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

Make the fast inverse test throughput-limited rather than latency-limited #7958

Merged
merged 2 commits into from
Nov 28, 2023
Merged

Make the fast inverse test throughput-limited rather than latency-limited #7958

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2fb0ccd
Make the fast inverse test throughput-limited rather than latency-lim…
abadams Nov 21, 2023
6dfdc2f
Merge branch 'main' into abadams/make_fast_inverse_test_throughput_li…
steven-johnson Nov 28, 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
11 changes: 8 additions & 3 deletions test/performance/fast_inverse.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,8 @@ int main(int argc, char **argv) {

if (target.arch == Target::ARM &&
target.os == Target::OSX) {
// vrecpe, vrecps, fmul have inverse throughputs of 1, 0.25, 0.25
// respectively, while fdiv has inverse throughput of 1.
printf("[SKIP] Apple M1 chips have division performance roughly on par with the reciprocal instruction\n");
return 0;
}
Expand All @@ -31,13 +33,16 @@ int main(int argc, char **argv) {
slow(x) = p / (slow(x) + 1) + 0 * r;
fast(x) = fast_inverse((fast(x) + 1) + 0 * r);

slow.update().vectorize(x, 4);
fast.update().vectorize(x, 4);
// Use wide vectors to ensure we're throughput-limited rather than latency-limited.
const int vec = 32;

slow.update().vectorize(x, vec);
fast.update().vectorize(x, vec);

slow.compile_jit();
fast.compile_jit();

Buffer<float> out_fast(8), out_slow(8);
Buffer<float> out_fast(vec), out_slow(vec);

double slow_time = benchmark([&]() { slow.realize(out_slow); });
double fast_time = benchmark([&]() { fast.realize(out_fast); });
Expand Down