Daniel/master-thesis
1
0
Fork 0
Code Issues 3 Pull Requests Actions 1 Packages Projects 1 Releases Wiki Activity

benchmarking: prepared tests for using actual data
Some checks are pending
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, 1.10) (push) Waiting to run
Details
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, 1.6) (push) Waiting to run
Details
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, pre) (push) Waiting to run
Details

Browse Source
This commit is contained in:
Daniel
2025-05-09 13:58:10 +02:00
parent 2c8a9cd2d8
commit 327e4ebf1b
3 changed files with 46 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
package/src/Transpiler.jl
View File

@ -56,8 +56,6 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
formattedExpr = ExpressionProcessing.expr_to_postfix(expressions[i], cacheFrontend)
kernel = transpile(formattedExpr, varRows, Utils.get_max_inner_length(parameters), variableCols, i-1) # i-1 because julia is 1-based but PTX needs 0-based indexing
# println(kernel)
linker = CuLink()
add_data!(linker, "ExpressionProcessing", kernel)
@ -77,7 +75,7 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
# execute each kernel (also try doing this with Threads.@threads. Since we can have multiple grids, this might improve performance)
for kernel in kernels
# config = launch_configuration(kernels[i])
threads = min(variableCols, 96)
threads = min(variableCols, 256)
blocks = cld(variableCols, threads)
cudacall(kernel, (CuPtr{Float32},CuPtr{Float32},CuPtr{Float32}), cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
@ -99,7 +97,7 @@ function transpile(expression::ExpressionProcessing.PostfixType, varSetSize::Int
ptxBuffer = IOBuffer()
regManager = Utils.RegisterManager(Dict(), Dict())
# TODO: Suboptimal solution
# TODO: Suboptimal solution. get_kernel_signature should also return the name of the registers used for the parameters, so further below, we do not have to hard-code them
signature, paramLoading = get_kernel_signature("ExpressionProcessing", [Float32, Float32, Float32], regManager) # Vars, Params, Results
guardClause, threadId64Reg = get_guard_clause(exitJumpLocationMarker, nrOfVariableSets, regManager)
@ -123,7 +121,7 @@ function transpile(expression::ExpressionProcessing.PostfixType, varSetSize::Int
return generatedCode
end
# TODO: Make version, target and address_size configurable; also see what address_size means exactly
# TODO: Make version, target and address_size configurable
function get_cuda_header()::String
return "
.version 8.5