[ET-VK] High dim tensor support for view, unsqueeze, squeeze, clone (#13597)

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at
bottom):
* __->__ #13597
* #13596
* #13595
* #13594
* #13593
* #13600
* #13599
* #13598

Differential Revision:
[D80800084](https://our.internmc.facebook.com/intern/diff/D80800084)

Co-authored-by: ssjia <[email protected]>

Author: SS-JIA

backends/vulkan/op_registry.py

@@ -489,10 +489,8 @@ def register_rotary_emb_op():
 
 @update_features(
     [
-        exir_ops.edge.aten.clone.default,
         exir_ops.edge.aten.permute.default,
         exir_ops.edge.aten.permute_copy.default,
-        exir_ops.edge.aten.view_copy.default,
     ]
 )
 def register_view_ops():
@@ -502,6 +500,21 @@ def register_view_ops():
     )
 
 
+@update_features(
+    [
+        exir_ops.edge.aten.view_copy.default,
+        exir_ops.edge.aten.squeeze_copy.dims,
+        exir_ops.edge.aten.unsqueeze_copy.default,
+        exir_ops.edge.aten.clone.default,
+    ]
+)
+def register_view_ops_with_buffer_meta():
+    return OpFeatures(
+        inputs_storage=utils.ANY_STORAGE,
+        supports_resize=True,
+    )
+
+
 # Fully featured transfer operators (i.e. operators that copy data from the input
 # tensor(s) to the output tensor(s)), which have memory layout agnostic implementations
 # for both texture and buffer storage types.
@@ -562,9 +575,6 @@ def register_ported_op():
 # Ops ported from PyTorch Vulkan backend. These ops are in a separate registry because they support all packed dimensions
 @update_features(
     [
-        # Shape Manipulation
-        exir_ops.edge.aten.squeeze_copy.dims,
-        exir_ops.edge.aten.unsqueeze_copy.default,
         # Tensor combination
         exir_ops.edge.aten.repeat.default,
         exir_ops.edge.aten.split_with_sizes_copy.default,

backends/vulkan/runtime/graph/ops/glsl/view_buffer.glsl

@@ -0,0 +1,44 @@
+#version 450 core
+
+#define PRECISION ${PRECISION}
+
+#define T ${buffer_scalar_type(DTYPE)}
+
+${define_required_extensions(DTYPE)}
+
+layout(std430) buffer;
+
+#include "indexing.glslh"
+
+${layout_declare_tensor(B, "w", "t_outp", DTYPE, STORAGE)}
+${layout_declare_tensor(B, "r", "t_inp", DTYPE, STORAGE)}
+
+${layout_declare_ubo(B, "BufferMetadata", "outp")}
+${layout_declare_ubo(B, "BufferMetadata", "inp")}
+
+layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
+
+/*
+ * The insight behind the view operation is that the contiguous index of each
+ * tensor element in the input and output tensors are the same.
+ */
+void main() {
+  const uint outp_bufi = gl_GlobalInvocationID.x;
+  if (outp_bufi >= numel(outp)) {
+    return;
+  }
+
+  TensorIndex outp_tidx;
+  linear_idx_to_tensor_idx(outp, outp_bufi, outp_tidx);
+
+  // To map the output to the input, find the input element that has the same
+  // contiguous index as the output element.
+  const uint contig_idx = tensor_idx_to_contiguous_idx(outp, outp_tidx);
+
+  TensorIndex inp_tidx;
+  contiguous_idx_to_tensor_idx(inp, contig_idx, inp_tidx);
+
+  const uint inp_bufi = tensor_idx_to_linear_idx(inp, inp_tidx);
+
+  t_outp[outp_bufi] = t_inp[inp_bufi];
+}

backends/vulkan/runtime/graph/ops/glsl/view_buffer.yaml

@@ -0,0 +1,20 @@
+# 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.
+
+view_buffer:
+  parameter_names_with_default_values:
+    DTYPE: float
+    STORAGE: buffer
+  generate_variant_forall:
+    DTYPE:
+      - VALUE: half
+      - VALUE: float
+      - VALUE: double
+      - VALUE: int8
+      - VALUE: uint8
+      - VALUE: int32
+  shader_variants:
+    - NAME: view_buffer

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

@@ -143,7 +143,11 @@ void clone(ComputeGraph& graph, const std::vector<ValueRef>& args) {
   if (src_storage == utils::kBuffer && dst_storage == utils::kTexture3D) {
     return add_buffer_to_image_node(graph, src, dst);
   }
-  VK_THROW("Buffer to buffer memory layout transition not supported yet!");
+
+  std::vector<ValueRef> extra_args = {};
+  // Buffer to buffer copy
+  return add_view_copy_buffer_node(
+      graph, src, dst, extra_args, resize_clone_node);
 }
 
 // Clone node is not the most efficient implementation for the aten.clone

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

@@ -10,6 +10,7 @@
 
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Clone.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Permute.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/View.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/KernelUtils.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h>
 
@@ -55,8 +56,49 @@ void add_squeeze_copy_dims_node(
   }
 }
 
+void resize_squeeze_node(
+    ComputeGraph* graph,
+    const std::vector<ArgGroup>& args,
+    const std::vector<ValueRef>& extra_args) {
+  const ValueRef out = args.at(0).refs.at(0);
+  const ValueRef in = args.at(1).refs.at(0);
+  const ValueRef dims_ref = extra_args.at(0);
+
+  const IntListPtr dims = graph->get_int_list(dims_ref);
+
+  std::vector<int64_t> out_sizes = graph->sizes_of(in);
+
+  // Remove the dimensions specified in dims if their size is 1
+  for (int64_t dim : *dims) {
+    if (dim >= 0 && dim < static_cast<int64_t>(out_sizes.size()) &&
+        out_sizes[dim] == 1) {
+      out_sizes.erase(out_sizes.begin() + dim);
+      // After erasing, all subsequent dims shift left by one
+      // So we need to decrement all subsequent dims in dims
+      for (auto& d : *dims) {
+        if (d > dim) {
+          --d;
+        }
+      }
+    }
+  }
+
+  graph->virtual_resize(out, out_sizes);
+}
+
 void squeeze_copy_dims(ComputeGraph& graph, const std::vector<ValueRef>& args) {
-  return add_squeeze_copy_dims_node(graph, args[0], args[1], args[2]);
+  int idx = 0;
+  const ValueRef in = args.at(idx++);
+  const ValueRef dims = args.at(idx++);
+  const ValueRef out = args.at(idx++);
+
+  std::vector<ValueRef> resize_args = {dims};
+
+  if (graph.is_buffer_storage(in)) {
+    return add_view_copy_buffer_node(
+        graph, in, out, resize_args, resize_squeeze_node);
+  }
+  return add_squeeze_copy_dims_node(graph, in, dims, out);
 }
 
 REGISTER_OPERATORS {

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

@@ -9,6 +9,7 @@
 #include <executorch/backends/vulkan/runtime/graph/ops/OperatorRegistry.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/Permute.h>
+#include <executorch/backends/vulkan/runtime/graph/ops/impl/View.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/KernelUtils.h>
 #include <executorch/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h>
 
@@ -45,8 +46,42 @@ void add_unsqueeze_node(
   add_permute_node(graph, in, permute_dims_ref, out);
 }
 
+void resize_unsqueeze_node(
+    ComputeGraph* graph,
+    const std::vector<ArgGroup>& args,
+    const std::vector<ValueRef>& extra_args) {
+  const ValueRef out = args.at(0).refs.at(0);
+  const ValueRef in = args.at(1).refs.at(0);
+  const ValueRef dims_ref = extra_args.at(0);
+
+  const IntListPtr dims = graph->get_int_list(dims_ref);
+
+  std::vector<int64_t> out_sizes = graph->sizes_of(in);
+
+  // Insert singleton dimensions at the specified positions
+  for (auto dim : *dims) {
+    int64_t d = dim;
+    if (d < 0) {
+      d += static_cast<int64_t>(out_sizes.size()) + 1;
+    }
+    out_sizes.insert(out_sizes.begin() + d, 1);
+  }
+
+  graph->virtual_resize(out, out_sizes);
+}
+
 void unsqueeze(ComputeGraph& graph, const std::vector<ValueRef>& args) {
-  return add_unsqueeze_node(graph, args[0], args[1], args[2]);
+  int idx = 0;
+  const ValueRef in = args.at(idx++);
+  const ValueRef dims = args.at(idx++);
+  const ValueRef out = args.at(idx++);
+
+  std::vector<ValueRef> resize_args = {dims};
+  if (graph.is_buffer_storage(in)) {
+    return add_view_copy_buffer_node(
+        graph, in, out, resize_args, resize_unsqueeze_node);
+  }
+  return add_unsqueeze_node(graph, in, dims, out);
 }
 
 REGISTER_OPERATORS {

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

@@ -89,8 +89,47 @@ void add_view_node(
       resize_view_node));
 }
 
+void add_view_copy_buffer_node(
+    ComputeGraph& graph,
+    ValueRef in,
+    ValueRef out,
+    const std::vector<ValueRef>& resize_args,
+    const ExecuteNode::ResizeFunction& resize_fn) {
+  std::string kernel_name = "view_buffer";
+  add_dtype_suffix(kernel_name, graph.dtype_of(out));
+
+  graph.execute_nodes().emplace_back(new DynamicDispatchNode(
+      graph,
+      VK_KERNEL_FROM_STR(kernel_name),
+      default_pick_global_wg_size,
+      default_pick_local_wg_size,
+      // Inputs and Outputs
+      {{out, vkapi::kWrite}, {in, vkapi::kRead}},
+      // Parameter Buffers
+      {graph.buffer_meta_ubo(out), graph.buffer_meta_ubo(in)},
+      // Push Constants
+      {},
+      // Specialization Constants
+      {},
+      // Resize Args
+      resize_args,
+      // Resizing Logic
+      resize_fn));
+}
+
 void view(ComputeGraph& graph, const std::vector<ValueRef>& args) {
-  return add_view_node(graph, args[0], args[1], args[2]);
+  int idx = 0;
+  const ValueRef in = args.at(idx++);
+  const ValueRef sizes = args.at(idx++);
+  const ValueRef out = args.at(idx++);
+
+  std::vector<ValueRef> resize_args = {sizes};
+
+  if (graph.is_buffer_storage(out)) {
+    return add_view_copy_buffer_node(
+        graph, in, out, resize_args, resize_view_node);
+  }
+  return add_view_node(graph, in, sizes, out);
 }
 
 REGISTER_OPERATORS {

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

@@ -12,6 +12,18 @@
 
 namespace vkcompute {
 
+/*
+ * Dispatches the view_copy compute shader. This can be used to implement ops
+ * that preserve the "contiguous" indexes of elements between the input and
+ * output such as view_copy, squeeze_copy, unsqueeze_copy, etc.
+ */
+void add_view_copy_buffer_node(
+    ComputeGraph& graph,
+    ValueRef in,
+    ValueRef out,
+    const std::vector<ValueRef>& resize_args,
+    const ExecuteNode::ResizeFunction& resize_fn);
+
 void add_view_node(
     ComputeGraph& graph,
     ValueRef in,