1，找一只活麻雀，下载编译 ucx

git clone https://github.com/openucx/ucx.git
cd ucx/
git checkout v1.16.0
./autogen.sh
./autogen.sh
mkdir build
cd build
../contrib/configure-devel --with-cuda=/usr/local/cuda --without-rocm --without-java --prefix=${PWD}/../../local_d_nv

make -j
make install

2，运行 ucx 普通示例

ls build/test/apps/

示例解析：

test_tcmalloc

未完待续 。。。。

3，运行 ucx cuda 相关示例

 $ ./test_cuda_hook_static

示例解析：

test_cuda_hook_static

未完待续 。。。。

4，挖掘更深入的 ucx cuda 功能

5，剖析 ucx cuda 功能的达成

6，ucx 使用了那些 cuda API

cuda 开头的：

cudaEventQuery

cudaEventCreateWithFlags

cudaEventDestroy

cudaStreamCreateWithFlags

cudaStreamDestroy

cudaMemcpyDefault

cudaMemcpyAsync

cudaFreeHost

cudaMallocFromPoolAsync

cudaGetErrorString

cudaSetDevice

cudaDeviceSynchronize

cudaGetDeviceCount

cudaFree

cudaMallocManaged

cudaMalloc

cudaMemcpy

cudaMemset

cudaFreeAsync

cudaMallocAsync

cudaMallocPitch

cudaEventDestroy

cudaHostUnregister

cudaHostRegister

cudaHostRegisterPortable

cudaStreamSynchronize

cudaEventRecord

cudaErrorUnsupportedPtxVersion???????????

cu开头的driver api：

cuDeviceGetName

cuPointerGetAttribute

cuMemGetAddress

cuPointerGetAttributes

cuMemRangeGetAttribute

cuMemGetHandleForAddressRange

cuPointerSetAttribute

cuDeviceGetCount

cuEventQuery

cuLaunchHostFunc

cuStreamAddCallback

cuEventCreate

cuStreamCreate

cuStreamDestroy

cuMemcpyDtoDAsync(dst, src, iov[0].length,

cuEventRecord(cuda_ipc_event->event,

cuIpcOpenMemHandle(mapped_addr, key->ph,

cuIpcCloseMemHandle((CUdeviceptr)region->mapped_addr));

cuIpcGetMemHandle(&key->ph, (CUdeviceptr)addr)


cuDeviceGetAttribute(&attrib, CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID,

cuGetErrorString(result, &error_str)

cuCtxGetDevice(&cuda_device)

cuMemGetAddressRange(&pbase, &length, ptr)

cuMemAlloc

cuMemAlloc_v2

cuMemAllocManaged

cuMemAllocPitch

cuMemAllocPitch_v2

cuMemAllocAsync

cuMemAllocFromPoolAsync

cuMemFree_v2

cuMemFreeHost_v2

cuDeviceTotalMem

cuDeviceTotalMem_v2

cuCtxDestroy(m_context)

cuMemAllocHost(&ptr, 64)

cuMemFreeHost(ptr)

cuMemAllocManaged(&dptr, 64, CU_MEM_ATTACH_GLOBAL)

cuMemFree(dptr)

cuMemAllocPitch(&dptr, &pitch, width, height, element_size)

cuMemAllocAsync(&dptr, 64, 0)

cuMemFreeAsync(dptr, 0);

cuDeviceGet(&device, 0)

cuCtxCreate(&context, 0, device)

cuMemAlloc(&dptr, 4096)

cuMemFree(dptr)

cuCtxDetach(context)

cuGetErrorString(_cu_result, &_error_string)

cuInit(0)

cuCtxGetCurrent(&cu_context)

cuDeviceGetUuid ();

7，ucx cuda VS cux rocm