try to speed up

tty_protocol_writer.cpp

@@ -77,6 +77,8 @@ namespace {
 
 constexpr std::size_t kFrameWordCount = 4U;
 constexpr std::size_t kFrameByteCount = kFrameWordCount * sizeof(uint16_t);
+constexpr std::size_t kWriteBatchFrames = 256U;
+constexpr std::size_t kWriteBatchBytes = kWriteBatchFrames * kFrameByteCount;
 
 using EncodedFrame = std::array<std::uint8_t, kFrameByteCount>;
 
@@ -376,8 +378,10 @@ void TtyProtocolWriter::enqueue_frame(uint16_t word0, uint16_t word1, uint16_t w
 }
 
 void TtyProtocolWriter::worker_loop() {
+    std::array<std::uint8_t, kWriteBatchBytes> batch_bytes {};
+
     for (;;) {
-        EncodedFrame frame {};
+        std::size_t batch_frames = 0;
         {
             std::unique_lock<std::mutex> lock(impl_->mutex);
             impl_->data_ready_cv.wait(lock, [this]() {
@@ -388,13 +392,18 @@ void TtyProtocolWriter::worker_loop() {
                 return;
             }
 
-            frame = impl_->ring[impl_->head];
-            impl_->head = (impl_->head + 1U) % impl_->capacity_frames;
-            --impl_->size;
+            batch_frames = std::min<std::size_t>(kWriteBatchFrames, impl_->size);
+            for (std::size_t i = 0; i < batch_frames; ++i) {
+                std::memcpy(batch_bytes.data() + (i * kFrameByteCount),
+                            impl_->ring[impl_->head].data(),
+                            kFrameByteCount);
+                impl_->head = (impl_->head + 1U) % impl_->capacity_frames;
+            }
+            impl_->size -= batch_frames;
         }
 
-        const std::uint8_t* bytes = frame.data();
-        std::size_t remaining = frame.size();
+        const std::uint8_t* bytes = batch_bytes.data();
+        std::size_t remaining = batch_frames * kFrameByteCount;
         while (remaining != 0U) {
             const ssize_t written = ::write(impl_->fd, bytes, remaining);
             if (written < 0) {
@@ -436,7 +445,7 @@ void TtyProtocolWriter::worker_loop() {
 
         {
             std::lock_guard<std::mutex> lock(impl_->mutex);
-            ++impl_->stats.frames_written;
+            impl_->stats.frames_written += static_cast<std::uint64_t>(batch_frames);
         }
     }
 }