using LinearAlgebra
using BenchmarkTools
using DelimitedFiles
using GZip
using CUDA

using .Transpiler
using .Interpreter
using .ExpressionProcessing

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)
# Variable-Sets: 1000 can be considered the minimum; 100.000 can be considered the maximum

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}}()
varnames = ["x$i" for i in 1:10]
paramnames = ["p$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 
	for line in eachline(io)
		expr, p = parse_infix(line, varnames, paramnames)

		push!(exprs, expr)
		push!(parameters, randn(Float32, length(p)))
	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

	# Make array const:
	# https://cuda.juliagpu.org/v2.6/api/kernel/#Device-arrays

	# Memory management like in C++ might help with performance improvements
	# https://cuda.juliagpu.org/v2.6/lib/driver/#Memory-Management

# https://cuda.juliagpu.org/stable/development/profiling/#NVIDIA-Nsight-Systems
# Systems and Compute installable via WSL. Compute UI can even be used inside wsl
# Add /usr/local/cuda/bin in .bashrc to PATH to access ncu and nsys (do the tests on FH PCs)
# University setup at 10.20.1.7 and 10.20.1.13

compareWithCPU = false

suite = BenchmarkGroup()
suite["CPU"] = BenchmarkGroup(["CPUInterpreter"])
suite["GPUI"] = BenchmarkGroup(["GPUInterpreter"])
suite["GPUT"] = BenchmarkGroup(["GPUTranspiler"])

if compareWithCPU
	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

# cacheInterpreter = Dict{Expr, PostfixType}()
# 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)

tune!(suite)
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

if compareWithCPU
	medianCPU = median(results["CPU"])
	stdCPU = std(results["CPU"])
	
	medianInterpreter = median(results["GPUI"])
	stdInterpreter = std(results["GPUI"])
	
	medianTranspiler = median(results["GPUT"])
	stdTranspiler = std(results["GPUT"])
	
	cpuVsGPUI_median = judge(medianInterpreter, medianCPU) # is interpreter better than cpu?
	cpuVsGPUT_median = judge(medianTranspiler, medianCPU) # is transpiler better than cpu?
	gpuiVsGPUT_median = judge(medianTranspiler, medianInterpreter) # is tranpiler better than interpreter?
	
	cpuVsGPUI_std = judge(stdInterpreter, stdCPU) # is interpreter better than cpu?
	cpuVsGPUT_std = judge(stdTranspiler, stdCPU) # is transpiler better than cpu?
	gpuiVsGPUT_std = judge(stdTranspiler, stdInterpreter) # is tranpiler better than interpreter?
	
	println()
	println("Is the interpreter better than the CPU implementation:")
	println(cpuVsGPUI_median)
	println(cpuVsGPUI_std)
	
	println()
	println("Is the transpiler better than the CPU implementation:")
	println(cpuVsGPUT_median)
	println(cpuVsGPUT_std)
	
	println()
	println("Is the transpiler better than the interpreter:")
	println(gpuiVsGPUT_median)
	println(gpuiVsGPUT_std)
	
	BenchmarkTools.save("$BENCHMARKS_RESULTS_PATH/0_initial.json", results)
else
	resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/3-tuned-blocksize_I128_T96.json")[1]
	# resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/3-tuned-blocksize_I128_T96.json")[1]
	
	medianGPUI_old = median(resultsOld["GPUI"])
	stdGPUI_old = std(resultsOld["GPUI"])
	
	medianGPUT_old = median(resultsOld["GPUT"])
	stdGPUT_old = std(resultsOld["GPUT"])
	
	medianInterpreter = median(results["GPUI"])
	stdInterpreter = std(results["GPUI"])
	
	medianTranspiler = median(results["GPUT"])
	stdTranspiler = std(results["GPUT"])
	
	oldVsGPUI_median = judge(medianInterpreter, medianGPUI_old) # is interpreter better than old?
	oldVsGPUI_std = judge(stdInterpreter, stdGPUI_old) # is interpreter better than old?
	
	oldVsGPUT_median = judge(medianTranspiler, medianGPUT_old) # is transpiler better than old?
	oldVsGPUT_std = judge(stdTranspiler, stdGPUT_old) # is transpiler better than old?
	
	
	println()
	println("Is the interpreter better than the old implementation:")
	println(oldVsGPUI_median)
	println(oldVsGPUI_std)
	
	println()
	println("Is the transpiler better than the old implementation:")
	println(oldVsGPUT_median)
	println(oldVsGPUT_std)
end