Integrate CAN-aware CMS risk pipeline and relocatable Orin runtime

accelerator/include/gpu_preprocess.hpp

@@ -0,0 +1,19 @@
+#pragma once
+#include "common.hpp"
+#include <npp.h>
+#include <opencv2/core.hpp>
+namespace cms {
+inline void nppCheck(NppStatus status) {if(status<0)throw std::runtime_error("NPP error "+std::to_string(int(status)));}
+inline NppStreamContext nppContext(cudaStream_t stream) {
+    NppStreamContext c{};c.hStream=stream;CUDA_CHECK(cudaGetDevice(&c.nCudaDeviceId));cudaDeviceProp p{};CUDA_CHECK(cudaGetDeviceProperties(&p,c.nCudaDeviceId));
+    c.nMultiProcessorCount=p.multiProcessorCount;c.nMaxThreadsPerMultiProcessor=p.maxThreadsPerMultiProcessor;c.nMaxThreadsPerBlock=p.maxThreadsPerBlock;c.nSharedMemPerBlock=p.sharedMemPerBlock;c.nCudaDevAttrComputeCapabilityMajor=p.major;c.nCudaDevAttrComputeCapabilityMinor=p.minor;CUDA_CHECK(cudaStreamGetFlags(stream,&c.nStreamFlags));return c;
+}
+class YoloPreprocess {
+public:
+    YoloPreprocess(int width,int height):width_(width),height_(height) {CUDA_CHECK(cudaMalloc(&letter_,size_t(width)*height*3));}
+    ~YoloPreprocess(){cudaFree(source_);cudaFree(letter_);}
+    void enqueue(cv::Mat const& frame,float* destination,cudaStream_t stream);
+private:
+    int width_,height_;size_t capacity_{};unsigned char *source_{},*letter_{};
+};
+}