How to get below 10ms latency using WASAPI shared mode?

Question

According to Microsoft, starting with Windows 10, applications using shared-mode WASAPI can request buffer sizes smaller than 10ms (see https://msdn.microsoft.com/en-us/library/windows/hardware/mt298187%28v=vs.85%29.aspx).

According to the article, achieving such low latencies requires some driver updates, which I did. Using an exclusive-mode render and capture stream, I measured a total round-trip latency (using a hardware loopback cable) of around 13ms. This suggests to me that at least one of the endpoints successfully achieves a latency of < 10ms. (Is this assumption correct?)

The article mentions that applications can use the new IAudioClient3 interface to query the minimum buffer size supported by the Windows audio engine using IAudioClient3::GetSharedModeEnginePeriod(). However, this function always returns 10ms on my system, and any attempt to initialize an audio stream using either IAudioClient::Initialize() or IAudioClient3::InitializeSharedAudioStream() with a period lower than 10ms always results in AUDCLNT_E_INVALID_DEVICE_PERIOD.

Just to be sure, I also disabled any effects processing in the audio drivers. What am I missing? Is it even possible to get low latency from shared mode? See below for some sample code.

#include <windows.h>
#include <atlbase.h>
#include <mmdeviceapi.h>
#include <audioclient.h>
#include <iostream>

#define VERIFY(hr) do {                                    \
  auto temp = (hr);                                        \
  if(FAILED(temp)) {                                       \
    std::cout << "Error: " << #hr << ": " << temp << "\n"; \
    goto error;                                            \
  }                                                        \
} while(0)


int main(int argc, char** argv) {

  HRESULT hr;
  CComPtr<IMMDevice> device;
  AudioClientProperties props;
  CComPtr<IAudioClient> client;
  CComPtr<IAudioClient2> client2;
  CComPtr<IAudioClient3> client3;
  CComHeapPtr<WAVEFORMATEX> format;
  CComPtr<IMMDeviceEnumerator> enumerator; 

  REFERENCE_TIME minTime, maxTime, engineTime;
  UINT32 min, max, fundamental, default_, current;

  VERIFY(CoInitializeEx(nullptr, COINIT_APARTMENTTHREADED));
  VERIFY(enumerator.CoCreateInstance(__uuidof(MMDeviceEnumerator)));
  VERIFY(enumerator->GetDefaultAudioEndpoint(eRender, eMultimedia, &device));
  VERIFY(device->Activate(__uuidof(IAudioClient), CLSCTX_ALL, nullptr, reinterpret_cast<void**>(&client)));
  VERIFY(client->QueryInterface(&client2));
  VERIFY(client->QueryInterface(&client3));

  VERIFY(client3->GetCurrentSharedModeEnginePeriod(&format, &current));

  // Always fails with AUDCLNT_E_OFFLOAD_MODE_ONLY.
  hr = client2->GetBufferSizeLimits(format, TRUE, &minTime, &maxTime);
  if(hr == AUDCLNT_E_OFFLOAD_MODE_ONLY)
    std::cout << "GetBufferSizeLimits returned AUDCLNT_E_OFFLOAD_MODE_ONLY.\n";
  else if(SUCCEEDED(hr))
    std::cout << "hw min = " << (minTime / 10000.0) << " hw max = " << (maxTime / 10000.0) << "\n";
  else
    VERIFY(hr);

  // Correctly? reports a minimum hardware period of 3ms and audio engine period of 10ms.
  VERIFY(client->GetDevicePeriod(&engineTime, &minTime));
  std::cout << "hw min = " << (minTime / 10000.0) << " engine = " << (engineTime / 10000.0) << "\n";

  // All values are set to a number of frames corresponding to 10ms.
  // This does not change if i change the device's sampling rate in the control panel.
  VERIFY(client3->GetSharedModeEnginePeriod(format, &default_, &fundamental, &min, &max));
  std::cout << "default = " << default_ 
            << " fundamental = " << fundamental 
            << " min = " << min 
            << " max = " << max 
            << " current = " << current << "\n";

  props.bIsOffload = FALSE;
  props.cbSize = sizeof(props);
  props.eCategory = AudioCategory_ForegroundOnlyMedia;
  props.Options = AUDCLNT_STREAMOPTIONS_RAW | AUDCLNT_STREAMOPTIONS_MATCH_FORMAT;

  // Doesn't seem to have any effect regardless of category/options values.
  VERIFY(client2->SetClientProperties(&props));

  format.Free();
  VERIFY(client3->GetCurrentSharedModeEnginePeriod(&format, &current));
  VERIFY(client3->GetSharedModeEnginePeriod(format, &default_, &fundamental, &min, &max));
  std::cout << "default = " << default_ 
            << " fundamental = " << fundamental 
            << " min = " << min 
            << " max = " << max 
            << " current = " << current << "\n";

error:
  CoUninitialize();
  return 0;
}

Answer 1

Per Hans in the comment above, double-check that you've followed the instructions for Low Latency Audio here.

I'd reboot the machine just to be sure; Windows can be a bit finicky with that kind of thing.