I found obs-ndi has been increasing the time to render frames about 8ms on two of my systems whenever I use it as a filter on a source. After some research, I found this greatly reduces it and my render times have gone from 10ms down to 2ms:

It appears void ndi_filter_offscreen_render(void *data, uint32_t cx, uint32_t cy) calls gs_stagesurface_map constantly when it is not necessary. I changed this section of obs-ndi-filter.cpp as follows. No idea if this is "correct", but it is working great for me and seems to be more efficient.

struct video_frame output_frame;
	if (video_output_lock_frame(s->video_output, &output_frame, 1,
				    os_gettime_ns())) {

// removed the unmap call and reset of video_data to NULL that was here...constant map/unmap seems to be a big hit

		gs_stage_texture(
			s->stagesurface,
			gs_texrender_get_texture(s->texrender));

// Only map once...mapping multiple times results in higher CPU
if (s->video_data == NULL)
{
gs_stagesurface_map(s->stagesurface, &s->video_data,
&s->video_linesize);

			gs_stagesurface_unmap(s->stagesurface);
        };

		uint32_t linesize = output_frame.linesize[0];
		for (uint32_t i = 0; i < s->known_height; ++i) {
			uint32_t dst_offset = linesize * i;
			uint32_t src_offset = s->video_linesize * i;
			memcpy(output_frame.data[0] + dst_offset,
			       s->video_data + src_offset, linesize);
		}

		video_output_unlock_frame(s->video_output);
	}