Porting/LLVMpipe

LLVMpipe (source code) runs on POWER, but performance has room for improvement.

Testing Notes

If your machine has a discrete accelerated GPU, then you'll probably need to set these environment variables in order to test LLVMpipe:

export LIBGL_ALWAYS_SOFTWARE=true
export GALLIUM_DRIVER=llvmpipe

Potentially useful benchmarking tools

• glmark2 (seems to be not very useful at noticing small LLVMpipe optimizations)
• OpenArena (seems to work okay for testing LLVMpipe optimizations).
  • Download this benchmarking script. To compare FPS with different thread counts, run ./benchmark.sh; to measure which functions are potential bottlenecks, run ./perf.sh.
• Xonotic (haven't tried it yet)
• list of desktop graphics tests run by Phoronix
• list of games that run on POWER9
• Application traces replay
  • Running traces on a local machine
  • CI traces for LLVMpipe

Building patched Mesa from source

On Debian (might also work for derivatives such as Devuan and Ubuntu):

• apt source mesa
• cd to the source directory that was created by apt source.
• Install build-essential, and all packages listed in Build-Depends and Build-Depends-indep fields of debian/control.
• Apply whatever patches you like to the source.
• Add an entry to the top of debian/changelog with a new version number (incrementing the last number is an okay approach), so that apt install will know it's a new version.
• dpkg-buildpackage -us -uc
• sudo apt install ../*.deb

Thread Count

LLVMpipe is limited to 16 threads. The only easily findable justification for this limit is in a commit from 2013, where it was increased from 8 because a user reported on a mailing list that 16 was faster for them. Given that POWER9 systems often have much higher thread counts than this, this limit may be suboptimal for POWER9.

luke-jr reports that bumping the limit to 128 threads noticeably improved performance in 3D games, e.g. Jedi Academy (detail set to High, Texture to Very High, Texture Filter BILINEAR, Detailed Shaders ON, Video Sync OFF, resolution 800x600) went from ~15 fps with 16 threads to ~25 fps with 64 threads (on a 2x 8-core Talos II). However, it also had the side effect that most GUI applications spawned more LLVMpipe threads, which was annoying in gdb/top. Luke worked around this issue by setting the environment variable LP_NUM_THREADS=2 globally, and overriding for 3D applications that needed more threads. Luke's patch is:

diff -ur mesa-17.3.9.orig/src/gallium/drivers/llvmpipe/lp_limits.h mesa-17.3.9/src/gallium/drivers/llvmpipe/lp_limits.h
--- mesa-17.3.9.orig/src/gallium/drivers/llvmpipe/lp_limits.h	2018-04-18 04:44:00.000000000 -0400
+++ mesa-17.3.9/src/gallium/drivers/llvmpipe/lp_limits.h	2018-05-02 05:20:57.586000000 -0400
@@ -61,7 +61,7 @@
 #define LP_MAX_WIDTH  (1 << (LP_MAX_TEXTURE_LEVELS - 1))
 
 
-#define LP_MAX_THREADS 16
+#define LP_MAX_THREADS 128
 
 
 /**
Only in mesa-17.3.9/src/gallium/drivers/llvmpipe: lp_limits.h~

To 32 Threads (Merged)

JeremyRand benchmarked OpenArena with LLVMpipe, and found that for 1920x1200 resolution on a 2x 4-core Talos II running Debian Bullseye, the following Mesa patch improved performance:

diff -ur stock/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_limits.h 32-threads/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_limits.h
--- stock/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_limits.h  2021-03-24 14:10:48.744070300 -0500
+++ 32-threads/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_limits.h     2022-07-02 04:08:46.880000000 -0500
@@ -66,7 +66,9 @@
 
 #define LP_MAX_SAMPLES 4
 
-#define LP_MAX_THREADS 16
+// Bumped by Jeremy
+//#define LP_MAX_THREADS 16
+#define LP_MAX_THREADS 32
 
 
 /**

Before the 32-thread patch:

3398 frames 487.6 seconds 7.0 fps 83.0/143.5/7186.0/25.1 ms
3398 frames 472.7 seconds 7.2 fps 86.0/139.1/2242.0/22.9 ms
3398 frames 466.5 seconds 7.3 fps 86.0/137.3/943.0/21.1 ms
3398 frames 467.5 seconds 7.3 fps 83.0/137.6/846.0/22.4 ms
3398 frames 474.8 seconds 7.2 fps 86.0/139.7/779.0/22.8 ms

After the 32-thread patch:

3398 frames 417.7 seconds 8.1 fps 77.0/122.9/1748.0/18.9 ms
3398 frames 417.9 seconds 8.1 fps 76.0/123.0/997.0/19.6 ms
3398 frames 419.9 seconds 8.1 fps 76.0/123.6/806.0/19.8 ms
3398 frames 422.7 seconds 8.0 fps 75.0/124.4/758.0/21.1 ms
3398 frames 419.1 seconds 8.1 fps 75.0/123.3/730.0/20.4 ms

The 32-thread patch has been merged to Mesa on 2022 October 4.

To More Than 32 Threads

Nashimus tested #define LP_MAX_THREADS 144 on a 2x 18-core Talos II, with the following results:

16 threads:                 3398 frames 312.5 seconds 10.9 fps 55.0/92.0/8375.0/21.2 ms
64 threads:                 3398 frames 221.5 seconds 15.3 fps 33.0/65.2/8390.0/21.4 ms
144 threads:                3398 frames 208.5 seconds 16.3 fps 30.0/61.4/8364.0/21.8 ms

It would be desirable to compare 32 threads to 64 threads on the same setup so that Jeremy's results and Nashimus's results can be more directly compared.

Thum tested #define LP_MAX_THREADS 176 with 2 RAM modules, with the following results:

thum@tls0:~$ cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor | tail -n1
performance
thum@tls0:~$ uname -r
5.19.0-2-powerpc64le
thum@tls0:~/llvmpipe-multithreaded-openarena-benchmark$ ./benchmark.sh 
Benchmarks should take around 4 hours to run, be patient!

CPU(s):                          176
OpenGL version string: 4.5 (Compatibility Profile) Mesa 22.2.0
MODE: 3, 640 x 480 windowed hz:N/A

Frames  TotalTime  averageFPS  minimum/average/maximum/std deviation

16 Threads:
3398 frames 107.0 seconds 31.7 fps 13.0/31.5/4209.0/16.7 ms
3398 frames 107.0 seconds 31.8 fps 12.0/31.5/4174.0/16.4 ms
3398 frames 107.3 seconds 31.7 fps 12.0/31.6/4214.0/16.5 ms
3398 frames 109.5 seconds 31.0 fps 14.0/32.2/4269.0/16.3 ms
3398 frames 106.6 seconds 31.9 fps 12.0/31.4/4192.0/16.8 ms

32 Threads:
3398 frames 104.3 seconds 32.6 fps 11.0/30.7/4192.0/17.2 ms
3398 frames 106.8 seconds 31.8 fps 12.0/31.4/4182.0/17.0 ms
3398 frames 106.6 seconds 31.9 fps 11.0/31.4/4176.0/16.9 ms
3398 frames 105.3 seconds 32.3 fps 11.0/31.0/4206.0/16.9 ms
3398 frames 105.3 seconds 32.3 fps 11.0/31.0/4305.0/16.8 ms

48 Threads:
3398 frames 109.2 seconds 31.1 fps 12.0/32.1/4186.0/17.0 ms
3398 frames 109.9 seconds 30.9 fps 12.0/32.4/4176.0/16.9 ms
3398 frames 112.2 seconds 30.3 fps 12.0/33.0/4230.0/16.9 ms
3398 frames 107.2 seconds 31.7 fps 11.0/31.6/4206.0/17.0 ms
3398 frames 107.4 seconds 31.7 fps 11.0/31.6/4194.0/16.9 ms

64 Threads:
3398 frames 113.6 seconds 29.9 fps 13.0/33.4/4140.0/17.1 ms
3398 frames 115.4 seconds 29.4 fps 14.0/34.0/4175.0/17.0 ms
3398 frames 116.7 seconds 29.1 fps 14.0/34.3/4233.0/17.0 ms
3398 frames 113.5 seconds 29.9 fps 13.0/33.4/4199.0/17.0 ms
3398 frames 113.5 seconds 30.0 fps 12.0/33.4/4204.0/17.0 ms

72 Threads:
3398 frames 117.1 seconds 29.0 fps 13.0/34.5/4201.0/17.0 ms
3398 frames 118.2 seconds 28.7 fps 14.0/34.8/4225.0/17.1 ms
3398 frames 116.9 seconds 29.1 fps 13.0/34.4/4239.0/17.1 ms
3398 frames 116.2 seconds 29.2 fps 13.0/34.2/4178.0/17.1 ms
3398 frames 116.2 seconds 29.2 fps 13.0/34.2/4181.0/17.1 ms

80 Threads:
3398 frames 117.9 seconds 28.8 fps 13.0/34.7/4190.0/17.1 ms
3398 frames 118.2 seconds 28.8 fps 14.0/34.8/4174.0/17.1 ms
3398 frames 117.3 seconds 29.0 fps 14.0/34.5/4216.0/17.1 ms
3398 frames 119.1 seconds 28.5 fps 13.0/35.0/4205.0/17.1 ms
3398 frames 117.9 seconds 28.8 fps 13.0/34.7/4186.0/17.2 ms

88 Threads:
3398 frames 119.2 seconds 28.5 fps 14.0/35.1/4224.0/17.2 ms
3398 frames 119.0 seconds 28.6 fps 13.0/35.0/4204.0/17.3 ms
3398 frames 118.5 seconds 28.7 fps 14.0/34.9/4179.0/17.3 ms
3398 frames 119.3 seconds 28.5 fps 14.0/35.1/4212.0/17.2 ms
3398 frames 119.0 seconds 28.6 fps 13.0/35.0/4223.0/17.2 ms

96 Threads:
3398 frames 120.5 seconds 28.2 fps 14.0/35.4/4210.0/17.4 ms
3398 frames 120.2 seconds 28.3 fps 14.0/35.4/4223.0/17.3 ms
3398 frames 119.5 seconds 28.4 fps 12.0/35.2/4251.0/17.3 ms
3398 frames 120.6 seconds 28.2 fps 14.0/35.5/4169.0/17.4 ms
3398 frames 119.6 seconds 28.4 fps 14.0/35.2/4213.0/17.2 ms

128 Threads:
3398 frames 122.3 seconds 27.8 fps 14.0/36.0/4211.0/17.5 ms
3398 frames 122.6 seconds 27.7 fps 14.0/36.1/4231.0/17.5 ms
3398 frames 122.9 seconds 27.7 fps 14.0/36.2/4224.0/17.5 ms
3398 frames 123.0 seconds 27.6 fps 15.0/36.2/4199.0/17.6 ms
3398 frames 122.4 seconds 27.8 fps 14.0/36.0/4194.0/17.4 ms

144 Threads:
3398 frames 123.7 seconds 27.5 fps 14.0/36.4/4144.0/17.6 ms
3398 frames 124.2 seconds 27.4 fps 14.0/36.6/4211.0/17.5 ms
3398 frames 124.1 seconds 27.4 fps 14.0/36.5/4200.0/17.6 ms
3398 frames 124.6 seconds 27.3 fps 14.0/36.7/4189.0/17.6 ms
3398 frames 123.9 seconds 27.4 fps 14.0/36.5/4203.0/17.6 ms

160 Threads:
3398 frames 125.6 seconds 27.1 fps 14.0/37.0/4262.0/17.7 ms
3398 frames 125.9 seconds 27.0 fps 14.0/37.1/4224.0/17.7 ms
3398 frames 125.6 seconds 27.1 fps 15.0/37.0/4200.0/17.7 ms
3398 frames 125.9 seconds 27.0 fps 14.0/37.1/4253.0/17.8 ms
3398 frames 126.3 seconds 26.9 fps 15.0/37.2/4284.0/17.7 ms

176 Threads:
3398 frames 128.2 seconds 26.5 fps 15.0/37.7/4207.0/17.8 ms
3398 frames 128.0 seconds 26.6 fps 14.0/37.7/4278.0/17.7 ms
3398 frames 128.0 seconds 26.6 fps 14.0/37.7/4256.0/17.9 ms
3398 frames 128.3 seconds 26.5 fps 14.0/37.8/4300.0/17.9 ms
3398 frames 127.7 seconds 26.6 fps 15.0/37.6/4211.0/17.8 ms

It would be useful to get test results with 8 RAM modules for maximum memory bandwidth.

Improving Thread Utilization

MR's:

• llvmpipe/lavapipe: add support for overlapping vertex and fragment processing.
  • optimizing llvmpipe vertex/fragment processing.
  • Merged Feb 21, 2022.
  • First released in Mesa 22.1.0.
  • First packaged in Debian Bookworm, Fedora 36, Ubuntu 22.10.

From #dri-devel :

23:49 Jeremy_Rand_Talos: In https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18415#note_1542451 , sroland said "there's just bottlenecks which makes using high number of threads a bit questionable".
23:49 Jeremy_Rand_Talos: Are those bottlenecks documented anywhere? I see a nontrivial amount of llvmpipe CPU usage (e.g. triangle_ccw) in the main thread (as opposed to the spawned threads) via perf; is that what he meant?
23:50 airlied: Jeremy_Rand_Talos: I've removed some of those bottlenecks, but I think binning is probably the largest one remaining
23:51 Jeremy_Rand_Talos: airlied, how feasible is it to improve the remaining such bottlenecks?
23:52 Jeremy_Rand_Talos: airlied, and are there any docs (formal or informal) where I could read up on the current state of such things?
23:58 airlied: Jeremy_Rand_Talos: how much time you got? :-)
23:58 airlied: the main thing would be to find a benchmark where you care
23:58 airlied: then profile, profile, profile, and see what bottlenecks
23:59 airlied: most of the bottlenecks are memory bandwidth on fragment shader execution
23:59 airlied: those are hard to fix :-P
23:59 ajax: airlied: did we finally get overlapping vs/fs/present working?
23:59 airlied: ajax: yes it seems to definitely be bug free this time :-P
00:00 Jeremy_Rand_Talos: airlied, so, I'm from the Talos Workstation community; some of us are on up to 176 hardware threads; it would be nice to be able to actually leverage that with llvmpipe.
00:00 airlied: now the tests that have vertex heavy workloads stall out on binning a lot
00:00 airlied: Jeremy_Rand_Talos: it probably comes down to memory b
00:00 airlied: bw
00:00 ajax: then... yeah pretty much all you have left for llvmpipe perf is either smarter image layouts or reducing the cost of binning
00:01 airlied: ajax: threaded binning was a vague handwave I had
00:01 airlied: but I'm not so sure how to make that a thing
00:01 ajax: i think swr had a multiply-and-surrender approach to that
00:01 airlied: smarter image layouts would be something if you could figure out what would work on a modern CPU
00:02 airlied: like you'd have to know where things are having cacheline pains
00:02 ajax: 2x2 microtiles and page-sized macrotiles would go a long way
00:02 ajax: or rather: i don't think there's much to be gained beyond that
00:02 airlied:wonders if vmware had that at one point, and didn't see enough throughput changes
00:03 airlied: though someone suggested tiled framebuffers might be a better win
00:04 airlied: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6972 has some notes
00:10 Jeremy_Rand_Talos: airlied, has it been considered whether memory bandwidth is still the bottleneck regardless of architecture? My loose understanding is that POWER9 often has better mem bw than x86_64.
00:12 airlied: Jeremy_Rand_Talos: again it depends on the load being tested, adding more threads with a complex fragment shader might show where things stop scaling
00:12 airlied: like power9 might have more mem bw, but it still might get saturated
00:13 Jeremy_Rand_Talos: airlied, "memory bandwidth on fragment shader execution" <-- what function name(s) would this show up as in perf?
00:13 airlied: Jeremy_Rand_Talos: it will show up as some JIT code
00:13 airlied: unfortunately that is another problem, getting JIT debugging going properly is kinda not there
00:14 airlied: so it's not that easy to spot where the bottlenecks are
00:14 airlied: I've never really figured out the best way to close that gap
00:14 Jeremy_Rand_Talos: airlied, ah, ok. So if I can find some bottleneck function in perf that's not marked as JIT, that would be likely to be lower-hanging fruit?
00:14 airlied: Jeremy_Rand_Talos: yes
00:15 airlied: esp if it's in the main thread not one of the side threads
00:15 Jeremy_Rand_Talos: airlied, I see, that's good info to have.
00:17 airlied: but yeah digging into anything that could make jit more debuggable might be a useful task

Vector Optimizations

As of 2022 June 10, grepping main branch src/gallium/drivers/llvmpipe/ for altivec yields only 2 files (lp_rast_tri.c and lp_setup_tri.c, both of which are POWER8 LE), while grepping for PIPE_ARCH_SSE yields 11 files. This seems to suggest that a lot of POWER vector optimizations are missing from LLVMpipe. POWER9 vector optimizations (the LLVM power9-vector feature), and POWER8 BE optimizations, appear to be completely absent.

lp_rast_tri.c and lp_setup_tri.c

markos looked at the existing Altivec code in LLVMpipe (as of main 2022 October 6) and observed a lot of SSE-isms, probably because the LLVMpipe Altivec code was translated from the SSE code. It is likely that rewriting the Altivec code would yield performance improvements.

calc_fixed_position

JeremyRand ran the perf.sh benchmark for OpenArena (see above link) on Debian Bullseye (with 32 threads, see above patch), and found that 0.21% of CPU time (6th-highest-ranked function) was used by triangle_ccw, which is mostly a wrapper for the inline function calc_fixed_position in lp_setup_tri.c. This happens to be the only function that uses SSE in Debian Bullseye but is missing an Altivec implementation in main branch (other SSE-utilizing functions were added to main after Debian Bullseye).

Jeremy enabled build-time SSE intrinsics translation in the calc_fixed_position function via this patch:

diff -ur 32-threads/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_setup_tri.c 32-threads-simd-wip2/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_setup_tri.c
--- 32-threads/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_setup_tri.c  2021-03-24 14:10:48.746070400 -0500
+++ 32-threads-simd-wip2/mesa-20.3.5/src/gallium/drivers/llvmpipe/lp_setup_tri.c        2022-07-02 05:22:44.160000000 -0500
@@ -49,9 +49,14 @@
 #elif defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
 #include <altivec.h>
 #include "util/u_pwr8.h"
+// Emulate SSE
+#define NO_WARN_X86_INTRINSICS
+#define __m128i __x86__m128i
+#include <emmintrin.h>
+#undef __m128i
 #endif
 
-#if !defined(PIPE_ARCH_SSE)
+#if !(defined(PIPE_ARCH_SSE) || (defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN))
 
 static inline int
 subpixel_snap(float a)
@@ -1032,7 +1037,10 @@
     * otherwise nearest/away-from-zero).
     * Both should be acceptable, I think.
     */
-#if defined(PIPE_ARCH_SSE)
+#if defined(PIPE_ARCH_SSE) || (defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN)
+   #if defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
+      #define __m128i __x86__m128i
+   #endif
    __m128 v0r, v1r;
    __m128 vxy0xy2, vxy1xy0;
    __m128i vxy0xy2i, vxy1xy0i;
@@ -1061,6 +1069,9 @@
    y0120 = _mm_unpackhi_epi32(x0x2y0y2, x1x0y1y0);
    _mm_store_si128((__m128i *)&position->x[0], x0120);
    _mm_store_si128((__m128i *)&position->y[0], y0120);
+   #if defined(_ARCH_PWR8) && UTIL_ARCH_LITTLE_ENDIAN
+      #undef __m128i
+   #endif
 
 #else
    position->x[0] = subpixel_snap(v0[0][0] - pixel_offset);

When combined with the 32-thread patch from the above section, Jeremy obtained the following benchmarks in OpenArena:

32 threads without SIMD patch:

3398 frames 417.7 seconds 8.1 fps 77.0/122.9/1748.0/18.9 ms
3398 frames 417.9 seconds 8.1 fps 76.0/123.0/997.0/19.6 ms
3398 frames 419.9 seconds 8.1 fps 76.0/123.6/806.0/19.8 ms
3398 frames 422.7 seconds 8.0 fps 75.0/124.4/758.0/21.1 ms
3398 frames 419.1 seconds 8.1 fps 75.0/123.3/730.0/20.4 ms

32 threads with SIMD patch:

3398 frames 418.5 seconds 8.1 fps 76.0/123.2/865.0/19.7 ms
3398 frames 414.7 seconds 8.2 fps 74.0/122.1/805.0/19.7 ms
3398 frames 418.8 seconds 8.1 fps 74.0/123.2/701.0/20.0 ms
3398 frames 424.0 seconds 8.0 fps 77.0/124.8/524.0/21.7 ms
3398 frames 414.1 seconds 8.2 fps 74.0/121.9/621.0/19.2 ms

This seems like an improvement, though a quite small one.

Jeremy checked the function size using nm:

nm -S --size-sort -t d ./build/src/gallium/drivers/llvmpipe/libllvmpipe.a.p/lp_setup_tri.c.o | grep ' triangle_ccw'
# Without SIMD patch:
0000000000000000 0000000000000736 t triangle_ccw
# With SIMD patch:
0000000000000000 0000000000000604 t triangle_ccw

Nashimus tested the above lp_setup_tri.c patch in OpenArena with different thread counts:

16 threads:                 3398 frames 312.5 seconds 10.9 fps 55.0/92.0/8375.0/21.2 ms
16 threads and SIMD patch:  3398 frames 309.1 seconds 11.0 fps 54.0/91.0/8421.0/20.9 ms
64 threads:                 3398 frames 221.5 seconds 15.3 fps 33.0/65.2/8390.0/21.4 ms
64 threads and SIMD patch:  3398 frames 222.3 seconds 15.3 fps 35.0/65.4/8464.0/21.4 ms
144 threads:                3398 frames 208.5 seconds 16.3 fps 30.0/61.4/8364.0/21.8 ms
144 threads and SIMD patch: 3398 frames 208.7 seconds 16.3 fps 30.0/61.4/8327.0/21.8 ms

Other functions

It would be desirable to create and test similar patches for SSE-based non-Altivec functions added between Debian Bullseye and current main.

Nashimus tested File:LLVMpipe-SIMD-bookworm.patch in OpenArena with different thread counts. Starting from Mesa commit b91971c2 on Ubuntu 22.04:

   32 threads (14.92 avg fps):
       3398 frames 224.7 seconds 15.1 fps 36.0/66.1/2483.0/14.4 ms
       3398 frames 229.7 seconds 14.8 fps 38.0/67.6/1533.0/15.1 ms
       3398 frames 227.6 seconds 14.9 fps 39.0/67.0/1522.0/13.9 ms
       3398 frames 227.9 seconds 14.9 fps 39.0/67.1/1533.0/14.1 ms
       3398 frames 227.7 seconds 14.9 fps 39.0/67.0/1551.0/14.2 ms
   
   32 threads and SIMD patch (14.84 avg fps):
       3398 frames 230.6 seconds 14.7 fps 40.0/67.9/2080.0/18.1 ms
       3398 frames 230.1 seconds 14.8 fps 39.0/67.7/1711.0/14.7 ms
       3398 frames 227.0 seconds 15.0 fps 38.0/66.8/1485.0/14.7 ms
       3398 frames 230.2 seconds 14.8 fps 36.0/67.7/1482.0/14.8 ms
       3398 frames 228.7 seconds 14.9 fps 40.0/67.3/1522.0/14.6 ms
   
   144 threads (17.68 avg fps):
       3398 frames 193.9 seconds 17.5 fps 31.0/57.1/2141.0/16.1 ms
       3398 frames 191.3 seconds 17.8 fps 30.0/56.3/1743.0/14.4 ms
       3398 frames 191.5 seconds 17.7 fps 30.0/56.4/1526.0/14.6 ms
       3398 frames 192.3 seconds 17.7 fps 30.0/56.6/1528.0/14.0 ms
       3398 frames 191.8 seconds 17.7 fps 29.0/56.4/1545.0/13.7 ms
   
   144 threads and SIMD patch (17.42 avg fps):
       3398 frames 206.8 seconds 16.4 fps 29.0/60.9/8731.0/21.7 ms
       3398 frames 195.5 seconds 17.4 fps 30.0/57.5/1710.0/14.9 ms
       3398 frames 190.6 seconds 17.8 fps 30.0/56.1/1538.0/14.6 ms
       3398 frames 191.6 seconds 17.7 fps 31.0/56.4/1587.0/14.5 ms
       3398 frames 191.4 seconds 17.8 fps 31.0/56.3/1510.0/13.6 ms