benchmarking: started tuning benchmarking results. found some errors that need fixing

package/src/ExpressionExecutorCuda.jl

@@ -58,21 +58,34 @@ end
 
 
 # Evaluate Expressions on the CPU
-function interpret_cpu(exprs::Vector{Expr}, X::Matrix{Float32}, p::Vector{Vector{Float32}}; repetitions=1)::Matrix{Float32}
+function interpret_cpu(exprs::Vector{Expr}, X::Matrix{Float32}, p::Vector{Vector{Float32}}; repetitions=1, parallel=false)::Matrix{Float32}
 	@assert axes(exprs) == axes(p)
 	nrows = size(X, 1)
 	
 	# each column of the matrix has the result for an expr
 	res = Matrix{Float32}(undef, nrows, length(exprs))
 
-	for i in eachindex(exprs) 
-		# 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])
+	if parallel
+		Threads.@threads for i in eachindex(exprs) 
+			# 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
+	else
+		for i in eachindex(exprs) 
+			# 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
 

package/src/ExpressionProcessing.jl

@@ -100,7 +100,7 @@ function get_operator(op::Symbol)::Operator
 	elseif op == :sqrt
 		return SQRT
 	else
-		throw("Operator unknown")
+		throw("Operator unknown. Operator was $op")
 	end
 end
 

package/test/CpuInterpreterTests.jl

@@ -65,11 +65,6 @@ function test_cpu_interpreter_nikuradse()
 		i = 0
         for line in eachline(io)
             expr, p = parse_infix(line, varnames, paramnames)
-            
-			if i > 10
-				return
-			end
-			println(expr)
 
             push!(exprs, expr)
             push!(parameters, randn(Float32, length(p)))

package/test/PerformanceTests.jl

@@ -1,9 +1,13 @@
 using LinearAlgebra
 using BenchmarkTools
+using DelimitedFiles
+using GZip
 
 using .Transpiler
 using .Interpreter
 
+include("parser.jl") # to parse expressions from a file
+
 const BENCHMARKS_RESULTS_PATH = "./results-fh-new"
 
 # Number of expressions can get really big (into millions)
@@ -11,6 +15,7 @@ const BENCHMARKS_RESULTS_PATH = "./results-fh-new"
 
 data,varnames = readdlm("data/nikuradse_1.csv", ',', header=true);
 X = convert(Matrix{Float32}, data)
+X_t = permutedims(X) # for gpu
 
 exprs = Expr[]
 parameters = Vector{Vector{Float32}}()
@@ -19,24 +24,15 @@ paramnames = ["p$i" for i in 1:20]
 # data/esr_nvar2_len10.txt.gz_9.txt.gz has  ~250_000 exprs
 # data/esr_nvar2_len10.txt.gz_10.txt.gz has ~800_000 exrps
 GZip.open("data/esr_nvar2_len10.txt.gz_9.txt.gz") do io 
-	i = 0
 	for line in eachline(io)
 		expr, p = parse_infix(line, varnames, paramnames)
-		
-		if i > 10
-			return
-		end
-		println(expr)
 
 		push!(exprs, expr)
 		push!(parameters, randn(Float32, length(p)))
-
-		i += 1
 	end
 end
 expr_reps = 100 # 100 parameter optimisation steps (local search; sequentially; only p changes but not X)
 
-
 # TODO: Tipps for tuning:
 	# Put data in shared memory: 
 	# https://cuda.juliagpu.org/v2.6/api/kernel/#Shared-memory
@@ -60,18 +56,20 @@ suite["GPUI"] = BenchmarkGroup(["GPUInterpreter"])
 suite["GPUT"] = BenchmarkGroup(["GPUTranspiler"])
 
 if compareWithCPU
-	suite["CPU"]["nikuradse_1"] = @benchmarkable interpret_cpu(exprsCPU, X, parameters; repetitions=expr_reps)
+	suite["CPU"]["nikuradse_1"] = @benchmarkable interpret_cpu(exprs, X, parameters; repetitions=expr_reps)
+	suite["CPU"]["nikuradse_1_parallel"] = @benchmarkable interpret_cpu(exprs, X, parameters; repetitions=expr_reps, parallel=true)
 end
 
-# TODO: Most likely need to transpose X matrix here, as we are expecting a column major matrix for more efficient memory access
-suite["GPUI"]["nikuradse_1"] = @benchmarkable interpret_gpu(exprsGPU, X, parameters; repetitions=expr_reps)
-suite["GPUT"]["nikuradse_1"] = @benchmarkable evaluate_gpu(exprsGPU, X, parameters; repetitions=expr_reps)
+suite["GPUI"]["nikuradse_1"] = @benchmarkable interpret_gpu(exprs, X_t, parameters; repetitions=expr_reps)
+suite["GPUT"]["nikuradse_1"] = @benchmarkable evaluate_gpu(exprs, X_t, parameters; repetitions=expr_reps)
 
-for i in 1:10
+for i in 1:2
 	tune!(suite)
 end
 BenchmarkTools.save("params.json", params(suite))
 
+throw("finished tuning")
+
 loadparams!(suite, BenchmarkTools.load("params.json")[1], :samples, :evals, :gctrial, :time_tolerance, :evals_set, :gcsample, :seconds, :overhead, :memory_tolerance)
 
 results = run(suite, verbose=true, seconds=3600) # 1 hour because of CPU. lets see if more is needed

package/test/runtests.jl

@@ -10,9 +10,9 @@ include(joinpath(baseFolder, "src", "Interpreter.jl"))
 include(joinpath(baseFolder, "src", "Transpiler.jl"))
 
 @testset "Functionality tests" begin
-	include("ExpressionProcessingTests.jl")
-	include("InterpreterTests.jl")
-	include("TranspilerTests.jl")
+	# include("ExpressionProcessingTests.jl")
+	# include("InterpreterTests.jl")
+	# include("TranspilerTests.jl")
 end
 
 
@@ -22,5 +22,5 @@ end
 
 @testset "Performance tests" begin
 	# include("PerformanceTuning.jl")
-	# include("PerformanceTests.jl")
+	include("PerformanceTests.jl")
 end