Skip to content

Cap Broadcast Blocks #224

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
krasow wants to merge 3 commits into
base: develop
Choose a base branch
from
Open

Cap Broadcast Blocks #224

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
efeeb5e
amortize per arg setup [regression-ci]
krasow Oct 4, 2026
abf6d94
[regression-ci develop] [skip dev]
krasow Oct 4, 2026
199dbdf
Merge branch 'cap-broadcast-block' of github.com:JuliaLegate/cuNumeri…
krasow Oct 4, 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
27 changes: 27 additions & 0 deletions lib/cunumeric_jl_wrapper/src/cuda.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -335,6 +335,30 @@ static inline void align8(char *&ptr) {
// CuStridedDeviceArray) val < 0 → scalar at index -(val + 1) in
// trailing scalars

// Limit the grid to a few waves of resident blocks. Broadcast kernels
// grid-stride, so each thread then loops over many elements and amortizes its
// per-argument setup instead of paying it for a single element.
static void cap_broadcast_blocks(PTXLaunchParams &lp) {
constexpr std::uint64_t waves = 4;
int per_sm = 0, sms = 0;
CUdevice dev;
if (cuOccupancyMaxActiveBlocksPerMultiprocessor(
&per_sm, lp.func, lp.tx * lp.ty * lp.tz, 0) != CUDA_SUCCESS ||
cuCtxGetDevice(&dev) != CUDA_SUCCESS ||
cuDeviceGetAttribute(&sms, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
dev) != CUDA_SUCCESS ||
per_sm <= 0 || sms <= 0)
return;
const std::uint64_t target = waves * per_sm * sms;
const auto cap = [](std::uint32_t b, std::uint64_t limit) {
return static_cast<std::uint32_t>(
std::min<std::uint64_t>(b, std::max<std::uint64_t>(1, limit)));
};
lp.bx = cap(lp.bx, target);
lp.by = cap(lp.by, target / lp.bx);
lp.bz = cap(lp.bz, target / (std::uint64_t(lp.bx) * lp.by));
}

static void broadcast_launch_dims_from_tile(PTXLaunchParams &lp,
const legate::PhysicalArray &out) {
const std::uint32_t budget = std::max(lp.tx, 1u);
Expand Down Expand Up @@ -362,6 +386,7 @@ static void broadcast_launch_dims_from_tile(PTXLaunchParams &lp,
lp.bx = blocks(cols, lp.tx, 2147483647);
lp.by = blocks(rows, lp.ty, 65535);
lp.bz = 1;
cap_broadcast_blocks(lp);

#ifdef CUDA_DEBUG
std::cerr << "[RunPTXBroadcastTask] local shape=" << rows << "x" << cols
Expand Down Expand Up @@ -390,6 +415,7 @@ static void broadcast_launch_dims_from_tile(PTXLaunchParams &lp,
lp.bx = blocks(dim3, lp.tx, 2147483647);
lp.by = blocks(dim2, lp.ty, 65535);
lp.bz = blocks(dim1, lp.tz, 65535);
cap_broadcast_blocks(lp);

#ifdef CUDA_DEBUG
std::cerr << "[RunPTXBroadcastTask] local shape=" << dim1 << "x" << dim2
Expand Down Expand Up @@ -436,6 +462,7 @@ static void broadcast_launch_dims_from_tile(PTXLaunchParams &lp,
lp.tx = threads;
lp.ty = 1;
lp.tz = 1;
cap_broadcast_blocks(lp);

#ifdef CUDA_DEBUG
std::cerr << "[RunPTXBroadcastTask] local volume=" << volume << " dim=" << dim
Expand Down
Loading