aoti_torch_copy_

backends/aoti/utils.h

@@ -36,6 +36,8 @@ inline executorch::aten::ScalarType dtype_to_scalar_type(int32_t dtype) {
   switch (dtype) {
     case 6: // PyTorch's float32 dtype code
       return executorch::aten::ScalarType::Float;
+    case 15: // PyTorch's bfloat16 dtype code
+      return executorch::aten::ScalarType::BFloat16;
     // Future support for additional dtypes can be added here
     default:
       ET_LOG(Error, "Unsupported dtype: %d for ScalarType conversion", dtype);
@@ -71,6 +73,23 @@ inline AOTITorchError validate_storage_offset(int64_t storage_offset) {
   return Error::Ok;
 }
 
+// Check if tensor is in contiguous memory format (NCHW for 4D tensors)
+// Contiguous format means strides decrease from left to right:
+// For NCHW: strides = [C*H*W, H*W, W, 1]
+inline bool is_tensor_contiguous(
+    int64_t ndim,
+    const int64_t* sizes,
+    const int64_t* strides) {
+  int64_t expected_stride = 1;
+  for (int64_t i = ndim - 1; i >= 0; i--) {
+    if (strides[i] != expected_stride) {
+      return false;
+    }
+    expected_stride *= sizes[i];
+  }
+  return true;
+}
+
 } // extern "C"
 
 } // namespace aoti

backends/cuda/runtime/TARGETS

@@ -0,0 +1,32 @@
+load("@fbsource//xplat/executorch/build:runtime_wrapper.bzl", "runtime")
+
+oncall("executorch")
+
+runtime.cxx_library(
+    name = "runtime_shims",
+    srcs = [
+        "shims/memory.cpp",
+        "shims/tensor_attribute.cpp",
+    ],
+    headers = [
+        "shims/memory.h",
+        "shims/tensor_attribute.h",
+        "shims/utils.h",
+    ],
+    # @lint-ignore BUCKLINT: Avoid `link_whole=True` (https://fburl.com/avoid-link-whole)
+    link_whole = True,
+    supports_python_dlopen = True,
+    # Constructor needed for backend registration.
+    compiler_flags = ["-Wno-global-constructors"],
+    visibility = ["@EXECUTORCH_CLIENTS"],
+    deps = [
+        "//executorch/backends/aoti:common_shims",
+        "//executorch/extension/tensor:tensor",
+        "//executorch/runtime/core:core",
+        "//executorch/runtime/core/exec_aten:lib",
+        "//executorch/runtime/platform:platform",
+    ],
+    external_deps = [
+        ("cuda", None, "cuda-lazy"),
+    ],
+)