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

Improve threading in _set_interpolated_values_device #2159

Merged
merged 1 commit into from
Jan 27, 2025
Merged

Improve threading in _set_interpolated_values_device #2159

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
223 changes: 109 additions & 114 deletions ext/cuda/remapping_distributed.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,26 +18,27 @@ function _set_interpolated_values_device!(
# FIXME: Avoid allocation of tuple
field_values = tuple(map(f -> Fields.field_values(f), fields)...)

purely_vertical_space = isnothing(interpolation_matrix)
num_horizontal_points =
purely_vertical_space ? 1 : prod(size(local_horiz_indices))
num_points = num_horizontal_points * length(vert_interpolation_weights)
max_threads = 256
nthreads = min(num_points, max_threads)
nblocks = cld(num_points, nthreads)
purely_vert_space = isnothing(interpolation_matrix)
nitems = length(out)

# For purely vertical spaces, `Nq` is not used, so we pass in -1 here.
_, Nq = purely_vert_space ? (-1, -1) : size(interpolation_matrix[1])
args = (
out,
interpolation_matrix,
local_horiz_indices,
vert_interpolation_weights,
vert_bounding_indices,
field_values,
Val(Nq),
)
threads = threads_via_occupancy(set_interpolated_values_kernel!, args)
p = linear_partition(nitems, threads)
auto_launch!(
set_interpolated_values_kernel!,
args;
threads_s = (nthreads),
blocks_s = (nblocks),
threads_s = (p.threads),
blocks_s = (p.blocks),
)
call_post_op_callback() && post_op_callback(
out,
Expand All @@ -60,41 +61,37 @@ function set_interpolated_values_kernel!(
vert_interpolation_weights,
vert_bounding_indices,
field_values,
)
# TODO: Check the memory access pattern. This was not optimized and likely inefficient!
::Val{Nq},
) where {Nq}
num_horiz = length(local_horiz_indices)
num_vert = length(vert_bounding_indices)
num_fields = length(field_values)

hindex = (blockIdx().x - Int32(1)) * blockDim().x + threadIdx().x
vindex = (blockIdx().y - Int32(1)) * blockDim().y + threadIdx().y
findex = (blockIdx().z - Int32(1)) * blockDim().z + threadIdx().z
@inbounds begin
i_thread =
(CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x +
CUDA.threadIdx().x
inds = (num_vert, num_horiz, num_fields)

1 ≤ i_thread ≤ prod(inds) || return nothing

totalThreadsX = gridDim().x * blockDim().x
totalThreadsY = gridDim().y * blockDim().y
totalThreadsZ = gridDim().z * blockDim().z
# TODO: Check the memory access pattern, we should maximize coalesced memory
(j, i, k) = CartesianIndices(map(x -> Base.OneTo(x), inds))[i_thread].I

_, Nq = size(I1)
CI = CartesianIndex
for i in hindex:totalThreadsX:num_horiz
CI = CartesianIndex
h = local_horiz_indices[i]
for j in vindex:totalThreadsY:num_vert
v_lo, v_hi = vert_bounding_indices[j]
A, B = vert_interpolation_weights[j]
for k in findex:totalThreadsZ:num_fields
if i ≤ num_horiz && j ≤ num_vert && k ≤ num_fields
out[i, j, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, j, k] +=
I1[i, t] *
I2[i, s] *
(
A * field_values[k][CI(t, s, 1, v_lo, h)] +
B * field_values[k][CI(t, s, 1, v_hi, h)]
)
end
end
end
v_lo, v_hi = vert_bounding_indices[j]
A, B = vert_interpolation_weights[j]
fvals = field_values[k]
out[i, j, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, j, k] +=
I1[i, t] *
I2[i, s] *
(
A * fvals[CI(t, s, 1, v_lo, h)] +
B * fvals[CI(t, s, 1, v_hi, h)]
)
end
end
return nothing
Expand All @@ -108,41 +105,37 @@ function set_interpolated_values_kernel!(
vert_interpolation_weights,
vert_bounding_indices,
field_values,
)
::Val{Nq},
) where {Nq}
# TODO: Check the memory access pattern. This was not optimized and likely inefficient!
num_horiz = length(local_horiz_indices)
num_vert = length(vert_bounding_indices)
num_fields = length(field_values)

hindex = (blockIdx().x - Int32(1)) * blockDim().x + threadIdx().x
vindex = (blockIdx().y - Int32(1)) * blockDim().y + threadIdx().y
findex = (blockIdx().z - Int32(1)) * blockDim().z + threadIdx().z
@inbounds begin
i_thread =
(CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x +
CUDA.threadIdx().x
inds = (num_vert, num_horiz, num_fields)

1 ≤ i_thread ≤ prod(inds) || return nothing

totalThreadsX = gridDim().x * blockDim().x
totalThreadsY = gridDim().y * blockDim().y
totalThreadsZ = gridDim().z * blockDim().z
# TODO: Check the memory access pattern, we should maximize coalesced memory
(j, i, k) = CartesianIndices(map(x -> Base.OneTo(x), inds))[i_thread].I

_, Nq = size(I)
CI = CartesianIndex
for i in hindex:totalThreadsX:num_horiz
CI = CartesianIndex
h = local_horiz_indices[i]
for j in vindex:totalThreadsY:num_vert
v_lo, v_hi = vert_bounding_indices[j]
A, B = vert_interpolation_weights[j]
for k in findex:totalThreadsZ:num_fields
if i ≤ num_horiz && j ≤ num_vert && k ≤ num_fields
out[i, j, k] = 0
for t in 1:Nq
out[i, j, k] +=
I[i, t] *
I[i, s] *
(
A * field_values[k][CI(t, 1, 1, v_lo, h)] +
B * field_values[k][CI(t, 1, 1, v_hi, h)]
)
end
end
end
v_lo, v_hi = vert_bounding_indices[j]
A, B = vert_interpolation_weights[j]
out[i, j, k] = 0
for t in 1:Nq
out[i, j, k] +=
I[i, t] *
I[i, s] *
(
A * field_values[k][CI(t, 1, 1, v_lo, h)] +
B * field_values[k][CI(t, 1, 1, v_hi, h)]
)
end
end
return nothing
Expand All @@ -156,29 +149,30 @@ function set_interpolated_values_kernel!(
vert_interpolation_weights,
vert_bounding_indices,
field_values,
)
::Val{Nq},
) where {Nq}
# TODO: Check the memory access pattern. This was not optimized and likely inefficient!
num_fields = length(field_values)
num_vert = length(vert_bounding_indices)

vindex = (blockIdx().y - Int32(1)) * blockDim().y + threadIdx().y
findex = (blockIdx().z - Int32(1)) * blockDim().z + threadIdx().z
@inbounds begin
i_thread =
(CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x +
CUDA.threadIdx().x
inds = (num_vert, num_fields)

1 ≤ i_thread ≤ prod(inds) || return nothing

totalThreadsY = gridDim().y * blockDim().y
totalThreadsZ = gridDim().z * blockDim().z
# TODO: Check the memory access pattern, we should maximize coalesced memory
(j, k) = CartesianIndices(map(x -> Base.OneTo(x), inds))[i_thread].I

CI = CartesianIndex
for j in vindex:totalThreadsY:num_vert
CI = CartesianIndex
v_lo, v_hi = vert_bounding_indices[j]
A, B = vert_interpolation_weights[j]
for k in findex:totalThreadsZ:num_fields
if j ≤ num_vert && k ≤ num_fields
out[j, k] = (
A * field_values[k][CI(1, 1, 1, v_lo, 1)] +
B * field_values[k][CI(1, 1, 1, v_hi, 1)]
)
end
end
out[j, k] = (
A * field_values[k][CI(1, 1, 1, v_lo, 1)] +
B * field_values[k][CI(1, 1, 1, v_hi, 1)]
)
end
return nothing
end
Expand All @@ -205,18 +199,21 @@ function _set_interpolated_values_device!(
# FIXME: Avoid allocation of tuple
field_values = tuple(map(f -> Fields.field_values(f), fields)...)
nitems = length(out)
nthreads, nblocks = _configure_threadblock(nitems)

args = (
out,
local_horiz_interpolation_weights,
local_horiz_indices,
field_values,
Val(Nq),
)
threads = threads_via_occupancy(set_interpolated_values_kernel!, args)
p = linear_partition(nitems, threads)
auto_launch!(
set_interpolated_values_kernel!,
args;
threads_s = (nthreads),
blocks_s = (nblocks),
threads_s = p.threads,
blocks_s = p.blocks,
)
end

Expand All @@ -225,31 +222,30 @@ function set_interpolated_values_kernel!(
(I1, I2)::NTuple{2},
local_horiz_indices,
field_values,
)
::Val{Nq},
) where {Nq}
# TODO: Check the memory access pattern. This was not optimized and likely inefficient!
num_horiz = length(local_horiz_indices)
num_fields = length(field_values)

hindex = (blockIdx().x - Int32(1)) * blockDim().x + threadIdx().x
findex = (blockIdx().z - Int32(1)) * blockDim().z + threadIdx().z
@inbounds begin
i_thread =
(CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x +
CUDA.threadIdx().x
inds = (num_horiz, num_fields)

totalThreadsX = gridDim().x * blockDim().x
totalThreadsZ = gridDim().z * blockDim().z
1 ≤ i_thread ≤ prod(inds) || return nothing

_, Nq = size(I1)
# TODO: Check the memory access pattern, we should maximize coalesced memory
(i, k) = CartesianIndices(map(x -> Base.OneTo(x), inds))[i_thread].I

for i in hindex:totalThreadsX:num_horiz
h = local_horiz_indices[i]
for k in findex:totalThreadsZ:num_fields
if i ≤ num_horiz && k ≤ num_fields
out[i, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, k] +=
I1[i, t] *
I2[i, s] *
field_values[k][CartesianIndex(t, s, 1, 1, h)]
end
end
out[i, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, k] +=
I1[i, t] *
I2[i, s] *
field_values[k][CartesianIndex(t, s, 1, 1, h)]
end
end
return nothing
Expand All @@ -260,29 +256,28 @@ function set_interpolated_values_kernel!(
(I,)::NTuple{1},
local_horiz_indices,
field_values,
)
::Val{Nq},
) where {Nq}
# TODO: Check the memory access pattern. This was not optimized and likely inefficient!
num_horiz = length(local_horiz_indices)
num_fields = length(field_values)

hindex = (blockIdx().x - Int32(1)) * blockDim().x + threadIdx().x
findex = (blockIdx().z - Int32(1)) * blockDim().z + threadIdx().z
@inbounds begin
i_thread =
(CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x +
CUDA.threadIdx().x
inds = (num_horiz, num_fields)

totalThreadsX = gridDim().x * blockDim().x
totalThreadsZ = gridDim().z * blockDim().z
1 ≤ i_thread ≤ prod(inds) || return nothing

_, Nq = size(I)
# TODO: Check the memory access pattern, we should maximize coalesced memory
(i, k) = CartesianIndices(map(x -> Base.OneTo(x), inds))[i_thread].I

for i in hindex:totalThreadsX:num_horiz
h = local_horiz_indices[i]
for k in findex:totalThreadsZ:num_fields
if i ≤ num_horiz && k ≤ num_fields
out[i, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, k] +=
I[i, i] * field_values[k][CartesianIndex(t, 1, 1, 1, h)]
end
end
out[i, k] = 0
for t in 1:Nq, s in 1:Nq
out[i, k] +=
I[i, i] * field_values[k][CartesianIndex(t, 1, 1, 1, h)]
end
end
return nothing
Expand Down
Loading