Add Metal swapchain presentation options

Engine/gapi/abstractgraphicsapi.cpp

@@ -131,3 +131,8 @@ std::shared_ptr<AbstractGraphicsApi::Fence> AbstractGraphicsApi::submit(Device*
 NonUniqResId AbstractGraphicsApi::Swapchain::syncId() const {
   return NonUniqResId(0x1);
   }
+
+AbstractGraphicsApi::Swapchain* AbstractGraphicsApi::createSwapchain(SystemApi::Window* w, Device* d,
+                                                                   const Swapchain::Options&) {
+  return createSwapchain(w,d);
+  }

Engine/gapi/abstractgraphicsapi.h

@@ -520,6 +520,16 @@ namespace Tempest {
         virtual void setPayload(std::vector<Detail::DSharedPtr<const AbstractGraphicsApi::Shared*>>&&) {}
         };
       struct Swapchain:NoCopy {
+        enum class RenderMode:uint8_t {
+          Copy,
+          Direct,
+          };
+
+        struct Options {
+          uint32_t   bufferCount = 0;
+          RenderMode renderMode  = RenderMode::Copy;
+          };
+
         virtual ~Swapchain()=default;
         virtual void          reset()=0;
         virtual uint32_t      currentBackBufferIndex()=0;
@@ -636,6 +646,7 @@ namespace Tempest {
       virtual Device*    createDevice(std::string_view gpuName) = 0;
 
       virtual Swapchain* createSwapchain(SystemApi::Window* w,AbstractGraphicsApi::Device *d) = 0;
+      virtual Swapchain* createSwapchain(SystemApi::Window* w,Device* d,const Swapchain::Options& options);
 
       virtual PPipeline  createPipeline(Device* d, const RenderState &st, Topology tp, const Shader* const* sh, size_t cnt)=0;
 

Engine/gapi/metal/mtcommandbuffer.cpp

@@ -11,6 +11,8 @@
 #include "mtswapchain.h"
 #include "mtaccelerationstructure.h"
 
+#include <algorithm>
+
 using namespace Tempest;
 using namespace Tempest::Detail;
 
@@ -65,6 +67,7 @@ void MtCommandBuffer::end() {
   }
 
 void MtCommandBuffer::reset() {
+  swapchainFrames.clear();
   auto pool = NsPtr<NS::AutoreleasePool>::init();
   auto desc = NsPtr<MTL::CommandBufferDescriptor>::init();
   desc->setRetainedReferences(false);
@@ -94,7 +97,10 @@ void MtCommandBuffer::beginRendering(const FrameBufferDesc& fbo, size_t fboSize,
     auto clr = desc->colorAttachments()->object(i);
     if(fbo.sw[i]!=nullptr) {
       auto& s = *reinterpret_cast<MtSwapchain*>(fbo.sw[i]);
-      clr->setTexture(s.img[fbo.imgId[i]].tex.get());
+      auto frame = s.acquireFrame(fbo.imgId[i]);
+      clr->setTexture(frame==nullptr ? s.img[fbo.imgId[i]].tex.get() : frame->texture.get());
+      if(frame!=nullptr && std::find(swapchainFrames.begin(),swapchainFrames.end(),frame)==swapchainFrames.end())
+        swapchainFrames.push_back(std::move(frame));
       curFbo.colorFormat[curFbo.numColors] = s.format();
       } else {
       auto& t = *reinterpret_cast<MtTexture*>(fbo.att[i]);

Engine/gapi/metal/mtcommandbuffer.h

@@ -20,6 +20,7 @@ class MtPipeline;
 class MtCompPipeline;
 class MtDescriptorArray;
 class MtTopAccelerationStructure;
+struct MtSwapchainFrame;
 
 class MtCommandBuffer : public AbstractGraphicsApi::CommandBuffer {
   public:
@@ -113,6 +114,7 @@ class MtCommandBuffer : public AbstractGraphicsApi::CommandBuffer {
     NsPtr<MTL::BlitCommandEncoder>    encBlit;
 
     std::vector<const void*>          usedResources;
+    std::vector<std::shared_ptr<MtSwapchainFrame>> swapchainFrames;
 
     MtFboLayout                       curFbo;
     Push                              pushData;

Engine/gapi/metal/mtswapchain.h

@@ -1,10 +1,10 @@
 #pragma once
 
 #include <Tempest/AbstractGraphicsApi>
-#include "utility/spinlock.h"
 #include "nsptr.h"
 
 #include <Metal/Metal.hpp>
+#include <mutex>
 
 namespace CA
 {
@@ -16,9 +16,19 @@ namespace Detail {
 
 class MtDevice;
 
+struct MtSwapchainFrame {
+  MtSwapchainFrame(MTL::Texture* texture, CA::MetalDrawable* drawable);
+  ~MtSwapchainFrame();
+
+  NsPtr<MTL::Texture>      texture;
+  NsPtr<CA::MetalDrawable> drawable;
+  };
+
 class MtSwapchain : public AbstractGraphicsApi::Swapchain {
   public:
-    MtSwapchain(MtDevice& dev, SystemApi::Window* w);
+    using Frame = std::shared_ptr<MtSwapchainFrame>;
+
+    MtSwapchain(MtDevice& dev, SystemApi::Window* w, const Options& options);
     ~MtSwapchain();
 
     void          reset() override;
@@ -30,6 +40,7 @@ class MtSwapchain : public AbstractGraphicsApi::Swapchain {
     NonUniqResId  syncId() const override { return NonUniqResId::I_None; }
 
     MTL::PixelFormat format() const;
+    Frame         acquireFrame(uint32_t image);
 
     struct Image {
       NsPtr<MTL::Texture> tex;
@@ -40,12 +51,14 @@ class MtSwapchain : public AbstractGraphicsApi::Swapchain {
     struct Impl;
     std::unique_ptr<Impl> pimpl;
 
-    SpinLock              sync;
+    std::mutex            sync;
     MtDevice&             dev;
     Tempest::Size         sz;
 
     uint32_t              imgCount   = 0;
     uint32_t              currentImg = 0;
+    bool                  direct = false;
+    Frame                 activeFrame;
 
     NsPtr<MTL::Texture>   mkTexture();
     void                  nextDrawable();

Engine/gapi/metal/mtswapchain.mm

@@ -63,6 +63,17 @@ - (CALayer *)makeBackingLayer {
     }
   };
 
+MtSwapchainFrame::MtSwapchainFrame(MTL::Texture* tex, CA::MetalDrawable* dr)
+  :texture(tex), drawable(dr) {
+  if(tex!=nullptr)
+    tex->retain();
+  if(dr!=nullptr)
+    dr->retain();
+  }
+
+MtSwapchainFrame::~MtSwapchainFrame() {
+  }
+
 static float backingScaleFactor(SysWindow* w) {
 #if defined(__OSX__)
   return [w screen].backingScaleFactor;
@@ -92,8 +103,8 @@ static CGRect windowRect(UIWindow* wnd) {
 #endif
 
 // note : MoltenVK supports NSView, UIView, CAMetalLayer, so we should align to it
-MtSwapchain::MtSwapchain(MtDevice& dev, SystemApi::Window *w)
-  :dev(dev), pimpl(new Impl()) {
+MtSwapchain::MtSwapchain(MtDevice& dev, SystemApi::Window *w, const Options& options)
+  :pimpl(new Impl()), dev(dev), direct(options.renderMode==RenderMode::Direct) {
   NSObject* obj = reinterpret_cast<NSObject*>(w);
   if([obj isKindOfClass : [SysWindow class]])
     pimpl->wnd = reinterpret_cast<SysWindow*>(w);
@@ -117,23 +128,28 @@ static CGRect windowRect(UIWindow* wnd) {
   [lay setContentsScale:dpi];
 #if defined(__IOS__)
   // Swapchain takes too much memory on 2GB iPhone
-  lay.maximumDrawableCount      = 2;
+  lay.maximumDrawableCount      = options.bufferCount==0 ? 2 : options.bufferCount;
+#elif defined(__OSX__)
+  if(options.bufferCount!=0)
+    lay.maximumDrawableCount = options.bufferCount;
 #endif
   lay.pixelFormat               = MTLPixelFormatBGRA8Unorm;
-  lay.allowsNextDrawableTimeout = NO;
+  lay.allowsNextDrawableTimeout = direct ? YES : NO;
   lay.framebufferOnly           = NO;
 
   reset();
   }
 
 MtSwapchain::~MtSwapchain() {
+  dev.waitIdle();
   if(pimpl->view!=nil)
     [pimpl->view release];
   }
 
 void MtSwapchain::reset() {
+  std::lock_guard<std::mutex> guard(sync);
   dev.waitIdle(); // pending commands
-  std::lock_guard<SpinLock> guard(sync);
+  activeFrame.reset();
 
   // https://developer.apple.com/documentation/quartzcore/cametallayer?language=objc
   CAMetalLayer* lay = pimpl->metalLayer();
@@ -146,8 +162,10 @@ static CGRect windowRect(UIWindow* wnd) {
   img.resize(imgCount);
   for(size_t i=0; i<imgCount; ++i)
     img[i].tex = nullptr;
-  for(size_t i=0; i<imgCount; ++i)
-    img[i].tex = mkTexture();
+  if(!direct) {
+    for(size_t i=0; i<imgCount; ++i)
+      img[i].tex = mkTexture();
+    }
 
   currentImg = 0;
   }
@@ -156,54 +174,75 @@ static CGRect windowRect(UIWindow* wnd) {
   return currentImg;
   }
 
+MtSwapchain::Frame MtSwapchain::acquireFrame(uint32_t image) {
+  // Keep the existing Copy path free of per-frame ownership allocations.
+  if(!direct)
+    return nullptr;
+  std::lock_guard<std::mutex> guard(sync);
+  if(image!=currentImg)
+    throw SwapchainSuboptimal();
+  if(activeFrame!=nullptr)
+    return activeFrame;
+
+  auto pool = NsPtr<NS::AutoreleasePool>::init();
+  auto* lay = reinterpret_cast<CA::MetalLayer*>(pimpl->metalLayer());
+  auto* dr  = lay->nextDrawable();
+  if(dr!=nullptr && dr->texture()->width()==size_t(sz.w) && dr->texture()->height()==size_t(sz.h)) {
+    activeFrame = std::make_shared<MtSwapchainFrame>(dr->texture(),dr);
+    } else {
+    // Allocate the private Copy fallback only when drawable acquisition fails.
+    if(img[image].tex==nullptr)
+      img[image].tex = mkTexture();
+    activeFrame = std::make_shared<MtSwapchainFrame>(img[image].tex.get(),nullptr);
+    }
+  return activeFrame;
+  }
+
 void MtSwapchain::present() {
+  std::lock_guard<std::mutex> guard(sync);
   auto pool = NsPtr<NS::AutoreleasePool>::init();
+  auto frame = activeFrame;
+  if(direct && frame==nullptr)
+    throw SwapchainSuboptimal();
 
   CA::MetalLayer* lay      = reinterpret_cast<CA::MetalLayer*>(pimpl->metalLayer());
   uint32_t        i        = currentImg;
-  auto            drawable = lay->nextDrawable();
+  auto            drawable = frame==nullptr ? nullptr : frame->drawable.get();
+  if(drawable==nullptr)
+    drawable = lay->nextDrawable();
   if(drawable==nullptr)
     throw SwapchainSuboptimal();
 
-  std::lock_guard<SpinLock> guard(sync);
+  auto src = frame==nullptr ? img[i].tex.get() : frame->texture.get();
   auto dr = drawable->texture();
-  if(dr->width()!=img[i].tex->width() || dr->height()!=img[i].tex->height()) {
+  if(dr->width()!=src->width() || dr->height()!=src->height()) {
     throw SwapchainSuboptimal();
     }
 
   auto desc = NsPtr<MTL::CommandBufferDescriptor>::init();
-  //desc->setRetainedReferences(true);
+  desc->setRetainedReferences(true);
   desc->setErrorOptions(MTL::CommandBufferErrorOptionEncoderExecutionStatus);
 
   auto cmd = dev.queue->commandBuffer(desc.get());
-  auto enc = cmd->blitCommandEncoder();
-
-  enc->copyFromTexture(img[i].tex.get(), 0, 0,
-                       dr, 0, 0,
-                       1, 1);
-  enc->endEncoding();
+  if(src!=dr) {
+    auto enc = cmd->blitCommandEncoder();
+    enc->copyFromTexture(src, 0, 0,
+                         dr, 0, 0,
+                         1, 1);
+    enc->endEncoding();
+    }
   cmd->presentDrawable(drawable);
 
+  auto* device = &dev;
   dev.onSubmit();
   cmd->addCompletedHandler(^(MTL::CommandBuffer* c){
-    MTL::CommandBufferStatus s = c->status();
-    if(s==MTL::CommandBufferStatusNotEnqueued ||
-       s==MTL::CommandBufferStatusEnqueued ||
-       s==MTL::CommandBufferStatusCommitted ||
-       s==MTL::CommandBufferStatusScheduled)
-      return;
-
-    if(s!=MTL::CommandBufferStatusCompleted) {
+    if(c->status()!=MTL::CommandBufferStatusCompleted)
       Log::e("swapchain fatal error");
-      dev.onFinish();
-      dev.waitIdle();
-      return;
-      }
-
-    dev.onFinish();
+    device->onFinish();
     });
   cmd->commit();
 
+  activeFrame.reset();
   nextDrawable();
   }
 

Engine/gapi/metalapi.cpp

@@ -23,6 +23,8 @@
 
 #include <Metal/Metal.hpp>
 
+#include <stdexcept>
+
 using namespace Tempest;
 using namespace Tempest::Detail;
 
@@ -68,8 +70,17 @@ AbstractGraphicsApi::Device* MetalApi::createDevice(std::string_view gpuName) {
 
 AbstractGraphicsApi::Swapchain *MetalApi::createSwapchain(SystemApi::Window *w,
                                                           AbstractGraphicsApi::Device* d) {
+  return createSwapchain(w,d,Swapchain::Options{});
+  }
+
+AbstractGraphicsApi::Swapchain* MetalApi::createSwapchain(SystemApi::Window* w, Device* d,
+                                                         const Swapchain::Options& options) {
+  if(options.bufferCount!=0 && options.bufferCount!=2 && options.bufferCount!=3)
+    throw std::invalid_argument("Metal swapchain buffer count must be 0, 2, or 3");
+  if(options.renderMode!=Swapchain::RenderMode::Copy && options.renderMode!=Swapchain::RenderMode::Direct)
+    throw std::invalid_argument("Unknown Metal swapchain render mode");
   auto& dev = *reinterpret_cast<MtDevice*>(d);
-  return new MtSwapchain(dev,w);
+  return new MtSwapchain(dev,w,options);
   }
 
 AbstractGraphicsApi::PPipeline MetalApi::createPipeline(AbstractGraphicsApi::Device *d,
@@ -203,8 +214,15 @@ std::shared_ptr<AbstractGraphicsApi::Fence> MetalApi::submit(Device* d, CommandB
     throw DeviceLostException();
 
   MTL::CommandBuffer& cmd = *cx.impl;
+  std::shared_ptr<std::vector<MtSwapchain::Frame>> frames;
+  if(!cx.swapchainFrames.empty()) {
+    frames = std::make_shared<std::vector<MtSwapchain::Frame>>(std::move(cx.swapchainFrames));
+    }
   dx->onSubmit();
   cmd.addCompletedHandler(^(MTL::CommandBuffer* c){
+    // A completed command buffer may itself remain alive until the next frame.
+    if(frames!=nullptr)
+      frames->clear();
     const MTL::CommandBufferStatus s = c->status();
     dx->signalFence(*pfence, s, MTL::CommandBufferError(c->error()->code()), c->error());
     if(s==MTL::CommandBufferStatusCompleted || s==MTL::CommandBufferStatusError)

Engine/gapi/metalapi.h

@@ -14,6 +14,7 @@ class MetalApi : public AbstractGraphicsApi {
   protected:
     Device*        createDevice(std::string_view gpuName) override;
     Swapchain*     createSwapchain(SystemApi::Window* w, Device *d) override;
+    Swapchain*     createSwapchain(SystemApi::Window* w, Device* d, const Swapchain::Options& options) override;
 
     PPipeline      createPipeline(Device* d, const RenderState &st, Topology tp,
                                   const Shader*const* sh, size_t cnt) override;

Engine/graphics/device.cpp

@@ -160,6 +160,10 @@ Swapchain Device::swapchain(SystemApi::Window* w) const {
   return Swapchain(api.createSwapchain(w,impl.dev));
   }
 
+Swapchain Device::swapchain(SystemApi::Window* w, const Swapchain::Options& options) const {
+  return Swapchain(api.createSwapchain(w,impl.dev,options));
+  }
+
 const Device::Props& Device::properties() const {
   return devProps;
   }

Engine/graphics/device.h

@@ -50,6 +50,7 @@ class Device {
     void                  present(Swapchain& sw);
 
     Swapchain             swapchain(SystemApi::Window* w) const;
+    Swapchain             swapchain(SystemApi::Window* w, const Swapchain::Options& options) const;
 
     Shader                shader(RFile&          file);
     Shader                shader(const char*     filename);

Engine/graphics/swapchain.cpp

@@ -21,6 +21,10 @@ Swapchain::Swapchain(Device& dev, SystemApi::Window* w) {
   *this = dev.swapchain(w);
   }
 
+Swapchain::Swapchain(Device& dev, SystemApi::Window* w, const Options& options) {
+  *this = dev.swapchain(w,options);
+  }
+
 Swapchain::~Swapchain() {
   delete impl.handler;
   }