layed groundwork for implementing interpretation

package/src/ExpressionProcessing.jl

@@ -2,16 +2,16 @@ module ExpressionProcessing
 
 export expr_to_postfix
 export PostfixType
-export Operator, Add, Subtract, Multiply, Divide, Power, Abs, Log, Exp, Sqrt
+export Operator, ADD, SUBTRACT, MULTIPLY, DIVIDE, POWER, ABS, LOG, EXP, SQRT
-export ElementType, FLOAT64, OPERATOR, INT64
+export ElementType, EMPTY, FLOAT64, OPERATOR, INT64
 export ExpressionElement
 
-@enum Operator::Int64 Add=1 Subtract=2 Multiply=3 Divide=4 Power=5 Abs=6 Log=7 Exp=8 Sqrt=9
+@enum Operator::Int64 ADD=1 SUBTRACT=2 MULTIPLY=3 DIVIDE=4 POWER=5 ABS=6 LOG=7 EXP=8 SQRT=9
-@enum ElementType FLOAT64=1 OPERATOR=2 INT64=3
+@enum ElementType EMPTY=0 FLOAT64=1 OPERATOR=2 INT64=3
 
 struct ExpressionElement
 	Type::ElementType
-	value::Int64 # Reinterpret the stored value to type "ElementType" when using it
+	Value::Int64 # Reinterpret the stored value to type "ElementType" when using it
 end
 
 const PostfixType = Vector{ExpressionElement}
@@ -45,23 +45,23 @@ end
 
 function get_operator(op::Symbol)::Operator
 	if op == :+
-		return Add
+		return ADD
 	elseif op == :-
-		return Subtract
+		return SUBTRACT
 	elseif op == :*
-		return Multiply
+		return MULTIPLY
 	elseif op == :/
-		return Divide
+		return DIVIDE
 	elseif op == :^
-		return Power
+		return POWER
 	elseif op == :abs
-		return Abs
+		return ABS
 	elseif op == :log
-		return Log
+		return LOG
 	elseif op == :exp
-		return Exp
+		return EXP
 	elseif op == :sqrt
-		return Sqrt
+		return SQRT
 	end
 end
 
@@ -80,13 +80,13 @@ function convert_var_to_int(var::Symbol)::Int
 end
 
 function convert_to_ExpressionElement(element)::ExpressionElement
-	value = reinterpret(Int64, element)
+	Value = reinterpret(Int64, element)
 	if element isa Float64
-		return ExpressionElement(FLOAT64, value)
+		return ExpressionElement(FLOAT64, Value)
 	elseif element isa Int64
-		return ExpressionElement(INT64, value)
+		return ExpressionElement(INT64, Value)
 	elseif element isa Operator
-		return ExpressionElement(OPERATOR, value)
+		return ExpressionElement(OPERATOR, Value)
 	else
 		error("Element was of type '$(typeof(element))', which is not supported.")
 	end
@@ -100,17 +100,17 @@ end
 
 const SymbolTable64 = Dict{Tuple{Expr, Symbol},Float64}
 
-"Replaces all the variables and parameters of the given expression with their corresponding value stored in the symtable
+"Replaces all the variables and parameters of the given expression with their corresponding Value stored in the symtable
 # Arguments
  - `symtable::SymbolTable64`: Contains the values of all variables for each expression
- - `originalExpr::Expr`: Contains a deep copy of the original expression. It is used to link the expression and variables to their according value stored in the symtable
+ - `originalExpr::Expr`: Contains a deep copy of the original expression. It is used to link the expression and variables to their according Value stored in the symtable
 "
 function replace_variables!(ex::Expr, symtable::SymbolTable64, originalExpr::Expr)
 	for i in 1:length(ex.args)
 		arg = ex.args[i]
 		if typeof(arg) === Expr
 			replace_variables!(arg, symtable, originalExpr)
-		elseif haskey(symtable, (originalExpr,arg)) # We found a variable/parameter and can replace it with the actual value
+		elseif haskey(symtable, (originalExpr,arg)) # We found a variable/parameter and can replace it with the actual Value
 			ex.args[i] = symtable[(originalExpr,arg)]
 		end
 	end

package/src/Interpreter.jl

@@ -12,59 +12,93 @@ export interpret
 "
 function interpret(expressions::Vector{ExpressionProcessing.PostfixType}, variables::Matrix{Float64}, parameters::Vector{Vector{Float64}})
 	# TODO: 
-	#      create CUDA array and fill it with the variables and parameters
 	#      create CUDA array for calculation results
 	variableRows = size(variables, 1)
 	cudaVars = CuArray(variables)
+	cudaParams = create_cuda_array(parameters, NaN64)
+	cudaExprs = create_cuda_array(expressions, ExpressionElement(EMPTY, 0))
+	cudaStepsize = CuArray([get_max_inner_length(expressions), get_max_inner_length(parameters)]) # put into seperate cuArray, as this is static and would be inefficient to send seperatly to every kernel
 
-	paramRows = get_max_parameter_rows(parameters)
+	# Start kernel for each expression to ensure that no warp is working on different expressions
-	p1aramCols = length(parameters)
-	cudaParams = CuArray{Float64}(undef, p1aramCols, paramRows) # length(parameters) == number of expressions
-	# TODO: Fill cudaParams
-
-	# TODO: Move CuArray(expression[i]) outside the loop for a more efficient transfer to GPU but leave kernel signature as is
 	for i in eachindex(expressions)
-		cudaExpr = CuArray(expressions[i])
+		kernel = @cuda launch=false interpret_expression(cudaExprs, cudaVars, cudaParams, cudaStepsize, i)
-		kernel = @cuda launch=false interpret_expression(cudaExpr, cudaVars, cudaParams, i)
 		config = launch_configuration(kernel.fun)
 		threads = min(variableRows, config.threads)
 		blocks = cld(variableRows, threads)
 
-		kernel(cudaExpr, cudaVars, cudaParams, i; threads, blocks)
+		kernel(cudaExprs, cudaVars, cudaParams, cudaStepsize, i; threads, blocks)
 	end
 end
 
-function interpret_expression(expression, variables, parameters, exprIndex::Int)
+function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, variables::CuDeviceArray{Float64}, parameters::CuDeviceArray{Float64}, stepsize::CuDeviceArray{Int}, exprIndex::Int)
+	firstExprIndex = (exprIndex - 1 * stepsize[1]) + 1 # Inclusive
+	lastExprIndex = firstExprIndex + stepsize[1] # Exclusive
+	firstParamIndex = (exprIndex - 1 * stepsize[2]) + 1 # Inclusive
+	# lastParamIndex = firstParamIndex + stepsize[2] # Exclusive (probably not needed)
+	
+	for i in firstExprIndex:lastExprIndex
 		# TODO Implement interpreter
+		#      - start at firstExprIndex and interpret until the first ExpressionElement is "Empty" or we reached lastExprIndex
 	end
 
+	return
+end
+
+
 "Retrieves the number of entries for the largest inner vector"
-function get_max_parameter_rows(params::Vector{Vector{T}})::Int where T
+function get_max_inner_length(vec::Vector{Vector{T}})::Int where T
 	maxLength = 0
-	for i in eachindex(params)
+	@inbounds for i in eachindex(vec)
-		if length(params) > maxLength
+		if length(vec[i]) > maxLength
-			maxLength = length(params)
+			maxLength = length(vec[i])
 		end
 	end
 
 	return maxLength
 end
 
+"Returns a CuArray filed with the data provided. The inner vectors do not have to have the same length. All missing elements will be the value ```invalidElement```"
+function create_cuda_array(data::Vector{Vector{T}}, invalidElement::T)::CuArray{T} where T
+	dataCols = get_max_inner_length(data)
+	dataRows = length(data)
+	dataMat = convert_to_matrix(data, invalidElement)
+	cudaArr = CuArray{T}(undef, dataCols, dataRows) # length(parameters) == number of expressions
+	copyto!(cudaArr, dataMat)
+
+	return cudaArr
+end
+
+"Converts a vector of vectors into a matrix. The inner vectors do not need to have the same length.
+
+All entries that cannot be filled have ```invalidElement``` as their value
+"
+function convert_to_matrix(vec::Vector{Vector{T}}, invalidElement::T)::Matrix{T} where T
+	vecCols = get_max_inner_length(vec)
+	vecRows = length(vec)
+	vecMat = fill(invalidElement, vecCols, vecRows)
+	
+	for i in eachindex(vec)
+		vecMat[:,i] = copyto!(vecMat[:,i], vec[i])
+	end
+
+	return vecMat
+end
 
 
-@deprecate InterpretExplicit!(op::Operator, x, y) interpret_expression(expression, variables, parameters, exprIndex::Int)
+
+# @deprecate InterpretExplicit!(op::Operator, x, y) interpret_expression(expression, variables, parameters, exprIndex::Int)
 # Kernel
 function InterpretExplicit!(op::Operator, x, y)
 	index = (blockIdx().x - 1) * blockDim().x + threadIdx().x
 	stride = gridDim().x * blockDim().x
 
-	if op == Add
+	if op == ADD
 		# @cuprintln("Performing Addition") # Will only be displayed when the GPU is synchronized
 		for i = index:stride:length(y)
 			@inbounds y[i] += x[i]
 		end
 		return
-	elseif op == Subtract
+	elseif op == SUBTRACT
 		# @cuprintln("Performing Subtraction") # Will only be displayed when the GPU is synchronized
 		for i = index:stride:length(y)
 			@inbounds y[i] -= x[i]

package/test/ExpressionProcessingTests.jl

@@ -14,18 +14,18 @@ parameters[1][1] = 5
 @testset "Test conversion expression element" begin
 	reference1 = ExpressionElement(FLOAT64, reinterpret(Int64, 1.0))
 	reference2 = ExpressionElement(INT64, reinterpret(Int64, 1))
-	reference3 = ExpressionElement(OPERATOR, reinterpret(Int64, Add))
+	reference3 = ExpressionElement(OPERATOR, reinterpret(Int64, ADD))
 	
 	@test isequal(reference1, ExpressionProcessing.convert_to_ExpressionElement(1.0))
 	@test isequal(reference2, ExpressionProcessing.convert_to_ExpressionElement(1))
-	@test isequal(reference3, ExpressionProcessing.convert_to_ExpressionElement(Add))
+	@test isequal(reference3, ExpressionProcessing.convert_to_ExpressionElement(ADD))
 end
 
 @testset "Test conversion to postfix" begin
 	reference = PostfixType()
 	
-	append!(reference, [ExpressionProcessing.convert_to_ExpressionElement(1), ExpressionProcessing.convert_to_ExpressionElement(1.0), ExpressionProcessing.convert_to_ExpressionElement(2), ExpressionProcessing.convert_to_ExpressionElement(Multiply), 
+	append!(reference, [ExpressionProcessing.convert_to_ExpressionElement(1), ExpressionProcessing.convert_to_ExpressionElement(1.0), ExpressionProcessing.convert_to_ExpressionElement(2), ExpressionProcessing.convert_to_ExpressionElement(MULTIPLY), 
-	ExpressionProcessing.convert_to_ExpressionElement(Add), ExpressionProcessing.convert_to_ExpressionElement(-1), ExpressionProcessing.convert_to_ExpressionElement(Add)])
+	ExpressionProcessing.convert_to_ExpressionElement(ADD), ExpressionProcessing.convert_to_ExpressionElement(-1), ExpressionProcessing.convert_to_ExpressionElement(ADD)])
 	postfix = expr_to_postfix(expressions[1])
 
 	@test isequal(reference, postfix)

package/test/InterpreterTests.jl

@@ -1,21 +1,42 @@
+using CUDA
 using .ExpressionProcessing
 using .Interpreter
 
-expressions = Vector{Expr}(undef, 1)
+expressions = Vector{Expr}(undef, 2)
-variables = Matrix{Float64}(undef, 1,2)
+variables = Matrix{Float64}(undef, 2,2)
-parameters = Vector{Vector{Float64}}(undef, 1)
+parameters = Vector{Vector{Float64}}(undef, 2)
 
 # Resulting value should be 10
 expressions[1] = :(x1 + 1 * x2 + p1)
-variables[1,1] = 2
+expressions[2] = :(5 + x1 + 1 * x2 + p1 + p2)
-variables[1,2] = 3
+variables[1,1] = 2.0
+variables[1,2] = 3.0
+variables[2,1] = 2.0
+variables[2,2] = 3.0
 parameters[1] = Vector{Float64}(undef, 1)
-parameters[1][1] = 5
+parameters[2] = Vector{Float64}(undef, 2)
+parameters[1][1] = 5.0
+parameters[2][1] = 5.0
+parameters[2][2] = 0.0
 
 @testset "Test interpretation" begin
 	postfixExpr = expr_to_postfix(expressions[1])
-	postfixExprs = Vector{PostfixType}(undef, 1)
+	postfixExprs = Vector([postfixExpr])
-	postfixExprs[1] = postfixExpr
+	push!(postfixExprs, expr_to_postfix(expressions[2]))
 
-	Interpret(postfixExprs, variables, parameters)
+	CUDA.@sync interpret(postfixExprs, variables, parameters)
+end
+
+@testset "Test conversion to matrix" begin
+	reference = Matrix{Float64}(undef, 2, 2)
+	reference[1,1] = 5.0
+	reference[2,1] = NaN64
+	reference[1,2] = 5.0
+	reference[2,2] = 0.0
+	# fill!(reference, [[], [5.0, 0.0]])
+	# reference = Matrix([5.0, NaN],
+	# 				   [5.0, 0.0])
+	result = Interpreter.convert_to_matrix(parameters, NaN64)
+
+	@test isequal(result, reference)
 end