Skip to content

Fix DivideError in 2D parallel_for for imbalanced dimensions #385

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

Merged
PhilipFackler merged 4 commits into from
Jul 14, 2026
Merged

Fix DivideError in 2D parallel_for for imbalanced dimensions #385

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0bf5ef3
Fix DivideError in 2D parallel_for for imbalanced dimensions
riteshbhirud Jul 9, 2026
d99c636
Handle n > m case in 2D parallel_for thread split
riteshbhirud Jul 9, 2026
bd8f40a
Fix 2d dimension swapping comparison
PhilipFackler Jul 9, 2026
9ca4126
Merge branch 'main' into fix-2d-parallel-for-divideerror
williamfgc Jul 13, 2026
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
8 changes: 4 additions & 4 deletions ext/AMDGPUExt/AMDGPUExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,7 +105,7 @@ function _parallel_for(indexer::TI, f, (m, n), (M, N), x...) where {TI}
kernel, shmem_size = _kernel_maxshmem(_parallel_for_amdgpu_MN, kargs)
config = AMDGPU.launch_configuration(kernel; shmem = shmem_size)
maxThreadsX = sqrt(config.groupsize)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, config.groupsize)
x_thr = fld(config.groupsize, y_thr)
threads = (x_thr, y_thr)
blocks = (cld(m, x_thr), cld(n, y_thr))
Expand All @@ -118,7 +118,7 @@ function JACC.parallel_for(
dev = AMDGPU.device()
props = AMDGPU.HIP.properties(dev)
maxBlocks = (x = props.maxGridSize[1], y = props.maxGridSize[2])
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, (M, N), (M, N), x...)
Expand All @@ -137,7 +137,7 @@ function _parallel_for(indexer::TI, f, spec::LaunchSpec{AMDGPUBackend}, (m, n),
if spec.threads == 0
config = AMDGPU.launch_configuration(kernel; shmem = spec.shmem_size)
maxThreadsX = sqrt(config.groupsize)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, config.groupsize)
x_thr = fld(config.groupsize, y_thr)
spec.threads = (x_thr, y_thr)
end
Expand All @@ -158,7 +158,7 @@ function JACC.parallel_for(
dev = AMDGPU.device()
props = AMDGPU.HIP.properties(dev)
maxBlocks = (x = props.maxGridSize[1], y = props.maxGridSize[2])
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, spec, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, spec, (M, N), (M, N), x...)
Expand Down
8 changes: 4 additions & 4 deletions ext/CUDAExt/CUDAExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,7 +100,7 @@ function _parallel_for(indexer::TI, f, (m, n), (M, N), x...) where {TI}
kernel, shmem_size = _kernel_maxshmem(_parallel_for_cuda_MN, kargs)
config = CUDA.launch_configuration(kernel.fun; shmem = shmem_size)
maxThreadsX = sqrt(config.threads)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, config.threads)
x_thr = fld(config.threads, y_thr)
threads = (x_thr, y_thr)
blocks = (cld(m, x_thr), cld(n, y_thr))
Expand All @@ -114,7 +114,7 @@ function JACC.parallel_for(f, ::CUDABackend, (M, N)::NTuple{2, Integer}, x...)
x = attribute(dev, CUDA.DEVICE_ATTRIBUTE_MAX_GRID_DIM_X),
y = attribute(dev, CUDA.DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y)
)
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, (M, N), (M, N), x...)
Expand All @@ -133,7 +133,7 @@ function _parallel_for(indexer::TI, f, spec::LaunchSpec{CUDABackend}, (m, n),
if spec.threads == 0
config = CUDA.launch_configuration(kernel.fun; shmem = spec.shmem_size)
maxThreadsX = sqrt(config.threads)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, config.threads)
x_thr = fld(config.threads, y_thr)
spec.threads = (x_thr, y_thr)
end
Expand All @@ -156,7 +156,7 @@ function JACC.parallel_for(
x = attribute(dev, CUDA.DEVICE_ATTRIBUTE_MAX_GRID_DIM_X),
y = attribute(dev, CUDA.DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y)
)
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, spec, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, spec, (M, N), (M, N), x...)
Expand Down
8 changes: 4 additions & 4 deletions ext/oneAPIExt/oneAPIExt.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,7 +71,7 @@ function _parallel_for(indexer::TI, f, (m, n), (M, N), x...) where {TI}
kernel = @oneapi launch=false _parallel_for_oneapi_MN(indexer, (M, N), f, x...)
maxThreads = div(oneAPI.launch_configuration(kernel), 2)
maxThreadsX = sqrt(maxThreads)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, maxThreads)
x_thr = fld(maxThreads, y_thr)
items = (x_thr, y_thr)
groups = (cld(m, items[1]), cld(n, items[2]))
Expand All @@ -85,7 +85,7 @@ function JACC.parallel_for(
dev = oneAPI.device()
props = oneAPI.compute_properties(dev)
maxBlocks = (x = props.maxGroupCountX, y = props.maxGroupCountY)
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, (M, N), (M, N), x...)
Expand All @@ -99,7 +99,7 @@ function _parallel_for(indexer::TI, f, spec::LaunchSpec{oneAPIBackend}, (m, n),
if spec.threads == 0
maxThreads = oneAPI.launch_configuration(kernel)
maxThreadsX = sqrt(maxThreads)
y_thr = floor(Int, (n / m) * maxThreadsX)
y_thr = clamp(floor(Int, (n / m) * maxThreadsX), 1, maxThreads)
x_thr = fld(maxThreads, y_thr)
spec.threads = (x_thr, y_thr)
end
Expand All @@ -120,7 +120,7 @@ function JACC.parallel_for(
dev = oneAPI.device()
props = oneAPI.compute_properties(dev)
maxBlocks = (x = props.maxGroupCountX, y = props.maxGroupCountY)
if M < N && maxBlocks.x > maxBlocks.y
if M < N && maxBlocks.x >= maxBlocks.y
_parallel_for(BlockIndexerSwapped(), f, spec, (N, M), (M, N), x...)
else
_parallel_for(BlockIndexerBasic(), f, spec, (M, N), (M, N), x...)
Expand Down
21 changes: 21 additions & 0 deletions test/unittests.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -571,6 +571,27 @@ end
@test JACC.to_host(C)≈C_expected rtol=1e-5
end

@testset "Add-2D imbalanced" begin
function add!(i, j, A, B, C)
@inbounds C[i, j] = A[i, j] + B[i, j]
end

for (M, N) in ((1024, 16), (16, 1024))
C_expected = Float32(2.0) .* ones(Float32, M, N)

A = JACC.ones(Float32, M, N)
B = JACC.ones(Float32, M, N)

C = JACC.zeros(Float32, M, N)
JACC.parallel_for((M, N), add!, A, B, C)
@test JACC.to_host(C)≈C_expected rtol=1e-5

C_spec = JACC.zeros(Float32, M, N)
JACC.parallel_for(JACC.launch_spec(), (M, N), add!, A, B, C_spec)
@test JACC.to_host(C_spec)≈C_expected rtol=1e-5
end
end

@testset "Add-3D" begin
function add!(i, j, k, A, B, C)
@inbounds C[i, j, k] = A[i, j, k] + B[i, j, k]
Expand Down
Loading