Replace std::stringstream with std::string for Shader names (#2964)

SS-JIA · facebook-github-bot · commit a983ebc9e4f7 · 2024-04-10T08:43:43.000-07:00
Summary: Pull Request resolved: #2964 ## Context Some research into efficient string concatenation suggests that streams in C++ are not quite efficient. The best way to concatenate strings seems to be creating a `std::string` and reserving sufficient capacity for the `std::string`. This diff deprecates the usage of `std::stringstream` when constructing kernel names in favor of using `std::string` directly. Reviewed By: copyrightly Differential Revision: D55951475 fbshipit-source-id: a1a584669e80984b85d11b7d6d4f7593290e562b
diff --git a/backends/vulkan/runtime/graph/ops/PrepackNode.cpp b/backends/vulkan/runtime/graph/ops/PrepackNode.cpp
@@ -17,11 +17,10 @@
 namespace vkcompute {
 
 api::ShaderInfo get_noop_shader(ComputeGraph& graph, const ValueRef packed) {
-  std::stringstream noop_shader_name;
-  noop_shader_name << "no_op";
-  apply_ndim_suffix(noop_shader_name, graph.get_val(packed).toTensor());
-  apply_dtype_suffix(noop_shader_name, graph.get_val(packed).toTensor());
-  return VK_KERNEL_FROM_STR(noop_shader_name.str());
+  std::string noop_shader_name("no_op");
+  add_ndim_suffix(noop_shader_name, graph.get_val(packed).toTensor());
+  add_dtype_suffix(noop_shader_name, graph.get_val(packed).toTensor());
+  return VK_KERNEL_FROM_STR(noop_shader_name);
 }
 
 PrepackNode::PrepackNode(
diff --git a/backends/vulkan/runtime/graph/ops/impl/BinaryOp.cpp b/backends/vulkan/runtime/graph/ops/impl/BinaryOp.cpp
@@ -75,14 +75,15 @@ void add_binary_op_node(
   const api::utils::ivec2 broadcast_params =
       create_broadcast_params(t_in1, t_in2);
 
-  std::stringstream kernel_name;
-  kernel_name << "binary_" << op_name;
-  apply_memory_layout_suffix(kernel_name, t_out);
-  apply_dtype_suffix(kernel_name, t_out);
+  std::string kernel_name("binary_");
+  kernel_name.reserve(kShaderNameReserve);
+  kernel_name += op_name;
+  add_memory_layout_suffix(kernel_name, t_out);
+  add_dtype_suffix(kernel_name, t_out);
 
   graph.execute_nodes().emplace_back(new ExecuteNode(
       graph,
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       global_size,
       local_size,
       // Inputs and Outputs
diff --git a/backends/vulkan/runtime/graph/ops/impl/Conv2d.cpp b/backends/vulkan/runtime/graph/ops/impl/Conv2d.cpp
@@ -94,40 +94,41 @@ api::ShaderInfo get_conv2d_shader(
     const bool prepack_weights,
     const Conv2dMethod method,
     const ValueRef weight) {
-  std::stringstream kernel_name;
+  std::string kernel_name;
+  kernel_name.reserve(kShaderNameReserve);
   switch (method) {
     case Conv2dMethod::Depthwise:
-      kernel_name << "conv2d_dw";
+      kernel_name = "conv2d_dw";
       if (!prepack_weights) {
         const auto& weight_sizes = graph.get_val(weight).toTensorRef().sizes;
         if (weight_sizes.at(2) == 3 && weight_sizes.at(3) == 3) {
-          kernel_name << "_output_tile_3x3";
+          kernel_name += "_output_tile_3x3";
         }
         if (weight_sizes.at(2) == 5 && weight_sizes.at(3) == 5) {
-          kernel_name << "_output_tile_5x5";
+          kernel_name += "_output_tile_5x5";
         }
       }
       break;
     case Conv2dMethod::Pointwise:
       if (prepack_weights) {
-        kernel_name << "conv2d";
+        kernel_name = "conv2d";
       } else {
-        kernel_name << "conv2d_pw";
+        kernel_name = "conv2d_pw";
       }
       break;
     case Conv2dMethod::SlidingWindow:
-      kernel_name << "conv2d";
+      kernel_name = "conv2d";
       break;
     case Conv2dMethod::Transposed:
-      kernel_name << "conv_transpose2d";
+      kernel_name = "conv_transpose2d";
       break;
   }
   if (prepack_weights) {
-    kernel_name << "_prepack_weights";
+    kernel_name += "_prepack_weights";
   }
-  apply_dtype_suffix(kernel_name, t_out);
+  add_dtype_suffix(kernel_name, t_out);
 
-  return VK_KERNEL_FROM_STR(kernel_name.str());
+  return VK_KERNEL_FROM_STR(kernel_name);
 }
 
 std::vector<int64_t> get_final_sizes(
diff --git a/backends/vulkan/runtime/graph/ops/impl/MatMul.cpp b/backends/vulkan/runtime/graph/ops/impl/MatMul.cpp
@@ -78,15 +78,15 @@ void add_matmul_node(
   api::utils::uvec3 global_size = t_out.virtual_extents();
   api::utils::uvec3 local_size = adaptive_work_group_size(global_size);
 
-  std::stringstream kernel_name;
-  kernel_name << "matmul";
-  apply_memory_layout_suffix(kernel_name, t_mat1);
-  apply_memory_layout_suffix(kernel_name, t_mat2);
-  apply_dtype_suffix(kernel_name, t_out);
+  std::string kernel_name("matmul");
+  kernel_name.reserve(kShaderNameReserve);
+  add_memory_layout_suffix(kernel_name, t_mat1);
+  add_memory_layout_suffix(kernel_name, t_mat2);
+  add_dtype_suffix(kernel_name, t_out);
 
   graph.execute_nodes().emplace_back(new ExecuteNode(
       graph,
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       global_size,
       local_size,
       // Inputs and Outputs
diff --git a/backends/vulkan/runtime/graph/ops/impl/Pool.cpp b/backends/vulkan/runtime/graph/ops/impl/Pool.cpp
@@ -73,9 +73,8 @@ void add_max_pool2d_node(
   api::utils::uvec3 global_size = t_out.virtual_extents();
   api::utils::uvec3 local_size = adaptive_work_group_size(global_size);
 
-  std::stringstream kernel_name;
-  kernel_name << "max_pool2d";
-  apply_dtype_suffix(kernel_name, t_out);
+  std::string kernel_name("max_pool2d");
+  add_dtype_suffix(kernel_name, t_out);
 
   KernelParams kernel_params = create_kernel_params(
       graph,
@@ -87,7 +86,7 @@ void add_max_pool2d_node(
 
   graph.execute_nodes().emplace_back(new ExecuteNode(
       graph,
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       global_size,
       local_size,
       // Inputs and Outputs
diff --git a/backends/vulkan/runtime/graph/ops/impl/Sum.cpp b/backends/vulkan/runtime/graph/ops/impl/Sum.cpp
@@ -71,17 +71,17 @@ void add_sum_dim_node(
   api::utils::uvec3 global_size = t_out.virtual_extents();
   api::utils::uvec3 local_size = adaptive_work_group_size(global_size);
 
-  std::stringstream kernel_name;
-  kernel_name << "sum_dim";
+  std::string kernel_name("sum_dim");
+  kernel_name.reserve(kShaderNameReserve);
   if (keepdim) {
-    kernel_name << "_keepdim";
+    kernel_name += "_keepdim";
   }
 
-  apply_dtype_suffix(kernel_name, t_out);
+  add_dtype_suffix(kernel_name, t_out);
 
   graph.execute_nodes().emplace_back(new ExecuteNode(
       graph,
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       global_size,
       local_size,
       // Inputs and Outputs
diff --git a/backends/vulkan/runtime/graph/ops/impl/UnaryOp.cpp b/backends/vulkan/runtime/graph/ops/impl/UnaryOp.cpp
@@ -44,13 +44,12 @@ void add_unary_op_node(
   api::utils::uvec3 global_size = t_out.virtual_extents();
   api::utils::uvec3 local_size = adaptive_work_group_size(global_size);
 
-  std::stringstream kernel_name;
-  kernel_name << op_name;
-  apply_dtype_suffix(kernel_name, t_out);
+  std::string kernel_name(op_name);
+  add_dtype_suffix(kernel_name, t_out);
 
   graph.execute_nodes().emplace_back(new ExecuteNode(
       graph,
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       global_size,
       local_size,
       // Inputs and Outputs
diff --git a/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.cpp b/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.cpp
@@ -10,47 +10,45 @@
 
 namespace vkcompute {
 
-void apply_dtype_suffix(std::stringstream& kernel_name, const vTensor& tensor) {
+void add_dtype_suffix(std::string& kernel_name, const vTensor& tensor) {
   switch (tensor.image().format()) {
     case VK_FORMAT_R32G32B32A32_SFLOAT:
-      kernel_name << "_float";
+      kernel_name += "_float";
       break;
     case VK_FORMAT_R16G16B16A16_SFLOAT:
-      kernel_name << "_half";
+      kernel_name += "_half";
       break;
     case VK_FORMAT_R32G32B32A32_SINT:
-      kernel_name << "_int";
+      kernel_name += "_int";
       break;
     default:
       break;
   }
 }
 
-void apply_ndim_suffix(std::stringstream& kernel_name, const vTensor& tensor) {
+void add_ndim_suffix(std::string& kernel_name, const vTensor& tensor) {
   switch (tensor.storage_type()) {
     case api::kTexture3D:
-      kernel_name << "_3d";
+      kernel_name += "_3d";
       break;
     case api::kTexture2D:
-      kernel_name << "_2d";
+      kernel_name += "_2d";
       break;
     default:
       break;
   }
 }
 
-void apply_memory_layout_suffix(
-    std::stringstream& kernel_name,
-    const vTensor& tensor) {
+void add_memory_layout_suffix(std::string& kernel_name, const vTensor& tensor) {
   switch (tensor.gpu_memory_layout()) {
     case api::kChannelsPacked:
-      kernel_name << "_C_packed";
+      kernel_name += "_C_packed";
       break;
     case api::kHeightPacked:
-      kernel_name << "_H_packed";
+      kernel_name += "_H_packed";
       break;
     case api::kWidthPacked:
-      kernel_name << "_W_packed";
+      kernel_name += "_W_packed";
       break;
     default:
       break;
diff --git a/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h b/backends/vulkan/runtime/graph/ops/utils/ShaderNameUtils.h
@@ -10,16 +10,16 @@
 
 #include <executorch/backends/vulkan/runtime/api/api.h>
 
-#include <sstream>
+#include <string>
 
 namespace vkcompute {
 
-void apply_dtype_suffix(std::stringstream& kernel_name, const vTensor& tensor);
+constexpr size_t kShaderNameReserve = 64u;
 
-void apply_ndim_suffix(std::stringstream& kernel_name, const vTensor& tensor);
+void add_dtype_suffix(std::string& kernel_name, const vTensor& tensor);
 
-void apply_memory_layout_suffix(
-    std::stringstream& kernel_name,
-    const vTensor& tensor);
+void add_ndim_suffix(std::string& kernel_name, const vTensor& tensor);
+
+void add_memory_layout_suffix(std::string& kernel_name, const vTensor& tensor);
 
 } // namespace vkcompute
diff --git a/backends/vulkan/runtime/graph/ops/utils/StagingUtils.cpp b/backends/vulkan/runtime/graph/ops/utils/StagingUtils.cpp
@@ -100,47 +100,49 @@ api::ShaderInfo get_nchw_to_image_shader(const vTensor& v_dst) {
     VK_THROW("Quantized Tensors are currently not supported!");
   }
 
-  std::stringstream kernel_name;
+  std::string kernel_name;
+  kernel_name.reserve(kShaderNameReserve);
 
   switch (v_dst.storage_type()) {
     case api::kTexture3D:
-      kernel_name << "nchw_to_image3d";
+      kernel_name = "nchw_to_image3d";
       break;
     case api::kTexture2D:
-      kernel_name << "nchw_to_image2d";
+      kernel_name = "nchw_to_image2d";
       break;
     default:
       VK_THROW("No kernel available!");
   }
 
-  apply_memory_layout_suffix(kernel_name, v_dst);
-  apply_dtype_suffix(kernel_name, v_dst);
+  add_memory_layout_suffix(kernel_name, v_dst);
+  add_dtype_suffix(kernel_name, v_dst);
 
-  return VK_KERNEL_FROM_STR(kernel_name.str());
+  return VK_KERNEL_FROM_STR(kernel_name);
 }
 
 api::ShaderInfo get_image_to_nchw_shader(const vTensor& v_src) {
   if (v_src.is_quantized()) {
     VK_THROW("Quantized Tensors are currently not supported!");
   }
 
-  std::stringstream kernel_name;
+  std::string kernel_name;
+  kernel_name.reserve(kShaderNameReserve);
 
   switch (v_src.storage_type()) {
     case api::kTexture3D:
-      kernel_name << "image3d_to_nchw";
+      kernel_name = "image3d_to_nchw";
       break;
     case api::kTexture2D:
-      kernel_name << "image2d_to_nchw";
+      kernel_name = "image2d_to_nchw";
       break;
     default:
       VK_THROW("No kernel available!");
   }
 
-  apply_memory_layout_suffix(kernel_name, v_src);
-  apply_dtype_suffix(kernel_name, v_src);
+  add_memory_layout_suffix(kernel_name, v_src);
+  add_dtype_suffix(kernel_name, v_src);
 
-  return VK_KERNEL_FROM_STR(kernel_name.str());
+  return VK_KERNEL_FROM_STR(kernel_name);
 }
 
 } // namespace vkcompute
diff --git a/backends/vulkan/test/utils/test_utils.cpp b/backends/vulkan/test/utils/test_utils.cpp
@@ -63,15 +63,15 @@ void record_conv2d_prepack_weights_op(
     const bool transposed) {
   api::PipelineBarrier pipeline_barrier{};
 
-  std::stringstream kernel_name;
+  std::string kernel_name;
   if (transposed) {
-    kernel_name << "conv_transpose2d";
+    kernel_name = "conv_transpose2d";
   } else {
-    kernel_name << "conv2d";
+    kernel_name = "conv2d";
   }
-  kernel_name << "_prepack_weights";
-  apply_dtype_suffix(kernel_name, v_dst);
-  api::ShaderInfo shader = VK_KERNEL_FROM_STR(kernel_name.str());
+  kernel_name += "_prepack_weights";
+  add_dtype_suffix(kernel_name, v_dst);
+  api::ShaderInfo shader = VK_KERNEL_FROM_STR(kernel_name);
 
   api::UniformParamsBuffer original_sizes_ubo(
       context, api::utils::make_ivec4(original_sizes, /*reverse = */ true));
@@ -100,13 +100,12 @@ void record_binary_op(
     vTensor& v_in1,
     vTensor& v_in2,
     vTensor& v_dst) {
-  std::stringstream kernel_name;
-  kernel_name << "binary_" << op_name << "_nobroadcast__test";
-  apply_dtype_suffix(kernel_name, v_dst);
+  std::string kernel_name = "binary_" + op_name + "_nobroadcast__test";
+  add_dtype_suffix(kernel_name, v_dst);
 
   api::PipelineBarrier pipeline_barrier{};
   context->submit_compute_job(
-      VK_KERNEL_FROM_STR(kernel_name.str()),
+      VK_KERNEL_FROM_STR(kernel_name),
       pipeline_barrier,
       v_dst.virtual_extents(),
       adaptive_work_group_size(v_dst.virtual_extents()),
diff --git a/backends/vulkan/test/vulkan_compute_api_test.cpp b/backends/vulkan/test/vulkan_compute_api_test.cpp
@@ -49,9 +49,8 @@ TEST_F(VulkanComputeAPITest, update_params_between_submit) {
   std::vector<int64_t> sizes = {4, 4, 2};
   vTensor a = CREATE_FLOAT_TEXTURE(sizes, /*allocate_memory = */ true);
 
-  std::stringstream kernel_name;
-  kernel_name << "fill_texture__test";
-  apply_dtype_suffix(kernel_name, a);
+  std::string kernel_name("fill_texture__test");
+  add_dtype_suffix(kernel_name, a);
 
   struct Params final {
     api::utils::ivec3 size;
@@ -70,7 +69,7 @@ TEST_F(VulkanComputeAPITest, update_params_between_submit) {
   {
     api::PipelineBarrier pipeline_barrier{};
     api::context()->submit_compute_job(
-        VK_KERNEL_FROM_STR(kernel_name.str()),
+        VK_KERNEL_FROM_STR(kernel_name),
         pipeline_barrier,
         {4, 4, 4},
         {4, 4, 4},
@@ -748,15 +747,14 @@ void run_from_gpu_test(
   vTensor vten =
       vTensor(api::context(), sizes, api::kFloat, storage_type, memory_layout);
 
-  std::stringstream kernel_name;
-  kernel_name << "idx_fill_texture";
-  apply_memory_layout_suffix(kernel_name, vten);
-  apply_dtype_suffix(kernel_name, vten);
+  std::string kernel_name("idx_fill_texture");
+  add_memory_layout_suffix(kernel_name, vten);
+  add_dtype_suffix(kernel_name, vten);
 
   {
     api::PipelineBarrier pipeline_barrier{};
     api::context()->submit_compute_job(
-        VK_KERNEL_FROM_STR(kernel_name.str()),
+        VK_KERNEL_FROM_STR(kernel_name),
         pipeline_barrier,
         vten.virtual_extents(),
         {4, 4, 4},
