From f4f39ec47c170bbc29bc13ee4c2d24df979cf687 Mon Sep 17 00:00:00 2001
From: Gabriel Kronberger <gabriel.kronberger@heuristiclab.com>
Date: Wed, 19 Feb 2025 17:18:43 +0100
Subject: [PATCH] Improvements / fixes.

---
 package/src/ExpressionExecutorCuda.jl | 11 +++++++++--
 package/test/CpuInterpreterTests.jl   | 12 ++++++++----
 2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/package/src/ExpressionExecutorCuda.jl b/package/src/ExpressionExecutorCuda.jl
index 3296e7c..2192dcf 100644
--- a/package/src/ExpressionExecutorCuda.jl
+++ b/package/src/ExpressionExecutorCuda.jl
@@ -29,7 +29,7 @@ end
 
 
 # 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)
 	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))
 
 	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
 
 	res
diff --git a/package/test/CpuInterpreterTests.jl b/package/test/CpuInterpreterTests.jl
index 591ef21..356a4a6 100644
--- a/package/test/CpuInterpreterTests.jl
+++ b/package/test/CpuInterpreterTests.jl
@@ -1,3 +1,4 @@
+using LinearAlgebra
 
 function test_cpu_interpreter(nrows; parallel = false)
     exprs = [
@@ -13,12 +14,15 @@ function test_cpu_interpreter(nrows; parallel = false)
     
     # warmup
     interpret_cpu(exprs, X, p)
+    expr_reps = 100 # for each expr
+    reps= 100
+
     if parallel 
-        t_sec = @elapsed Threads.@threads :static for i in 1:100 interpret_cpu(exprs, X, p) end
-        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))")
+        t_sec = @elapsed fetch.([Threads.@spawn interpret_cpu(exprs, X, p; repetitions=expr_reps) for i in 1:reps])
+        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
-        t_sec = @elapsed for i in 1:100 interpret_cpu(exprs, X, p) 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))")
+        t_sec = @elapsed for i in 1:reps interpret_cpu(exprs, X, p; repetitions=expr_reps) end
+        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
     true
 end