Fixup

tracel-ai · Nov 28, 2024 · a483aa3 · a483aa3
1 parent 29e5557
commit a483aa3
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 5 deletions.
diff --git a/crates/cubecl-core/src/runtime_tests/cmma.rs b/crates/cubecl-core/src/runtime_tests/cmma.rs
@@ -220,7 +220,7 @@ pub fn test_simple_tf32<R: Runtime>(
         b: Elem::Float(FloatKind::TF32),
         c: Elem::Float(FloatKind::F32),
         m: 16,
-        k: 16,
+        k: 8,
         n: 16,
     }) {
         // We can't execute the test, skip.

diff --git a/crates/cubecl-linalg/src/matmul/components/tile/accelerated.rs b/crates/cubecl-linalg/src/matmul/components/tile/accelerated.rs
@@ -16,6 +16,7 @@ pub trait CmmaValid<I: Numeric, O: Numeric> {}
 impl CmmaValid<f16, f16> for (f16, f16) {}
 impl CmmaValid<f16, f32> for (f16, f32) {}
 impl CmmaValid<bf16, f32> for (bf16, f32) {}
+impl CmmaValid<tf32, f32> for (tf32, f32) {}
 
 macro_rules! instruction {
     ($name:ident, $m:expr, $n:expr, $k:expr) => {

diff --git a/crates/cubecl-reduce/src/test.rs b/crates/cubecl-reduce/src/test.rs
@@ -295,10 +295,10 @@ impl TestCase {
             naive_reduce_dim_kernel::launch_unchecked::<I, O, K, R>(
                 &client,
                 self.cube_count.clone(),
-                self.cube_dim.clone(),
+                self.cube_dim,
                 input_tensor,
                 output_tensor,
-                ScalarArg::new(self.reduce_dim.clone()),
+                ScalarArg::new(self.reduce_dim),
             );
         }
 
@@ -311,6 +311,7 @@ impl TestCase {
 
     fn cpu_sum_dim<F: Float>(&self, values: &[F]) -> Vec<F> {
         let mut expected = vec![F::new(0.0); self.num_output_values()];
+        #[allow(clippy::needless_range_loop)]
         for input_index in 0..values.len() {
             let output_index = self.to_output_index(input_index);
             expected[output_index] += values[input_index];
@@ -320,6 +321,7 @@ impl TestCase {
 
     fn cpu_prod_dim<F: Float>(&self, values: &[F]) -> Vec<F> {
         let mut expected = vec![F::new(1.0); self.num_output_values()];
+        #[allow(clippy::needless_range_loop)]
         for value_index in 0..values.len() {
             let output_index = self.to_output_index(value_index);
             expected[output_index] *= values[value_index];
@@ -336,6 +338,7 @@ impl TestCase {
 
     fn cpu_argmax_dim<F: Float>(&self, values: &[F]) -> Vec<u32> {
         let mut expected = vec![(F::MIN, 0_u32); self.num_output_values()];
+        #[allow(clippy::needless_range_loop)]
         for input_index in 0..values.len() {
             let output_index = self.to_output_index(input_index);
             let (best, _) = expected[output_index];
@@ -350,6 +353,7 @@ impl TestCase {
 
     fn cpu_argmin_dim<F: Float>(&self, values: &[F]) -> Vec<u32> {
         let mut expected = vec![(F::MAX, 0_u32); self.num_output_values()];
+        #[allow(clippy::needless_range_loop)]
         for input_index in 0..values.len() {
             let output_index = self.to_output_index(input_index);
             let (best, _) = expected[output_index];
@@ -382,6 +386,7 @@ impl TestCase {
             .collect()
     }
 
+    #[allow(clippy::wrong_self_convention)]
     fn from_output_coordinate(&self, coordinate: Vec<usize>) -> usize {
         coordinate
             .into_iter()

diff --git a/crates/cubecl-spirv/src/cmma.rs b/crates/cubecl-spirv/src/cmma.rs
@@ -55,7 +55,7 @@ impl<T: SpirvTarget> SpirvCompiler<T> {
             let shift = stride_item.const_u32(self, line_size.trailing_zeros());
             let stride_ty = stride_item.id(self);
             stride = self
-                .shift_left_logical(stride_ty, None, stride, shift)
+                .shift_right_logical(stride_ty, None, stride, shift)
                 .unwrap();
         }
 
@@ -121,7 +121,7 @@ impl<T: SpirvTarget> SpirvCompiler<T> {
             let shift = stride_item.const_u32(self, line_size.trailing_zeros());
             let stride_ty = stride_item.id(self);
             stride = self
-                .shift_left_logical(stride_ty, None, stride, shift)
+                .shift_right_logical(stride_ty, None, stride, shift)
                 .unwrap();
         }