[Impeller] Encode render passes concurrently on iOS. (#42028)

Allows pushing encoding of command buffers to a worker thread, relying on the fact that these buffers are always scheduled in the order that they are enqueued. This follows the guidelines from https://developer.apple.com/documentation/metal/mtlcommandbuffer?language=objc

Author: Jonah Williams

impeller/entity/contents/content_context.cc

@@ -375,11 +375,7 @@ std::shared_ptr<Texture> ContentContext::MakeSubpass(
     return nullptr;
   }
 
-  if (!sub_renderpass->EncodeCommands()) {
-    return nullptr;
-  }
-
-  if (!sub_command_buffer->SubmitCommands()) {
+  if (!sub_command_buffer->SubmitCommandsAsync(std::move(sub_renderpass))) {
     return nullptr;
   }
 

impeller/entity/inline_pass_context.cc

@@ -54,15 +54,9 @@ bool InlinePassContext::EndPass() {
     return true;
   }
 
-  if (!pass_->EncodeCommands()) {
-    VALIDATION_LOG
-        << "Failed to encode commands while ending the current render pass.";
-    return false;
-  }
-
-  if (!command_buffer_->SubmitCommands()) {
-    VALIDATION_LOG
-        << "Failed to submit command buffer while ending render pass.";
+  if (!command_buffer_->SubmitCommandsAsync(std::move(pass_))) {
+    VALIDATION_LOG << "Failed to encode and submit command buffer while ending "
+                      "render pass.";
     return false;
   }
 

impeller/playground/backend/metal/playground_impl_mtl.h

@@ -6,7 +6,9 @@
 
 #include <memory>
 
+#include "flutter/fml/concurrent_message_loop.h"
 #include "flutter/fml/macros.h"
+#include "flutter/fml/synchronization/sync_switch.h"
 #include "impeller/playground/playground_impl.h"
 
 namespace impeller {
@@ -27,6 +29,8 @@ class PlaygroundImplMTL final : public PlaygroundImpl {
   // To ensure that ObjC stuff doesn't leak into C++ TUs.
   std::unique_ptr<Data> data_;
   std::shared_ptr<Context> context_;
+  std::shared_ptr<fml::ConcurrentMessageLoop> concurrent_loop_;
+  std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch_;
 
   // |PlaygroundImpl|
   std::shared_ptr<Context> GetContext() const override;

impeller/playground/backend/metal/playground_impl_mtl.mm

@@ -63,16 +63,20 @@
 PlaygroundImplMTL::PlaygroundImplMTL(PlaygroundSwitches switches)
     : PlaygroundImpl(switches),
       handle_(nullptr, &DestroyWindowHandle),
-      data_(std::make_unique<Data>()) {
+      data_(std::make_unique<Data>()),
+      concurrent_loop_(fml::ConcurrentMessageLoop::Create()),
+      is_gpu_disabled_sync_switch_(new fml::SyncSwitch(false)) {
   ::glfwDefaultWindowHints();
   ::glfwWindowHint(GLFW_CLIENT_API, GLFW_NO_API);
   ::glfwWindowHint(GLFW_VISIBLE, GLFW_FALSE);
   auto window = ::glfwCreateWindow(1, 1, "Test", nullptr, nullptr);
   if (!window) {
     return;
   }
-  auto context = ContextMTL::Create(ShaderLibraryMappingsForPlayground(),
-                                    "Playground Library");
+  auto worker_task_runner = concurrent_loop_->GetTaskRunner();
+  auto context = ContextMTL::Create(
+      ShaderLibraryMappingsForPlayground(), worker_task_runner,
+      is_gpu_disabled_sync_switch_, "Playground Library");
   if (!context) {
     return;
   }

impeller/renderer/backend/metal/command_buffer_mtl.h

@@ -36,6 +36,9 @@ class CommandBufferMTL final : public CommandBuffer {
   // |CommandBuffer|
   void OnWaitUntilScheduled() override;
 
+  // |CommandBuffer|
+  bool SubmitCommandsAsync(std::shared_ptr<RenderPass> render_pass) override;
+
   // |CommandBuffer|
   std::shared_ptr<RenderPass> OnCreateRenderPass(RenderTarget target) override;
 

impeller/renderer/backend/metal/command_buffer_mtl.mm

@@ -4,8 +4,13 @@
 
 #include "impeller/renderer/backend/metal/command_buffer_mtl.h"
 
+#include "flutter/fml/make_copyable.h"
+#include "flutter/fml/synchronization/semaphore.h"
+#include "flutter/fml/trace_event.h"
+
 #include "impeller/renderer/backend/metal/blit_pass_mtl.h"
 #include "impeller/renderer/backend/metal/compute_pass_mtl.h"
+#include "impeller/renderer/backend/metal/context_mtl.h"
 #include "impeller/renderer/backend/metal/render_pass_mtl.h"
 
 namespace impeller {
@@ -171,6 +176,62 @@ static bool LogMTLCommandBufferErrorIfPresent(id<MTLCommandBuffer> buffer) {
   return true;
 }
 
+bool CommandBufferMTL::SubmitCommandsAsync(
+    std::shared_ptr<RenderPass> render_pass) {
+  TRACE_EVENT0("impeller", "CommandBufferMTL::SubmitCommandsAsync");
+  if (!IsValid() || !render_pass->IsValid()) {
+    return false;
+  }
+  auto context = context_.lock();
+  if (!context) {
+    return false;
+  }
+  [buffer_ enqueue];
+  auto buffer = buffer_;
+  buffer_ = nil;
+
+  auto worker_task_runner = ContextMTL::Cast(*context).GetWorkerTaskRunner();
+  auto mtl_render_pass = static_cast<RenderPassMTL*>(render_pass.get());
+
+  // Render command encoder creation has been observed to exceed the stack size
+  // limit for worker threads, and therefore is intentionally constructed on the
+  // raster thread.
+  auto render_command_encoder =
+      [buffer renderCommandEncoderWithDescriptor:mtl_render_pass->desc_];
+  if (!render_command_encoder) {
+    return false;
+  }
+
+  auto task = fml::MakeCopyable([render_pass, buffer, render_command_encoder,
+                                 weak_context = context_]() {
+    auto context = weak_context.lock();
+    if (!context) {
+      return;
+    }
+    auto is_gpu_disabled_sync_switch =
+        ContextMTL::Cast(*context).GetIsGpuDisabledSyncSwitch();
+    is_gpu_disabled_sync_switch->Execute(fml::SyncSwitch::Handlers().SetIfFalse(
+        [&render_pass, &render_command_encoder, &buffer, &context] {
+          auto mtl_render_pass = static_cast<RenderPassMTL*>(render_pass.get());
+          if (!mtl_render_pass->label_.empty()) {
+            [render_command_encoder
+                setLabel:@(mtl_render_pass->label_.c_str())];
+          }
+
+          auto result = mtl_render_pass->EncodeCommands(
+              context->GetResourceAllocator(), render_command_encoder);
+          [render_command_encoder endEncoding];
+          if (result) {
+            [buffer commit];
+          } else {
+            VALIDATION_LOG << "Failed to encode command buffer";
+          }
+        }));
+  });
+  worker_task_runner->PostTask(task);
+  return true;
+}
+
 void CommandBufferMTL::OnWaitUntilScheduled() {}
 
 std::shared_ptr<RenderPass> CommandBufferMTL::OnCreateRenderPass(

impeller/renderer/backend/metal/context_mtl.h

@@ -9,7 +9,9 @@
 #include <string>
 #include <vector>
 
+#include "flutter/fml/concurrent_message_loop.h"
 #include "flutter/fml/macros.h"
+#include "flutter/fml/synchronization/sync_switch.h"
 #include "impeller/base/backend_cast.h"
 #include "impeller/core/sampler.h"
 #include "impeller/renderer/backend/metal/allocator_mtl.h"
@@ -26,10 +28,14 @@ class ContextMTL final : public Context,
                          public std::enable_shared_from_this<ContextMTL> {
  public:
   static std::shared_ptr<ContextMTL> Create(
-      const std::vector<std::string>& shader_library_paths);
+      const std::vector<std::string>& shader_library_paths,
+      std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+      std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch);
 
   static std::shared_ptr<ContextMTL> Create(
       const std::vector<std::shared_ptr<fml::Mapping>>& shader_libraries_data,
+      std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+      std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch,
       const std::string& label);
 
   // |Context|
@@ -66,6 +72,10 @@ class ContextMTL final : public Context,
 
   id<MTLCommandBuffer> CreateMTLCommandBuffer() const;
 
+  const std::shared_ptr<fml::ConcurrentTaskRunner>& GetWorkerTaskRunner() const;
+
+  std::shared_ptr<const fml::SyncSwitch> GetIsGpuDisabledSyncSwitch() const;
+
  private:
   id<MTLDevice> device_ = nullptr;
   id<MTLCommandQueue> command_queue_ = nullptr;
@@ -74,9 +84,15 @@ class ContextMTL final : public Context,
   std::shared_ptr<SamplerLibrary> sampler_library_;
   std::shared_ptr<AllocatorMTL> resource_allocator_;
   std::shared_ptr<const Capabilities> device_capabilities_;
+  std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner_;
+  std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch_;
   bool is_valid_ = false;
 
-  ContextMTL(id<MTLDevice> device, NSArray<id<MTLLibrary>>* shader_libraries);
+  ContextMTL(
+      id<MTLDevice> device,
+      NSArray<id<MTLLibrary>>* shader_libraries,
+      std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+      std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch);
 
   std::shared_ptr<CommandBuffer> CreateCommandBufferInQueue(
       id<MTLCommandQueue> queue) const;

impeller/renderer/backend/metal/context_mtl.mm

@@ -65,9 +65,14 @@ static bool DeviceSupportsComputeSubgroups(id<MTLDevice> device) {
       .Build();
 }
 
-ContextMTL::ContextMTL(id<MTLDevice> device,
-                       NSArray<id<MTLLibrary>>* shader_libraries)
-    : device_(device) {
+ContextMTL::ContextMTL(
+    id<MTLDevice> device,
+    NSArray<id<MTLLibrary>>* shader_libraries,
+    std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+    std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch)
+    : device_(device),
+      worker_task_runner_(std::move(worker_task_runner)),
+      is_gpu_disabled_sync_switch_(std::move(is_gpu_disabled_sync_switch)) {
   // Validate device.
   if (!device_) {
     VALIDATION_LOG << "Could not setup valid Metal device.";
@@ -200,10 +205,13 @@ static bool DeviceSupportsComputeSubgroups(id<MTLDevice> device) {
 }
 
 std::shared_ptr<ContextMTL> ContextMTL::Create(
-    const std::vector<std::string>& shader_library_paths) {
+    const std::vector<std::string>& shader_library_paths,
+    std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+    std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch) {
   auto device = CreateMetalDevice();
   auto context = std::shared_ptr<ContextMTL>(new ContextMTL(
-      device, MTLShaderLibraryFromFilePaths(device, shader_library_paths)));
+      device, MTLShaderLibraryFromFilePaths(device, shader_library_paths),
+      std::move(worker_task_runner), std::move(is_gpu_disabled_sync_switch)));
   if (!context->IsValid()) {
     FML_LOG(ERROR) << "Could not create Metal context.";
     return nullptr;
@@ -213,11 +221,14 @@ static bool DeviceSupportsComputeSubgroups(id<MTLDevice> device) {
 
 std::shared_ptr<ContextMTL> ContextMTL::Create(
     const std::vector<std::shared_ptr<fml::Mapping>>& shader_libraries_data,
+    std::shared_ptr<fml::ConcurrentTaskRunner> worker_task_runner,
+    std::shared_ptr<const fml::SyncSwitch> is_gpu_disabled_sync_switch,
     const std::string& label) {
   auto device = CreateMetalDevice();
   auto context = std::shared_ptr<ContextMTL>(new ContextMTL(
       device,
-      MTLShaderLibraryFromFileData(device, shader_libraries_data, label)));
+      MTLShaderLibraryFromFileData(device, shader_libraries_data, label),
+      worker_task_runner, std::move(is_gpu_disabled_sync_switch)));
   if (!context->IsValid()) {
     FML_LOG(ERROR) << "Could not create Metal context.";
     return nullptr;
@@ -257,6 +268,16 @@ static bool DeviceSupportsComputeSubgroups(id<MTLDevice> device) {
   return CreateCommandBufferInQueue(command_queue_);
 }
 
+const std::shared_ptr<fml::ConcurrentTaskRunner>&
+ContextMTL::GetWorkerTaskRunner() const {
+  return worker_task_runner_;
+}
+
+std::shared_ptr<const fml::SyncSwitch> ContextMTL::GetIsGpuDisabledSyncSwitch()
+    const {
+  return is_gpu_disabled_sync_switch_;
+}
+
 std::shared_ptr<CommandBuffer> ContextMTL::CreateCommandBufferInQueue(
     id<MTLCommandQueue> queue) const {
   if (!IsValid()) {

impeller/renderer/backend/metal/render_pass_mtl.mm

@@ -6,11 +6,13 @@
 
 #include "flutter/fml/closure.h"
 #include "flutter/fml/logging.h"
+#include "flutter/fml/make_copyable.h"
 #include "flutter/fml/trace_event.h"
 #include "impeller/base/backend_cast.h"
 #include "impeller/core/formats.h"
 #include "impeller/core/host_buffer.h"
 #include "impeller/core/shader_types.h"
+#include "impeller/renderer/backend/metal/context_mtl.h"
 #include "impeller/renderer/backend/metal/device_buffer_mtl.h"
 #include "impeller/renderer/backend/metal/formats_mtl.h"
 #include "impeller/renderer/backend/metal/pipeline_mtl.h"

impeller/renderer/command_buffer.cc

@@ -36,6 +36,21 @@ void CommandBuffer::WaitUntilScheduled() {
   return OnWaitUntilScheduled();
 }
 
+bool CommandBuffer::SubmitCommandsAsync(
+    std::shared_ptr<RenderPass>
+        render_pass  // NOLINT(performance-unnecessary-value-param)
+) {
+  TRACE_EVENT0("impeller", "CommandBuffer::SubmitCommandsAsync");
+  if (!render_pass->IsValid() || !IsValid()) {
+    return false;
+  }
+  if (!render_pass->EncodeCommands()) {
+    return false;
+  }
+
+  return SubmitCommands(nullptr);
+}
+
 std::shared_ptr<RenderPass> CommandBuffer::CreateRenderPass(
     const RenderTarget& render_target) {
   auto pass = OnCreateRenderPass(render_target);