Skip to content

armgit / uzu

Public Imported from GitHub
0

A high-performance inference engine for AI models

Fix overflow in matmul benchmark throughput (#1111)

GreatHatof52ac1de3d7c
diff --git a/crates/uzu-engine/unit/backends/common/kernel/matmul/quant_gemv_bench.rs b/crates/uzu-engine/unit/backends/common/kernel/matmul/quant_gemv_bench.rs
index c65d6af48..450179b6b 100644
--- a/crates/uzu-engine/unit/backends/common/kernel/matmul/quant_gemv_bench.rs
+++ b/crates/uzu-engine/unit/backends/common/kernel/matmul/quant_gemv_bench.rs
@@ -43,7 +43,7 @@ fn bench_gemv_typed<B: Backend, T: DataTypeRepr + Float>(
)
.unwrap();
- group.throughput(Throughput::Elements((m * n * k) as u64));
+ group.throughput(Throughput::Elements(u64::from(m) * u64::from(n) * u64::from(k)));
group.bench_function(BenchmarkId::from_parameter(shape.to_string()), |b| {
let benchmark_path = format!("{group_path}/{shape}");
iter_encode_loop_named::<B, _>(context, b, &benchmark_path, |command_buffer| {
diff --git a/crates/uzu-engine/unit/backends/metal/kernel/matmul/a8w_bench.rs b/crates/uzu-engine/unit/backends/metal/kernel/matmul/a8w_bench.rs
index 86a3ed3fa..2d35d64b8 100644
--- a/crates/uzu-engine/unit/backends/metal/kernel/matmul/a8w_bench.rs
+++ b/crates/uzu-engine/unit/backends/metal/kernel/matmul/a8w_bench.rs
@@ -259,7 +259,7 @@ fn bench_bits(
let paths = [BenchPath::A8GemmMxu, BenchPath::Bf16GemmMxu, BenchPath::Bf16Routed];
- group.throughput(Throughput::Elements((m * k * n) as u64));
+ group.throughput(Throughput::Elements(u64::from(m) * u64::from(k) * u64::from(n)));
for path in paths {
group.bench_function(BenchmarkId::new(path.label(), &shape_label), |bench| {
let benchmark_path =
diff --git a/crates/uzu-engine/unit/backends/metal/kernel/matmul/quant_gemm_bench.rs b/crates/uzu-engine/unit/backends/metal/kernel/matmul/quant_gemm_bench.rs
index 47d46e237..4da311d68 100644
--- a/crates/uzu-engine/unit/backends/metal/kernel/matmul/quant_gemm_bench.rs
+++ b/crates/uzu-engine/unit/backends/metal/kernel/matmul/quant_gemm_bench.rs
@@ -53,7 +53,7 @@ fn bench_unified_quant_typed<T: DataTypeRepr + Float>(
path_label,
label
));
- group.throughput(Throughput::Elements((m * n * k) as u64));
+ group.throughput(Throughput::Elements(u64::from(m) * u64::from(n) * u64::from(k)));
group.bench_function(BenchmarkId::from_parameter(shape.to_string()), |b| {
iter_encode_loop::<Metal, _>(context, b, |command_buffer| {
matmul
diff --git a/crates/uzu-engine/unit/backends/metal/kernel/matmul/qwen3_bench.rs b/crates/uzu-engine/unit/backends/metal/kernel/matmul/qwen3_bench.rs
index fd258eccc..7d6f014de 100644
--- a/crates/uzu-engine/unit/backends/metal/kernel/matmul/qwen3_bench.rs
+++ b/crates/uzu-engine/unit/backends/metal/kernel/matmul/qwen3_bench.rs
@@ -42,7 +42,7 @@ fn bench_qwen3_layers_typed<T: DataTypeRepr + Float>(
)
.unwrap();
- group.throughput(Throughput::Elements((m * n * k) as u64));
+ group.throughput(Throughput::Elements(u64::from(m) * u64::from(n) * u64::from(k)));
group.bench_function(BenchmarkId::from_parameter(format!("{layer}_{shape}")), |b| {
iter_encode_loop::<Metal, _>(context, b, |command_buffer| {
matmul.encode(quant_arguments(buffers.next_mut(), &input), command_buffer).expect("encode qwen3 layer");