benchmarking: changes to not fill up all of the RAM

package/src/ExpressionProcessing.jl

@@ -27,6 +27,13 @@ NOTE: This function is not thread save, especially cache access is not thread sa
 function expr_to_postfix(expression::Expr)::PostfixType
 	expr = expression
 	if expression.head === :->
+		if typeof(expression.args[2]) == Float64
+			println()
+			println("Expression: $expression")
+			println("Expr: $expr")
+			println()
+			dump(expression; maxdepth=10)
+		end
 		# if the expression equals (x, p) -> (...) then the below statement extracts the expression to evaluate
 		if expression.args[2].head == :block # expressions that are not generated with the parser (./test/parser.jl) contain this extra "block" node, which needs to be skipped
 			expr = expression.args[2].args[2] 
@@ -41,6 +48,8 @@ function expr_to_postfix(expression::Expr)::PostfixType
 
 	postfix = PostfixType()
 
+	
+
 	# Special handling in the case where the expression is an array access
 	# This can happen if the token is a variable/parameter of the form x[n]/p[n]
 	if expr.head == :ref

package/src/Transpiler.jl

@@ -15,7 +15,7 @@ const Operand = Union{Float32, String} # Operand is either fixed value or regist
 function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, parameters::Vector{Vector{Float32}})::Matrix{Float32}
 	varRows = size(variables, 1)
 	variableCols = size(variables, 2)
-	kernels = Vector{CuFunction}(undef, length(expressions))
+	# kernels = Vector{CuFunction}(undef, length(expressions))
 
 	# TODO: test this again with multiple threads. The first time I tried, I was using only one thread
 	# Test this parallel version again when doing performance tests. With the simple "functionality" tests this took 0.03 seconds while sequential took "0.00009" seconds
@@ -48,6 +48,16 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 	# 	@lock cacheLock transpilerCache[expressions[i]] = kernels[i]
 	# end
 
+	cudaVars = CuArray(variables) # maybe put in shared memory (see PerformanceTests.jl for more info)
+	cudaParams = Utils.create_cuda_array(parameters, NaN32) # maybe make constant (see PerformanceTests.jl for more info)
+
+	# each expression has nr. of variable sets (nr. of columns of the variables) results and there are n expressions
+	cudaResults = CuArray{Float32}(undef, variableCols, length(expressions))
+
+	threads = min(variableCols, 256)
+	blocks = cld(variableCols, threads)
+
+	# TODO: Implement batching as a middleground between "transpile everything and then run" and "tranpile one run one" even though cudacall is async
 	@inbounds for i in eachindex(expressions)
 		# if haskey(resultCache, expressions[i])
 		# 	kernels[i] = resultCache[expressions[i]]
@@ -64,7 +74,11 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 			image = complete(linker)
 			
 			mod = CuModule(image)
-			kernels[i] = CuFunction(mod, "ExpressionProcessing")
+
+			compiledKernel = CuFunction(mod, "ExpressionProcessing")
+			cudacall(compiledKernel, (CuPtr{Float32},CuPtr{Float32},CuPtr{Float32}), cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
+
+			# kernels[i] = CuFunction(mod, "ExpressionProcessing")
 		# 	resultCache[expressions[i]] = kernels[i]
 		# catch
 		# 	dump(expressions[i]; maxdepth=10)
@@ -78,20 +92,9 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 
 	end
 
-	cudaVars = CuArray(variables) # maybe put in shared memory (see PerformanceTests.jl for more info)
-	cudaParams = Utils.create_cuda_array(parameters, NaN32) # maybe make constant (see PerformanceTests.jl for more info)
-
-	# each expression has nr. of variable sets (nr. of columns of the variables) results and there are n expressions
-	cudaResults = CuArray{Float32}(undef, variableCols, length(expressions))
-
-	# 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, 256)
-		blocks = cld(variableCols, threads)
-
-		cudacall(kernel, (CuPtr{Float32},CuPtr{Float32},CuPtr{Float32}), cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
-	end
+	# for kernel in kernels
+	# 	cudacall(kernel, (CuPtr{Float32},CuPtr{Float32},CuPtr{Float32}), cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
+	# end
 	
 	return cudaResults
 end

package/test/PerformanceTests.jl

@@ -63,15 +63,13 @@ if compareWithCPU
 end
 
 # cacheInterpreter = Dict{Expr, PostfixType}()
-suite["GPUI"]["nikuradse_1"] = @benchmarkable interpret_gpu(exprs, X_t, parameters; repetitions=expr_reps)
+# suite["GPUI"]["nikuradse_1"] = @benchmarkable interpret_gpu(exprs, X_t, parameters; repetitions=expr_reps)
 
 # cacheTranspilerFront = Dict{Expr, PostfixType}()
 # cacheTranspilerRes = Dict{Expr, CuFunction}()
 suite["GPUT"]["nikuradse_1"] = @benchmarkable evaluate_gpu(exprs, X_t, parameters; repetitions=expr_reps)
 
-for i in 1:1
 tune!(suite)
-end
 BenchmarkTools.save("params.json", params(suite))
 
 throw("finished tuning")