RemoteIO: use a pinned bounce buffer

cpp/include/kvikio/bounce_buffer.hpp

@@ -36,6 +36,7 @@ class AllocRetain {
   // The size of each allocation in `_free_allocs`
   std::size_t _size{defaults::bounce_buffer_size()};
 
+ public:
   /**
    * @brief An host memory allocation
    */
@@ -56,6 +57,7 @@ class AllocRetain {
     Alloc& operator=(Alloc&& o)    = delete;
     ~Alloc() noexcept { _manager->put(_alloc, _size); }
     void* get() noexcept { return _alloc; }
+    void* get(std::ptrdiff_t offset) noexcept { return static_cast<char*>(_alloc) + offset; }
     std::size_t size() noexcept { return _size; }
   };
 
@@ -67,6 +69,7 @@ class AllocRetain {
   // <https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#initialization>
   ~AllocRetain() noexcept = default;
 
+ private:
   /**
    * @brief Free all retained allocations
    *

cpp/include/kvikio/remote_handle.hpp

@@ -15,8 +15,10 @@
  */
 #pragma once
 
+#include <cassert>
 #include <cstddef>
 #include <cstring>
+#include <iostream>
 #include <memory>
 #include <optional>
 #include <regex>
@@ -34,6 +36,96 @@
 namespace kvikio {
 namespace detail {
 
+/**
+ * @brief Bounce buffer in pinned host memory.
+ */
+class BounceBufferH2D {
+  CUstream _stream;                 // The CUDA steam to use.
+  CUdeviceptr _dev;                 // The output device buffer.
+  AllocRetain::Alloc _host_buffer;  // The host buffer to bounce data on.
+  std::ptrdiff_t _dev_offset{0};    // Number of bytes written to `_dev`.
+  std::ptrdiff_t _host_offset{0};   // Number of bytes written to `_host` (resets on flush).
+
+ public:
+  /**
+   * @brief Create a bounce buffer for an output device buffer.
+   *
+   * @param stream The CUDA stream used throughout the lifetime of the bounce buffer.
+   * @param device_buffer The output device buffer (final destination of the data).
+   */
+  BounceBufferH2D(CUstream stream, void* device_buffer)
+    : _stream{stream},
+      _dev{convert_void2deviceptr(device_buffer)},
+      _host_buffer{AllocRetain::instance().get()}
+  {
+  }
+
+  /**
+   * @brief The bounce buffer if flushed to device on destruction.
+   */
+  ~BounceBufferH2D() noexcept
+  {
+    try {
+      flush();
+    } catch (CUfileException const& e) {
+      std::cerr << "BounceBufferH2D error on final flush: ";
+      std::cerr << e.what();
+      std::cerr << std::endl;
+    }
+  }
+
+ private:
+  /**
+   * @brief Write host memory to the output device buffer.
+   *
+   * @param src The host memory source.
+   * @param size Number of bytes to write.
+   */
+  void write_to_device(void const* src, std::size_t size)
+  {
+    if (size > 0) {
+      CUDA_DRIVER_TRY(cudaAPI::instance().MemcpyHtoDAsync(_dev + _dev_offset, src, size, _stream));
+      CUDA_DRIVER_TRY(cudaAPI::instance().StreamSynchronize(_stream));
+      _dev_offset += size;
+    }
+  }
+
+  /**
+   * @brief FLush the bounce buffer by writing everything to the output device buffer.
+   */
+  void flush()
+  {
+    write_to_device(_host_buffer.get(), _host_offset);
+    _host_offset = 0;
+  }
+
+ public:
+  /**
+   * @brief Write host memory to the bounce buffer (also host memory).
+   *
+   * Only when the bounce buffer has been filled up is data copied to the output device buffer.
+   *
+   * @param data The host memory source.
+   * @param size Number of bytes to write.
+   */
+  void write(char const* data, std::size_t size)
+  {
+    if (_host_buffer.size() - _host_offset < size) {  // Not enough space left on the bounce buffer
+      flush();
+      assert(_host_offset == 0);
+    }
+    if (_host_buffer.size() < size) {
+      // If still not enough space, we just copy the data to the device. This only happens when
+      // `defaults::bounce_buffer_size()` is smaller than 16kb thus no need to performance
+      // optimize for this case.
+      write_to_device(data, size);
+    } else if (size > 0) {
+      std::memcpy(_host_buffer.get(_host_offset), data, size);
+      _host_offset += size;
+    }
+  }
+};
+
 /**
  * @brief Context used by the "CURLOPT_WRITEFUNCTION" callbacks.
  */
@@ -46,6 +138,7 @@ struct CallbackContext {
     : buf{static_cast<char*>(buf)}, size{size}, offset{0}, overflow_error{0}
   {
   }
+  BounceBufferH2D* bounce_buffer{nullptr};  // Only used by callback_device_memory
 };
 
 /**
@@ -98,12 +191,7 @@ inline std::size_t callback_device_memory(char* data,
   }
   KVIKIO_NVTX_FUNC_RANGE("RemoteHandle - callback_device_memory()", nbytes);
 
-  CUstream stream = detail::StreamsByThread::get();
-  CUDA_DRIVER_TRY(cudaAPI::instance().MemcpyHtoDAsync(
-    convert_void2deviceptr(ctx->buf + ctx->offset), data, nbytes, stream));
-  // We have to sync since curl might overwrite or free `data`.
-  CUDA_DRIVER_TRY(cudaAPI::instance().StreamSynchronize(stream));
-
+  ctx->bounce_buffer->write(data, nbytes);
   ctx->offset += nbytes;
   return nbytes;
 }
@@ -450,6 +538,10 @@ class RemoteHandle {
         curl.perform();
       } else {
         PushAndPopContext c(get_context_from_pointer(buf));
+        // We use a bounce buffer to avoid many small memory copies to device. Libcurl has a
+        // maximum chunk size of 16kb (`CURL_MAX_WRITE_SIZE`) but chunks are often much smaller.
+        detail::BounceBufferH2D bounce_buffer(detail::StreamsByThread::get(), buf);
+        ctx.bounce_buffer = &bounce_buffer;
         curl.perform();
       }
     } catch (std::runtime_error const& e) {