How to use concurrent kernel with nvjpeg？

I am using nvjpeg to encode a series of pictures into JPEG. The size of each array of bmp is 240 * 240 * 3.

        for (int i = 0; i < cur_ti - bat_ti; i+=params.num_threads) {
            for (int j = 0; j < params.num_threads; j++) {
                auto& per_thread_params = params.enjpeg_per_thread_data[j];
                CHECK_CUDA(cudaMemcpyAsync(
                    imgdesc[i].channel[0],
                    batch_data + (i + bat_ti) * params.tile_size * params.tile_size * 3,
                    params.tile_size * params.tile_size * 3,
                    cudaMemcpyHostToDevice,
                    per_thread_params.stream));
            }
            for (int j = 0; j < params.num_threads; j++) {
                auto& per_thread_params = params.enjpeg_per_thread_data[j];
                CHECK_NVJPEG(nvjpegEncodeImage(
                    params.nvjpeg_handle,
                    per_thread_params.nv_enc_state,
                    per_thread_params.nv_enc_params,
                    &imgdesc[i],
                    params.fmt,
                    params.tile_size,
                    params.tile_size,
                    per_thread_params.stream));
            }
            for (int j = 0; j < params.num_threads; j++) {
                auto& per_thread_params = params.enjpeg_per_thread_data[j];
                size_t length;
                CHECK_NVJPEG(nvjpegEncodeRetrieveBitstream(
                    params.nvjpeg_handle,
                    per_thread_params.nv_enc_state,
                    NULL,
                    &length,
                    per_thread_params.stream));
                output[i].resize(length);
                CHECK_NVJPEG(nvjpegEncodeRetrieveBitstream(
                    params.nvjpeg_handle,
                    per_thread_params.nv_enc_state,
                    (unsigned char*)output[i].data(),
                    &length,
                    per_thread_params.stream));
            }
        }

It seems that the encoding time is too long, and the use of multi stream is ineffective for its acceleration. I tried to use multithreading, but it didn't work. Can I find a way that can be concurrent kernel with the function of nvjpegEncodeImage. This is the function of nvjpeg encoder.

https://docs.nvidia.com/cuda/nvjpeg/index.html#nvjpeg-encode-examples