Update base for Update on "[ET-VK] Add support for binary symint ops"

## Changes

* Add an implementation for binary operators which add symbolic integers.

## Motivation

Support executing llama models with dynamic shapes. This operator shows up when exporting with dynamic shapes.

Differential Revision: [D75238029](https://our.internmc.facebook.com/intern/diff/D75238029/)

[ghstack-poisoned]

Author: SS-JIA

.ci/scripts/test_model.sh

@@ -49,14 +49,24 @@ prepare_artifacts_upload() {
 }
 
 build_cmake_executor_runner() {
+  local backend_string_select="${1:-}"
   echo "Building executor_runner"
   rm -rf ${CMAKE_OUTPUT_DIR}
-  cmake -DCMAKE_BUILD_TYPE=Debug \
-      -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
-      -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
-      -B${CMAKE_OUTPUT_DIR} .
-
-  cmake --build ${CMAKE_OUTPUT_DIR} -j4 --config Debug
+  mkdir ${CMAKE_OUTPUT_DIR}
+  if [[ "$backend_string_select" == "XNNPACK" ]]; then
+    echo "Backend $backend_string_select selected"
+    (cd ${CMAKE_OUTPUT_DIR} \
+      && cmake -DCMAKE_BUILD_TYPE=Release \
+        -DEXECUTORCH_BUILD_XNNPACK=ON \
+        -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" ..)
+    cmake --build ${CMAKE_OUTPUT_DIR} -j4
+  else
+    cmake -DCMAKE_BUILD_TYPE=Debug \
+        -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON \
+        -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" \
+        -B${CMAKE_OUTPUT_DIR} .
+    cmake --build ${CMAKE_OUTPUT_DIR} -j4 --config Debug
+  fi
 }
 
 run_portable_executor_runner() {
@@ -111,19 +121,6 @@ test_model() {
   run_portable_executor_runner
 }
 
-build_cmake_xnn_executor_runner() {
-  echo "Building xnn_executor_runner"
-
-  (rm -rf ${CMAKE_OUTPUT_DIR} \
-    && mkdir ${CMAKE_OUTPUT_DIR} \
-    && cd ${CMAKE_OUTPUT_DIR} \
-    && retry cmake -DCMAKE_BUILD_TYPE=Release \
-      -DEXECUTORCH_BUILD_XNNPACK=ON \
-      -DPYTHON_EXECUTABLE="$PYTHON_EXECUTABLE" ..)
-
-  cmake --build ${CMAKE_OUTPUT_DIR} -j4
-}
-
 test_model_with_xnnpack() {
   WITH_QUANTIZATION=$1
   WITH_DELEGATION=$2
@@ -148,12 +145,11 @@ test_model_with_xnnpack() {
 
   # Run test model
   if [[ "${BUILD_TOOL}" == "buck2" ]]; then
+    # TODO eventually buck should also use consolidated executor runners
     buck2 run //examples/xnnpack:xnn_executor_runner -- --model_path "${OUTPUT_MODEL_PATH}"
   elif [[ "${BUILD_TOOL}" == "cmake" ]]; then
-    if [[ ! -f ${CMAKE_OUTPUT_DIR}/backends/xnnpack/xnn_executor_runner ]]; then
-      build_cmake_xnn_executor_runner
-    fi
-    ./${CMAKE_OUTPUT_DIR}/backends/xnnpack/xnn_executor_runner --model_path "${OUTPUT_MODEL_PATH}"
+    build_cmake_executor_runner "XNNPACK"
+    ./${CMAKE_OUTPUT_DIR}/executor_runner --model_path "${OUTPUT_MODEL_PATH}"
   else
     echo "Invalid build tool ${BUILD_TOOL}. Only buck2 and cmake are supported atm"
     exit 1

backends/vulkan/runtime/graph/ops/glsl/nchw_to_bitw8_image_nobitw8buffer.glsl

@@ -22,7 +22,13 @@ layout(std430) buffer;
 
 ${layout_declare_tensor(B, "w", "t_out", DTYPE, STORAGE)}
 ${layout_declare_buffer(B, "r", "nchw_in", "int")}
-${layout_declare_ubo(B, "ivec4", "sizes")}
+
+$if USE_PUSH_CONST:
+  layout(push_constant) uniform restrict Block {
+    ivec4 sizes;
+  };
+$else:
+  ${layout_declare_ubo(B, "ivec4", "sizes")}
 
 layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
 

backends/vulkan/runtime/graph/ops/glsl/nchw_to_bitw8_image_nobitw8buffer.yaml

@@ -8,6 +8,7 @@ nchw_to_bitw8_image_nobitw8buffer:
   parameter_names_with_default_values:
     STORAGE: texture3d
     DTYPE: int8
+    USE_PUSH_CONST: True
   generate_variant_forall:
     STORAGE:
       - VALUE: texture2d
@@ -17,3 +18,5 @@ nchw_to_bitw8_image_nobitw8buffer:
       - VALUE: uint8
   shader_variants:
     - NAME: nchw_to_bitw8_image_nobitw8buffer
+    - NAME: nchw_to_bitw8_image_nobitw8buffer_no_pc
+      USE_PUSH_CONST: False

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

@@ -12,9 +12,17 @@ layout(std430) buffer;
 
 ${layout_declare_tensor(0, "w", "t_out", DTYPE, STORAGE)}
 ${layout_declare_tensor(1, "r", "nchw_in", DTYPE, STORAGE)}
-${layout_declare_ubo(2, "ivec4", "out_sizes")}
-${layout_declare_ubo(3, "ivec4", "out_strides")}
-${layout_declare_ubo(4, "int", "numel")}
+
+$if USE_PUSH_CONST:
+  layout(push_constant) uniform restrict Block {
+    ivec4 out_sizes;
+    ivec4 out_strides;
+    int numel;
+  };
+$else:
+  ${layout_declare_ubo(2, "ivec4", "out_sizes")}
+  ${layout_declare_ubo(3, "ivec4", "out_strides")}
+  ${layout_declare_ubo(4, "int", "numel")}
 
 layout(local_size_x_id = 0, local_size_y_id = 1, local_size_z_id = 2) in;
 

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

@@ -8,6 +8,7 @@ nchw_to_buffer:
   parameter_names_with_default_values:
     DTYPE: float
     STORAGE: buffer
+    USE_PUSH_CONST: True
   generate_variant_forall:
     DTYPE:
       - VALUE: half
@@ -17,3 +18,5 @@ nchw_to_buffer:
       - VALUE: uint8
   shader_variants:
     - NAME: nchw_to_buffer
+    - NAME: nchw_to_buffer_no_pc
+      USE_PUSH_CONST: False

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

@@ -21,9 +21,17 @@ layout(std430) buffer;
 
 ${layout_declare_tensor(B, "w", "t_out", DTYPE, STORAGE)}
 ${layout_declare_buffer(B, "r", "buf_in", DTYPE)}
-${layout_declare_ubo(B, "ivec4", "sizes")}
-$if not FROM_STAGING:
-  ${layout_declare_ubo(B, "ivec4", "buf_strides")}
+
+$if USE_PUSH_CONST:
+  layout(push_constant) uniform restrict Block {
+    ivec4 sizes;
+  $if not FROM_STAGING:
+    ivec4 buf_strides;
+  };
+$else:
+  ${layout_declare_ubo(B, "ivec4", "sizes")}
+  $if not FROM_STAGING:
+    ${layout_declare_ubo(B, "ivec4", "buf_strides")}
 
 #include "indexing_utils.h"
 

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

@@ -9,6 +9,7 @@ nchw_to_image:
     STORAGE: texture3d
     DTYPE: float
     FROM_STAGING: True
+    USE_PUSH_CONST: True
   generate_variant_forall:
     DTYPE:
       - VALUE: half
@@ -22,3 +23,11 @@ nchw_to_image:
       STORAGE: texture2d
     - NAME: clone_buffer_to_image
       FROM_STAGING: False
+    - NAME: nchw_to_image_no_pc_texture3d
+      USE_PUSH_CONST: False
+    - NAME: nchw_to_image_no_pc_texture2d
+      STORAGE: texture2d
+      USE_PUSH_CONST: False
+    - NAME: clone_buffer_to_image_no_pc
+      FROM_STAGING: False
+      USE_PUSH_CONST: False

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

@@ -105,9 +105,9 @@ void add_buffer_to_image_node(
       // Input and Outputs
       {{image, vkapi::kWrite}, {buffer, vkapi::kRead}},
       // Parameter Buffers
-      {graph.sizes_ubo(image), graph.strides_ubo(buffer)},
-      // Push Constants
       {},
+      // Push Constants
+      {graph.sizes_pc_of(image), graph.strides_pc_of(buffer)},
       // Specialization Constants
       {graph.hashed_layout_of(image)},
       // Resize Args

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

@@ -106,9 +106,10 @@ ValueRef prepack_biases(
       graph.create_local_wg_size(v),
       vref,
       v,
-      {t->sizes_ubo()},
+      {},
       // Specialization constants
-      {t->hashed_layout()}));
+      {t->hashed_layout()},
+      {graph.sizes_pc_of(v)}));
 
   return v;
 }

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

@@ -28,14 +28,14 @@ void add_staging_to_tensor_node(
   vkapi::ShaderInfo shader = get_nchw_to_tensor_shader(
       *graph.get_tensor(out_tensor), graph.int8_buffers_enabled());
 
-  vkapi::ParamsBindList ubos;
+  std::vector<PushConstantDataInfo> pcs;
   if (graph.is_buffer_storage(out_tensor)) {
-    ubos.append(
-        {graph.sizes_ubo(out_tensor),
-         graph.strides_ubo(out_tensor),
-         graph.numel_ubo(out_tensor)});
+    pcs = {
+        graph.sizes_pc_of(out_tensor),
+        graph.strides_pc_of(out_tensor),
+        graph.numel_pc_of(out_tensor)};
   } else {
-    ubos.append({graph.sizes_ubo(out_tensor)});
+    pcs = {graph.sizes_pc_of(out_tensor)};
   }
 
   graph.execute_nodes().emplace_back(new DispatchNode(
@@ -46,9 +46,9 @@ void add_staging_to_tensor_node(
       // Input and Outputs
       {{out_tensor, vkapi::kWrite}, {in_staging, vkapi::kRead}},
       // Parameter Buffers
-      ubos,
-      // Push Constants
       {},
+      // Push Constants
+      pcs,
       // Specialization Constants
       {graph.hashed_layout_of(out_tensor)},
       // Resize Args
@@ -127,14 +127,14 @@ void add_prepack_standard_node(
   vkapi::ShaderInfo shader = get_nchw_to_tensor_shader(
       *graph.get_tensor(tensor), graph.int8_buffers_enabled());
 
-  vkapi::ParamsBindList ubos;
+  std::vector<PushConstantDataInfo> pcs;
   if (graph.is_buffer_storage(tensor)) {
-    ubos.append(
-        {graph.sizes_ubo(tensor),
-         graph.strides_ubo(tensor),
-         graph.numel_ubo(tensor)});
+    pcs = {
+        graph.sizes_pc_of(tensor),
+        graph.strides_pc_of(tensor),
+        graph.numel_pc_of(tensor)};
   } else {
-    ubos.append({graph.sizes_ubo(tensor)});
+    pcs = {graph.sizes_pc_of(tensor)};
   }
 
   int transpose_hw_spec = transpose_hw ? 1 : 0;
@@ -148,9 +148,10 @@ void add_prepack_standard_node(
       tensor_data,
       tensor,
       // Parameter Buffers
-      ubos,
+      {},
       // Specialization Constants
-      {graph.hashed_layout_of(tensor), transpose_hw_spec}));
+      {graph.hashed_layout_of(tensor), transpose_hw_spec},
+      pcs));
 }
 
 ValueRef prepack_standard(