[ET-VK] Fix output offset calculation and add symint support to ComputeGraph

Fix output argument indexing in VulkanBackend::execute() and extend
ComputeGraph to transparently handle symint values.

The output loop previously computed the args index as `i + num_inputs`,
which breaks when non-tensor arguments (e.g. symints) sit between the
tensor inputs and outputs in the args array. Fix by computing the offset
from the end: `args.size() - num_outputs`.

ComputeGraph changes add symint support so that operators can read symint
values uniformly:
- `extract_scalar<T>()` now handles SymInt values, allowing operators to
  call extract_scalar on arguments that may be either plain ints or
  symints without special-casing.
- `read_symint()` falls back to reading plain Int values, so values
  stored as Int (rather than SymInt objects) can be read uniformly.

Pull Request resolved: pytorch#18050
ghstack-source-id: 353546683
@exported-using-ghexport

Differential Revision: [D95970167](https://our.internmc.facebook.com/intern/diff/D95970167/)

Author: ssjia

backends/vulkan/runtime/VulkanBackend.cpp

@@ -671,6 +671,7 @@ class VulkanBackend final : public ::executorch::runtime::BackendInterface {
     ComputeGraph* compute_graph = static_cast<ComputeGraph*>(handle);
 
     const size_t num_inputs = compute_graph->inputs().size();
+    const size_t num_outputs = compute_graph->outputs().size();
     bool should_propagate_resize = false;
 #ifdef ET_EVENT_TRACER_ENABLED
     runtime::EventTracer* event_tracer = context.event_tracer();
@@ -770,14 +771,13 @@ class VulkanBackend final : public ::executorch::runtime::BackendInterface {
             "ETVK_COPY_OUTPUTS",
             /* delegate_debug_id = */ -1);
 #endif // ET_EVENT_TRACER_ENABLED
-    for (size_t i = 0; i < compute_graph->outputs().size(); i++) {
-      const size_t o = i + num_inputs;
+    const size_t output_offset = args.size() - num_outputs;
+    for (size_t i = 0; i < num_outputs; i++) {
+      const size_t o = output_offset + i;
       const ValueRef oref = compute_graph->outputs()[i].value;
       if (compute_graph->val_is_tensor(oref)) {
         VK_CHECK_COND(args[o]->isTensor());
         maybe_resize_output(compute_graph, i, args[o]->toTensor());
-        // args holds inputs directly followed by outputs, so the i'th output
-        // for compute_graph corresponds to the o'th arg
         compute_graph->maybe_cast_and_copy_from_staging(
             compute_graph->outputs()[i].staging,
             args[o]->toTensor().mutable_data_ptr(),

backends/vulkan/runtime/graph/ComputeGraph.cpp

@@ -725,6 +725,9 @@ void ComputeGraph::set_symint(const ValueRef idx, const int32_t val) {
 }
 
 int32_t ComputeGraph::read_symint(const ValueRef idx) {
+  if (values_.at(idx).isInt()) {
+    return static_cast<int32_t>(values_.at(idx).toInt());
+  }
   return get_symint(idx)->get();
 }
 

backends/vulkan/runtime/graph/ComputeGraph.h

@@ -573,6 +573,9 @@ class ComputeGraph final {
     if (value.isBool()) {
       return static_cast<T>(value.toBool());
     }
+    if (value.isSymInt()) {
+      return utils::safe_downcast<T>(read_symint(idx));
+    }
     VK_THROW("Cannot extract scalar from Value with type ", value.type());
   }