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

Refactor/client #17

Merged
merged 2 commits into from
Jul 19, 2024
Merged

Refactor/client #17

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e14ae42
Refactor: use the client instead of the device
nathanielsimard Jul 19, 2024
59587ac
Cleanup
nathanielsimard Jul 19, 2024
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
142 changes: 95 additions & 47 deletions crates/cubecl-linalg/src/matmul/cmma/launch.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,44 +1,116 @@
use std::cmp::max;

use cubecl_core::{
frontend::{Float, TensorArg, F16},
Compiler, Runtime,
client::ComputeClient,
frontend::{Float, TensorArg, TensorHandleRef, F16},
ir::{Elem, FloatKind},
Compiler, Feature, Runtime,
};

use crate::{
matmul::cmma::{
base::cmma_kernel,
config::{cmma_cube_count, cmma_cube_dim, CmmaConfig, CmmaLaunchConfig},
},
tensor::{MatrixLayout, TensorHandle},
tensor::{matrix_layout, MatrixLayout, TensorHandle},
};

/// Matrix multiplication using tiling 2d algorithm
/// Matrix multiplication using [cooperative matrix-multiply and accumulate operations](cubecl_core::cmma).
pub fn matmul_cmma<R: Runtime, F: Float>(
client: &ComputeClient<R::Server, R::Channel>,
lhs: TensorHandle<R, F>,
rhs: TensorHandle<R, F>,
out: TensorHandle<R, F>,
device: &R::Device,
) -> TensorHandle<R, F> {
let rank = lhs.rank();
let m = lhs.shape[rank - 2];
let k = lhs.shape[rank - 1];
let n = rhs.shape[rank - 1];
matmul_cmma_ref::<R, F>(client, lhs.as_ref(), rhs.as_ref(), out.as_ref());
out
}

let client = R::client(device);
#[derive(Debug)]
pub enum UnavailabilityReason {
TransposedInput, // TODO: Support that case.
NotMultipleOf4, // TODO: Support that case.
HiglyPermutatedInput,
ShapeMemoryLimitBusted,
InvalidConfig(String),
CmmaInstructionsUnsupported,
}

let check_layout = |tensor: &TensorHandle<R, F>| match tensor.matrix_layout() {
MatrixLayout::Contiguous => {}
/// Checks if the matmul cmma can be used.
pub fn check_cmma_availability<R: Runtime>(
client: &ComputeClient<R::Server, R::Channel>,
lhs: &TensorHandleRef<'_, R>,
rhs: &TensorHandleRef<'_, R>,
config: Option<&CmmaLaunchConfig>,
) -> Result<(), UnavailabilityReason> {
let check_layout = |tensor: &TensorHandleRef<'_, R>| match matrix_layout(tensor.strides) {
MatrixLayout::Contiguous => Ok(()),
MatrixLayout::MildlyPermuted {
transposed: _,
batch_swap: _,
} => panic!("Transposed input not supported yet."),
MatrixLayout::HighlyPermuted => {
panic!("Can't run on highly permuted tensor.")
}
} => Err(UnavailabilityReason::TransposedInput),
MatrixLayout::HighlyPermuted => Err(UnavailabilityReason::HiglyPermutatedInput),
};
check_layout(&lhs);
check_layout(&rhs);

if !client.features().enabled(Feature::Cmma {
a: Elem::Float(FloatKind::F16),
b: Elem::Float(FloatKind::F16),
c: Elem::Float(FloatKind::F32),
m: 16,
k: 16,
n: 16,
}) {
return Err(UnavailabilityReason::CmmaInstructionsUnsupported);
}

check_layout(lhs)?;
check_layout(rhs)?;

let rank = lhs.shape.len();
let m = lhs.shape[rank - 2];
let k = lhs.shape[rank - 1];
let n = rhs.shape[rank - 1];

if !(m % 4 == 0 && k % 4 == 0 && n % 4 == 0) {
return Err(UnavailabilityReason::NotMultipleOf4);
}

if let Some(config) = config {
let (b_m, b_k, b_n) = (
config.block_size_m,
config.block_size_k,
config.block_size_n,
);

if b_k * max(b_m, b_n) > <R::Compiler as Compiler>::max_shared_memory_size() {
return Err(UnavailabilityReason::ShapeMemoryLimitBusted);
}

if b_m * b_n > <R::Compiler as Compiler>::max_shared_memory_size() {
return Err(UnavailabilityReason::ShapeMemoryLimitBusted);
}

if b_k != 2 * config.tile_size {
return Err(UnavailabilityReason::InvalidConfig(
"Variable tile number per coop_units not supported".to_string(),
));
}
}

Ok(())
}
/// Matrix multiplication using [cooperative matrix-multiply and accumulate operations](cubecl_core::cmma).
pub fn matmul_cmma_ref<R: Runtime, F: Float>(
client: &ComputeClient<R::Server, R::Channel>,
lhs: TensorHandleRef<'_, R>,
rhs: TensorHandleRef<'_, R>,
out: TensorHandleRef<'_, R>,
) {
let rank = lhs.strides.len();

let m = lhs.shape[rank - 2];
let k = lhs.shape[rank - 1];
let n = rhs.shape[rank - 1];

let vectorization = |shape: usize| {
[4, 2]
Expand All @@ -53,41 +125,17 @@ pub fn matmul_cmma<R: Runtime, F: Float>(
let rhs_vectorization = vectorization(n);
let out_vectorization = vectorization(n);

let cube_count = cmma_cube_count::<R>(&out.shape, 64, 64);
let cube_count = cmma_cube_count::<R>(out.shape, 64, 64);
let cube_dim = cmma_cube_dim();
let launch_config = CmmaLaunchConfig::default();
let (b_m, b_k, b_n) = (
launch_config.block_size_m,
launch_config.block_size_k,
launch_config.block_size_n,
);

assert!(
lhs_vectorization == 4 && rhs_vectorization == 4 && out_vectorization == 4,
"Only vec4 is supported"
);
assert!(
b_k * max(b_m, b_n) <= <R::Compiler as Compiler>::max_shared_memory_size(),
"Shared memory limit will be busted. "
);
assert!(
b_m * b_n <= <R::Compiler as Compiler>::max_shared_memory_size(),
"Shared memory limit will be busted. "
);
assert!(
b_k == 2 * launch_config.tile_size,
"Variable tile number per coop_units not supported"
);

cmma_kernel::launch::<F, F16, R>(
&client,
client,
cube_count,
cube_dim,
TensorArg::vectorized(lhs_vectorization, &lhs.handle, &lhs.strides, &lhs.shape),
TensorArg::vectorized(rhs_vectorization, &rhs.handle, &rhs.strides, &rhs.shape),
TensorArg::vectorized(out_vectorization, &out.handle, &out.strides, &out.shape),
TensorArg::vectorized(lhs_vectorization, lhs.handle, lhs.strides, lhs.shape),
TensorArg::vectorized(rhs_vectorization, rhs.handle, rhs.strides, rhs.shape),
TensorArg::vectorized(out_vectorization, out.handle, out.strides, out.shape),
CmmaConfig::new(m, k, n, launch_config),
);

out
}
4 changes: 3 additions & 1 deletion crates/cubecl-linalg/src/matmul/cmma/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,4 +7,6 @@ mod launch;
pub(crate) mod load_shared_memory;
pub(crate) mod write_output;

pub use launch::matmul_cmma;
pub use launch::check_cmma_availability as is_available;
pub use launch::matmul_cmma as launch;
pub use launch::matmul_cmma_ref as launch_ref;
16 changes: 16 additions & 0 deletions crates/cubecl-linalg/src/matmul/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
use cubecl_core::prelude::*;

/// Contains algorithms for cooperative matrix multiplication.
pub mod cmma;

Expand All @@ -7,3 +9,17 @@ pub mod tiling2d;

#[cfg(feature = "export_tests")]
pub mod tests;

/// Launch a matrix multiplication kernel.
pub fn launch_ref<R: Runtime, F: Float>(
client: &ComputeClient<R::Server, R::Channel>,
lhs: TensorHandleRef<'_, R>,
rhs: TensorHandleRef<'_, R>,
out: TensorHandleRef<'_, R>,
) {
if cmma::is_available(client, &lhs, &rhs, None).is_ok() {
cmma::launch_ref::<R, F>(client, lhs, rhs, out);
} else {
tiling2d::launch_ref::<R, F>(client, lhs, rhs, out, Default::default());
}
}
29 changes: 16 additions & 13 deletions crates/cubecl-linalg/src/matmul/tests/cmma/compute_loop.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,9 +66,10 @@ pub fn compute_loop_k_test<R: Runtime>(device: &R::Device) {
let m = 16;
let k = 32;
let n = 16;
let lhs = range_tensor_f16::<R>(m, k, device);
let rhs = range_tensor_f16::<R>(k, n, device);
let results = create_empty::<R>(m, n, device);
let client = R::client(device);
let lhs = range_tensor_f16::<R>(&client, m, k);
let rhs = range_tensor_f16::<R>(&client, k, n);
let results = create_empty::<R>(&client, m, n);
let cube_dim = CubeDim::new(32, 1, 1);
let cube_count = CubeCount::Static(1, 1, 1);

Expand Down Expand Up @@ -128,7 +129,7 @@ pub fn compute_loop_k_test<R: Runtime>(device: &R::Device) {
3659328., 3671344., 3683360., 3695376.,
];

assert_equals::<R>(results, expected, device);
assert_equals::<R>(&client, results, expected);
}

/// Exported test
Expand All @@ -141,9 +142,10 @@ pub fn compute_loop_warp_test<R: Runtime>(device: &R::Device) {
let m = 16;
let k = 32;
let n = 32;
let lhs = range_tensor_f16::<R>(m, k, device);
let rhs = range_tensor_f16::<R>(k, n, device);
let results = create_empty::<R>(m, n, device);
let client = R::client(device);
let lhs = range_tensor_f16::<R>(&client, m, k);
let rhs = range_tensor_f16::<R>(&client, k, n);
let results = create_empty::<R>(&client, m, n);
let cube_dim = CubeDim::new(32, 1, 1);
let cube_count = CubeCount::Static(1, 1, 1);

Expand Down Expand Up @@ -231,7 +233,7 @@ pub fn compute_loop_warp_test<R: Runtime>(device: &R::Device) {
9763456., 9775472., 9787488., 9799504., 9811520., 9823536., 9835552., 9847568.,
];

assert_equals::<R>(results, expected, device);
assert_equals::<R>(&client, results, expected);
}

/// Exported test
Expand All @@ -244,10 +246,11 @@ pub fn cmma_compute_loop_two_warps_same_tile_row_test<R: Runtime>(device: &R::De
let m = 16;
let k = 32;
let n = 64;
let client = R::client(device);

let lhs = range_tensor_f16::<R>(m, k, device);
let rhs = range_tensor_f16::<R>(k, n, device);
let results = create_empty::<R>(m, n, device);
let lhs = range_tensor_f16::<R>(&client, m, k);
let rhs = range_tensor_f16::<R>(&client, k, n);
let results = create_empty::<R>(&client, m, n);
let cube_dim = CubeDim::new(32, 2, 1);
let cube_count = CubeCount::Static(1, 1, 1);

Expand All @@ -263,7 +266,7 @@ pub fn cmma_compute_loop_two_warps_same_tile_row_test<R: Runtime>(device: &R::De
};

compute_loop_test::launch::<F32, F16, R>(
&R::client(device),
&client,
cube_count,
cube_dim,
TensorArg::new(&lhs.handle, &lhs.strides, &lhs.shape),
Expand Down Expand Up @@ -413,5 +416,5 @@ pub fn cmma_compute_loop_two_warps_same_tile_row_test<R: Runtime>(device: &R::De
22103888.0, 22115904.0, 22127920.0, 22139936.0, 22151952.0,
];

assert_equals::<R>(results, expected, device);
assert_equals::<R>(&client, results, expected);
}
Loading
Loading