#import namespace { struct FlowOutput {float x,y,error,valid;}; struct PyramidInfo {int width[5],height[5],offset[5],levels;}; class MetalFlow final : public dy::FlowBackend { id device_=nil; id queue_=nil; id pyramid_=nil,flow_=nil; id frames_[2]={nil,nil},points_=nil,results_=nil; PyramidInfo info_{}; int previous_=0; bool finish(id b) { [b commit];[b waitUntilCompleted];return b.status==MTLCommandBufferStatusCompleted; } void pyramid(id b,id frame) { for(int i=1;ie=[b computeCommandEncoder];[e setComputePipelineState:pyramid_]; [e setBuffer:frame offset:0 atIndex:0];[e setBytes:&info_ length:sizeof(info_) atIndex:1];[e setBytes:&i length:sizeof(i) atIndex:2]; [e dispatchThreads:MTLSizeMake(info_.width[i],info_.height[i],1) threadsPerThreadgroup:MTLSizeMake(16,8,1)];[e endEncoding]; } } public: MetalFlow(NSURL* url,NSError** error) { device_=MTLCreateSystemDefaultDevice();queue_=[device_ newCommandQueue]; id l=[device_ newLibraryWithURL:url error:error];if(!l)return; id p=[l newFunctionWithName:@"dyTrackPyramid"],f=[l newFunctionWithName:@"dyTrackFlow"]; pyramid_=[device_ newComputePipelineStateWithFunction:p error:error];flow_=[device_ newComputePipelineStateWithFunction:f error:error];[p release];[f release];[l release]; } ~MetalFlow(){for(int i=0;i<2;i++)[frames_[i] release];[points_ release];[results_ release];[pyramid_ release];[flow_ release];[queue_ release];[device_ release];} bool ready()const{return device_&&queue_&&pyramid_&&flow_&&flow_.threadExecutionWidth==32;} bool reset(const dy::GrayFrame& f) override { if(!ready()||!f.valid())return false; info_={};info_.levels=1;info_.width[0]=f.width;info_.height[0]=f.height; int total=f.width*f.height; while(info_.levels<5&&info_.width[info_.levels-1]>=80&&info_.height[info_.levels-1]>=80) { int i=info_.levels++;info_.width[i]=info_.width[i-1]/2;info_.height[i]=info_.height[i-1]/2; info_.offset[i]=total;total+=info_.width[i]*info_.height[i]; } for(int i=0;i<2;i++){[frames_[i] release];frames_[i]=nil;}[points_ release];points_=nil;[results_ release];results_=nil; for(int i=0;i<2;i++)frames_[i]=[device_ newBufferWithLength:total*sizeof(float) options:MTLResourceStorageModeShared]; points_=[device_ newBufferWithLength:180*sizeof(dy::FlowPoint) options:MTLResourceStorageModeShared]; results_=[device_ newBufferWithLength:180*sizeof(FlowOutput) options:MTLResourceStorageModeShared]; if(!frames_[0]||!frames_[1]||!points_||!results_)return false; previous_=0;memcpy(frames_[0].contents,f.pixels.data(),f.pixels.size()*sizeof(float)); idb=[queue_ commandBuffer];pyramid(b,frames_[0]);return finish(b); } bool advance(const dy::GrayFrame& f,const std::vector& points,std::vector& matches) override { if(!f.valid()||f.width!=info_.width[0]||f.height!=info_.height[0]||points.empty()||points.size()>180)return false; int next=1-previous_;memcpy(frames_[next].contents,f.pixels.data(),f.pixels.size()*sizeof(float)); memcpy(points_.contents,points.data(),points.size()*sizeof(dy::FlowPoint)); idb=[queue_ commandBuffer];pyramid(b,frames_[next]); ide=[b computeCommandEncoder];[e setComputePipelineState:flow_]; [e setBuffer:frames_[previous_] offset:0 atIndex:0];[e setBuffer:frames_[next] offset:0 atIndex:1]; [e setBuffer:points_ offset:0 atIndex:2];[e setBuffer:results_ offset:0 atIndex:3];[e setBytes:&info_ length:sizeof(info_) atIndex:4]; [e dispatchThreads:MTLSizeMake(points.size()*32,1,1) threadsPerThreadgroup:MTLSizeMake(flow_.threadExecutionWidth,1,1)];[e endEncoding]; if(!finish(b))return false;auto*r=static_cast(results_.contents); for(size_t i=0;i0)matches.push_back({points[i],{r[i].x,r[i].y},r[i].error}); previous_=next;return true; } }; } std::shared_ptr DYToolCreateMetalFlow(NSURL* url,NSError** error) { auto backend=std::make_shared(url,error);if(backend->ready())return backend; if(error&&!*error)*error=[NSError errorWithDomain:@"com.dayu.color.tracking" code:1 userInfo:@{NSLocalizedDescriptionKey:@"Metal 跟踪引擎不可用"}];return {}; }