Daniel/master-thesis
1
0
Fork 0
Code Issues 3 Pull Requests Actions Packages Projects 1 Releases Wiki Activity

benchmarking: reverted previous; made interpreter use fast math
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 13:26:35 +02:00
parent 6d6874c7ba
commit a5c34a53b7
7 changed files with 32 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
package/src/Interpreter.jl
View File

@ -1,6 +1,5 @@
module Interpreter
using CUDA
using CUDA: i32
using StaticArrays
using ..ExpressionProcessing
using ..Utils
@ -25,14 +24,14 @@ function interpret(expressions::Vector{Expr}, variables::Matrix{Float32}, parame
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
# put into seperate cuArray, as this is static and would be inefficient to send seperatly to every kernel
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
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
# 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))
# Start kernel for each expression to ensure that no warp is working on different expressions
@inbounds for i in eachindex(exprs)
kernel = @cuda launch=false interpret_expression(cudaExprs, cudaVars, cudaParams, cudaResults, cudaStepsize, convert(Int32, i))
kernel = @cuda launch=false fastmath=true interpret_expression(cudaExprs, cudaVars, cudaParams, cudaResults, cudaStepsize, i)
# config = launch_configuration(kernel.fun)
threads = min(variableCols, 128)
blocks = cld(variableCols, threads)
@ -45,8 +44,8 @@ end
#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
function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, variables::CuDeviceArray{Float32}, parameters::CuDeviceArray{Float32}, results::CuDeviceArray{Float32}, stepsize::CuDeviceArray{Int32}, exprIndex::Int32)
varSetIndex = (blockIdx().x - 1i32) * blockDim().x + threadIdx().x # ctaid.x * ntid.x + tid.x (1-based)
function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, variables::CuDeviceArray{Float32}, parameters::CuDeviceArray{Float32}, results::CuDeviceArray{Float32}, stepsize::CuDeviceArray{Int}, exprIndex::Int)
varSetIndex = (blockIdx().x - 1) * blockDim().x + threadIdx().x # ctaid.x * ntid.x + tid.x (1-based)
@inbounds variableCols = length(variables) / stepsize[2]
if varSetIndex > variableCols
@ -55,19 +54,19 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
# firstExprIndex = ((exprIndex - 1) * stepsize[1]) + 1 # Inclusive
# lastExprIndex = firstExprIndex + stepsize[1] - 1 # Inclusive
@inbounds firstParamIndex = ((exprIndex - 1i32) * stepsize[1]) # Exclusive
@inbounds firstParamIndex = ((exprIndex - 1) * 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
operationStackTop = 0i32 # stores index of the last defined/valid value
operationStackTop = 0 # stores index of the last defined/valid value
@inbounds firstVariableIndex = ((varSetIndex-1i32) * stepsize[2]) # Exclusive
@inbounds firstVariableIndex = ((varSetIndex-1) * stepsize[2]) # Exclusive
@inbounds for expr in expressions
if expr.Type == EMPTY
break
elseif expr.Type == INDEX
val = expr.Value
operationStackTop += 1i32
operationStackTop += 1
if val > 0
operationStack[operationStackTop] = variables[firstVariableIndex + val]
@ -76,25 +75,25 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
operationStack[operationStackTop] = parameters[firstParamIndex + val]
end
elseif expr.Type == FLOAT32
operationStackTop += 1i32
operationStackTop += 1
operationStack[operationStackTop] = reinterpret(Float32, expr.Value)
elseif expr.Type == OPERATOR
type = reinterpret(Operator, expr.Value)
if type == ADD
operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] + operationStack[operationStackTop + 1i32]
operationStackTop -= 1
operationStack[operationStackTop] = operationStack[operationStackTop] + operationStack[operationStackTop + 1]
elseif type == SUBTRACT
operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] - operationStack[operationStackTop + 1i32]
operationStackTop -= 1
operationStack[operationStackTop] = operationStack[operationStackTop] - operationStack[operationStackTop + 1]
elseif type == MULTIPLY
operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] * operationStack[operationStackTop + 1i32]
operationStackTop -= 1
operationStack[operationStackTop] = operationStack[operationStackTop] * operationStack[operationStackTop + 1]
elseif type == DIVIDE
operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] / operationStack[operationStackTop + 1i32]
operationStackTop -= 1
operationStack[operationStackTop] = operationStack[operationStackTop] / operationStack[operationStackTop + 1]
elseif type == POWER
operationStackTop -= 1i32
operationStack[operationStackTop] = operationStack[operationStackTop] ^ operationStack[operationStackTop + 1i32]
operationStackTop -= 1
operationStack[operationStackTop] = operationStack[operationStackTop] ^ operationStack[operationStackTop + 1]
elseif type == ABS
operationStack[operationStackTop] = abs(operationStack[operationStackTop])
elseif type == LOG
@ -112,7 +111,7 @@ function interpret_expression(expressions::CuDeviceArray{ExpressionElement}, var
# "(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 - 1i32) * variableCols + varSetIndex) # Inclusive
resultIndex = convert(Int, (exprIndex - 1) * variableCols + varSetIndex) # Inclusive
@inbounds results[resultIndex] = operationStack[operationStackTop]
return