Daniel/master-thesis
1
0
Fork 0
Code Issues 5 Pull Requests Actions 1 Packages Projects 1 Releases Wiki Activity

Improvements / fixes.

Browse Source
This commit is contained in:
Gabriel Kronberger 2025-02-19 17:18:43 +01:00
parent 942adb8612
commit f4f39ec47c
2 changed files with 17 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
package/src/ExpressionExecutorCuda.jl
View File

@ -29,7 +29,7 @@ end
# Evaluate Expressions on the CPU # Evaluate Expressions on the CPU
function interpret_cpu(exprs::Vector{Expr}, X::Matrix{Float32}, p::Vector{Vector{Float32}})::Matrix{Float32} function interpret_cpu(exprs::Vector{Expr}, X::Matrix{Float32}, p::Vector{Vector{Float32}}; repetitions=1)::Matrix{Float32}
@assert axes(exprs) == axes(p) @assert axes(exprs) == axes(p)
nrows = size(X, 1) nrows = size(X, 1)
@ -37,7 +37,14 @@ function interpret_cpu(exprs::Vector{Expr}, X::Matrix{Float32}, p::Vector{Vector
res = Matrix{Float32}(undef, nrows, length(exprs)) res = Matrix{Float32}(undef, nrows, length(exprs))
for i in eachindex(exprs) for i in eachindex(exprs)
CpuInterpreter.interpret!((@view res[:,i]), exprs[i], X, p[i]) # The interpreter holds the postfix code and buffers for evaluation. It is costly to create
interpreter = CpuInterpreter.Interpreter{Float32}(exprs[i], length(p[i]))
# If an expression has to be evaluated multiple times (e.g. for different parameters),
# it is worthwhile to reuse the interpreter to reduce the number of allocations
for rep in 1:repetitions
CpuInterpreter.interpret!((@view res[:,i]), interpreter, X, p[i])
end
end end
res res

12
package/test/CpuInterpreterTests.jl
View File

@ -1,3 +1,4 @@
using LinearAlgebra
function test_cpu_interpreter(nrows; parallel = false) function test_cpu_interpreter(nrows; parallel = false)
exprs = [ exprs = [
@ -13,12 +14,15 @@ function test_cpu_interpreter(nrows; parallel = false)
# warmup # warmup
interpret_cpu(exprs, X, p) interpret_cpu(exprs, X, p)
expr_reps = 100 # for each expr
reps= 100
if parallel if parallel
t_sec = @elapsed Threads.@threads :static for i in 1:100 interpret_cpu(exprs, X, p) end t_sec = @elapsed fetch.([Threads.@spawn interpret_cpu(exprs, X, p; repetitions=expr_reps) for i in 1:reps])
println("~ $(round(30 * 100 * nrows / 1e9 / t_sec, digits=2)) GFLOPS (single-core) ($(round(peakflops(1000, eltype=Float32, ntrials=1, parallel=false) / 1e9, digits=2)) GFLOPS (peak, single-core))") println("~ $(round(30 * reps * expr_reps * nrows / 1e9 / t_sec, digits=2)) GFLOPS ($(Threads.nthreads()) threads) ($(round(peakflops(1000, eltype=Float32, ntrials=1) / 1e9, digits=2)) GFLOPS (peak, single-core))")
else else
t_sec = @elapsed for i in 1:100 interpret_cpu(exprs, X, p) end t_sec = @elapsed for i in 1:reps interpret_cpu(exprs, X, p; repetitions=expr_reps) end
println("~ $(round(30 * 100 * nrows / 1e9 / t_sec, digits=2)) GFLOPS ($(Threads.nthreads()) threads) ($(round(peakflops(1000, eltype=Float32, ntrials=1, parallel=false) / 1e9, digits=2)) GFLOPS (peak, single-core))") println("~ $(round(30 * reps * expr_reps * nrows / 1e9 / t_sec, digits=2)) GFLOPS (single-core) ($(round(peakflops(1000, eltype=Float32, ntrials=1) / 1e9, digits=2)) GFLOPS (peak, single-core))")
end end
true true
end end