Rust-related/luminal
1
0
Fork 1
mirror of https://github.com/jafioti/luminal.git synced 2026-06-01 21:49:47 +09:00
Code Issues Packages Projects Releases Wiki Activity

fmt

Browse Source
This commit is contained in:
Joe Fioti
2026-01-05 21:16:50 -08:00
parent a1af80c677
commit 33133ad7a8
2 changed files with 61 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
crates/luminal_cuda/src/kernel/ops.rs
View File

@@ -17,8 +17,13 @@ use luminal::{
};
pub type Ops = (
KernelAdd, KernelMul, KernelIota, KernelGather,
KernelSumReduce, KernelMaxReduce, KernelMeanReduce,
KernelAdd,
KernelMul,
KernelIota,
KernelGather,
KernelSumReduce,
KernelMaxReduce,
KernelMeanReduce,
);
#[derive(Default, Debug, Clone)]
@@ -77,10 +82,8 @@ impl EgglogOp for KernelMaxReduce {
vec![children[2]],
)
}
}
impl KernelOp for KernelMaxReduce {
fn compile(
&self,
@@ -195,7 +198,6 @@ extern \"C\" {{
}
}
#[derive(Default, Debug, Clone)]
pub struct KernelMeanReduce {
@@ -256,10 +258,8 @@ impl EgglogOp for KernelMeanReduce {
vec![children[2]],
)
}
}
impl KernelOp for KernelMeanReduce {
fn compile(
&self,
@@ -373,7 +373,6 @@ extern \"C\" {{
}
}
#[derive(Default, Debug, Clone)]
pub struct KernelSumReduce {
@@ -430,10 +429,8 @@ impl EgglogOp for KernelSumReduce {
vec![children[2]],
)
}
}
impl KernelOp for KernelSumReduce {
fn compile(
&self,
@@ -547,7 +544,6 @@ extern \"C\" {{
}
}
#[derive(Default, Debug, Clone)]
pub struct KernelAdd {
out_shape: Vec<Expression>,

72
crates/luminal_cuda/src/tests.rs
View File

@@ -2,7 +2,7 @@ use cudarc::driver::CudaContext;
use luminal::prelude::*;
use proptest::prelude::*;
use crate::runtime::{CudaRuntime};
use crate::runtime::CudaRuntime;
proptest! {
#![proptest_config(ProptestConfig::with_cases(5))]
@@ -32,13 +32,12 @@ proptest! {
}
}
#[test]
pub fn cuda_sum_reduce_test() {
let mut cx = Graph::default();
let input = cx.tensor((1000, 1000));
let sum_dim0 = input.sum(0).output(); // row sum
let sum_dim1 = input.sum(1).output(); // col sum
let sum_dim0 = input.sum(0).output(); // row sum
let sum_dim1 = input.sum(1).output(); // col sum
let data: Vec<f32> = (0..1_000_000).map(|i| (i % 100) as f32 * 0.01).collect();
@@ -65,8 +64,18 @@ pub fn cuda_sum_reduce_test() {
for i in 0..1000 {
let rel_err_0 = (out_dim0[i] - expected_dim0[i]).abs() / expected_dim0[i].abs().max(1.0);
let rel_err_1 = (out_dim1[i] - expected_dim1[i]).abs() / expected_dim1[i].abs().max(1.0);
assert!(rel_err_0 < 0.001, "dim0 mismatch at {i}: got {}, expected {}", out_dim0[i], expected_dim0[i]);
assert!(rel_err_1 < 0.001, "dim1 mismatch at {i}: got {}, expected {}", out_dim1[i], expected_dim1[i]);
assert!(
rel_err_0 < 0.001,
"dim0 mismatch at {i}: got {}, expected {}",
out_dim0[i],
expected_dim0[i]
);
assert!(
rel_err_1 < 0.001,
"dim1 mismatch at {i}: got {}, expected {}",
out_dim1[i],
expected_dim1[i]
);
}
}
@@ -74,16 +83,24 @@ pub fn cuda_sum_reduce_test() {
pub fn cuda_max_reduce_test() {
let mut cx = Graph::default();
let input = cx.tensor((1000, 1000));
let max_dim0 = input.max(0).output(); // row max
let max_dim1 = input.max(1).output(); // col max
let max_dim0 = input.max(0).output(); // row max
let max_dim1 = input.max(1).output(); // col max
let data: Vec<f32> = (0..1_000_000).map(|i| (i % 100) as f32 * 0.01).collect();
let expected_dim0: Vec<f32> = (0..1000)
.map(|col| (0..1000).map(|row| data[row * 1000 + col]).fold(f32::NEG_INFINITY, f32::max))
.map(|col| {
(0..1000)
.map(|row| data[row * 1000 + col])
.fold(f32::NEG_INFINITY, f32::max)
})
.collect();
let expected_dim1: Vec<f32> = (0..1000)
.map(|row| (0..1000).map(|col| data[row * 1000 + col]).fold(f32::NEG_INFINITY, f32::max))
.map(|row| {
(0..1000)
.map(|col| data[row * 1000 + col])
.fold(f32::NEG_INFINITY, f32::max)
})
.collect();
let ctx = CudaContext::new(0).unwrap();
@@ -102,18 +119,27 @@ pub fn cuda_max_reduce_test() {
for i in 0..1000 {
let rel_err_0 = (out_dim0[i] - expected_dim0[i]).abs() / expected_dim0[i].abs().max(1.0);
let rel_err_1 = (out_dim1[i] - expected_dim1[i]).abs() / expected_dim1[i].abs().max(1.0);
assert!(rel_err_0 < 0.001, "dim0 mismatch at {i}: got {}, expected {}", out_dim0[i], expected_dim0[i]);
assert!(rel_err_1 < 0.001, "dim1 mismatch at {i}: got {}, expected {}", out_dim1[i], expected_dim1[i]);
assert!(
rel_err_0 < 0.001,
"dim0 mismatch at {i}: got {}, expected {}",
out_dim0[i],
expected_dim0[i]
);
assert!(
rel_err_1 < 0.001,
"dim1 mismatch at {i}: got {}, expected {}",
out_dim1[i],
expected_dim1[i]
);
}
}
#[test]
pub fn cuda_mean_reduce_test() {
let mut cx = Graph::default();
let input = cx.tensor((1000, 1000));
let mean_dim0 = input.mean(0).output(); // mean along rows
let mean_dim1 = input.mean(1).output(); // mean along cols
let mean_dim0 = input.mean(0).output(); // mean along rows
let mean_dim1 = input.mean(1).output(); // mean along cols
let data: Vec<f32> = (0..1_000_000).map(|i| (i % 100) as f32 * 0.01).collect();
@@ -140,7 +166,17 @@ pub fn cuda_mean_reduce_test() {
for i in 0..1000 {
let rel_err_0 = (out_dim0[i] - expected_dim0[i]).abs() / expected_dim0[i].abs().max(1.0);
let rel_err_1 = (out_dim1[i] - expected_dim1[i]).abs() / expected_dim1[i].abs().max(1.0);
assert!(rel_err_0 < 0.001, "dim0 mismatch at {i}: got {}, expected {}", out_dim0[i], expected_dim0[i]);
assert!(rel_err_1 < 0.001, "dim1 mismatch at {i}: got {}, expected {}", out_dim1[i], expected_dim1[i]);
assert!(
rel_err_0 < 0.001,
"dim0 mismatch at {i}: got {}, expected {}",
out_dim0[i],
expected_dim0[i]
);
assert!(
rel_err_1 < 0.001,
"dim1 mismatch at {i}: got {}, expected {}",
out_dim1[i],
expected_dim1[i]
);
}
}
}