Skip to content

performance improvement with manual cholesky mapping #1064

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

Open
rohany wants to merge 2 commits into
base: branch-24.03
Choose a base branch
from
Open

performance improvement with manual cholesky mapping #1064

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
eb638e8
performance improvement with manual cholesky mapping
Oct 31, 2023
defdcf4
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Oct 31, 2023
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
63 changes: 43 additions & 20 deletions cunumeric/linalg/cholesky.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,23 @@
from ..runtime import Runtime


def get_gpu_lower_triangular(point):
assert len(point) == 2
from legate.core import get_machine
from legate.core.machine import ProcessorKind

machine = get_machine()
num_gpus = machine.count(ProcessorKind.GPU)
gpus = machine.only(ProcessorKind.GPU)

# The linearized block-cyclic lower-triangular-blocked decomposition
# mapping of 2-d points to blocks is:
def mapping(i, j):
return (j + (i * (i + 1)) // 2) % num_gpus

return gpus[mapping(point[0], point[1])]


def transpose_copy_single(
context: Context, input: Store, output: Store
) -> None:
Expand Down Expand Up @@ -83,7 +100,8 @@ def potrf(context: Context, p_output: StorePartition, i: int) -> None:
task.throws_exception(LinAlgError)
task.add_output(p_output)
task.add_input(p_output)
task.execute()
with get_gpu_lower_triangular((i, i)):
task.execute()


def trsm(
Expand All @@ -95,14 +113,16 @@ def trsm(
rhs = p_output.get_child_store(i, i)
lhs = p_output

launch_domain = Rect(lo=(lo, i), hi=(hi, i + 1))
task = context.create_manual_task(
CuNumericOpCode.TRSM, launch_domain=launch_domain
)
task.add_output(lhs)
task.add_input(rhs)
task.add_input(lhs)
task.execute()
for point in Rect(lo=(lo, i), hi=(hi, i + 1)):
task = context.create_manual_task(
CuNumericOpCode.TRSM,
launch_domain=Rect(lo=point, hi=point, exclusive=False),
)
task.add_output(lhs)
task.add_input(rhs)
task.add_input(lhs)
with get_gpu_lower_triangular(point):
task.execute()


def syrk(context: Context, p_output: StorePartition, k: int, i: int) -> None:
Expand All @@ -116,7 +136,8 @@ def syrk(context: Context, p_output: StorePartition, k: int, i: int) -> None:
task.add_output(lhs)
task.add_input(rhs)
task.add_input(lhs)
task.execute()
with get_gpu_lower_triangular((k, k)):
task.execute()


def gemm(
Expand All @@ -134,18 +155,20 @@ def gemm(
lhs = p_output
rhs1 = p_output

launch_domain = Rect(lo=(lo, k), hi=(hi, k + 1))
task = context.create_manual_task(
CuNumericOpCode.GEMM, launch_domain=launch_domain
)
task.add_output(lhs)
task.add_input(rhs1, proj=lambda p: (p[0], i))
task.add_input(rhs2)
task.add_input(lhs)
task.execute()
for point in Rect(lo=(lo, k), hi=(hi, k + 1)):
task = context.create_manual_task(
CuNumericOpCode.GEMM,
launch_domain=Rect(lo=point, hi=point, exclusive=False),
)
task.add_output(lhs)
task.add_input(rhs1, proj=lambda p: (p[0], i))
task.add_input(rhs2)
task.add_input(lhs)
with get_gpu_lower_triangular(point):
task.execute()


MIN_CHOLESKY_TILE_SIZE = 2048
MIN_CHOLESKY_TILE_SIZE = 4096
MIN_CHOLESKY_MATRIX_SIZE = 8192


Expand Down