MKL: Level-1 (scal) primitive support #1
Conversation
There was a problem hiding this comment.
There is a 'testf' function defined in test/setup.jl, which uses the TestSuite.compare function in
https://github.com/JuliaGPU/GPUArrays.jl/blob/master/test/testsuite.jl#L38. It copies the input data to CPU and GPU arrays and compare the results. Please modify the test using the 'testf' function for more concise code.
Thanks for your inputs. I have implemented rmul! and testf based tests they fail for one case commonly. |
1. Input array is converted to non-gpu array 2. alpha is rand instead of based on T typewq
non-gpuarray flow
| @@ -49,6 +49,13 @@ function gemm_dispatch!(C::oneStridedVecOrMat, A, B, alpha::Number=true, beta::N | |||
| end | |||
| end | |||
|
|
|||
| LinearAlgebra.rmul!(x::oneStridedVecOrMat{<:onemklFloat}, k::Number) = | |||
There was a problem hiding this comment.
I am not sure if these two rules will properly dispatch the combination of alpha types of ComplexF32, ComplexF64, F32 and F64 to the ComplexF32 and Complex64 scal functions.
Please write specific tests for all the combinations to make sure they are dispatched properly.
If not, we will want to use AMDGPU.jl's type based dispatching rules at
https://github.com/JuliaGPU/AMDGPU.jl/blob/master/src/blas/wrappers.jl#L85, and
https://github.com/JuliaGPU/AMDGPU.jl/blob/master/src/blas/wrappers.jl#L106
This PR consists of Level-1 (scal) primitive support for oneJulia