benchmarking: improved performance with @inbounds. still slower in most cases

package/src/Interpreter.jl

@@ -15,7 +15,7 @@ export interpret
 function interpret(expressions::Vector{Expr}, variables::Matrix{Float32}, parameters::Vector{Vector{Float32}})::Matrix{Float32}
 	
 	exprs = Vector{ExpressionProcessing.PostfixType}(undef, length(expressions))
-	for i in eachindex(expressions)
+	@inbounds for i in eachindex(expressions)
 		exprs[i] = ExpressionProcessing.expr_to_postfix(expressions[i])
 	end
 	
@@ -30,7 +30,7 @@ function interpret(expressions::Vector{Expr}, variables::Matrix{Float32}, parame
 	cudaResults = CuArray{Float32}(undef, variableCols, length(exprs))
 
 	# Start kernel for each expression to ensure that no warp is working on different expressions
-	for i in eachindex(exprs)
+	@inbounds for i in eachindex(exprs)
 		kernel = @cuda launch=false interpret_expression(cudaExprs, cudaVars, cudaParams, cudaResults, cudaStepsize, i)
 		# config = launch_configuration(kernel.fun)
 		threads = min(variableCols, 256)
@@ -61,11 +61,11 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
 	
 	firstVariableIndex = ((varSetIndex-1) * stepsize[3]) # Exclusive
 	
-	for i in firstExprIndex:lastExprIndex
+	@inbounds for expr in expressions
-		if expressions[i].Type == EMPTY
+		if expr.Type == EMPTY
 			break
-		elseif expressions[i].Type == INDEX
+		elseif expr.Type == INDEX
-			val = expressions[i].Value
+			val = expr.Value
 			operationStackTop += 1
 
 			if val > 0
@@ -74,11 +74,11 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
 				val = abs(val)
 				operationStack[operationStackTop] = parameters[firstParamIndex + val]
 			end
-		elseif expressions[i].Type == FLOAT32
+		elseif expr.Type == FLOAT32
 			operationStackTop += 1
-			operationStack[operationStackTop] = reinterpret(Float32, expressions[i].Value)
+			operationStack[operationStackTop] = reinterpret(Float32, expr.Value)
-		elseif expressions[i].Type == OPERATOR
+		elseif expr.Type == OPERATOR
-			type = reinterpret(Operator, expressions[i].Value)
+			type = reinterpret(Operator, expr.Value)
 			if type == ADD
 				operationStackTop -= 1
 				operationStack[operationStackTop] = operationStack[operationStackTop] + operationStack[operationStackTop + 1]
@@ -108,10 +108,11 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
 			break
 		end
 	end
+
 	# "(exprIndex - 1) * variableCols" -> calculates the column in which to insert the result (expression = column)
 	# "+ varSetIndex" -> to get the row inside the column at which to insert the result of the variable set (variable set = row)
 	resultIndex = convert(Int, (exprIndex - 1) * variableCols + varSetIndex) # Inclusive
-	results[resultIndex] = operationStack[operationStackTop]
+	@inbounds results[resultIndex] = operationStack[operationStackTop]
 
 	return
 end

package/src/Transpiler.jl

@@ -14,10 +14,25 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 	variableCols = size(variables, 2)
 	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
 	# Threads.@threads for i in eachindex(expressions)
-	#   TODO: Use cache
+	# 	cacheLock = ReentrantLock()
-	# 	kernel = transpile(expressions[i], varRows, Utils.get_max_inner_length(parameters))
+	# 	cacheHit = false
+	# 	lock(cacheLock) do 
+	# 		if haskey(cache, expressions[i])
+	# 			kernels[i] = cache[expressions[i]]
+	# 			cacheHit = true
+	# 		end
+	# 	end
+
+	# 	if cacheHit
+	# 		continue
+	# 	end
+
+	# 	formattedExpr = ExpressionProcessing.expr_to_postfix(expressions[i])
+
+	# 	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
 
 	# 	linker = CuLink()
 	# 	add_data!(linker, "ExpressionProcessing", kernel)
@@ -26,9 +41,11 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 	
 	# 	mod = CuModule(image)
 	# 	kernels[i] = CuFunction(mod, "ExpressionProcessing")
+
+	# 	@lock cacheLock cache[expressions[i]] = kernels[i]
 	# end
 
-	for i in eachindex(expressions)
+	@inbounds for i in eachindex(expressions)
 		if haskey(cache, expressions[i])
 			kernels[i] = cache[expressions[i]]
 			continue
@@ -54,7 +71,7 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
 	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 i in eachindex(kernels)
+	@inbounds for i in eachindex(kernels)
 		# config = launch_configuration(kernels[i])
 		threads = min(variableCols, 256)
 		blocks = cld(variableCols, threads)

package/test/PerformanceTests.jl

@@ -105,10 +105,6 @@ loadparams!(suite, BenchmarkTools.load("params.json")[1], :samples, :evals, :gct
 
 results = run(suite, verbose=true, seconds=180)
 
-BenchmarkTools.save("$BENCHMARKS_RESULTS_PATH/256_blocksize.json", results)
-
-
-
 if compareWithCPU
 	medianCPU = median(results["CPU"])
 	stdCPU = std(results["CPU"])
@@ -142,8 +138,9 @@ if compareWithCPU
 	println(gpuiVsGPUT_median)
 	println(gpuiVsGPUT_std)
 	
+	BenchmarkTools.save("$BENCHMARKS_RESULTS_PATH/using_inbounds.json", results)
 else
-	resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/initial_results.json")[1]
+	resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/256_blocksize.json")[1]
 	
 	medianGPUI_old = median(resultsOld["GPUI"])
 	stdGPUI_old = std(resultsOld["GPUI"])