[ET-VK][8/n] Unsqueeze

Pull Request resolved: #3172

Exploit the fact that, we reduce the unsqueeze operation to permute.

```
torch.all(torch.permute(x.unsqueeze(0), [1, 0, 2, 3]) == x.unsqueeze(1))
torch.all(torch.permute(x.unsqueeze(0), [1, 2, 0, 3]) == x.unsqueeze(2))
torch.all(torch.permute(x.unsqueeze(0), [1, 2, 3, 0]) == x.unsqueeze(3))
```

This diff introduce a minor change to the Permute implementation that it no longer requires the input dimension length to match the length of the permute array. This allows the `unsqueeze` operation to achieve a no-op `unsqueeze(0)` and then apply a permute.

Differential Revision: [D56347734](https://our.internmc.facebook.com/intern/diff/D56347734/)
ghstack-source-id: 223528485

Author: yipjustin

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

@@ -6,8 +6,11 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/Permute.h>
+
 #include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
 
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/DimUtils.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/KernelUtils.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/TensorUtils.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h>
@@ -20,53 +23,51 @@ using api::utils::uvec4;
 
 void check_args(
     const vTensor& in,
-    const IntListPtr& permute_dims,
+    const std::vector<int64_t>& permute_dims,
     const vTensor& out) {
   VK_CHECK_COND(check_memory_layout_is(in, api::kChannelsPacked));
   VK_CHECK_COND(check_memory_layout_is(out, api::kChannelsPacked));
 
-  int64_t in_dim = in.dim();
+  // This implementation doesn't not requires the input tensor to have the same
+  // dim size as the argument. The code will work as long as the input tensor's
+  // dim size is shorter than the permute dim array. In this case, the code
+  // assume size of 1 at the higher dimensions.
+
+  int64_t out_dim = out.dim();
   VK_CHECK_COND(
-      in_dim == permute_dims->size(),
-      "Input tensor dim size must match argument");
+      out_dim == permute_dims.size(),
+      "Output tensor dim size must match argument");
 }
 
 void add_permute_node(
     ComputeGraph& graph,
     ValueRef in,
-    ValueRef permute_dims_ref,
+    const std::vector<int64_t>& permute_dims,
     ValueRef out) {
   vTensorPtr t_in = graph.get_tensor(in);
   vTensorPtr t_out = graph.get_tensor(out);
 
-  IntListPtr permute_dims = graph.get_int_list(permute_dims_ref);
-
   check_args(*t_in, permute_dims, *t_out);
 
-  uvec4 in_size{1u, 1u, 1u, 1u}, out_size{1u, 1u, 1u, 1u};
   uvec4 out_dims{0u, 1u, 2u, 3u};
 
-  int64_t in_dim = t_in->dim();
-
-  std::vector<bool> seen(in_dim);
-  for (int i = 0; i < in_dim; i++) {
-    int64_t permute_dim = (*permute_dims)[i];
+  int64_t out_dim = t_out->dim();
+  std::vector<bool> seen(out_dim);
+  for (int i = 0; i < t_out->dim(); i++) {
+    int64_t permute_dim = permute_dims[i];
     VK_CHECK_COND(
         !seen[permute_dim], "Argument dim ", permute_dim, "  is repeated");
     seen[permute_dim] = true;
 
-    // Map to 4D tensor dims.
-    in_size.data[(4u - in_dim) + i] = t_in->size(i);
-    out_size.data[(4u - in_dim) + i] = t_in->size(permute_dim);
-    out_dims.data[(4u - in_dim) + i] = permute_dim + (4u - in_dim);
+    out_dims.data[(4u - out_dim) + i] = permute_dim + (4u - out_dim);
   }
 
   std::string kernel_name = "permute";
   kernel_name.reserve(kShaderNameReserve);
   add_dtype_suffix(kernel_name, *t_out);
 
-  uint32_t out_channels = out_size.data[1u];
-  uint32_t in_channels = in_size.data[1u];
+  uint32_t out_channels = dim_at<Dim4D::Channel>(t_out->sizes());
+  uint32_t in_channels = dim_at<Dim4D::Channel>(t_in->sizes());
 
   uint32_t out_c_aligned = api::utils::align_up(out_channels, 4u);
   uint32_t in_c_aligned = api::utils::align_up(in_channels, 4u);
@@ -98,6 +99,16 @@ void add_permute_node(
       {}));
 }
 
+void add_permute_node(
+    ComputeGraph& graph,
+    ValueRef in,
+    ValueRef permute_dims_ref,
+    ValueRef out) {
+  IntListPtr permute_dims = graph.get_int_list(permute_dims_ref);
+
+  add_permute_node(graph, in, *permute_dims, out);
+}
+
 void permute(ComputeGraph& graph, const std::vector<ValueRef>& args) {
   return add_permute_node(graph, args[0], args[1], args[2]);
 }

backends/vulkan/runtime/graph/ops/impl/Permute.h

@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/backends/vulkan/runtime/api/api.h>
+
+#include <executorch/backends/vulkan/runtime/graph/ComputeGraph.h>
+
+#include <vector>
+
+namespace vkcompute {
+
+void add_permute_node(
+    ComputeGraph& graph,
+    ValueRef in,
+    const std::vector<int64_t>& permute_dims,
+    ValueRef out);
+
+} // namespace vkcompute

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

@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
+
+#include <executorch/backends/vulkan/runtime/api/api.h>
+
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/Permute.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/KernelUtils.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/TensorUtils.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h>
+
+namespace vkcompute {
+
+void add_unsqueeze_node(
+    ComputeGraph& graph,
+    ValueRef in,
+    ValueRef dim_ref,
+    ValueRef out) {
+  vTensorPtr t_in = graph.get_tensor(in);
+  vTensorPtr t_out = graph.get_tensor(out);
+
+  VK_CHECK_COND(
+      t_in->dim() < 4, "Cannot unsqueeze a tensor with more than 3 dimensions");
+
+  int64_t dim = graph.extract_scalar<int64_t>(dim_ref);
+  int64_t out_dim = t_out->dim();
+
+  std::vector<int64_t> permute_dims(out_dim);
+  for (int i = 1; i <= dim; i++) {
+    permute_dims[i - 1] = i;
+  }
+  permute_dims[dim] = 0;
+
+  for (int i = dim + 1; i < out_dim; i++) {
+    permute_dims[i] = i;
+  }
+
+  add_permute_node(graph, in, permute_dims, out);
+}
+
+void unsqueeze(ComputeGraph& graph, const std::vector<ValueRef>& args) {
+  return add_unsqueeze_node(graph, args[0], args[1], args[2]);
+}
+
+REGISTER_OPERATORS {
+  VK_REGISTER_OP(aten.unsqueeze_copy.default, unsqueeze);
+}
+
+} // namespace vkcompute

backends/vulkan/test/op_tests/cases.py

@@ -214,6 +214,7 @@ def get_permute_inputs():
             ((9, 2), [1, 0]),
         ]
     )
+
     test_suite.layouts = ["api::kChannelsPacked"]
     return test_suite
 
@@ -312,6 +313,32 @@ def get_slice_inputs():
     return test_suite
 
 
+def get_unsqueeze_inputs():
+    test_suite = VkTestSuite(
+        [
+            ((2, 3, 4), 0),
+            ((1, 1, 1), 0),
+            ((1, 1, 1), 1),
+            ((1, 1, 1), 2),
+            ((1, 1, 1), 3),
+            ((9, 9, 9), 0),
+            ((9, 9, 9), 1),
+            ((9, 9, 9), 2),
+            ((9, 9, 9), 3),
+            ((9, 9), 0),
+            ((9, 9), 1),
+            ((9, 9), 2),
+            ((9,), 0),
+            ((9,), 1),
+        ]
+    )
+    test_suite.layouts = [
+        "api::kChannelsPacked",
+    ]
+    test_suite.data_gen = "make_seq_tensor"
+    return test_suite
+
+
 test_suites = {
     "aten.add.Tensor": get_binary_elementwise_inputs(),
     "aten.sub.Tensor": get_binary_elementwise_inputs(),
@@ -328,4 +355,5 @@ def get_slice_inputs():
     "aten.permute_copy.default": get_permute_inputs(),
     "aten.view_copy.default": get_view_inputs(),
     "aten.slice_copy.Tensor": get_slice_inputs(),
+    "aten.unsqueeze_copy.default": get_unsqueeze_inputs(),
 }