Make example compile again

Author: stijnh

examples/vector_add/main.cu

@@ -21,7 +21,7 @@ __global__ void my_kernel(
     int i = blockIdx.x * blockDim.x + threadIdx.x;
 
     if (i * N < length) {
-        output[i] = kf::fma(input[i], input[i], kf::cast<__half>(constant));
+        output(i) = kf::fma(input[i], input[i], kf::cast<__half>(constant));
     }
 }
 
@@ -82,7 +82,7 @@ int main() {
 
     run_kernel<1>(n);
     run_kernel<2>(n);
-    run_kernel<3>(n);
+    //    run_kernel<3>(n);
     run_kernel<4>(n);
     run_kernel<8>(n);
 

include/kernel_float/memory.h

@@ -383,7 +383,7 @@ struct vector_ref<T, N, const U, Align> {
 
 #define KERNEL_FLOAT_VECTOR_REF_ASSIGN_OP(OP, OP_ASSIGN)                 \
     template<typename T, size_t N, typename U, size_t Align, typename V> \
-    KERNEL_FLOAT_INLINE vector_ref<T, N> operator OP_ASSIGN(             \
+    KERNEL_FLOAT_INLINE vector_ref<T, N, U, Align> operator OP_ASSIGN(   \
         vector_ref<T, N, U, Align> ptr,                                  \
         const V& value) {                                                \
         ptr.write(ptr.read() OP value);                                  \

single_include/kernel_float.h

@@ -16,8 +16,8 @@
 
 //================================================================================
 // this file has been auto-generated, do not modify its contents!
-// date: 2024-02-22 10:29:30.717359
-// git hash: c4fae9db85e61372cfefc53d4755b418beb2d93e
+// date: 2024-03-18 16:06:55.100306
+// git hash: 06e08f55399e148d96070afd0ac36dd414045f04
 //================================================================================
 
 #ifndef KERNEL_FLOAT_MACROS_H
@@ -2229,7 +2229,7 @@ struct concat_impl<V, Vs...> {
 template<>
 struct concat_impl<> {
     using value_type = void;
-    static constexpr size_t size = 1;
+    static constexpr size_t size = 0;
 
     template<typename U>
     KERNEL_FLOAT_INLINE static void call(U* output) {}