Warning, /eic-opticks/ana/rainbow_cfg4_notes.rst is written in an unsupported language. File is not indexed.
0001 rainbow_cfg4.py performance notes
0002 =====================================
0003
0004 Seconds for rainbow 1M
0005 -----------------------
0006
0007 CAVEAT: using laptop GPU with only 384 cores, desktop GPUs expected x20-30 faster
0008
0009 Disabling step-by-step recording has large improvement
0010 factor for Opticks of about x3 but not much impact on cfg4-.
0011 The result match between G4 and Op remains unchanged.
0012
0013
0014
0015 Seems like reducing the number and size of
0016 buffers in context is a big win for Opticks.
0017
0018 With step by step and sequence recording::
0019
0020 Op 4.6 5.8 # Opticks timings rather fickle to slight code changes, maybe stack
0021 G4 56.8 55.9
0022
0023 Just final photon recording::
0024
0025 Op 1.8
0026 G4 47.9
0027
0028
0029
0030 CAVEAT: above Op is Op/INTEROP
0031 --------------------------------
0032
0033 Actually this behavior is for Opticks INTEROP mode using OpenGL
0034 buffers, in compute mode with OptiX buffers there is almost no
0035 difference between enabling step-by-step recording and not.
0036 It seems like OpenGL constrains performance once
0037 total buffer size gets too big.
0038
0039
0040
0041 Matching curand buffer to requirement
0042 ---------------------------------------
0043
0044 * tried using 1M cuRAND buffer matching the requirement rather than using default 3M all the time,
0045 saw no change in propagation time
0046
0047 ::
0048
0049 # change ggeoview-rng-max value down to 1M
0050
0051 ggeoview-rng-prep # create the states cache
0052
0053 # opticks-/OpticksCfg.hh accordingly
0054
0055
0056 Compute Mode, ie no OpenGL
0057 -----------------------------
0058
0059 Revived "--compute" mode of ggv binary which uses OptiX owned buffers
0060 as opposed to the usual interop approach of using OpenGL buffers.
0061 Both with and without step recording is giving similar times in
0062 compute mode. This is very different from interop mode where
0063 cutting down on buffers gives big wins.
0064
0065 ::
0066
0067 Op 0.75 0.65
0068 G4 57. 56.
0069
0070 A related cmp mode controlled by "--cmp" option uses different computeTest binary,
0071 is not operational and little motivation now that "--compute" mode works.
0072 Could create package without OpenGL dependencies if there is a need.
0073
0074 ::
0075
0076 ggv-;ggv-rainbow --compute
0077 ggv-;ggv-rainbow --compute --nostep
0078 ggv-;ggv-rainbow --compute --nostep --dbg
0079
0080
0081 * look at how time scales with photon count
0082
0083
0084 Split the prelaunch from launch timings
0085 -----------------------------------------
0086
0087 Kernel validation, compilation and prelaunch does not
0088 need to be done for each event so can exclude it from
0089 timings.
0090
0091 Doing this get::
0092
0093 Op (interop mode) 1.509
0094 Op (--compute) 0.290
0095 Op (--compute --nostep) 0.294 # skipping step recording not advantageous
0096 Op (--compute) 0.1416 # hmm some performance instability
0097
0098 In ">console" login mode "ggv-rainbow" gives error that no GPU available
0099
0100 Immediately after login getting::
0101
0102 Op (--compute) 0.148
0103
0104
0105 Testing in Console Mode
0106 -------------------------
0107
0108 ::
0109
0110 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_170136/t_delta.ini:propagate=0.14798854396212846
0111
0112 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171121/t_delta.ini:propagate=0.44531063502654433 # try >console mode
0113 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171142/t_delta.ini:propagate=0.45501201006118208
0114 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171156/t_delta.ini:propagate=0.33855076995678246
0115 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171213/t_delta.ini:propagate=0.46851423906628042
0116 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171226/t_delta.ini:propagate=0.33861030195839703
0117
0118 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171527/t_delta.ini:propagate=1.5933509200112894 # GUI interop mode
0119 /usr/local/env/opticks/rainbow/mdtorch/5/20160102_171548/t_delta.ini:propagate=0.27229616406839341 # GUI --compute mode
0120
0121 Immediately after switching back to automatic graphics switching, then shortly after that::
0122
0123 0.142
0124 0.293
0125
0126
0127
0128 To do the nostep check
0129 ------------------------
0130
0131 After standard comparison::
0132
0133 ggv-;ggv-rainbow
0134 ggv-;ggv-rainbow --cfg4
0135
0136 * recompile optixrap- without RECORD define
0137 * run with --nostep option::
0138
0139 ggv-;ggv-rainbow --nostep
0140 ggv-;ggv-rainbow --cfg4 --nostep
0141
0142
0143
0144
