[client/host/idd/lgmp] add additional timing metrics

client/include/interface/transport.h

@@ -92,6 +92,14 @@ enum
 
 typedef uint32_t LG_TransportFrameFlags;
 
+typedef struct LG_TransportFrameTiming
+{
+  uint64_t captureTime;
+  uint64_t postProcessTime;
+  uint64_t copyTime;
+}
+LG_TransportFrameTiming;
+
 typedef struct LG_TransportFrameFormat
 {
   uint32_t version;
@@ -200,6 +208,11 @@ typedef struct LG_TransportOps
 
   LG_TransportStatus (*nextFrame)(LG_Transport * transport, bool useDMA,
       LG_TransportFrame * frame);
+  /* Read producer timings after the renderer has consumed the frame. Some
+   * transports publish the frame while its asynchronous copy is in progress,
+   * so these values are intentionally sampled late. */
+  void (*getFrameTiming)(LG_Transport * transport,
+      const LG_TransportFrame * frame, LG_TransportFrameTiming * timing);
   void (*releaseFrame)(LG_Transport * transport, LG_TransportFrame * frame);
   /* Called by the frame consumer as it exits. A backend may release transient
    * stream resources; nextFrame must reacquire them when the consumer restarts. */

client/src/main.c

@@ -199,11 +199,21 @@ static bool tickTimerFn(void * unused)
   return true;
 }
 
+struct RenderTiming
+{
+  uint64_t renderStart;
+  uint64_t producerTime;
+};
+
 static void preSwapCallback(void * udata)
 {
-  const uint64_t * renderStart = (const uint64_t *)udata;
+  const struct RenderTiming * timing = (const struct RenderTiming *)udata;
-  ringbuffer_push(g_state.renderDuration,
+  const uint64_t renderTime = nanotime() - timing->renderStart;
-      &(float) {(nanotime() - *renderStart) * 1e-6f});
+  ringbuffer_push(g_state.renderDuration, &(float) {renderTime * 1e-6f});
+
+  if (timing->producerTime)
+    ringbuffer_push(g_state.endToEndDuration,
+        &(float) {(timing->producerTime + renderTime) * 1e-6f});
 
 #ifdef ENABLE_TESTS
   if (!l_testCapture.enabled || l_testCapture.complete)
@@ -397,13 +407,17 @@ static int renderThread(void * unused)
 
     const bool invalidate = atomic_exchange(&g_state.invalidateWindow, false);
 
-    const uint64_t renderStart = nanotime();
+    const struct RenderTiming renderTiming = {
+      .renderStart  = nanotime(),
+      .producerTime = newFrame ? atomic_load_explicit(
+          &g_state.producerFrameTime, memory_order_acquire) : 0,
+    };
     LG_LOCK(g_state.lgrLock);
 
     renderQueue_process();
 
     if (unlikely(!RENDERER(render, g_params.winRotate, newFrame, invalidate,
-          preSwapCallback, (void *)&renderStart)))
+          preSwapCallback, (void *)&renderTiming)))
     {
       LG_UNLOCK(g_state.lgrLock);
       break;
@@ -778,6 +792,21 @@ int main_frameThread(void * unused)
       break;
     }
 
+    LG_TransportFrameTiming timing = {};
+    if (g_state.transportOps->getFrameTiming)
+      g_state.transportOps->getFrameTiming(
+          g_state.transport, &frame, &timing);
+
+    ringbuffer_push(g_state.captureDuration,
+        &(float) {timing.captureTime * 1e-6f});
+    ringbuffer_push(g_state.postProcessDuration,
+        &(float) {timing.postProcessTime * 1e-6f});
+    ringbuffer_push(g_state.copyDuration,
+        &(float) {timing.copyTime * 1e-6f});
+    atomic_store_explicit(&g_state.producerFrameTime,
+        timing.captureTime + timing.postProcessTime + timing.copyTime,
+        memory_order_release);
+
     overlaySplash_show(false);
     if ((frame.flags & LG_TRANSPORT_FRAME_REQUEST_ACTIVATION) &&
         g_params.requestActivation)
@@ -1305,12 +1334,20 @@ static int lg_run(void)
   DEBUG_INFO("Using font: %s", g_state.fontName);
 
   // initialize metrics ringbuffers
-  g_state.renderTimings  = ringbuffer_new(256, sizeof(float));
+  g_state.renderTimings       = ringbuffer_new(256, sizeof(float));
-  g_state.uploadTimings  = ringbuffer_new(256, sizeof(float));
+  g_state.uploadTimings       = ringbuffer_new(256, sizeof(float));
-  g_state.renderDuration = ringbuffer_new(256, sizeof(float));
+  g_state.renderDuration      = ringbuffer_new(256, sizeof(float));
-  overlayGraph_register("FRAME" , g_state.renderTimings , 0.0f, 50.0f, NULL);
+  g_state.captureDuration     = ringbuffer_new(256, sizeof(float));
-  overlayGraph_register("UPLOAD", g_state.uploadTimings , 0.0f, 50.0f, NULL);
+  g_state.postProcessDuration = ringbuffer_new(256, sizeof(float));
-  overlayGraph_register("RENDER", g_state.renderDuration, 0.0f, 10.0f, NULL);
+  g_state.copyDuration        = ringbuffer_new(256, sizeof(float));
+  g_state.endToEndDuration    = ringbuffer_new(256, sizeof(float));
+  overlayGraph_register("FRAME"       , g_state.renderTimings      , 0.0f, 50.0f, NULL);
+  overlayGraph_register("UPLOAD"      , g_state.uploadTimings      , 0.0f, 50.0f, NULL);
+  overlayGraph_register("RENDER"      , g_state.renderDuration     , 0.0f, 10.0f, NULL);
+  overlayGraph_register("CAPTURE"     , g_state.captureDuration    , 0.0f, 20.0f, NULL);
+  overlayGraph_register("POST PROCESS", g_state.postProcessDuration, 0.0f, 20.0f, NULL);
+  overlayGraph_register("COPY"        , g_state.copyDuration       , 0.0f, 20.0f, NULL);
+  overlayGraph_register("END TO END"  , g_state.endToEndDuration   , 0.0f, 50.0f, NULL);
 
   // unknown guest OS at this time
   g_state.guestOS = LG_TRANSPORT_OS_OTHER;
@@ -1795,6 +1832,10 @@ static void lg_shutdown(void)
   ringbuffer_free(&g_state.renderTimings);
   ringbuffer_free(&g_state.uploadTimings);
   ringbuffer_free(&g_state.renderDuration);
+  ringbuffer_free(&g_state.captureDuration);
+  ringbuffer_free(&g_state.postProcessDuration);
+  ringbuffer_free(&g_state.copyDuration);
+  ringbuffer_free(&g_state.endToEndDuration);
 
   free(g_state.fontName);
   igDestroyContext(NULL);

client/src/main.h

@@ -150,8 +150,13 @@ struct AppState
   RingBuffer            renderTimings;
   RingBuffer            renderDuration;
   RingBuffer            uploadTimings;
+  RingBuffer            captureDuration;
+  RingBuffer            postProcessDuration;
+  RingBuffer            copyDuration;
+  RingBuffer            endToEndDuration;
 
   atomic_uint_least64_t pendingCount;
+  atomic_uint_least64_t producerFrameTime;
   atomic_uint_least64_t renderCount, frameCount;
   _Atomic(float)        fps, ups;
 

client/tests/lgmp_transport_test.c

@@ -122,17 +122,22 @@ int main(void)
   wireFrame->formatVer   = 1;
   wireFrame->frameSerial = 1;
   wireFrame->type        = FRAME_TYPE_BGRA;
-  wireFrame->screenWidth = 1;
+  wireFrame->screenWidth     = 1;
-  wireFrame->screenHeight = 1;
+  wireFrame->screenHeight    = 1;
-  wireFrame->dataWidth   = 1;
+  wireFrame->dataWidth       = 1;
-  wireFrame->dataHeight  = 1;
+  wireFrame->dataHeight      = 1;
-  wireFrame->frameWidth  = 1;
+  wireFrame->frameWidth      = 1;
-  wireFrame->frameHeight = 1;
+  wireFrame->frameHeight     = 1;
-  wireFrame->rotation    = FRAME_ROT_0;
+  wireFrame->rotation        = FRAME_ROT_0;
-  wireFrame->stride      = 1;
+  wireFrame->stride          = 1;
-  wireFrame->pitch       = sizeof(uint32_t);
+  wireFrame->pitch           = sizeof(uint32_t);
-  wireFrame->offset      = sizeof(*wireFrame);
+  wireFrame->offset          = sizeof(*wireFrame);
-  wireFrame->sdrWhiteLevel = KVMFR_SDR_WHITE_LEVEL_DEFAULT;
+  wireFrame->sdrWhiteLevel   = KVMFR_SDR_WHITE_LEVEL_DEFAULT;
+  wireFrame->captureTime     = 100;
+  wireFrame->postProcessTime = 200;
+  wireFrame->copyTime        = 300;
+  wireFrame->timingSerial    = wireFrame->frameSerial;
+  wireFrame->timingValid     = 1;
   FrameBuffer * framebuffer = (FrameBuffer *)((uint8_t *)wireFrame +
       wireFrame->offset);
   atomic_store(&framebuffer->wp, sizeof(uint32_t));
@@ -172,6 +177,11 @@ int main(void)
   CHECK(lgmpHostQueuePost(pointerQueue, CURSOR_FLAG_POSITION,
         pointerMemory) == LGMP_OK);
   CHECK(LGT_LGMP.nextFrame(transport, false, &frame) == LG_TRANSPORT_OK);
+  LG_TransportFrameTiming timing;
+  LGT_LGMP.getFrameTiming(transport, &frame, &timing);
+  CHECK(timing.captureTime == wireFrame->captureTime);
+  CHECK(timing.postProcessTime == wireFrame->postProcessTime);
+  CHECK(timing.copyTime == wireFrame->copyTime);
   LGT_LGMP.releaseFrame(transport, &frame);
   CHECK(LGT_LGMP.nextPointer(transport, &pointer) == LG_TRANSPORT_OK);
   LGT_LGMP.releasePointer(transport, &pointer);
@@ -212,7 +222,8 @@ int main(void)
    * Also recover when the host has already marked the cached handles bad.
    * LGMP leaves a timed-out handle non-NULL when unsubscribe fails.
    */
-  wireFrame->frameSerial = 2;
+  wireFrame->frameSerial  = 2;
+  wireFrame->timingSerial = wireFrame->frameSerial;
   CHECK(lgmpHostQueuePost(frameQueue, 0, frameMemory) == LGMP_OK);
   CHECK(lgmpHostQueuePost(pointerQueue, CURSOR_FLAG_POSITION,
         pointerMemory) == LGMP_OK);

client/transports/LGMP/lgmp.c

@@ -56,6 +56,7 @@ struct LG_Transport
   bool     allowDMA;
   bool     connected;
   bool     framePending;
+  const KVMFRFrame * pendingFrame;
   uint32_t frameSerial;
   bool     formatValid;
   LG_TransportFrameFormat format;
@@ -214,6 +215,7 @@ static void lgmp_stopFrame(struct LG_Transport * this)
           lgmpStatusString(status));
   }
   this->framePending = false;
+  this->pendingFrame = NULL;
   LGMP_STATUS status = lgmpClientUnsubscribe(&this->frameQueue);
   if (status != LGMP_OK)
   {
@@ -602,14 +604,47 @@ static LG_TransportStatus lgmp_nextFrame(LG_Transport * this, bool useDMA,
     DEBUG_WARN("Invalid damage rectangles, forcing a full update");
 
   this->framePending = true;
+  this->pendingFrame = frame;
   return LG_TRANSPORT_OK;
 }
 
+static bool lgmp_frameTimingReady(const KVMFRFrame * frame)
+{
+  return __atomic_load_n(&frame->timingValid, __ATOMIC_ACQUIRE) &&
+    frame->timingSerial == frame->frameSerial;
+}
+
+static void lgmp_getFrameTiming(LG_Transport * this,
+    const LG_TransportFrame * frame, LG_TransportFrameTiming * timing)
+{
+  memset(timing, 0, sizeof(*timing));
+  if (!this->framePending || !this->pendingFrame ||
+      this->pendingFrame->frameSerial != frame->serial)
+    return;
+
+  /* The producer writes these immediately before completing the framebuffer.
+   * nextFrame can observe the header earlier, so sample them only after the
+   * renderer's onFrame call has consumed the framebuffer. */
+  for (unsigned i = 0;
+       !lgmp_frameTimingReady(this->pendingFrame) &&
+       i < 1000;
+       ++i)
+    usleep(1);
+
+  if (!lgmp_frameTimingReady(this->pendingFrame))
+    return;
+
+  timing->captureTime     = this->pendingFrame->captureTime;
+  timing->postProcessTime = this->pendingFrame->postProcessTime;
+  timing->copyTime        = this->pendingFrame->copyTime;
+}
+
 static void lgmp_releaseFrame(LG_Transport * this, LG_TransportFrame * frame)
 {
   if (this->framePending && this->frameQueue)
     lgmpClientMessageDone(this->frameQueue);
   this->framePending = false;
+  this->pendingFrame = NULL;
   memset(frame, 0, sizeof(*frame));
 }
 
@@ -800,6 +835,7 @@ const LG_TransportOps LGT_LGMP =
   .attachRenderer = lgmp_attachRenderer,
   .detachRenderer = lgmp_detachRenderer,
   .nextFrame      = lgmp_nextFrame,
+  .getFrameTiming = lgmp_getFrameTiming,
   .releaseFrame   = lgmp_releaseFrame,
   .stopFrame      = lgmp_stopFrame,
   .nextPointer    = lgmp_nextPointer,

common/include/common/KVMFR.h

@@ -25,11 +25,12 @@
 
 #include <stdint.h>
 #include <stdbool.h>
+#include <stddef.h>
 #include "types.h"
 #include "LGMPConfig.h"
 
 #define KVMFR_MAGIC   "KVMFR---"
-#define KVMFR_VERSION 23
+#define KVMFR_VERSION 24
 
 // Fallback used by producers that cannot report the source display's SDR
 // white level. IDD frames override this with IDDCX_METADATA2::SdrWhiteLevel.
@@ -165,6 +166,11 @@ typedef uint32_t KVMFRFrameFlags;
 
 typedef struct KVMFRFrame
 {
+  /*
+   * Keep the fields consumed for every frame in the first cache line. The
+   * timing and HDR metadata below are diagnostic/conditional, while the large
+   * damage rectangle array is deliberately last.
+   */
   uint32_t        formatVer;          // the frame format version number
   uint32_t        frameSerial;        // the unique frame number
   FrameType       type;               // the frame data type
@@ -178,9 +184,28 @@ typedef struct KVMFRFrame
   uint32_t        stride;             // the row stride (zero if compressed data)
   uint32_t        pitch;              // the row pitch  (stride in bytes or the compressed frame size)
   uint32_t        offset;             // offset from the start of this header to the FrameBuffer header
-  uint32_t        damageRectsCount;   // the number of damage rectangles (zero for full-frame damage)
-  FrameDamageRect damageRects[KVMFR_MAX_DAMAGE_RECTS];
   KVMFRFrameFlags flags;              // bit field combination of FRAME_FLAG_*
+  uint32_t        damageRectsCount;   // the number of damage rectangles (zero for full-frame damage)
+
+  // White level in nits for SDR content composited into an HDR frame, such as
+  // the hardware cursor. Valid for all frames; SDR modes normally report 80.
+  uint32_t        sdrWhiteLevel;
+
+  /*
+   * Producer stage durations in nanoseconds. These are durations rather than
+   * absolute timestamps because the producer runs in a VM whose monotonic
+   * clock has a different epoch from the client. Durations can safely be
+   * combined with the client's local render duration for an end-to-end
+   * processing time.
+   */
+  uint64_t        captureTime;
+  uint64_t        postProcessTime;
+  uint64_t        copyTime;
+
+  // Published after the timing fields and matched against frameSerial by the
+  // client. timingValid is written last by the producer.
+  uint32_t        timingSerial;
+  uint32_t        timingValid;
 
   // HDR static metadata (valid when FRAME_FLAG_HDR_METADATA is set)
   // Display color primaries in 0.00002 units (SMPTE ST 2086 format)
@@ -195,12 +220,22 @@ typedef struct KVMFRFrame
   uint32_t hdrMaxContentLightLevel;      // MaxCLL (cd/m²)
   uint32_t hdrMaxFrameAverageLightLevel; // MaxFALL (cd/m²)
 
-  // White level in nits for SDR content composited into an HDR frame, such as
+  FrameDamageRect damageRects[KVMFR_MAX_DAMAGE_RECTS];
-  // the hardware cursor. Valid for all frames; SDR modes normally report 80.
-  uint32_t sdrWhiteLevel;
 }
 KVMFRFrame;
 
+#if defined(__cplusplus)
+static_assert(offsetof(KVMFRFrame, captureTime) == 64,
+    "KVMFRFrame hot fields must fit in one cache line");
+static_assert(offsetof(KVMFRFrame, damageRects) == 128,
+    "KVMFRFrame damage rectangles must be cache-line aligned");
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
+_Static_assert(offsetof(KVMFRFrame, captureTime) == 64,
+    "KVMFRFrame hot fields must fit in one cache line");
+_Static_assert(offsetof(KVMFRFrame, damageRects) == 128,
+    "KVMFRFrame damage rectangles must be cache-line aligned");
+#endif
+
 typedef struct KVMFRMessage
 {
   KVMFRMessageType type;

host/include/interface/capture.h

@@ -97,6 +97,12 @@ typedef struct CaptureFrame
 
   uint32_t        damageRectsCount;
   FrameDamageRect damageRects[KVMFR_MAX_DAMAGE_RECTS];
+
+  // Producer-local durations. Absolute timestamps must not cross KVMFR as the
+  // client and producer use different monotonic clock domains.
+  uint64_t        captureTime;
+  uint64_t        postProcessTime;
+  uint64_t        copyTime;
 }
 CaptureFrame;
 
@@ -146,8 +152,9 @@ typedef struct CaptureInterface
     CaptureFrame * frame,
     const size_t maxFrameSize);
   CaptureResult (*getFrame  )(
-    unsigned frameBufferIndex,
+    unsigned       frameBufferIndex,
     FrameBuffer  * frame,
-    const size_t maxFrameSize);
+    const size_t   maxFrameSize,
+    CaptureFrame * captureFrame);
 }
 CaptureInterface;

host/platform/Linux/capture/XCB/src/xcb.c

@@ -299,10 +299,12 @@ static CaptureResult xcb_waitFrame(
 }
 
 static CaptureResult xcb_getFrame(
-  unsigned frameBufferIndex,
+  unsigned       frameBufferIndex,
   FrameBuffer  * frame,
-  const size_t maxFrameSize)
+  const size_t   maxFrameSize,
+  CaptureFrame * captureFrame)
 {
+  (void)captureFrame;
   DEBUG_ASSERT(this);
   DEBUG_ASSERT(this->initialized);
 

host/platform/Linux/capture/pipewire/src/pipewire.c

@@ -463,10 +463,12 @@ static CaptureResult pipewire_waitFrame(
 }
 
 static CaptureResult pipewire_getFrame(
-  unsigned frameBufferIndex,
+  unsigned       frameBufferIndex,
   FrameBuffer  * frame,
-  const size_t maxFrameSize)
+  const size_t   maxFrameSize,
+  CaptureFrame * captureFrame)
 {
+  (void)captureFrame;
   if (this->stop || !this->frameData)
     return CAPTURE_RESULT_REINIT;
 

host/platform/Windows/capture/D12/d12.c

@@ -29,6 +29,7 @@
 #include "common/rects.h"
 #include "common/vector.h"
 #include "common/display.h"
+#include "common/time.h"
 #include "com_ref.h"
 
 #include "backend.h"
@@ -771,8 +772,11 @@ exit:
   return result;
 }
 
-static CaptureResult d12_getFrame(unsigned frameBufferIndex,
+static CaptureResult d12_getFrame(
-  FrameBuffer * frameBuffer, const size_t maxFrameSize)
+  unsigned       frameBufferIndex,
+  FrameBuffer  * frameBuffer,
+  const size_t   maxFrameSize,
+  CaptureFrame * captureFrame)
 {
   CaptureResult result = CAPTURE_RESULT_ERROR;
   comRef_scopePush(3);
@@ -808,6 +812,7 @@ static CaptureResult d12_getFrame(unsigned frameBufferIndex,
   if (result != CAPTURE_RESULT_OK)
     goto exit;
 
+  const uint64_t postProcessStart = nanotime();
   ID3D12Resource * next = *src;
   D12Effect * effect;
   vector_forEach(effect, &this->effects)
@@ -934,6 +939,8 @@ static CaptureResult d12_getFrame(unsigned frameBufferIndex,
     ID3D12CommandQueue_Wait(*this->copyQueue,
       *this->computeCommand.fence, this->computeCommand.fenceValue);
   }
+  captureFrame->postProcessTime = this->effectsActive ?
+    nanotime() - postProcessStart : 0;
 
   // execute the copy commands
   DEBUG_TRACE("Execute copy commands");

host/platform/Windows/capture/DXGI/src/dxgi.c

@@ -1504,9 +1504,13 @@ static CaptureResult dxgi_waitFrame(unsigned frameBufferIndex,
   return CAPTURE_RESULT_OK;
 }
 
-static CaptureResult dxgi_getFrame(unsigned frameBufferIndex,
+static CaptureResult dxgi_getFrame(
-  FrameBuffer * frame, const size_t maxFrameSize)
+  unsigned       frameBufferIndex,
+  FrameBuffer  * frame,
+  const size_t   maxFrameSize,
+  CaptureFrame * captureFrame)
 {
+  (void)captureFrame;
   DEBUG_ASSERT(this);
   DEBUG_ASSERT(this->initialized);
 

host/platform/Windows/capture/NVFBC/src/nvfbc.c

@@ -718,9 +718,13 @@ static CaptureResult nvfbc_waitFrame(unsigned frameBufferIndex,
   return CAPTURE_RESULT_OK;
 }
 
-static CaptureResult nvfbc_getFrame(unsigned frameBufferIndex,
+static CaptureResult nvfbc_getFrame(
-    FrameBuffer * frame, const size_t maxFrameSize)
+  unsigned       frameBufferIndex,
+  FrameBuffer  * frame,
+  const size_t   maxFrameSize,
+  CaptureFrame * captureFrame)
 {
+  (void)captureFrame;
   const unsigned int h = DIFF_MAP_DIM(this->grabHeight, this->diffShift);
   const unsigned int w = DIFF_MAP_DIM(this->grabWidth,  this->diffShift);
   uint8_t * frameData = framebuffer_get_data(frame);

host/src/app.c

@@ -109,6 +109,7 @@ struct app
   bool           frameValid;
   uint32_t       frameSerial;
   uint32_t       formatVer;
+  uint64_t       captureTime[LGMP_Q_FRAME_LEN];
   unsigned int   captureFormatVer;
   bool           hdr;
   bool           hdrPQ;
@@ -291,6 +292,7 @@ static bool sendFrame(CaptureResult result, bool * restart)
   }
 
   KVMFRFrame * fi = app.frame[app.captureIndex];
+  frame.captureTime = app.captureTime[app.captureIndex];
   const uint32_t sdrWhiteLevel = frame.sdrWhiteLevel ?
     frame.sdrWhiteLevel : KVMFR_SDR_WHITE_LEVEL_DEFAULT;
   const bool metadataChanged =
@@ -396,6 +398,11 @@ static bool sendFrame(CaptureResult result, bool * restart)
   // fi->offset is initialized at startup
   fi->flags             = flags;
   fi->sdrWhiteLevel     = sdrWhiteLevel;
+  fi->captureTime       = frame.captureTime;
+  fi->postProcessTime   = 0;
+  fi->copyTime          = 0;
+  fi->timingSerial      = 0;
+  __atomic_store_n(&fi->timingValid, 0, __ATOMIC_RELAXED);
   if (frame.hdrMetadata)
   {
     memcpy(fi->hdrDisplayPrimary, frame.hdrDisplayPrimary,
@@ -433,10 +440,20 @@ static bool sendFrame(CaptureResult result, bool * restart)
     return true;
   }
 
+  const uint64_t copyStart = nanotime();
   app.iface->getFrame(
     app.captureIndex,
     app.frameBuffer[app.captureIndex],
-    app.maxFrameSize);
+    app.maxFrameSize,
+    &frame);
+  const uint64_t copyEnd = nanotime();
+
+  fi->postProcessTime = frame.postProcessTime;
+  frame.copyTime      = copyEnd - copyStart > frame.postProcessTime ?
+    copyEnd - copyStart - frame.postProcessTime : 0;
+  fi->copyTime        = frame.copyTime;
+  fi->timingSerial    = fi->frameSerial;
+  __atomic_store_n(&fi->timingValid, 1, __ATOMIC_RELEASE);
 
   app.readIndex = app.captureIndex;
   if (++app.captureIndex == LGMP_Q_FRAME_LEN)
@@ -1108,13 +1125,15 @@ int app_main(int argc, char * argv[])
             nsleep(us * 1000);
         }
 
-        const uint64_t captureStartTime = microtime();
+        const uint64_t captureStartTime = nanotime();
 
         const CaptureResult result = app.iface->capture(
           app.captureIndex, app.frameBuffer[app.captureIndex]);
 
+        app.captureTime[app.captureIndex] = nanotime() - captureStartTime;
+
         if (likely(result == CAPTURE_RESULT_OK))
-          previousFrameTime = captureStartTime;
+          previousFrameTime = captureStartTime / 1000;
         else if (likely(result == CAPTURE_RESULT_TIMEOUT))
         {
           if (!app.iface->asyncCapture)

idd/LGIdd/CFrameBufferResource.h

@@ -33,13 +33,16 @@ using namespace Microsoft::WRL;
 class CFrameBufferResource
 {
   private:
-    bool                   m_valid      = false;
+    bool                   m_valid           = false;
-    unsigned               m_frameIndex = 0;
+    unsigned               m_frameIndex      = 0;
-    uint8_t              * m_base       = nullptr;
+    uint8_t              * m_base            = nullptr;
-    size_t                 m_size       = 0;
+    size_t                 m_size            = 0;
-    size_t                 m_frameSize  = 0;
+    size_t                 m_frameSize       = 0;
+    uint64_t               m_captureTime     = 0;
+    uint64_t               m_postProcessTime = 0;
+    uint64_t               m_copyStart       = 0;
     ComPtr<ID3D12Resource> m_res;
-    void                 * m_map        = nullptr;
+    void                 * m_map             = nullptr;
 
   public:
     bool Init(CSwapChainProcessor * swapChain, unsigned frameIndex, uint8_t * base, size_t size);
@@ -52,5 +55,16 @@ class CFrameBufferResource
     size_t    GetFrameSize()  { return m_frameSize;  }
     void *    GetMap()        { return m_map;        }
 
+    void SetTiming(uint64_t captureTime, uint64_t postProcessTime,
+      uint64_t copyStart)
+    {
+      m_captureTime     = captureTime;
+      m_postProcessTime = postProcessTime;
+      m_copyStart       = copyStart;
+    }
+    uint64_t GetCaptureTime    () const { return m_captureTime;     }
+    uint64_t GetPostProcessTime() const { return m_postProcessTime; }
+    uint64_t GetCopyStart      () const { return m_copyStart;       }
+
     ComPtr<ID3D12Resource> Get() { return m_res; }
 };

idd/LGIdd/CIndirectDeviceContext.cpp

@@ -212,7 +212,8 @@ void CIndirectDeviceContext::InitAdapter()
     return;
   }
 
-  if (InterlockedCompareExchange(&m_initInProgress, 1, 0) != 0)
+  LONG initExpected = 0;
+  if (!m_initInProgress.compare_exchange_strong(initExpected, 1))
   {
     DEBUG_TRACE("Adapter initialization skipped: initialization already in progress");
     return;
@@ -227,7 +228,7 @@ void CIndirectDeviceContext::InitAdapter()
     {
       DEBUG_WARN("IVSHMEM not available yet, scheduling init retry");
       ScheduleInitRetry();
-      InterlockedExchange(&m_initInProgress, 0);
+      m_initInProgress.store(0);
       return;
     }
     m_ivshmemOpened = true;
@@ -370,7 +371,7 @@ void CIndirectDeviceContext::InitAdapter()
   if (!NT_SUCCESS(status))
   {
     DEBUG_ERROR_HR(status, "IddCxAdapterInitAsync Failed");
-    InterlockedExchange(&m_initInProgress, 0);
+    m_initInProgress.store(0);
     return;
   }
 
@@ -378,7 +379,7 @@ void CIndirectDeviceContext::InitAdapter()
   if (!m_adapter)
   {
     DEBUG_ERROR("IddCxAdapterInitAsync succeeded without returning an adapter object");
-    InterlockedExchange(&m_initInProgress, 0);
+    m_initInProgress.store(0);
     return;
   }
 
@@ -401,7 +402,7 @@ void CIndirectDeviceContext::InitAdapter()
 
   // Adapter is up; no need to keep retrying.
   StopInitRetry();
-  InterlockedExchange(&m_initInProgress, 0);
+  m_initInProgress.store(0);
   DEBUG_INFO("Adapter initialization request complete; returning to IddCx");
 }
 
@@ -501,7 +502,7 @@ void CIndirectDeviceContext::ReplugMonitor()
     ReleaseSRWLockExclusive(&m_stateLock);
     // Either no monitor yet, or one is already pending; build it now and
     // cancel any queued rebuild so we do not create two.
-    InterlockedExchange(&m_finishInitQueued, 0);
+    m_finishInitQueued.store(0);
     FinishInit(0);
     return;
   }
@@ -537,7 +538,7 @@ void CIndirectDeviceContext::ReplugMonitor()
   // If there was no swap chain there will be no unassign callback to queue the
   // rebuild. Otherwise OnSwapChainReleased does so after teardown has drained.
   if (rebuild)
-    InterlockedExchange(&m_finishInitQueued, 1);
+    m_finishInitQueued.store(1);
 }
 
 void CIndirectDeviceContext::OnMonitorDestroyed(IDDCX_MONITOR monitor)
@@ -571,7 +572,7 @@ void CIndirectDeviceContext::OnSwapChainReleased()
   ReleaseSRWLockExclusive(&m_stateLock);
 
   if (rebuild)
-    InterlockedExchange(&m_finishInitQueued, 1);
+    m_finishInitQueued.store(1);
 }
 
 void CIndirectDeviceContext::OnSwapChainReady()
@@ -613,7 +614,7 @@ void CIndirectDeviceContext::OnSwapChainReady()
   g_pipe.SetDeviceContext(this);
 
   if (replug)
-    InterlockedExchange(&m_replugQueued, 1);
+    m_replugQueued.store(1);
   else if (doSetMode)
     g_pipe.SetDisplayMode(mode.width, mode.height, mode.refresh);
 }
@@ -1100,7 +1101,7 @@ bool CIndirectDeviceContext::SetupLGMP(size_t alignSize)
 
 void CIndirectDeviceContext::DeInitLGMP()
 {
-  InterlockedExchange(&m_publishedFrameIndex, -1);
+  m_publishedFrameIndex.store(-1);
 
   // The retry timer callback dereferences this context, so make sure it is
   // stopped and drained before we tear anything down. Wait for any in-flight
@@ -1135,13 +1136,13 @@ void CIndirectDeviceContext::DeInitLGMP()
 void CIndirectDeviceContext::LGMPTimer()
 {
   // Rebuild the monitor queued by ReplugMonitor, off the IddCx callback thread.
-  if (InterlockedExchange(&m_finishInitQueued, 0))
+  if (m_finishInitQueued.exchange(0))
   {
     FinishInit(0);
     return;
   }
 
-  if (InterlockedExchange(&m_replugQueued, 0))
+  if (m_replugQueued.exchange(0))
   {
     ReplugMonitor();
     return;
@@ -1188,8 +1189,7 @@ void CIndirectDeviceContext::LGMPTimer()
 
   if (lgmpHostQueueNewSubs(m_frameQueue) && m_monitor)
   {
-    const LONG frameIndex =
+    const LONG frameIndex = m_publishedFrameIndex.load();
-      InterlockedCompareExchange(&m_publishedFrameIndex, 0, 0);
     if (frameIndex >= 0)
       lgmpHostQueuePost(m_frameQueue, 0, m_frameMemory[frameIndex]);
   }
@@ -1299,6 +1299,11 @@ CIndirectDeviceContext::PreparedFrameBuffer CIndirectDeviceContext::PrepareFrame
   // fi->offset is initialized at startup
   fi->flags            = flags;
   fi->sdrWhiteLevel    = dstFormat.sdrWhiteLevel;
+  fi->captureTime      = 0;
+  fi->postProcessTime  = 0;
+  fi->copyTime         = 0;
+  fi->timingSerial     = 0;
+  InterlockedExchange((volatile LONG *)&fi->timingValid, 0);
   fi->rotation         = FRAME_ROT_0;
   fi->type             = dstFormat.format;
 
@@ -1351,7 +1356,7 @@ bool CIndirectDeviceContext::PublishFrameBuffer(unsigned frameIndex)
   /* Make resends select this submitted frame before posting it. This prevents
    * a new subscriber racing publication from receiving the previous frame
    * after the new one. */
-  InterlockedExchange(&m_publishedFrameIndex, (LONG)frameIndex);
+  m_publishedFrameIndex.store(static_cast<LONG>(frameIndex));
 
   const LGMP_STATUS status =
     lgmpHostQueuePost(m_frameQueue, 0, m_frameMemory[frameIndex]);
@@ -1364,6 +1369,20 @@ bool CIndirectDeviceContext::PublishFrameBuffer(unsigned frameIndex)
   return true;
 }
 
+void CIndirectDeviceContext::SetFrameTiming(unsigned frameIndex,
+  uint64_t captureTime, uint64_t postProcessTime, uint64_t copyTime)
+{
+  if (frameIndex >= LGMP_Q_FRAME_LEN)
+    return;
+
+  KVMFRFrame * frame      = m_frame[frameIndex];
+  frame->captureTime      = captureTime;
+  frame->postProcessTime  = postProcessTime;
+  frame->copyTime         = copyTime;
+  frame->timingSerial     = frame->frameSerial;
+  InterlockedExchange((volatile LONG *)&frame->timingValid, 1);
+}
+
 void CIndirectDeviceContext::WriteFrameBuffer(unsigned frameIndex, void* src, size_t offset, size_t len, bool setWritePos) const
 {
   FrameBuffer * fb = m_frameBuffer[frameIndex];

idd/LGIdd/CIndirectDeviceContext.h

@@ -23,6 +23,7 @@
 #include <Windows.h>
 #include <wdf.h>
 #include <IddCx.h>
+#include <atomic>
 #include <vector>
 
 #include "CIVSHMEM.h"
@@ -78,9 +79,9 @@ private:
 
   // Retry state for InitAdapter. At boot the IVSHMEM device may not have
   // enumerated yet; if so we re-attempt from a timer instead of giving up.
-  WDFTIMER      m_initTimer     = nullptr;
+  WDFTIMER          m_initTimer       = nullptr;
-  bool          m_ivshmemOpened = false;
+  bool              m_ivshmemOpened   = false;
-  volatile LONG m_initInProgress = 0;
+  std::atomic<LONG> m_initInProgress  = 0;
 
   CIVSHMEM m_ivshmem;
 
@@ -99,16 +100,16 @@ private:
   bool m_cursorVisible = false;
   int m_cursorX = 0, m_cursorY = 0;
 
-  size_t         m_alignSize           = 0;
+  size_t            m_alignSize           = 0;
-  size_t         m_frameMemoryOffset   = 0;
+  size_t            m_frameMemoryOffset   = 0;
-  size_t         m_maxFrameSize        = 0;
+  size_t            m_maxFrameSize        = 0;
-  int            m_frameIndex          = 0;
+  int               m_frameIndex          = 0;
-  volatile LONG  m_publishedFrameIndex = -1;
+  std::atomic<LONG> m_publishedFrameIndex = -1;
-  uint32_t       m_formatVer           = 0;
+  uint32_t          m_formatVer           = 0;
-  uint32_t       m_frameSerial         = 0;
+  uint32_t          m_frameSerial         = 0;
-  PLGMPMemory    m_frameMemory[LGMP_Q_FRAME_LEN] = {};
+  PLGMPMemory       m_frameMemory[LGMP_Q_FRAME_LEN] = {};
-  KVMFRFrame   * m_frame      [LGMP_Q_FRAME_LEN] = {};
+  KVMFRFrame      * m_frame      [LGMP_Q_FRAME_LEN] = {};
-  FrameBuffer  * m_frameBuffer[LGMP_Q_FRAME_LEN] = {};
+  FrameBuffer     * m_frameBuffer[LGMP_Q_FRAME_LEN] = {};
 
   unsigned    m_width    = 0;
   unsigned    m_height   = 0;
@@ -164,8 +165,8 @@ private:
 
   // Set by ReplugMonitor after a departure to rebuild the monitor from the LGMP
   // timer, off the IddCx callback thread.
-  volatile LONG m_finishInitQueued = 0;
+  std::atomic<LONG> m_finishInitQueued = 0;
-  volatile LONG m_replugQueued     = 0;
+  std::atomic<LONG> m_replugQueued     = 0;
 
 public:
   CIndirectDeviceContext(_In_ WDFDEVICE wdfDevice) :
@@ -216,6 +217,8 @@ public:
   bool FrameBufferAvailable() const;
   PreparedFrameBuffer PrepareFrameBuffer(unsigned pitch, const D12FrameFormat& srcFormat, const D12FrameFormat& dstFormat, const RECT * dirtyRects, unsigned nbDirtyRects);
   bool PublishFrameBuffer(unsigned frameIndex);
+  void SetFrameTiming(unsigned frameIndex, uint64_t captureTime,
+    uint64_t postProcessTime, uint64_t copyTime);
   void WriteFrameBuffer(unsigned frameIndex, void* src, size_t offset, size_t len, bool setWritePos) const;
   void FinalizeFrameBuffer(unsigned frameIndex) const;
 

idd/LGIdd/CSwapChainProcessor.cpp

@@ -28,6 +28,22 @@
 static const uint32_t HDR_PQ_MIN_LUMINANCE = 50;
 static const uint32_t HDR_PQ_MAX_LUMINANCE = 10000;
 
+static uint64_t Nanotime()
+{
+  static const uint64_t frequency = []()
+  {
+    LARGE_INTEGER value;
+    QueryPerformanceFrequency(&value);
+    return (uint64_t)value.QuadPart;
+  }();
+
+  LARGE_INTEGER counter;
+  QueryPerformanceCounter(&counter);
+  const uint64_t ticks = (uint64_t)counter.QuadPart;
+  return ticks / frequency * 1000000000ULL +
+    ticks % frequency * 1000000000ULL / frequency;
+}
+
 CSwapChainProcessor::CSwapChainProcessor(CIndirectMonitorContext * monitorContext,
     UINT64 assignmentGeneration, IDDCX_MONITOR monitor,
     CIndirectDeviceContext* devContext, IDDCX_SWAPCHAIN hSwapChain,
@@ -192,6 +208,7 @@ bool CSwapChainProcessor::SwapChainThreadCore()
     // path HDR is not available, so default to SDR.
     DXGI_COLOR_SPACE_TYPE colorSpace = DXGI_COLOR_SPACE_RGB_FULL_G22_NONE_P709;
     UINT sdrWhiteLevel = KVMFR_SDR_WHITE_LEVEL_DEFAULT;
+    const uint64_t captureStart = Nanotime();
 
 #ifdef HAS_IDDCX_110
     if (m_devContext->HasIddCx110DDIs())
@@ -252,7 +269,8 @@ bool CSwapChainProcessor::SwapChainThreadCore()
       if (frameNumber != lastFrameNumber)
       {
         lastFrameNumber = frameNumber;
-        if (!SwapChainNewFrame(surface, dirtyRectCount, colorSpace, sdrWhiteLevel))
+        if (!SwapChainNewFrame(surface, dirtyRectCount, colorSpace,
+              sdrWhiteLevel, Nanotime() - captureStart))
           DEBUG_WARN("Failed to submit frame");
       }
 
@@ -286,16 +304,21 @@ void CSwapChainProcessor::CompletionFunction(
   // fail gracefully
   if (!result)
   {
+    sc->m_devContext->SetFrameTiming(fbRes->GetFrameIndex(),
+      fbRes->GetCaptureTime(), fbRes->GetPostProcessTime(),
+      Nanotime() - fbRes->GetCopyStart());
     sc->m_devContext->FinalizeFrameBuffer(fbRes->GetFrameIndex());
     return;
   }
 
   if (sc->m_dx12Device->IsIndirectCopy())
     sc->m_devContext->WriteFrameBuffer(
-      fbRes->GetFrameIndex(),
+      fbRes->GetFrameIndex(), fbRes->GetMap(), 0, fbRes->GetFrameSize(), false);
-      fbRes->GetMap(), 0, fbRes->GetFrameSize(), true);
+
-  else
+  const uint64_t copyTime = Nanotime() - fbRes->GetCopyStart();
-    sc->m_devContext->FinalizeFrameBuffer(fbRes->GetFrameIndex());
+  sc->m_devContext->SetFrameTiming(fbRes->GetFrameIndex(),
+    fbRes->GetCaptureTime(), fbRes->GetPostProcessTime(), copyTime);
+  sc->m_devContext->FinalizeFrameBuffer(fbRes->GetFrameIndex());
 }
 
 
@@ -470,8 +493,9 @@ bool CSwapChainProcessor::GetContentHDRMetadata(D12FrameFormat& format) const
 #endif
 }
 
-bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer, unsigned dirtyRectCount,
+bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer,
-  DXGI_COLOR_SPACE_TYPE colorSpace, UINT sdrWhiteLevel)
+  unsigned dirtyRectCount, DXGI_COLOR_SPACE_TYPE colorSpace,
+  UINT sdrWhiteLevel, uint64_t captureTime)
 {
   // Preserve the fast drop path: never hold an IddCx frame while waiting for
   // a slow or disconnected client. We have not read its rectangles, so force
@@ -482,6 +506,8 @@ bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer
     return true;
   }
 
+  const uint64_t postProcessStart = Nanotime();
+
   ComPtr<ID3D11Texture2D> texture;
   HRESULT hr = acquiredBuffer.As(&texture);
   if (FAILED(hr))
@@ -708,6 +734,9 @@ bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer
     return false;
   }
 
+  const uint64_t copyStart = Nanotime();
+  fbRes->SetTiming(captureTime, copyStart - postProcessStart, copyStart);
+
   copyQueue->SetCompletionCallback(&CompletionFunction, this, fbRes);
 
   D3D12_TEXTURE_COPY_LOCATION srcLoc = {};

idd/LGIdd/CSwapChainProcessor.h

@@ -102,7 +102,7 @@ private:
 #endif
   bool GetContentHDRMetadata(D12FrameFormat& format) const;
   bool SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer, unsigned dirtyRectCount,
-    DXGI_COLOR_SPACE_TYPE colorSpace, UINT sdrWhiteLevel);
+    DXGI_COLOR_SPACE_TYPE colorSpace, UINT sdrWhiteLevel, uint64_t captureTime);
 
 public:
   CSwapChainProcessor(CIndirectMonitorContext * monitorContext, UINT64 assignmentGeneration,