Skip to content

i8 x i8 MaxSim Kernels #1429

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
Suryansh Gupta (suri-kumkaran) wants to merge 10 commits into
base: main
Choose a base branch
from
+2,489 −151
Open

i8 x i8 MaxSim Kernels #1429

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
59e22f3
PACK support in packed.rs
Sep 16, 2026
f7cf1a3
Merge branch 'main' into users/suryangupta/i8_x_i8_maxsim_kernel
Sep 17, 2026
6b5293d
Add v3 and Scalar kernel
Sep 21, 2026
12e04bd
Add Neon kernel
suri-kumkaran Sep 21, 2026
0aef974
Fix review comments
Sep 24, 2026
f043eee
Widen i8 to i16 for v3 kernel
Sep 25, 2026
a3af61e
Merge branch 'main' into users/suryangupta/i8_x_i8_maxsim_kernel
Sep 29, 2026
aa4e102
Implement v4 i8 x i8 kernels
Oct 1, 2026
4d0b7aa
Merge branch 'main' into users/suryangupta/i8_x_i8_maxsim_kernel
Oct 1, 2026
c91a7fa
Address Review Comments
Oct 3, 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
22 changes: 22 additions & 0 deletions diskann-benchmark/example/multi-vector.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,28 @@
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 2, "num_measurements": 1 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "auto",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 2, "num_measurements": 1 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 2, "num_measurements": 1 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "reference",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 2, "num_measurements": 1 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 2, "num_measurements": 1 }
]
}
}
]
}
90 changes: 90 additions & 0 deletions diskann-benchmark/perf_test_inputs/multi-vector.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,6 +144,96 @@
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "auto",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 500, "num_measurements": 50 },
{ "num_query_vectors": 16, "num_doc_vectors": 64, "dim": 256, "loops_per_measurement": 100, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 128, "dim": 384, "loops_per_measurement": 20, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 50, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 1250, "dim": 128, "loops_per_measurement": 10, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 1250, "dim": 512, "loops_per_measurement": 2, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "scalar",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 500, "num_measurements": 50 },
{ "num_query_vectors": 16, "num_doc_vectors": 64, "dim": 256, "loops_per_measurement": 100, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 128, "dim": 384, "loops_per_measurement": 20, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 50, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 1250, "dim": 128, "loops_per_measurement": 10, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 1250, "dim": 512, "loops_per_measurement": 2, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "x86-64-v3",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 500, "num_measurements": 50 },
{ "num_query_vectors": 16, "num_doc_vectors": 64, "dim": 256, "loops_per_measurement": 100, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 128, "dim": 384, "loops_per_measurement": 20, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 50, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 1250, "dim": 128, "loops_per_measurement": 10, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 1250, "dim": 512, "loops_per_measurement": 2, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "x86-64-v4",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 500, "num_measurements": 50 },
{ "num_query_vectors": 16, "num_doc_vectors": 64, "dim": 256, "loops_per_measurement": 100, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 128, "dim": 384, "loops_per_measurement": 20, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 50, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 1250, "dim": 128, "loops_per_measurement": 10, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 1250, "dim": 512, "loops_per_measurement": 2, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
},
{
"type": "multi-vector-op",
"content": {
"element_type": "int8",
"isa": "reference",
"runs": [
{ "num_query_vectors": 8, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 500, "num_measurements": 50 },
{ "num_query_vectors": 16, "num_doc_vectors": 64, "dim": 256, "loops_per_measurement": 100, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 128, "dim": 384, "loops_per_measurement": 20, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 16, "dim": 256, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 264, "loops_per_measurement": 50, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 1250, "dim": 128, "loops_per_measurement": 10, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 1250, "dim": 512, "loops_per_measurement": 2, "num_measurements": 50 },
{ "num_query_vectors": 64, "num_doc_vectors": 32, "dim": 128, "loops_per_measurement": 200, "num_measurements": 50 },
{ "num_query_vectors": 32, "num_doc_vectors": 32, "dim": 512, "loops_per_measurement": 50, "num_measurements": 50 }
]
}
}
]
}
8 changes: 5 additions & 3 deletions diskann-benchmark/src/multi_vector/driver.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,9 @@ use diskann_benchmark_runner::{
},
Checker, Input,
};
use diskann_quantization::multi_vector::{Mat, MatRef, MaxSimKernel, Overflow, Standard};
use diskann_quantization::multi_vector::{
Mat, MatRef, MaxSimElement, MaxSimKernel, Overflow, Standard,
};
use rand::{
distr::{Distribution, StandardUniform},
rngs::StdRng,
Expand Down Expand Up @@ -94,12 +96,12 @@ where
// Timing harness //
//////////////////////

pub(super) fn run_with_kernel<T: Copy>(
pub(super) fn run_with_kernel<T: MaxSimElement>(
run: &Run,
doc: MatRef<'_, Standard<T>>,
kernel: &dyn MaxSimKernel<T>,
) -> RunResult {
let mut scores = vec![0.0f32; run.num_query_vectors.get()];
let mut scores = vec![T::Score::default(); run.num_query_vectors.get()];
let mut latencies = Vec::with_capacity(run.num_measurements.get());

for _ in 0..run.num_measurements.get() {
Expand Down
1 change: 1 addition & 0 deletions diskann-benchmark/src/multi_vector/kernels.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,5 +147,6 @@ where
pub(super) fn register(registry: &mut Registry) -> anyhow::Result<()> {
registry.register_regression("multi-vector-op-f32", Kernel::<f32>::new())?;
registry.register_regression("multi-vector-op-f16", Kernel::<half::f16>::new())?;
registry.register_regression("multi-vector-op-i8", Kernel::<i8>::new())?;
Ok(())
}
Loading
Loading