@nemequ Do you remember that the Intel compiler generated bad code for some of my GFNI implementations with
I would like to remove as many of these #ifdefs as possible.
I am also wondering whether this may have been due to the extra complexity for the compiler caused by using simde prefixed intrinsics in the implementations and I would like to remove the prefixes (except obviously where it is necessary or offers real advantages).
Unfortunately I don't have the Intel compiler installed and I don't really want to install it on my old laptop which has very limited resources.
So this might entail many iterations through CI which could eat up valuable tokens, as I understand was the case with Travis.
What do you think?
SIMDE_VECTORIZE and Michael (or maybe you) conditionally #ifdefed out SIMDE_VECTORIZE everywhere?I would like to remove as many of these #ifdefs as possible.
I am also wondering whether this may have been due to the extra complexity for the compiler caused by using simde prefixed intrinsics in the implementations and I would like to remove the prefixes (except obviously where it is necessary or offers real advantages).
Unfortunately I don't have the Intel compiler installed and I don't really want to install it on my old laptop which has very limited resources.
So this might entail many iterations through CI which could eat up valuable tokens, as I understand was the case with Travis.
What do you think?
7 replies
hello, i was trying to implement _mm512_mask_or_pd function and while generating tests i encountered "SIMDE_MATH_NAN"in one of the test vectors. Also, another test vector had "Not enough space to write value (given 53 bytes, need 331 bytes)
SIMDE_FLOAT64_C(-14935775276824340548037239704596964,".
Can someone advise me about what to do with these ?
SIMDE_FLOAT64_C(-14935775276824340548037239704596964,".
Can someone advise me about what to do with these ?
@simba611, It sounds like you're trying to do bitwise operations of floats which is very tricky. Even if you assume IEEE 754 representation, you have to be very careful about crafting the values so they don't write tons of data to the test cases. It's generally better to just avoid this by creating tests which generate and check integer values, and use functions like
_mm512_castsi512_pd and _mm512_castsi512_ps to convert to floats, perform the bitwise operation, then use _mm512_castpd_si512 / _mm512_castps_si512 to convert the results back to integer types.
also, x86 svml test for intel-all-gcc-10 test is resulting in timeout for me, is that something i should be worried about ?
Are you running docker inside of a vm or something?
it ran fine indivisually, real 0m8.181s
user 0m6.678s
sys 0m1.397s
user 0m6.678s
sys 0m1.397s
Hi, I attempted to code _mm512_mask_or_pd and made the changes you suggested with respect to using integers for testing.
Can you please look at them once to make sure everything is fine since this is my first attempt at implementing a function ?
The changes I made are reflected in these two commit patches -
Can you please look at them once to make sure everything is fine since this is my first attempt at implementing a function ?
The changes I made are reflected in these two commit patches -
- simba611/simde@ae16975
and then to take care of the problem due to floats, - simba611/simde@5e1a1f1
Hi, I submitted the pull request but one required test failed, for emscripten, with errors of the type "constant expression evaluates to -123188372992032540 which cannot be narrowed to type 'int_fast32_t' (aka 'int') [-Wc++11-narrowing]. " I will make sure to test for emscripten in the future. Are there any specific compilers I should always test with or should I test all of them always ?
Also, can someone please explain the cause for this error ?
Also, can someone please explain the cause for this error ?
thanks for the comment
@simba611, here is an up-to-date list of missing functions: https://github.com/simd-everywhere/implementation-status/blob/main/x86.md
btw https://simd-everywhere.github.io doesn't go anywhere only https://simd-everywhere.github.io/blog works
@mr-c, I'm going to try to expand this (a lot) into something more reusable, but: https://paste.centos.org/view/e1245e50
3 replies
Hi, I implemented mm512_insertf32x8 with their mask(z) variants. They were slightly different from the other functions I have implemented so far.
Could you please review it once ? This is the commit patch -
simba611/simde@384831a
Could you please review it once ? This is the commit patch -
simba611/simde@384831a
@simba611, you can just file a PR to get it reviewed; it's easier to comment on the code that way. What you have looks mostly right, though you should be using
float instead of int32_t in the tests. Using integers instead of floats is really only for bitwise operations since they regularly create NaNs, but likely will not have exactly the same bit pattern as SIMDE_MATH_NAN / SIMDE_MATH_NANF, so there isn't a good way to recreate the exact bit pattern without using integer types instead. Also, instead of choosing a couple fixed values for the immediate mode parameters, you can use the SIMDE_CONSTIFY_* macros in the tests. For an example, see https://github.com/simd-everywhere/simde/blob/854f91319dacd1d91c5990c0334ff774eaf5c6b4/test/x86/xop.c#L7011
The CONSTIFY stuff is newer than a lot of the tests, especially the x86 ones (which are generally the oldest), so a lot of tests don't use them, but newer tests should. They are pretty straightforward, see simde-constify.h. Basically, they create a switch which chooses a constant value based on a non-constant one. So for simde_mm512_insertf32x8 the imm8 parameter should be one of two possible values (0 or 1), so
SIMDE_CONSTIFY_2_ would be appropriate.
There isn't, no, but it's pretty easy to go from a basic image to something which will build SIMDe; our requirements are pretty light (basically, a C compiler). The docker image is just large because we install lots of emulators and compilers, but if you just want to target one or two architectures you really just need meson, git, and either gcc or clang.
@refactormyself, if you start with a basic debian or ubuntu image, I think
apt-get install meson git gcc g++ will get you everything you need without wasting a ton of space. Or, if you just want to use SIMDe in your code you don't really need anything additional; just put a copy of simde in your source directory and include the header(s) you want, no build system necessary.
Also https://github.com/simd-everywhere/simde-no-tests is really neat to include as a git subtree/submodule. Makes everything super easy.