Fix: wgpu cat + copy ops (#477)

burn-wgpu/src/kernel/cat.rs

@@ -47,26 +47,35 @@ pub fn cat<E: WgpuElement, const D: usize>(
 #[cfg(test)]
 mod tests {
     use crate::tests::{ReferenceBackend, TestBackend};
-    use burn_tensor::{Distribution, Tensor};
+    use burn_tensor::{backend::Backend, Distribution, Tensor};
 
     #[test]
-    fn cat_should_support_multiple_invokations() {
+    fn cat_should_support_multiple_invocations_dim0() {
-        test_same_as_reference([6, 256]);
+        test_same_as_reference([6, 256], 2, 0);
+    }
+
+    #[test]
+    fn cat_should_support_multiple_invocations_dim1() {
+        test_same_as_reference([6, 256], 2, 1);
     }
 
     #[test]
     fn cat_should_support_uneven_launch() {
-        test_same_as_reference([1, 137]);
+        test_same_as_reference([1, 137], 2, 0);
     }
 
-    fn test_same_as_reference(shape: [usize; 2]) {
+    fn test_same_as_reference(shape: [usize; 2], num_tensors: usize, dim: usize) {
-        let tensor1 = Tensor::<TestBackend, 2>::random(shape, Distribution::Default);
+        TestBackend::seed(0);
-        let tensor2 = Tensor::<TestBackend, 2>::random(shape, Distribution::Default);
+        let tensors = (0..num_tensors)
-        let tensor1_ref = Tensor::<ReferenceBackend, 2>::from_data(tensor1.to_data());
+            .map(|_| Tensor::<TestBackend, 2>::random(shape, Distribution::Default))
-        let tensor2_ref = Tensor::<ReferenceBackend, 2>::from_data(tensor2.to_data());
+            .collect::<Vec<_>>();
+        let tensors_ref = tensors
+            .iter()
+            .map(|tensor| Tensor::<ReferenceBackend, 2>::from_data(tensor.to_data()))
+            .collect::<Vec<_>>();
 
-        let tensor = Tensor::<TestBackend, 2>::cat(vec![tensor1, tensor2], 0);
+        let tensor = Tensor::<TestBackend, 2>::cat(tensors, dim);
-        let tensor_ref = Tensor::<ReferenceBackend, 2>::cat(vec![tensor1_ref, tensor2_ref], 0);
+        let tensor_ref = Tensor::<ReferenceBackend, 2>::cat(tensors_ref, dim);
 
         tensor
             .into_data()

burn-wgpu/src/ops/base.rs

@@ -58,7 +58,6 @@ pub fn swap_dims<E: WgpuElement, const D: usize>(
     dim2: usize,
 ) -> WgpuTensor<E, D> {
     tensor.strides.swap(dim1, dim2);
-
     tensor.shape.dims.swap(dim1, dim2);
 
     tensor

burn-wgpu/src/template/cat.wgsl

@@ -23,8 +23,9 @@ fn main(
     let dim_cat = info[4u * dim + 1u];
     let dim_cat_index = info[4u * dim + 2u];
 
-    var index_input: u32 = 0u;
+    var num_elems = 1u;
-    var index_output: u32 = 0u;
+    var index_input = 0u;
+    var index_output = 0u;
 
     for (var i: u32 = 1u; i <= dim; i++) {
         let stride_input = info[i];
@@ -32,8 +33,9 @@ fn main(
         let shape_input = info[i + 2u * dim];
         let shape_output = info[i + 3u * dim];
 
-        let num_block_output = id / stride_output % shape_output;
+        let num_block_output = id / stride_input % shape_input;
         index_input += num_block_output * stride_input;
+        num_elems *= shape_input;
 
         if i - 1u == dim_cat {
             index_output += (num_block_output + dim_cat_index) * stride_output;
@@ -42,6 +44,8 @@ fn main(
         }
     }
 
+    if id < num_elems {
         output[index_output] = input[index_input];
+    }
 }
 

burn-wgpu/src/tensor/base.rs

@@ -72,7 +72,7 @@ impl<E: WgpuElement, const D: usize> WgpuTensor<E, D> {
         // slowdowns.
         //
         // The solution is just to use a simple unary compute shader.
-        unary!(CopyBuffer, body "output[global_id.x] = input[global_id.x];");
+        unary!(CopyBuffer, body "output[id] = input[id];");
         unary_default::<CopyBuffer, E, D>(self.clone())
     }
 

run-checks.sh

@@ -79,6 +79,16 @@ std_func() {
     # all features
     echo "Running all-features checks"
     build_and_test_all_features "burn-dataset"
+
+    cd burn-core || exit
+
+    echo "Test burn-core with tch backend"
+    cargo test --features test-tch
+
+    echo "Test burn-core with wgpu backend"
+    cargo test --features test-wgpu
+
+    cd .. || exit
 }
 
 # Run the checks for no_std