There is indeed a time increase, although it is much smaller than my measurement precision. A comparison for JIT: 1st run (empty disk cache) File Old New Timing difference jit.cpp:146 generate functionname generate functionname = jit.cpp:147 hash functionname hash functionname = jit.cpp:152 search mem with hash functionname search mem with hash functionname = jit.cpp:154 not found in mem cache not found in mem cache = jit.cpp:155 generate full code generate full code = jit.cpp:159 generate hash full code generate hash full code = jit.cpp:171 call getKernel call getkernel = kernel_cache.cpp:82 hash functionname hash functionname = kernel_cache.cpp:93 search mem with hash functionname search mem with hash functionname = kernel_cache.cpp:95 not found in mem cache not found in mem cache = kernel_cache.cpp:101 combine hash full code {jit_cl_src, jitKer_cl_src} 4Bytes kernel_cache.cpp:104 combine hash options " -D USE_DOUBLE" 14Bytes kernel_cache.cpp:112 search disk with hash functionname search disk with hash full code = kernel_cache.cpp:113 not found on disk not found on disk = kernel_cache.cpp:118 compile compile = kernel_cache.cpp:118 save on disk with hash functionname save on disk with hash full code = kernel_cache.cpp:129 save in mem with hash functionname save in mem with hash functionname = kernel_cache.cpp:136 return kernel return kernel = jit.cpp:175 return kernel return kernel = 1st run (cached on disk) Old New Timing difference jit.cpp:146 generate functionname generate functionname = jit.cpp:147 hash functionname hash functionname = jit.cpp:152 search mem with hash functionname search mem with hash functionname = jit.cpp:154 not found in mem cache not found in mem cache = jit.cpp:155 generate full code generate full code = jit.cpp:159 generate hash full code generate hash full code = jit.cpp:171 call getKernel call getkernel = kernel_cache.cpp:82 hash functionname hash functionname = kernel_cache.cpp:93 search mem with hash functionname search mem with hash functionname = kernel_cache.cpp:95 not found in mem cache not found in mem cache = kernel_cache.cpp:101 combine hash full code {jit_cl_src, jitKer_cl_src} 4Bytes kernel_cache.cpp:104 combine hash options " -D USE_DOUBLE" 14Bytes kernel_cache.cpp:112 search disk with hash functionname search disk with hash full code = kernel_cache.cpp:112 found on disk found on disk = kernel_cache.cpp:112 read from disk read from disk = kernel_cache.cpp:129 save in mem with hash functionname save in mem with hash functionname = kernel_cache.cpp:136 return kernel return kernel = jit.cpp:175 return kernel return kernel = 2nd run (independent if previous cached on disk) Old New Timing difference jit.cpp:146 generate functionname generate functionname = jit.cpp:147 hash functionname hash functionname = jit.cpp:152 search mem with hash functionname search mem with hash functionname = jit.cpp:154 found in mem found in mem = jit.cpp:175 return kernel return kernel = As you can see, ONLY AT 1st CALL the existing common::Source.hash (header, generated code & options) are combined --> 18Bytes (18x XOR, 18x MUL). For all following runs, no extra time is spent. af::timeit() will even not contain these extra operations, because the func is called in the sample_times before the real measurement starts. By the way, static code performs the above operation EACH time it gets a Kernel, since it calls getKernel each time. BR, Willy

@umar I first included the AF_REVISION, as proposed. l still had to clear the cache for each test run, because this version number is constant during development recompiles. I assume the version nr changes after a commit in GitHub, which is not the case during development. On my system they never changed. With the static kernels, I did not have this problem so I searched further to understand why, with this PR as result. Regarding the timings: the 1st run gives a theoretical time almost equal to the static kernel (any run), where the combining of the hashes is always performed. I did not give times, because as soon as you put it inside a loop, the extra steps are no longer executed. The measured times, before or after the change, are as result equal. (I believe we are talking here over microseconds, if not picoseconds). PS: Only the combining is performed because the hashing on the code text is already performed before the call, also before the change. These hashes are stored in common::Source struct and are present in the parameters of the getKernel call. (also for JIT)