decklink: Pass frames between threads using queues

Eliminates the stalls that have been seen when waiting for frame copies.

We shouldn't be copying at all ideally, but that would break the API.

plugins/decklink/decklink-device-instance.cpp

@@ -613,24 +613,31 @@ bool DeckLinkDeviceInstance::StartOutput(DeckLinkDeviceMode *mode_)
 		}
 	}
 
-	frameData.clear();
+	frameQueueDecklinkToObs.reset();
+	frameQueueObsToDecklink.reset();
+
+	const int rowSize = decklinkOutput->GetWidth() * 4;
+	const int frameSize = rowSize * decklinkOutput->GetHeight();
+	for (std::vector<uint8_t> &blob : frameBlobs) {
+		blob.assign(frameSize, 0);
+		frameQueueDecklinkToObs.push(blob.data());
+	}
+	activeBlob = nullptr;
+
 	const int64_t minimumPrerollFrames =
 		std::max(device->GetMinimumPrerollFrames(), INT64_C(3));
 	for (int64_t i = 0; i < minimumPrerollFrames; ++i) {
 		ComPtr<IDeckLinkMutableVideoFrame> decklinkOutputFrame;
 		HRESULT result = output_->CreateVideoFrame(
 			decklinkOutput->GetWidth(), decklinkOutput->GetHeight(),
-			decklinkOutput->GetWidth() * 4, bmdFormat8BitBGRA,
-			bmdFrameFlagDefault, &decklinkOutputFrame);
+			rowSize, bmdFormat8BitBGRA, bmdFrameFlagDefault,
+			&decklinkOutputFrame);
 		if (result != S_OK) {
 			blog(LOG_ERROR, "failed to create video frame 0x%X",
 			     result);
 			return false;
 		}
 
-		const long size = decklinkOutputFrame->GetRowBytes() *
-				  decklinkOutputFrame->GetHeight();
-		frameData.resize(size);
 		result = output_->ScheduleVideoFrame(decklinkOutputFrame,
 						     i * frameDuration,
 						     frameDuration,
@@ -669,6 +676,8 @@ bool DeckLinkDeviceInstance::StopOutput()
 	output->DisableAudioOutput();
 	output.Clear();
 	renderDelegate.Clear();
+	frameQueueDecklinkToObs.reset();
+	frameQueueObsToDecklink.reset();
 
 	return true;
 }
@@ -679,26 +688,36 @@ void DeckLinkDeviceInstance::UpdateVideoFrame(video_data *frame)
 	if (decklinkOutput == nullptr)
 		return;
 
-	std::lock_guard lock(frameDataMutex);
-	const uint8_t *const outData = frame->data[0];
-	frameData.assign(outData,
-			 outData + decklinkOutput->GetWidth() *
-					   decklinkOutput->GetHeight() * 4);
+	uint8_t *const blob = frameQueueDecklinkToObs.pop();
+	if (blob) {
+		memcpy(blob, frame->data[0],
+		       frame->linesize[0] * decklinkOutput->GetHeight());
+		frameQueueObsToDecklink.push(blob);
+	}
 }
 
 void DeckLinkDeviceInstance::ScheduleVideoFrame(IDeckLinkVideoFrame *frame)
 {
 	void *bytes;
 	if (SUCCEEDED(frame->GetBytes(&bytes))) {
-		{
-			std::lock_guard lock(frameDataMutex);
-			memcpy(bytes, frameData.data(),
-			       frame->GetRowBytes() * frame->GetHeight());
+		uint8_t *blob = frameQueueObsToDecklink.pop();
+		if (blob) {
+			if (activeBlob)
+				frameQueueDecklinkToObs.push(activeBlob);
+			activeBlob = blob;
+		} else {
+			blob = activeBlob;
 		}
 
-		output->ScheduleVideoFrame(
-			frame, (totalFramesScheduled * frameDuration),
-			frameDuration, frameTimescale);
+		const int frameSize = frame->GetRowBytes() * frame->GetHeight();
+		if (blob)
+			memcpy(bytes, blob, frameSize);
+		else
+			memset(bytes, 0, frameSize);
+
+		output->ScheduleVideoFrame(frame,
+					   totalFramesScheduled * frameDuration,
+					   frameDuration, frameTimescale);
 		++totalFramesScheduled;
 	}
 }

plugins/decklink/decklink-device-instance.hpp

@@ -8,7 +8,6 @@
 #include "decklink-device.hpp"
 #include "OBSVideoFrame.h"
 #include <atomic>
-#include <mutex>
 #include <vector>
 
 class AudioRepacker;
@@ -38,6 +37,104 @@ public:
 	virtual HRESULT STDMETHODCALLTYPE ScheduledPlaybackHasStopped();
 };
 
+/*
+Unbounded SPSC Queue with modifications:
+https://www.1024cores.net/home/lock-free-algorithms/queues/unbounded-spsc-queue
+- Convert to bounded. Fixed node cache is part of the class layout.
+- Queue doesn't handle push failure because it should never be full.
+- Templated type has been replaced with a hard-coded type.
+The license text has been copied below.
+Copyright (c) 2010-2011 Dmitry Vyukov. All rights reserved.
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+   1. Redistributions of source code must retain the above copyright notice,
+      this list of conditions and the following disclaimer.
+   2. Redistributions in binary form must reproduce the above copyright notice,
+      this list of conditions and the following disclaimer in the documentation
+      and/or other materials provided with the distribution.
+THIS SOFTWARE IS PROVIDED BY DMITRY VYUKOV "AS IS" AND ANY EXPRESS OR IMPLIED
+WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+SHALL DMITRY VYUKOV OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
+ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+The views and conclusions contained in the software and documentation are those
+of the authors and should not be interpreted as representing official policies,
+either expressed or implied, of Dmitry Vyukov.
+*/
+static constexpr size_t FrameQueueFrameCount = 3;
+class FrameQueue {
+private:
+	/* Intel may grab two 64-byte cache lines */
+	static constexpr size_t FalseSharingSize = 128;
+
+	struct Node {
+		std::atomic<Node *> next = nullptr;
+		uint8_t *frame = nullptr;
+	};
+
+	struct alignas(FalseSharingSize) PaddedNode {
+		Node node;
+		uint8_t padding[FalseSharingSize - sizeof(struct Node)]{};
+	};
+
+	PaddedNode cache[FrameQueueFrameCount + 1];
+
+	alignas(FalseSharingSize) Node *front;
+	alignas(FalseSharingSize) Node *back;
+	Node *cache_list;
+
+public:
+	FrameQueue() { reset(); }
+
+	void reset()
+	{
+		for (size_t i = 0; i < FrameQueueFrameCount; ++i) {
+			cache[i].node.next.store(&cache[i + 1].node,
+						 std::memory_order_relaxed);
+		}
+
+		Node &last = cache[FrameQueueFrameCount].node;
+		last.next.store(nullptr, std::memory_order_relaxed);
+		last.frame = nullptr;
+
+		front = &last;
+		back = &last;
+		cache_list = &cache[0].node;
+	}
+
+	void push(uint8_t *v)
+	{
+		Node *const n = cache_list;
+		cache_list = cache_list->next.load(std::memory_order_relaxed);
+
+		n->next.store(nullptr, std::memory_order_relaxed);
+		n->frame = v;
+
+		back->next.store(n, std::memory_order_release);
+
+		back = n;
+	}
+
+	uint8_t *pop()
+	{
+		uint8_t *frame = nullptr;
+
+		Node *const n_front =
+			front->next.load(std::memory_order_consume);
+		if (n_front != nullptr) {
+			frame = n_front->frame;
+			front = n_front;
+		}
+
+		return frame;
+	}
+};
+
 class DeckLinkDeviceInstance : public IDeckLinkInputCallback {
 protected:
 	ComPtr<IDeckLinkConfiguration> deckLinkConfiguration;
@@ -66,8 +163,10 @@ protected:
 	bool allow10Bit;
 
 	OBSVideoFrame *convertFrame = nullptr;
-	std::mutex frameDataMutex;
-	std::vector<uint8_t> frameData;
+	std::vector<uint8_t> frameBlobs[FrameQueueFrameCount];
+	FrameQueue frameQueueObsToDecklink;
+	FrameQueue frameQueueDecklinkToObs;
+	uint8_t *activeBlob = nullptr;
 	BMDTimeValue frameDuration;
 	BMDTimeScale frameTimescale;
 	BMDTimeScale totalFramesScheduled;