[Impeller] Migrate gaussian blur to half precision.

[jonahwilliams]

From bisecting through the shader performance, i've determined something odd which is that both the vector operations in IPHalfSampleDecal and the conversion of the input texture coordinates to half precision led to the regressed performance. No other changes cause a performance regression, at least locally - and should be safe to reland so that we can get the benefits of reduced CPU usage. I will update this with local benchmark numbers shortly.

BEFORE

{
  "success": true,
  "data": {
    "average_frame_build_time_millis": 0.27609999999999996,
    "90th_percentile_frame_build_time_millis": 0.341,
    "99th_percentile_frame_build_time_millis": 0.435,
    "worst_frame_build_time_millis": 0.435,
    "missed_frame_build_budget_count": 0,
    "average_frame_rasterizer_time_millis": 250.66653846153844,
    "90th_percentile_frame_rasterizer_time_millis": 251.383,
    "99th_percentile_frame_rasterizer_time_millis": 252.709,
    "worst_frame_rasterizer_time_millis": 252.709,
    "missed_frame_rasterizer_budget_count": 39,
    "frame_count": 40,
    "frame_rasterizer_count": 39,
    "new_gen_gc_count": 0,
    "old_gen_gc_count": 0,
    "average_vsync_transitions_missed": 16.075949367088608,
    "90th_percentile_vsync_transitions_missed": 16.0,
    "99th_percentile_vsync_transitions_missed": 17.0,
    "average_vsync_frame_lag": 0.0,
    "90th_percentile_vsync_frame_lag": 0.0,
    "99th_percentile_vsync_frame_lag": 0.0,
    "average_layer_cache_count": 0.0,
    "90th_percentile_layer_cache_count": 0.0,
    "99th_percentile_layer_cache_count": 0.0,
    "worst_layer_cache_count": 0.0,
    "average_layer_cache_memory": 0.0,
    "90th_percentile_layer_cache_memory": 0.0,
    "99th_percentile_layer_cache_memory": 0.0,
    "worst_layer_cache_memory": 0.0,
    "average_picture_cache_count": 0.0,
    "90th_percentile_picture_cache_count": 0.0,
    "99th_percentile_picture_cache_count": 0.0,
    "worst_picture_cache_count": 0.0,
    "average_picture_cache_memory": 0.0,
    "90th_percentile_picture_cache_memory": 0.0,
    "99th_percentile_picture_cache_memory": 0.0,
    "worst_picture_cache_memory": 0.0,
    "total_ui_gc_time": 0.0,
    "30hz_frame_percentage": 0.0,
    "60hz_frame_percentage": 100.0,
    "80hz_frame_percentage": 0.0,
    "90hz_frame_percentage": 0.0,
    "120hz_frame_percentage": 0.0,
    "illegal_refresh_rate_frame_count": 0,
    "average_cpu_usage": 69.94081630612244,
    "90th_percentile_cpu_usage": 72.5,
    "99th_percentile_cpu_usage": 75.4,
    "average_gpu_usage": 100.0,
    "90th_percentile_gpu_usage": 100.0,
    "99th_percentile_gpu_usage": 100.0,
    "average_memory_usage": 100.2139668367347,
    "90th_percentile_memory_usage": 103.59375,
    "99th_percentile_memory_usage": 104.171875
  },

AFTER

  "success": true,
  "data": {
    "average_frame_build_time_millis": 0.30504999999999993,
    "90th_percentile_frame_build_time_millis": 0.402,
    "99th_percentile_frame_build_time_millis": 0.442,
    "worst_frame_build_time_millis": 0.442,
    "missed_frame_build_budget_count": 0,
    "average_frame_rasterizer_time_millis": 250.5412820512821,
    "90th_percentile_frame_rasterizer_time_millis": 251.221,
    "99th_percentile_frame_rasterizer_time_millis": 251.948,
    "worst_frame_rasterizer_time_millis": 251.948,
    "missed_frame_rasterizer_budget_count": 39,
    "frame_count": 40,
    "frame_rasterizer_count": 39,
    "new_gen_gc_count": 0,
    "old_gen_gc_count": 0,
    "average_vsync_transitions_missed": 16.050632911392405,
    "90th_percentile_vsync_transitions_missed": 16.0,
    "99th_percentile_vsync_transitions_missed": 17.0,
    "average_vsync_frame_lag": 0.0,
    "90th_percentile_vsync_frame_lag": 0.0,
    "99th_percentile_vsync_frame_lag": 0.0,
    "average_layer_cache_count": 0.0,
    "90th_percentile_layer_cache_count": 0.0,
    "99th_percentile_layer_cache_count": 0.0,
    "worst_layer_cache_count": 0.0,
    "average_layer_cache_memory": 0.0,
    "90th_percentile_layer_cache_memory": 0.0,
    "99th_percentile_layer_cache_memory": 0.0,
    "worst_layer_cache_memory": 0.0,
    "average_picture_cache_count": 0.0,
    "90th_percentile_picture_cache_count": 0.0,
    "99th_percentile_picture_cache_count": 0.0,
    "worst_picture_cache_count": 0.0,
    "average_picture_cache_memory": 0.0,
    "90th_percentile_picture_cache_memory": 0.0,
    "99th_percentile_picture_cache_memory": 0.0,
    "worst_picture_cache_memory": 0.0,
    "total_ui_gc_time": 0.0,
    "30hz_frame_percentage": 0.0,
    "60hz_frame_percentage": 100.0,
    "80hz_frame_percentage": 0.0,
    "90hz_frame_percentage": 0.0,
    "120hz_frame_percentage": 0.0,
    "illegal_refresh_rate_frame_count": 0,
    "average_cpu_usage": 70.76600016000002,
    "90th_percentile_cpu_usage": 72.400001,
    "99th_percentile_cpu_usage": 74.099999,
    "average_gpu_usage": 100.0,
    "90th_percentile_gpu_usage": 100.0,
    "99th_percentile_gpu_usage": 100.0,
    "average_memory_usage": 100.56187499999999,
    "90th_percentile_memory_usage": 103.23437499999997,
    "99th_percentile_memory_usage": 104.67187499999997
  },

[zanderso]

Sorry, I've got the Friday afternoon brain fog. Could you spell out the conclusion from the before/after data?

[zanderso]

lgtm w/ above question.

I guess the effects of this will show up in the new blur benchmark?

[jonahwilliams]

I would not expect to see a significant difference except in power consumption. This change is the same as previous one, but it retains high precision in the input texture coordinates.