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

transpiler: invalid memory access error finally fixed
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
2025-03-27 22:32:24 +01:00
parent 561b37160b
commit 9df78ca72e
7 changed files with 133 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
package/src/Transpiler.jl
View File

@ -53,7 +53,7 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
cudaParams = Utils.create_cuda_array(parameters, NaN32) # maybe make constant (see runtests.jl for more info)
# 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(expressions))
cudaResults = CuArray{Float32}(undef, variableCols * length(expressions))
# cudaResults = CUDA.zeros(variableCols * length(expressions))
# ptr = CuPtr{Float32}(C_NULL)
# CUDA.cuMemAlloc(ptr, sizeof(Float32) * 10)
@ -68,8 +68,9 @@ function evaluate(expressions::Vector{Expr}, variables::Matrix{Float32}, paramet
blocks = cld(variableCols, threads)
# cudacall(kernels[i], (CuPtr{Float32},CuPtr{Float32},CuPtr{Float32}), cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
cudacall(kernels[i], (CuPtr{Float32},), cudaResults; threads=threads, blocks=blocks)
# launch(kernels[i], cudaVars, cudaParams, cudaResults; threads=threads, blocks=blocks)
launch(kernels[i], cudaResults; threads=threads, blocks=blocks)
# launch(kernels[i], cudaResults; threads=threads, blocks=blocks)
end
println(Array(cudaResults))
@ -120,9 +121,9 @@ end
# TODO: Make version, target and address_size configurable; also see what address_size means exactly
function get_cuda_header()::String
return "
.version 7.1
.version 8.5
.target sm_61
.address_size 32
.address_size 64
"
end
@ -137,11 +138,11 @@ function get_kernel_signature(kernelName::String, parameters::Vector{DataType}):
println(signatureBuffer, "(")
for i in eachindex(parameters)
print(signatureBuffer, " .param .u32", " ", "param_", i)
print(signatureBuffer, " .param .u64", " ", "param_", i)
parametersLocation = get_next_free_register("i")
println(paramLoadingBuffer, "ld.param.u32 $parametersLocation, [param_$i];")
println(paramLoadingBuffer, "cvta.to.global.u32 $(get_next_free_register("parameter")), $parametersLocation;")
println(paramLoadingBuffer, "ld.param.u64 $parametersLocation, [param_$i];")
println(paramLoadingBuffer, "cvta.to.global.u64 $(get_next_free_register("parameter")), $parametersLocation;")
if i != lastindex(parameters)
println(signatureBuffer, ",")
end
@ -169,12 +170,12 @@ function get_guard_clause(exitJumpLocation::String, nrOfVarSets::Integer)::Tuple
println(guardBuffer, "mov.u32 $currentThreadId, %tid.x;")
globalThreadId = get_next_free_register("r") # basically the index of the thread in the variable set
# breakCondition = get_next_free_register("p")
breakCondition = get_next_free_register("p")
println(guardBuffer, "mad.lo.s32 $globalThreadId, $threadIds, $threadsPerCTA, $currentThreadId;")
# println(guardBuffer, "setp.ge.s32 $breakCondition, $globalThreadId, $nrOfVarSets;") # guard clause = index > nrOfVariableSets
println(guardBuffer, "setp.gt.s32 $breakCondition, $globalThreadId, $nrOfVarSets;") # guard clause = index > nrOfVariableSets
# branch to end if breakCondition is true
# print(guardBuffer, "@$breakCondition bra $exitJumpLocation;")
print(guardBuffer, "@$breakCondition bra $exitJumpLocation;")
return (String(take!(guardBuffer)), globalThreadId)
end
@ -186,7 +187,7 @@ function generate_calculation_code(expression::ExpressionProcessing.PostfixType,
parametersLocation::String, parametersSetSize::Integer, resultsLocation::String,
threadIdReg::String, expressionIndex::Integer, nrOfVarSets::Integer)::String
return "st.global.f32 [$resultsLocation], 10.0;"
# return "st.global.f32 [$resultsLocation], 10.0;"
codeBuffer = IOBuffer()
operands = Vector{Operand}()
@ -360,9 +361,9 @@ let registers = Dict() # stores the count of the register already used.
elseif definition.first == "r"
regType = ".b32"
elseif definition.first == "parameter"
regType = ".u32"
regType = ".b64"
elseif definition.first == "i"
regType = ".u32"
regType = ".b64"
else
throw(ArgumentError("Unknown register name used. Name '$(definition.first)' cannot be mapped to a PTX type."))
end