fix: Fix packed fp4 casting from int (#890)

* Fix packed fp4 casting from int

* Fix matmul and conv2d test launch

Author: wingertge

crates/cubecl-convolution/src/tests/convolution_test_launcher.rs

@@ -61,7 +61,7 @@ pub fn test_convolution_algorithm<A, Args, P, R>(
     .pick_max()
     .unwrap();
 
-    let config = match A::setup::<R, (P::EG, P::EG, P::ES, P::ES, f32, P::EG)>(
+    let config = match A::setup::<R, (P::EG, P::EG, P::EG, P::ES, P::ES, f32)>(
         &client,
         &problem,
         &selection,

crates/cubecl-cpp/src/cuda/convert.rs

@@ -57,6 +57,21 @@ pub(crate) fn special_cast<D: Dialect>(
         current_in = out_var;
     }
 
+    // Broadcast scalars to packing factor
+    if out.item().packing_factor() > 1 && input.item().vectorization == 1 {
+        let tmp = Variable::tmp(Item {
+            elem: input.item().elem,
+            vectorization: out.item().packing_factor(),
+            native: input.item().native,
+        });
+        let assign = Instruction::Assign(UnaryInstruction {
+            input: current_in,
+            out: tmp,
+        });
+        writeln!(f, "{assign}")?;
+        current_in = tmp;
+    }
+
     if matches!(
         current_in.elem(),
         Elem::U8
@@ -72,8 +87,8 @@ pub(crate) fn special_cast<D: Dialect>(
         // Precision is irrelevant for int, so use bf16 for the range
         let tmp = Variable::tmp(Item {
             elem: Elem::BF16,
-            vectorization: input.item().vectorization,
-            native: input.item().native,
+            vectorization: current_in.item().vectorization,
+            native: current_in.item().native,
         });
         let assign = Instruction::Assign(UnaryInstruction {
             input: current_in,
@@ -137,7 +152,7 @@ fn cast_to_fp4_fp6<D: Dialect>(
         _ => unreachable!("Must be fp4 or fp6"),
     };
 
-    let in_ty = match input.elem() {
+    let in_ty = match input.elem().unpacked() {
         Elem::F64 => format!("double{pack_suffix}"),
         Elem::TF32 | Elem::F32 => format!("float{pack_suffix}"),
         Elem::F16 => format!("halfraw{pack_suffix}"),

crates/cubecl-cpp/src/shared/base.rs

@@ -1136,18 +1136,26 @@ impl<D: Dialect> CppCompiler<D> {
             gpu::Operator::Cast(op)
                 if is_fp4_fp6_fp8(op.input.elem_type()) || is_fp4_fp6_fp8(out.elem_type()) =>
             {
-                let inst = self.compile_unary(op, out);
-
                 // We may need these for intermediates
                 self.flags.elem_f16 = true;
                 self.flags.elem_bf16 = true;
-                let vec = inst.input.item().vectorization as u32;
+                let vec_in = op.input.ty.line_size();
+                let packing = out.storage_type().packing_factor();
+                self.compile_type(op.input.ty.line(packing));
+                self.compile_type(
+                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::F16)).line(vec_in),
+                );
                 self.compile_type(
-                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::F16)).line(vec),
+                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::BF16)).line(vec_in),
                 );
                 self.compile_type(
-                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::BF16)).line(vec),
+                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::F16)).line(packing),
                 );
+                self.compile_type(
+                    gpu::Type::scalar(gpu::ElemType::Float(FloatKind::BF16)).line(packing),
+                );
+
+                let inst = self.compile_unary(op, out);
 
                 instructions.push(Instruction::SpecialCast(inst));
             }

crates/cubecl-matmul/src/tests/layered/matmul_test_launcher.rs

@@ -60,12 +60,7 @@ pub fn test_matmul_algorithm<A, P, R>(
         .pick_max()
         .unwrap();
 
-    let config = match A::setup::<(P::EG, P::EG, P::ES, P::ES, P::EA, P::EG), R>(
-        &client,
-        &problem,
-        &selection,
-        &line_sizes,
-    ) {
+    let config = match A::setup::<P::MP, R>(&client, &problem, &selection, &line_sizes) {
         Ok(config) => config,
         Err(err) => {
             let msg = format!("Can't launch the test: {err}");