Skip to content

refactor: Migrate to usize indexing #4273

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
wingertge wants to merge 3 commits into
base: main
Choose a base branch
from
Open

refactor: Migrate to usize indexing #4273

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c5f0a50
Migrate to `usize` indexing
wingertge Jan 4, 2026
91bc2d3
Merge branch 'main' into feat/usize
wingertge Jan 4, 2026
e60a2d3
Fix merge
wingertge Jan 5, 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
373 changes: 195 additions & 178 deletions Cargo.lock
View file
Open in desktop

Large diffs are not rendered by default.

12 changes: 6 additions & 6 deletions Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,13 +180,13 @@ portable-atomic = { version = "1.11.1" }
portable-atomic-util = { version = "0.2.4", features = ["alloc"] }

### For the main burn branch. ###
cubecl = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "48ff83f19952d053b80ab5762baf387f451e5c63" }
cubecl-common = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "48ff83f19952d053b80ab5762baf387f451e5c63" }
cubek = { git = "https://github.com/tracel-ai/cubek", default-features = false, rev = "0d9a635229d3cabfa8297ddc967ff4e783be348c" }
# cubecl = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "48ff83f19952d053b80ab5762baf387f451e5c63" }
# cubecl-common = { git = "https://github.com/tracel-ai/cubecl", default-features = false, rev = "48ff83f19952d053b80ab5762baf387f451e5c63" }
# cubek = { git = "https://github.com/tracel-ai/cubek", default-features = false, rev = "0d9a635229d3cabfa8297ddc967ff4e783be348c" }
### For local development. ###
# cubecl = { path = "../../cubecl/crates/cubecl", default-features = false }
# cubecl-common = { path = "../../cubecl/crates/cubecl-common", default-features = false }
# cubek = { path = "../../cubek/crates/cubek", default-features = false }
cubecl = { path = "../cubecl/crates/cubecl", default-features = false }
cubecl-common = { path = "../cubecl/crates/cubecl-common", default-features = false }
cubek = { path = "../cubek/crates/cubek", default-features = false }
### For the release. ###
# cubecl = { version = "=0.9.0-pre.6", default-features = false }
# cubecl-common = { version = "=0.9.0-pre.6", default-features = false }
Expand Down
2 changes: 1 addition & 1 deletion crates/burn-backend-tests/tests/common/autodiff.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,5 +30,5 @@ test_float_elem_variant!(
bf16,
burn_tensor::bf16,
"../autodiff/mod.rs",
["vulkan", "metal"] // ["cuda", "rocm"] TODO
["metal"] // ["cuda", "rocm"] TODO, ["vulkan"] only supports bf16 for matmul
Copy link
Member

@laggui laggui Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Whoops, I added bf16 for vulkan but clearly the tests don't pass 😅 thanks for fixing.

[Unrelated to this PR]

I vaguely remember adding it when refactoring the tests since it was a supported global type for vulkan in cubecl. Also reflected by B::supports_dtype:

burn/crates/burn-wgpu/src/lib.rs

Lines 117 to 145 in 8682709

#[test]
fn should_support_dtypes() {
type B = Wgpu;
let device = Default::default();
assert!(B::supports_dtype(&device, DType::F32));
assert!(B::supports_dtype(&device, DType::I64));
assert!(B::supports_dtype(&device, DType::I32));
assert!(B::supports_dtype(&device, DType::U64));
assert!(B::supports_dtype(&device, DType::U32));
assert!(B::supports_dtype(
&device,
DType::QFloat(CubeTensor::<WgpuRuntime>::default_scheme())
));
// Registered as supported type but we don't actually use it?
assert!(B::supports_dtype(&device, DType::Bool));
#[cfg(feature = "vulkan")]
{
assert!(B::supports_dtype(&device, DType::F16));
assert!(B::supports_dtype(&device, DType::BF16));
assert!(B::supports_dtype(&device, DType::I16));
assert!(B::supports_dtype(&device, DType::I8));
assert!(B::supports_dtype(&device, DType::U16));
assert!(B::supports_dtype(&device, DType::U8));
assert!(!B::supports_dtype(&device, DType::F64));
assert!(!B::supports_dtype(&device, DType::Flex32));
}

maybe we should have a better way to represent the actual supported types?

burn/crates/burn-cubecl/src/backend.rs

Lines 83 to 88 in 8682709

fn supports_dtype(device: &Self::Device, dtype: DType) -> bool {
let client = R::client(device);
let ty: StorageType = dtype.into();
client.properties().supports_type(ty.elem_type())
}

That way, tested dtypes can actually reflected supported dtypes.

Copy link
Contributor Author

@wingertge wingertge Jan 5, 2026 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The supports_type method on the CubeCL side only checks if a type is supported in any way (in this case, it's supported for conversion, as a type for buffers, and for dot product on Intel, along with tensor core instructions). It's kinda tough though because there's no good way to express that in just a single boolean (hence why the TypeUsage enum set exists in CubeCL).

This is how it's registered for Vulkan

if let Some(bfloat16) = ext_feat.bfloat16 {
    if bfloat16.shader_b_float16_type == TRUE {
        register(
            ElemType::Float(FloatKind::BF16).into(),
            TypeUsage::Conversion | TypeUsage::Buffer,
        );
    }
    if bfloat16.shader_b_float16_dot_product == TRUE {
        register(
            ElemType::Float(FloatKind::BF16).into(),
            TypeUsage::DotProduct.into(),
        );
    }
}

Copy link
Contributor Author

@wingertge wingertge Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So it's supported for matmul and casting, but none of the other ops.

Copy link
Member

@laggui laggui Jan 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The supports_type method on the CubeCL side only checks if a type is supported in any way

Yeah and for the first draft I simply mirrored that for the backends, but I think it should be refined.

It's kinda tough though because there's no good way to express that in just a single boolean (hence why the TypeUsage enum set exists in CubeCL).

That's a good point. It's still useful to query backend supported types for burn, so maybe we should also define an enum similar to TypeUsage? (without atomics, and perhaps consolidate conversion / buffer into "storage" variant or similar).

);
2 changes: 1 addition & 1 deletion crates/burn-backend-tests/tests/common/tensor.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,5 +34,5 @@ test_float_elem_variant!(
bf16,
burn_tensor::bf16,
"../tensor/float/mod.rs",
["vulkan", "metal"] // ["cuda", "rocm"] TODO
["metal"] // ["cuda", "rocm"] TODO, ["vulkan"] only supports bf16 for matmul
);
2 changes: 1 addition & 1 deletion crates/burn-backend-tests/tests/cubecl/mask_fill.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@ use super::*;
use burn_cubecl::kernel::{MaskFillStrategy, mask_fill};
use burn_tensor::Tolerance;
use burn_tensor::{Bool, Distribution, Element, Tensor, TensorPrimitive, backend::Backend};
use cubecl::std::scalar::InputScalar;
use cubecl::prelude::InputScalar;

#[test]
fn mask_fill_should_match_reference_backend() {
Expand Down
13 changes: 10 additions & 3 deletions crates/burn-cubecl-fusion/src/base.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,12 +1,15 @@
use burn_fusion::stream::Context;
use burn_std::{DType, quantization::QParamTensor};
use cubecl::quant::scheme::{QuantParam, QuantScheme};
use cubecl::{
CubeElement, Runtime,
client::ComputeClient,
ir::ElemType,
prelude::{TensorArg, TensorHandleRef},
};
use cubecl::{
ir::LineSize,
quant::scheme::{QuantParam, QuantScheme},
};
use std::marker::PhantomData;

/// Defines a fallback operation when fusion isn't possible.
Expand Down Expand Up @@ -73,15 +76,19 @@ impl<R: Runtime> CubeFusionHandle<R> {
}
}
/// Return the reference to a tensor argument.
pub fn as_tensor_arg<'a>(&'a self, shape: &'a [usize], vectorisation: u8) -> TensorArg<'a, R> {
pub fn as_tensor_arg<'a>(
&'a self,
shape: &'a [usize],
line_size: LineSize,
) -> TensorArg<'a, R> {
let handle: TensorHandleRef<'a, R> = self.as_handle_ref(shape);

unsafe {
TensorArg::from_raw_parts_and_size(
handle.handle,
handle.strides,
handle.shape,
vectorisation,
line_size,
self.dtype.size(),
)
}
Expand Down
Loading
Loading