Daniel/master-thesis
1
0
Fork 0
Code Issues 5 Pull Requests Actions 1 Packages Projects 1 Releases Wiki Activity

benchmarking: used int32 wherever possible; resulted in noticeable performance drop
Some checks are pending
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, 1.10) (push) Waiting to run
Details
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, 1.6) (push) Waiting to run
Details
CI / Julia ${{ matrix.version }} - ${{ matrix.os }} - ${{ matrix.arch }} - ${{ github.event_name }} (x64, ubuntu-latest, pre) (push) Waiting to run
Details

Browse Source
This commit is contained in:
Daniel 2025-04-13 11:32:54 +02:00
parent 4c60331288
commit af3b72f196
8 changed files with 29 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
package/src/Interpreter.jl
View File

@ -1,5 +1,6 @@
module Interpreter module Interpreter
using CUDA using CUDA
using CUDA: i32
using StaticArrays using StaticArrays
using ..ExpressionProcessing using ..ExpressionProcessing
using ..Utils using ..Utils
@ -24,14 +25,14 @@ function interpret(expressions::Vector{Expr}, variables::Matrix{Float32}, parame
cudaParams = Utils.create_cuda_array(parameters, NaN32) # column corresponds to data for one expression cudaParams = Utils.create_cuda_array(parameters, NaN32) # column corresponds to data for one expression
cudaExprs = Utils.create_cuda_array(exprs, ExpressionElement(EMPTY, 0)) # column corresponds to data for one expression cudaExprs = Utils.create_cuda_array(exprs, ExpressionElement(EMPTY, 0)) # column corresponds to data for one expression
# put into seperate cuArray, as this is static and would be inefficient to send seperatly to every kernel # put into seperate cuArray, as this is static and would be inefficient to send seperatly to every kernel
cudaStepsize = CuArray([Utils.get_max_inner_length(parameters), size(variables, 1)]) # max num of values per expression; max nam of parameters per expression; number of variables per expression cudaStepsize::CuArray{Int32} = CuArray([Utils.get_max_inner_length(parameters), size(variables, 1)]) # max num of values per expression; max nam of parameters per expression; number of variables per expression
# each expression has nr. of variable sets (nr. of columns of the variables) results and there are n expressions # 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(exprs)) cudaResults = CuArray{Float32}(undef, variableCols, length(exprs))
# Start kernel for each expression to ensure that no warp is working on different expressions # Start kernel for each expression to ensure that no warp is working on different expressions
@inbounds for i in eachindex(exprs) @inbounds for i in eachindex(exprs)
kernel = @cuda launch=false interpret_expression(cudaExprs, cudaVars, cudaParams, cudaResults, cudaStepsize, i) kernel = @cuda launch=false interpret_expression(cudaExprs, cudaVars, cudaParams, cudaResults, cudaStepsize, convert(Int32, i))
# config = launch_configuration(kernel.fun) # config = launch_configuration(kernel.fun)
threads = min(variableCols, 128) threads = min(variableCols, 128)
blocks = cld(variableCols, threads) blocks = cld(variableCols, threads)
@ -44,8 +45,8 @@ end
#TODO: Add @inbounds to all indexing after it is verified that all works https://cuda.juliagpu.org/stable/development/kernel/#Bounds-checking #TODO: Add @inbounds to all indexing after it is verified that all works https://cuda.juliagpu.org/stable/development/kernel/#Bounds-checking
const MAX_STACK_SIZE = 25 # The depth of the stack to store the values and intermediate results const MAX_STACK_SIZE = 25 # The depth of the stack to store the values and intermediate results
function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, variables::CuDeviceArray{Float32}, parameters::CuDeviceArray{Float32}, results::CuDeviceArray{Float32}, stepsize::CuDeviceArray{Int}, exprIndex::Int) function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, variables::CuDeviceArray{Float32}, parameters::CuDeviceArray{Float32}, results::CuDeviceArray{Float32}, stepsize::CuDeviceArray{Int32}, exprIndex::Int32)
varSetIndex = (blockIdx().x - 1) * blockDim().x + threadIdx().x # ctaid.x * ntid.x + tid.x (1-based) varSetIndex = (blockIdx().x - 1i32) * blockDim().x + threadIdx().x # ctaid.x * ntid.x + tid.x (1-based)
@inbounds variableCols = length(variables) / stepsize[2] @inbounds variableCols = length(variables) / stepsize[2]
if varSetIndex > variableCols if varSetIndex > variableCols
@ -54,19 +55,19 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
# firstExprIndex = ((exprIndex - 1) * stepsize[1]) + 1 # Inclusive # firstExprIndex = ((exprIndex - 1) * stepsize[1]) + 1 # Inclusive
# lastExprIndex = firstExprIndex + stepsize[1] - 1 # Inclusive # lastExprIndex = firstExprIndex + stepsize[1] - 1 # Inclusive
@inbounds firstParamIndex = ((exprIndex - 1) * stepsize[1]) # Exclusive @inbounds firstParamIndex = ((exprIndex - 1i32) * stepsize[1]) # Exclusive
operationStack = MVector{MAX_STACK_SIZE, Float32}(undef) # Try to get this to function with variable size too, to allow better memory usage operationStack = MVector{MAX_STACK_SIZE, Float32}(undef) # Try to get this to function with variable size too, to allow better memory usage
operationStackTop = 0 # stores index of the last defined/valid value operationStackTop = 0i32 # stores index of the last defined/valid value
@inbounds firstVariableIndex = ((varSetIndex-1) * stepsize[2]) # Exclusive @inbounds firstVariableIndex = ((varSetIndex-1i32) * stepsize[2]) # Exclusive
@inbounds for expr in expressions @inbounds for expr in expressions
if expr.Type == EMPTY if expr.Type == EMPTY
break break
elseif expr.Type == INDEX elseif expr.Type == INDEX
val = expr.Value val = expr.Value
operationStackTop += 1 operationStackTop += 1i32
if val > 0 if val > 0
operationStack[operationStackTop] = variables[firstVariableIndex + val] operationStack[operationStackTop] = variables[firstVariableIndex + val]
@ -75,25 +76,25 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
operationStack[operationStackTop] = parameters[firstParamIndex + val] operationStack[operationStackTop] = parameters[firstParamIndex + val]
end end
elseif expr.Type == FLOAT32 elseif expr.Type == FLOAT32
operationStackTop += 1 operationStackTop += 1i32
operationStack[operationStackTop] = reinterpret(Float32, expr.Value) operationStack[operationStackTop] = reinterpret(Float32, expr.Value)
elseif expr.Type == OPERATOR elseif expr.Type == OPERATOR
type = reinterpret(Operator, expr.Value) type = reinterpret(Operator, expr.Value)
if type == ADD if type == ADD
operationStackTop -= 1 operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] + operationStack[operationStackTop + 1] operationStack[operationStackTop] = operationStack[operationStackTop] + operationStack[operationStackTop + 1i32]
elseif type == SUBTRACT elseif type == SUBTRACT
operationStackTop -= 1 operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] - operationStack[operationStackTop + 1] operationStack[operationStackTop] = operationStack[operationStackTop] - operationStack[operationStackTop + 1i32]
elseif type == MULTIPLY elseif type == MULTIPLY
operationStackTop -= 1 operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] * operationStack[operationStackTop + 1] operationStack[operationStackTop] = operationStack[operationStackTop] * operationStack[operationStackTop + 1i32]
elseif type == DIVIDE elseif type == DIVIDE
operationStackTop -= 1 operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] / operationStack[operationStackTop + 1] operationStack[operationStackTop] = operationStack[operationStackTop] / operationStack[operationStackTop + 1i32]
elseif type == POWER elseif type == POWER
operationStackTop -= 1 operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] ^ operationStack[operationStackTop + 1] operationStack[operationStackTop] = operationStack[operationStackTop] ^ operationStack[operationStackTop + 1i32]
elseif type == ABS elseif type == ABS
operationStack[operationStackTop] = abs(operationStack[operationStackTop]) operationStack[operationStackTop] = abs(operationStack[operationStackTop])
elseif type == LOG elseif type == LOG
@ -104,14 +105,14 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
operationStack[operationStackTop] = sqrt(operationStack[operationStackTop]) operationStack[operationStackTop] = sqrt(operationStack[operationStackTop])
end end
else else
operationStack[operationStackTop] = NaN operationStack[operationStackTop] = NaN32
break break
end end
end end
# "(exprIndex - 1) * variableCols" -> calculates the column in which to insert the result (expression = column) # "(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) # "+ 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 resultIndex = convert(Int, (exprIndex - 1i32) * variableCols + varSetIndex) # Inclusive
@inbounds results[resultIndex] = operationStack[operationStackTop] @inbounds results[resultIndex] = operationStack[operationStackTop]
return return

4
package/test/PerformanceTests.jl
View File

@ -4,7 +4,7 @@ using BenchmarkTools
using .Transpiler using .Transpiler
using .Interpreter using .Interpreter
const BENCHMARKS_RESULTS_PATH = "./results-fh" const BENCHMARKS_RESULTS_PATH = "./results"
exprsCPU = [ exprsCPU = [
# CPU interpreter requires an anonymous function and array ref s # CPU interpreter requires an anonymous function and array ref s
:(p[1] * x[1] + p[2]), # 5 op :(p[1] * x[1] + p[2]), # 5 op
@ -143,7 +143,7 @@ if compareWithCPU
println(gpuiVsGPUT_median) println(gpuiVsGPUT_median)
println(gpuiVsGPUT_std) println(gpuiVsGPUT_std)
BenchmarkTools.save("$BENCHMARKS_RESULTS_PATH/3-tuned-blocksize_I128_T96.json", results) BenchmarkTools.save("$BENCHMARKS_RESULTS_PATH/4-interpreter_using_int32.json", results)
else else
resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/2-using_inbounds.json")[1] resultsOld = BenchmarkTools.load("$BENCHMARKS_RESULTS_PATH/2-using_inbounds.json")[1]

0
package/test/results/initial_results.json → package/test/results/0-initial_results.json
View File

0
package/test/results/256_blocksize.json → package/test/results/1-256_blocksize.json
View File

0
package/test/results/using_inbounds.json → package/test/results/2-using_inbounds.json
View File

1
package/test/results/4-interpreter_using_int32.json Normal file
View File

File diff suppressed because one or more lines are too long

4
thesis/chapters/evaluation.tex
View File

@ -21,6 +21,8 @@ Initial: CPU-Side single-threaded; up to 1024 threads per block; bounds-checking
1.) Blocksize reduced to a maximum of 256 -> moderate improvement in medium and large 1.) Blocksize reduced to a maximum of 256 -> moderate improvement in medium and large
2.) Using @inbounds -> noticeable improvement in 2 out of 3 2.) Using @inbounds -> noticeable improvement in 2 out of 3
3.) Tuned blocksize with NSight compute -> slight improvement
4.) used int32 everywhere to reduce register usage -> significant performance drop (probably because a lot more waiting time, or more type conversions happening on GPU? would need to look at PTX)
\subsection{Transpiler} \subsection{Transpiler}
Results only for Transpiler (also contains final kernel configuration and probably quick overview/recap of the implementation used and described in Implementation section Results only for Transpiler (also contains final kernel configuration and probably quick overview/recap of the implementation used and described in Implementation section
@ -31,6 +33,8 @@ Initial: CPU-Side single-threaded; up to 1024 threads per block; bounds-checking
1.) Blocksize reduced to a maximum of 256 -> moderate improvement in medium and large 1.) Blocksize reduced to a maximum of 256 -> moderate improvement in medium and large
2.) Using @inbounds -> small improvement only on CPU side code 2.) Using @inbounds -> small improvement only on CPU side code
3.) Tuned blocksize with NSight compute -> slight improvement
4.) Only changed things on interpreter side
\subsection{Comparison} \subsection{Comparison}
Comparison of Interpreter and Transpiler as well as Comparing the two with CPU interpreter Comparison of Interpreter and Transpiler as well as Comparing the two with CPU interpreter

BIN
thesis/main.pdf
View File

Binary file not shown.