [7] macro expansion at /home/leo/.julia/dev/CUDA/src/compiler/execution.jl:200 [inlined]
 [8] call(::CUDA.HostKernel{GemmKernels.Kernel.matmul_pipelined,Tuple{CuDeviceArray{Float16,2,1},CuDeviceArray{Float16,2,1},CuDeviceArray{Float32,2,1},CuDeviceArray{Float32,2,1},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#1#3"{Int64,Int64}},GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#2#4"{Int64}},GemmKernels.Epilogue.Default,Type{GemmKernels.Config{(M = 64, N = 64, K = 64),(M = 128, N = 128, K = 64),8,(M = 128, K = 2),(M = 8, K = 1),(K = 64, N = 4),(K = 8, N = 1),(M = 128, N = 1),(M = 4, N = 1),(M = 32, N = 64, K = 16),(M = 16, N = 16, K = 16),GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},WMMAOp{16,16,16},true,true}}}}, ::CuDeviceArray{Float16,2,1}, ::CuDeviceArray{Float16,2,1}, ::CuDeviceArray{Float32,2,1}, ::CuDeviceArray{Float32,2,1}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{typeof(identity)}, ::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#1#3"{Int64,Int64}}, ::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#2#4"{Int64}}, ::GemmKernels.Epilogue.Default, ::Type{GemmKernels.Config{(M = 64, N = 64, K = 64),(M = 128, N = 128, K = 64),8,(M = 128, K = 2),(M = 8, K = 1),(K = 64, N = 4),(K = 8, N = 1),(M = 128, N = 1),(M = 4, N = 1),(M = 32, N = 64, K = 16),(M = 16, N = 16, K = 16),GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},WMMAOp{16,16,16},true,true}}; call_kwargs::Base.Iterators.Pairs{Symbol,Any,Tuple{Symbol,Symbol,Symbol},NamedTuple{(:threads, :blocks, :shmem),Tuple{Int64,Tuple{Int64,Int64},Int64}}}) at /home/leo/.julia/dev/CUDA/src/compiler/execution.jl:171
 [9] (::CUDA.HostKernel{GemmKernels.Kernel.matmul_pipelined,Tuple{CuDeviceArray{Float16,2,1},CuDeviceArray{Float16,2,1},CuDeviceArray{Float32,2,1},CuDeviceArray{Float32,2,1},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{typeof(identity)},GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#1#3"{Int64,Int64}},GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#2#4"{Int64}},GemmKernels.Epilogue.Default,Type{GemmKernels.Config{(M = 64, N = 64, K = 64),(M = 128, N = 128, K = 64),8,(M = 128, K = 2),(M = 8, K = 1),(K = 64, N = 4),(K = 8, N = 1),(M = 128, N = 1),(M = 4, N = 1),(M = 32, N = 64, K = 16),(M = 16, N = 16, K = 16),GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float16},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.Padded{GemmKernels.Layout.AlignedColMajor{Float16},8},GemmKernels.Layout.AlignedColMajor{Float32},GemmKernels.Layout.AlignedColMajor{Float32},WMMAOp{16,16,16},true,true}}}})(::CuDeviceArray{Float16,2,1}, ::Vararg{Any,N} where N; kwargs::Base.Iterators.Pairs{Symbol,Any,Tuple{Symbol,Symbol,Symbol},NamedTuple{(:threads, :blocks, :shmem),Tuple{Int64,Tuple{Int64,Int64},Int64}}}) at /home/leo/.julia/dev/CUDA/src/compiler/execution.jl:353
 [10] matmul(::CuArray{Float16,2}, ::CuArray{Float16,2}, ::CuArray{Float32,2}, ::CuArray{Float32,2}, ::Type{T} where T; transform_global_to_shared_a::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_global_to_shared_b::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_global_to_shared_c::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_shared_to_global_d::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_shared_to_regs_a::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_shared_to_regs_b::GemmKernels.Transform.Elementwise{typeof(identity)}, transform_shared_to_regs_c::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#1#3"{Int64,Int64}}, transform_regs_to_shared_d::GemmKernels.Transform.Elementwise{GemmKernels.BLAS.var"#2#4"{Int64}}, epilogue::GemmKernels.Epilogue.Default, kernel::typeof(GemmKernels.Kernel.matmul_pipelined)) at /home/leo/.julia/dev/GemmKernels/src/launch.jl:26
 [11] gemmEx!(::Char, ::Char, ::Int64, ::CuArray{Float16,2}, ::CuArray{Float16,2}, ::Int64, ::CuArray{Float32,2}) at /home/leo/.julia/dev/GemmKernels/src/blas.jl:64
 [12] top-level scope at /home/leo/.julia/dev/GemmKernels/matmul.jl:42
 [13] include_string(::Function, ::Module, ::String, ::String) at ./loading.jl:1091
 [14] invokelatest(::Any, ::Any, ::Vararg{Any,N} where N; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at ./essentials.jl:710
 [15] invokelatest(::Any, ::Any, ::Vararg{Any,N} where N) at ./essentials.jl:709
 [16] inlineeval(::Module, ::String, ::Int64, ::Int64, ::String; softscope::Bool) at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/eval.jl:83
 [17] (::VSCodeServer.var"#43#45"{VSCodeServer.ReplRunCodeRequestParams,String,Int64,Int64,String,Module,Bool})() at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/eval.jl:45
 [18] withpath(::VSCodeServer.var"#43#45"{VSCodeServer.ReplRunCodeRequestParams,String,Int64,Int64,String,Module,Bool}, ::String) at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/repl.jl:118
 [19] (::VSCodeServer.var"#42#44"{VSCodeServer.ReplRunCodeRequestParams,String,Int64,Int64,String,Module,Bool,Bool})() at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/eval.jl:43
 [20] hideprompt(::VSCodeServer.var"#42#44"{VSCodeServer.ReplRunCodeRequestParams,String,Int64,Int64,String,Module,Bool,Bool}) at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/repl.jl:36
 [21] repl_runcode_request(::VSCodeServer.JSONRPC.JSONRPCEndpoint{Base.PipeEndpoint,Base.PipeEndpoint}, ::VSCodeServer.ReplRunCodeRequestParams) at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/eval.jl:23
 [22] dispatch_msg(::VSCodeServer.JSONRPC.JSONRPCEndpoint{Base.PipeEndpoint,Base.PipeEndpoint}, ::VSCodeServer.JSONRPC.MsgDispatcher, ::Dict{String,Any}) at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/JSONRPC/src/typed.jl:66
 [23] macro expansion at /home/leo/.vscode/extensions/julialang.language-julia-1.0.8/scripts/packages/VSCodeServer/src/VSCodeServer.jl:95 [inlined]
 [24] (::VSCodeServer.var"#61#63"{Bool,String})() at ./task.jl:356
in expression starting at /home/leo/.julia/dev/GemmKernels/matmul.jl:42

Also, the computed value is not correct if the array size is not an exponent of

[thomasfaingnaert]

This is a limitation of the current implementation: only arrays whose size is a multiple of the threadblock size (e.g. (M = 128, N = 128, K = 64) for WMMA mixed-precision) are supported at the moment.
One way to support arbitrary matrix dimensions would be to predicate the loads from global memory to only access elements inside the bounds of the global matrix.