Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Native RNG fixes for very large arrays #2561

Merged
merged 5 commits into from
Dec 11, 2024
Merged

Native RNG fixes for very large arrays #2561

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
54bf5b0
Native RNG: Improve randn determinism by using fixed grid.
maleadt Nov 21, 2024
0a515fc
Native RNG: Support large arrays by avoiding counter overflowing.
maleadt Nov 21, 2024
0bbeb97
Native RNG: Fix randn window calculation.
maleadt Nov 21, 2024
c23e09f
Use host memory to prevent OOM.
maleadt Dec 10, 2024
150af15
Try using unified memory.
maleadt Dec 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 11 additions & 11 deletions src/random.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,8 +52,8 @@ function Random.rand!(rng::RNG, A::AnyCuArray)

# grid-stride loop
threadId = threadIdx().x
window = blockDim().x * gridDim().x
offset = (blockIdx().x - 1) * blockDim().x
window = widemul(blockDim().x, gridDim().x)
offset = widemul(blockIdx().x - 1i32, blockDim().x)
while offset < length(A)
i = threadId + offset
if i <= length(A)
Expand Down Expand Up @@ -96,8 +96,8 @@ function Random.randn!(rng::RNG, A::AnyCuArray{<:Union{AbstractFloat,Complex{<:A

# grid-stride loop
threadId = threadIdx().x
window = (blockDim().x - 1) * gridDim().x
offset = (blockIdx().x - 1) * blockDim().x
window = widemul(blockDim().x, gridDim().x)
offset = widemul(blockIdx().x - 1i32, blockDim().x)
while offset < length(A)
i = threadId + offset
j = threadId + offset + window
Expand Down Expand Up @@ -129,8 +129,8 @@ function Random.randn!(rng::RNG, A::AnyCuArray{<:Union{AbstractFloat,Complex{<:A

# grid-stride loop
threadId = threadIdx().x
window = (blockDim().x - 1) * gridDim().x
offset = (blockIdx().x - 1) * blockDim().x
window = widemul(blockDim().x, gridDim().x)
offset = widemul(blockIdx().x - 1i32, blockDim().x)
while offset < length(A)
i = threadId + offset
if i <= length(A)
Expand All @@ -150,11 +150,11 @@ function Random.randn!(rng::RNG, A::AnyCuArray{<:Union{AbstractFloat,Complex{<:A
return
end

kernel = @cuda launch=false name="rand!" kernel(A, rng.seed, rng.counter)
config = launch_configuration(kernel.fun; max_threads=64)
threads = max(32, min(config.threads, length(A)÷2))
blocks = min(config.blocks, cld(cld(length(A), 2), threads))
kernel(A, rng.seed, rng.counter; threads, blocks)
# see note in `rand!` about the launch configuration
threads = 32
blocks = cld(cld(length(A), 2), threads)

@cuda threads=threads blocks=blocks name="randn!" kernel(A, rng.seed, rng.counter)

new_counter = Int64(rng.counter) + length(A)
overflow, remainder = fldmod(new_counter, typemax(UInt32))
Expand Down
7 changes: 7 additions & 0 deletions test/base/random.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,3 +198,10 @@ end
end
end

@testset "counter overflow" begin
rng = CUDA.RNG()
# we may not be able to allocate over 4GB on the GPU, so use unified memory
c = CuArray{Float16, 5, CUDA.UnifiedMemory}(undef, 64, 32, 512, 32, 64)
rand!(rng, c)
randn!(rng, c)
end