[ET-VK] Avoid the near-square local work group for the tiled linear on Adreno

Adreno drivers miscompute linear_tiled for 8x8x1 and 4x16x1 local work
groups: entire work groups intermittently write garbage while every other
block is bit-exact. The shader has no shared memory and no barriers and its
invocations are independent, so the result cannot legitimately depend on the
group shape.

Use a 2:1 x:y shape (16x4x1 at the default 64 threads) on Adreno and keep the
square shape elsewhere.

Fixes #22327

Author: msluszniak

backends/vulkan/runtime/graph/ops/impl/Linear.cpp

@@ -65,6 +65,34 @@ vkapi::ShaderInfo pick_linear_shader(
   return VK_KERNEL_FROM_STR(kernel_name);
 }
 
+// Adreno drivers miscompute this shader for some local work group shapes. With
+// the default 8x8x1 group (and with 4x16x1), entire work groups intermittently
+// write garbage: at M = 1500, K = N = 384 roughly 10-20% of executions produce
+// a handful of wholly corrupted 32x32 output blocks while every other block is
+// bit-exact. The shader has no shared memory and no barriers and its
+// invocations are independent, so its result cannot legitimately depend on the
+// group shape. 16x4x1, 32x2x1, 64x1x1, 2x32x1 and 1x64x1 were all correct over
+// 60 executions each; only the two near-square shapes fail. See #22327.
+//
+// A 2:1 x:y shape yields 16x4x1 at the default 64 threads and costs ~1.5% on a
+// Whisper encoder compared to 8x8x1, far less than the wider shapes (64x1x1 is
+// +34%). Non-Adreno devices keep the square shape.
+LocalWorkGroup pick_linear_lwg(
+    ComputeGraph* graph,
+    const vkapi::ShaderInfo& shader,
+    const GlobalWorkGrid& gwg,
+    const std::vector<ArgGroup>& args,
+    const std::vector<ValueRef>& resize_args) {
+  if (!graph->device_is_adreno()) {
+    return pick_xy_square_lwg(graph, shader, gwg, args, resize_args);
+  }
+  LocalWorkGroup lwg(
+      LwgShape(2u, 1u, 0u),
+      graph->context()->adapter_ptr()->recommended_lwg_nthreads());
+  lwg.fit_to_global(gwg);
+  return lwg;
+}
+
 GlobalWorkGrid pick_linear_gwg(
     ComputeGraph* graph,
     const vkapi::ShaderInfo& shader,
@@ -124,7 +152,7 @@ void add_linear_tiled_node(
       graph,
       pick_linear_shader,
       pick_linear_gwg,
-      pick_xy_square_lwg,
+      pick_linear_lwg,
       // Inputs and Outputs
       {{out, vkapi::kWrite}, {read_inputs, vkapi::kRead}},
       // Shader params buffers