Method CudaKernelCommand

platforms/cuda/hal/src/common/cuda_command.cpp:54–118

}

CudaKernelCommand::CudaKernelCommand(CUfunction func, void **params, void **extra, bool deep_copy)
    : CudaCommand(preempt::kCommandPropertyDeactivatable)
    // cuXtraKernelGetFunction() will convert func to CUfunction if func is a CUkernel.
    , kFunc(func), kFuncHandle(cuXtraKernelGetFunction((CUkernel)func))
    , params_(params), extra_(extra)
{
    if (!deep_copy) return;

    param_cnt_ = cuXtraGetParamCount(kFuncHandle);
    /// @FIXME: even if param_cnt_ == 0, params or extra can be non-nullptr.
    if (param_cnt_ == 0) return;
    if (params == nullptr && extra == nullptr) {
        XWARN("params and extra of kernel (%p) are both nullptr", func);
        return;
    }
    if (params != nullptr && extra != nullptr) {
        XWARN("illegally using both params (%p) and extra (%p) of kernel %p", params, extra, func);
        return;
    }

    deep_copy_ = true;
    params_ = (void **)malloc(param_cnt_ * sizeof(void *));

    if (params != nullptr) {
        // Allocate a continuous buffer for all of the params
        // buffer size = last param offset + last param size
        size_t last_offset, last_size;
        cuXtraGetParamInfo(kFuncHandle, param_cnt_ - 1, &last_offset, &last_size, nullptr);
        size_t buffer_size = last_offset + last_size;
        param_data_ = (char *)malloc(buffer_size);

        for (size_t i = 0; i < param_cnt_; ++i) {
            size_t offset, size;
            cuXtraGetParamInfo(kFuncHandle, i, &offset, &size, nullptr);
            params_[i] = (void*)&param_data_[offset];
            memcpy(params_[i], params[i], size);
        }
    } else if (extra != nullptr) {
        // Get extra buffer and its size from extra map.
        void *extra_buffer = nullptr;
        cuXtraGetExtraBuffer(extra, &extra_buffer, &extra_buffer_size_);
        // We have checked param_cnt_ > 0.
        XASSERT(extra_buffer != nullptr && extra_buffer_size_ > 0,
                "invalid extra buffer (%p) and size (%zu)", extra_buffer, extra_buffer_size_);

        // Deep-copy extra buffer to extra_data_.
        extra_data_ = (char *)malloc(extra_buffer_size_);
        memcpy(extra_data_, extra_buffer, extra_buffer_size_);

        // Set extra map.
        extra_ = (void **)malloc(5 * sizeof(void *));
        extra_[0] = CU_LAUNCH_PARAM_BUFFER_POINTER;
        extra_[1] = extra_data_;
        extra_[2] = CU_LAUNCH_PARAM_BUFFER_SIZE;
        extra_[3] = (void *)&extra_buffer_size_;
        extra_[4] = CU_LAUNCH_PARAM_END;

        // Set params_ to point to extra_data_.