Method CorexKernelCommand

platforms/corex/hal/src/common/corex_command.cpp:44–89

}

CorexKernelCommand::CorexKernelCommand(const void *func, void **params, bool deep_copy)
    : CorexCommand(kCommandPropertyDeactivatable), func_(func), params_(params)
{
    if (!deep_copy) return;
    if (params == nullptr) {
        XWARN("kernel_args of %p is nullptr", func);
        return;
    }

    cudaKernel_t kernel = nullptr;
    cudaError_t res = RtDriver::GetKernel(&kernel, func);
    if (res != cudaSuccess) RtDriver::GetLastError();
    XASSERT(kernel != nullptr, "fail to get kernel for func %p", func);

    std::vector<size_t> offsets;
    std::vector<size_t> sizes;

    param_cnt_ = 0;
    while (true) {
        size_t offset = 0, size = 0;
        CUresult res = Driver::KernelGetParamInfo(CUkernel(kernel), param_cnt_, &offset, &size);
        if (res != CUDA_SUCCESS || size == 0) break;
        offsets.push_back(offset);
        sizes.push_back(size);
        ++param_cnt_;
    }

    if (param_cnt_ == 0) {
        XWARN("kernel %p has no params", func);
        return;
    }

    deep_copy_ = true;
    params_ = (void **)malloc(param_cnt_ * sizeof(void *));
    // Allocate a continuous buffer for all of the params
    // buffer size = last param offset + last param size
    size_t buffer_size = offsets.back() + sizes.back();
    param_data_ = (char *)malloc(buffer_size);

    for (size_t i = 0; i < param_cnt_; ++i) {
        size_t offset = offsets[i];
        size_t size = sizes[i];
        params_[i] = (void*)&param_data_[offset];
        memcpy(params_[i], params[i], size);
    }
}

CorexKernelCommand::~CorexKernelCommand()
{