.Net Foundation Design Reviews, 6 Nov 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: .NET Design Reviews: Intel Hardware Intrinsics

Description

We’ll do another pass on Intel’s proposal for HW intrinsics.

https://github.com/dotnet/corefx/issues/32721

A

B

C

Hey and I tried to summarize things.

B

You have to tell you I, think yeah. Just three topics we need is yourself today: yeah the first one is written to you that yeah we needed to D stated that needs to the class named after 64-bits online entry things right now. We call it X 64 yeah, and then you can see some interesting in a car's name: the x86 tall, sse2, dou X and dou X 64. That means ok.

B

This interesting is holding for it 64-bit platforms, janna some concerns of this name is that this name may be confusing, because we we use x86 for all the platforms, yeah yeah x64, for the system.

B

Let us sometimes we use x86 432 from Poland yeah. So maybe here some somebody suggested maybe x64 a ring of 64 bits already something maybe better. Oh yeah.

C

Basically, there were, there were some people that didn't like just x64. They said that it might be confusing and the rest of the people said: X 64 is 5, and so we need to decide whether what one name of one I think Carol I suggest suggested. X86, 64, 64-bit only art, 64-bit,.

C

X86 64, as invested costly.

B

D

Think that's a little confusing because the way that usually refers to the 64 gate I said this is not talking about.

D

C

Right I think just X 64 would be more than sufficient, especially in the nested context and being under the x86 namespace to differentiate that these are only available on the 64. That is a yeah.

E

F

Top levels of the issue is titled Intel a hardware but like amd64 support such as DSP yeah.

C

This is x86 and forensics, so it works on AMD platforms as well. If AMD supports those eyes, A's right, okay and.

E

Just to be clear, every 32-bit intrinsic will also work on a 64-bit platforms, long.

G

E

Is a support over empty net.

C

And then the secondary point for for this first point was.

C

How we wanted to do the nesting I think they in his original TR he put up. He had the x64 class inherit from the class. It was nested in so at SSE, 2 x64 didn't hear from SSE I suggested. We change that such that the x64 is a inherits from the x64 I say in the next lowest face class.

C

So if you scroll up a little bit, the Pope's just above it I give an example of how this would look in the hierarchy.

C

So BMI 1 x64 just inherit from nothing object it rather than the hair game to be in my 1 and then for the long chain. You have hop count of x64 carrots from SSE, 2 x64 and then for the 4 1 2 you've got a gap because there are no 64-bit specific SS 3.

B

Yeah yeah I agree this yeah.

C

And if they did change it to this, don't we wanted to confirm that this was the desired.

E

Shape of API, would you also use? Would you also expose like an is supported, boolean property on these and use method, hiding just like you do on the outer class? Yes,.

A

It's an API, a listing somewhere for those.

C

The PR has them right now. It's a fairly.

E

Small change: this means that this would be currently for people who are consumed consuming the Premium Package. This will be a breaking change for them right because all the 64-bit API we're.

C

Making a ton of other breaking changes ever owned.

A

That kids risk is lower than breaking everywhere. Then.

E

I'm, just like internally thinking of all the code, all.

A

And it's this PR here right.

A

What is this thing on? It's on yeah.

C

It's nested under BMI 1.

C

A

My only question is: if you go down that path, what happens for specifications where we want to establish inheritance later on, like what happens for BMI if I say BMI 2, dot, X 64? That means I can no longer see everything that was in BMI 1, dot, X well,.

C

A

Guess you could dinner, you know BMI.

C

1 & 2 don't inherit from each other, because there's no codependency at the cpu level sure.

A

Well, let's say that's it: what I'm concerned about like these kind of established patterns here right so like an egg meal I think even if this particular aspect does inherit egg eventually be, may find that we want to be able to do that right. And so, if we have a pattern, if we say X, 64 specialized instructions are nested time, then you want to pop into it in other places as well as we.

A

If we have similar issues right and then the question becomes well, what happens if you do want to have a BMI I guess you can just make the B and Y 2 x64 inherit from BMI 1 go take 64 rather than you have the same experience right.

C

And that's also why this is abstract. We took that way. You can have the inheritance well, yeah.

E

I guess differences for this. Well, it is a single inheritance model thing yeah.

C

E

C

All static methods non-virtual, so we can add in we can add inheritance without it being a breaking change right.

A

That certainly makes sense.

A

Yeah that works and.

C

Then this is BMI one again except the the actual fermentation.

A

Yeah, so how do we.

G

Real dumb question about that comment, that was in there, which.

A

One where there.

G

Was a comment that said this is only available on 64-bit platforms: yeah hook line 29, really, it means processes right yep, if I'm on an x86 processor, but my process is a 32-bit process. Can I call these.

C

A

Combination of you're running on x64 hardware and you're, the next 64 process.

G

A

G

The word platform really should be processed and.

B

Also so I could.

G

Get it straight in my head: okay,.

B

A

Yeah so given I don't have an issue for that. It's just moved on and say this one is approved. So this is the first item in your list and then I.

C

Think that the third bullet point is probably easier to do than the second one. The third one is mostly just approving API shape again.

F

We've already gone through three years, intrinsics along these lines, and we developed a pattern for them, and I was doing this. Just all follow and there's the same matters or are there new things here? That was sort of need to be reviewed categorically. No.

C

For for these ones, these are things that we reviewed and approved changes on in the last meeting, but that we didn't actually come sit down and say this name was okay and when we went to do the PR there were people that were like what about this name. What about this name.

C

Okay, yeah we're basically just confirming that the decisions that were made in the PR match what the design review committee thinks is correct, except for sse4, which is the second bullet point in phase paws, and that probably requires some more discussion.

A

So, which part of car is more discretion.

C

So I'm faced post he's got point. One follow point one: two and three we'd always.

G

C

Two requires more in-depth discussion and Figgy.

A

C

Mostly the same as what we didn't want so.

A

That's not okay, thought it was another two or somewhere else, but I mean when you just dismiss so.

A

What we want to do this is this an API review issue here or.

C

C

B

A

B

Already stated the flag tenor.

G

Are you thinking, let's skip this one for now, go to three cuz, it's simpler, yeah.

C

So for this one we basically decided move all the helper methods out of the is any specific classes and just move them onto the vector types of themselves.

C

So when I went to do that for the Constructors, specifically I decided to put these into a static class. That is non-generic. So this is a similar pattern to what we did with system collections, immutable yep, just to avoid the confusion of well. What happens if you say vectors, sixty forty for a fight and then you call the constructor that takes four ins right.

A

Answer yeah yeah I mean yeah, that's the problem with constructive climate.

C

Weather, it's not working the way I wanted to so just expose the cream the create methods which the first set there that just take a value. They set all elements to that value.

C

The second set sets each respective element to the matching element in the input list and then crate scaler, since the first element to that value in the rest of the values to zero. There's an open question, which one does a great.

E

Scaler, this is that the standard terminology for that scaler yeah.

C

E

You have the first element set everything else, zeroed out. It's.

C

When you're dealing with only the first element of the vector of the scaler there's, an open question with a corresponding API proposal, which says for create scalars, are a number of performance situations where you want to leave the upper bits uninitialized and what should we call the method that does that? My original proposal, which is a 3 2 8 3 4, suggests that we call these unsafe, great scaler yon commented in and said he did not like the term unsafe, because it's a bit overloaded and there was a bunch of back-and-forth there.

C

I think Morgan, Brown and Levi discussed a bunch a little bit about this, and it's still potentially unsafe from the perspective of the upper bits, could have any value and attempting to access that could expose data from a previous method call that was still in the register. So it could expose data you wouldn't want exposed. Otherwise, but basically these suggestions were crate. Scalar, unsafe, crate, scalar, knowing it and I think a couple of things were passed on, but we need determine so.

A

Create scalar there will be lickable to all of them. We wouldn't have both create scaler and create scaler and say. Is that fair I think.

C

We would have both yeah, you may want both yeah, you would want. You would want crate, crate scalar so that you can say I explicitly want the uh permit zero, and you would also want crate scaler unsafe, which says: leave the upper bits uninitialized for performance reasons.

A

My parents would say honestly because I mean I get where we go in with the uninitialized, but like I, don't think that's don't even want, because we have two conflicting technologies and I think I like unsafe, because it's basically means read the docs right, and so you can make whatever caveat you want to express you just take. Well the kid the caveat is actually in the dogs. It's just that this is. The API is only good what they really know. What they're doing yeah.

E

I, like I, do I do understand their concern about the term. You know overloaded, though, because normally the term deals with stuff like caste and flowing identity and thread context and stuff like that, but yeah all right. It comes down to read the docs I.

A

Mean I mean I would say they all come down to the same thing. The job.

E

A

May experience absolutely unintended behavior like I mean if you don't flow execution context on court. That's not unsafe right like because there's no Kazakh or it's just like well bad things might happen, and you may not want that right or oh yeah. You can't contra in the memory and you may be right, I mean they're. All I I mean this one I would say it's honestly actually in the unsafe territory and the traditional sense I think you have uninitialized layers of the stack like I mean absolutely bad.

A

Things can happen when you do that right, I mean you could argue that you never corrupt. This state, but I mean depends on what you use to fight for right. A few. That's the size of an array then maybe they're things that.

E

Happened, if you say, if you guys, would actually also be a good candidate for some theoretical like unsafe, attribute that we keep talking about a number doing or.

A

E

Code can only be called with an unsafe block. We.

A

Talked about this at the C sharp to him and like they, they don't seem to be very care to do that. I think they they have some concerns around like you know what you know, what doesn't it mean for the language and whether that makes sense or not I mean quite frankly, it doesn't have to be this way or we couldn't like it. You know also have just an analyzer, they won't ever flag those and then maybe you suppress them with being in an unsafe context.

A

But then you have to fight the probably compiled of one that says you're saying you won say, but you're not using any pointers. So maybe so that's I think there's a few concerns around there, but I think you know I. Think the poor man's version of we're keeping the unsafe suffix is not horrible.

G

What's wrong with what's wrong with a parameter a.

A

G

Parameter to treat scaler because.

A

That was the thing we discussed in some other context before the problems you look at the code of users create scalar, 42, comma fault. I know you're like well. Is that something that this never.

G

Happened parameters.

A

To methods such as that also the fact that it doesn't really mean- even if you say, I'll, initialize, :, true or uninitialized : false right I mean neither of them really convey the idea of like oh. This is something you only really want to review, because that's that might be unsafe, but if you just call a method called create an to create scalar unsafe, it's a bit more attention-seeking right.

A

That was the intent with the whole topics and I think the same is true, with a you're forced to be in an unsafe context where, if you're in a code review with somebody as an unsafe block, you probably spend a little more psycho comparing of what's going on there, then you know a regular for loop right.

C

For the crate methods, the other change that I made as part of the PR, the existing set vector methods that we have take the parameters in reverse order, so starting with e 15, for example, and then going down to e 0. This matches the native intrinsics for x86 I change that to be normal ordering, because that I think that makes more sense for a shared help. Intrinsic. That also applies to other platforms.

C

It also is worth mentioning that for the native intrinsics there is a set R, which means that before reverse, which does take it in this order- and that seems to be what people tend to use in a lot of their native code when companies, but.

E

We would only expose this particular ordering through the API, so right, yeah and I. Guess if, if people were for some reason, converting code from C to c-sharp, we could have a tool or something the host helps free, like those yeah.

C

And then, if you scroll down past this to like vector 128 of T,.

C

A

I was just trying to find another API name using the approve, move, your head and say.

C

A

Are all net new API, that's correct, or they these.

C

A

Clients but net new methods going yeah.

C

There are new methods, but will be deleting this so they're like, for example, for create, create 128. Here we already exposed SSE dot, set vector 128 with it. We're.

A

Getting rid of that in favor.

C

Of this, for you and.

A

Then, of course, instead of classes on exist a guy's a bit like that yeah yeah.

C

For the concept of extend, which is taking a vector 64 in turning into a vector 128, there was some discussion on whether we should expose it as a create method. Here, so I've got the create lower upper and the crate scalar or exposing it as an instance, method called like extend to 128 or something.

C

You also have the concept of now, if you, if you're doing, creating a scalar of a vector 128 from a 64, you also have the same concept of do. I want to 0 the upper bits board or I want to leave them uninitialized, which is why I opted to suppose it is create scalar, but.

A

Scalar will always use the lower order. Suppose yes,.

C

Alright, because.

B

Yeah but I prefer to another name like the extension, because yeah scalar or sometimes means just the first value of the vector is available, but here for extend actually the harm. The first half of the vector it's available and on the upper half is 0 or non deterministic yeah. It.

A

Could scanner seems a bit weird here, given it you're, not good at planning units Gator, they.

E

Said would it make more sense for us to remote creates they were entirely and just have that single, create method around two-thirds of the way down and say caller you're responsible for telling us what's part of it? He went to zero out that might make the call cited. They come for something well.

C

So that doesn't cover the case of you want to set the lower bits so like for x86 code, you don't have victor 64. You really only deal with 128 and 256. All right, you, you have a 128 and you want to turn it into a 256. But you want to leave the upper bits on initialized and.

D

You also get unsafe, yeah create unsafe and then passing the hundred thirty eight bits for 256 P right.

C

So if we did something like an instance method extend to vector 128, we would like we had better 64 X 10 to 128. We would also need in extend to 128 unsafe.

E

And that would be an instance method. Rather myn estat accompany is I'm, so I don't see a like a reinterpret cast. Nothing in here is I'm innocent about it as well. Yeah.

A

I mean that one seems reasonable. I mean none of the. My training is correct: they're all just yeah.

C

These are all helpers implemented in.

A

Yeah I think zeroing out manually is just annoying. That I mean it defeats the point of having other method to begin with, especially if you look like bites or Schwartz is like Willie. You want to want me to set like 128 individual items, not quite right but yeah.

A

So what did we decide? Then? We just leave it as an instance method, I call it extend and say any extended, extend and say I.

C

Think it needs to be extend to vector, 128 and extend to vector 256.

A

Yeah because you're yes.

C

We have to clarify the return type yeah.

B

All we can call them the zero extension for the sleep version and extended to for the concert version.

G

Unique three right: there's three methods here: one is extending it where they're, both from 64 to 128, extends to they're, both upper and lower right.

G

One word upper gets zeroed and the lower sets of a 64-bit one, and then one more uppers left uninitialized and.

C

The lowers do we want to it's for the create vector 128 from two vector 64's? Do we want to have that as a create method or as an extend method that takes that explicitly takes offer.

G

It just feels like the create map, the cream, if it seems cleaner to me for the upper and lower.

G

But then it's like, why do I know which one to go to right.

A

This extent to an order that we usually use because there's the thing on the other one we had widen and narrowing, where is that different I suppose, is that the same isn't widen.

E

Widening the individual elements, rather than just appending zeros to the others. Okay,.

A

So we can either say: go figure out the design of this one. Let go one way or the other, but uh we can make call here like do we do we feel like we can't make a call here. They want to go back to the drawing board and think about this. A bit more I think.

C

What probably makes the most sense is to have the explicit, lower and upper as a create method and then have the and then have a terrific.

A

All of them were put adjust for vector 64 to the vector 128 for each of them. Yeah.

C

So so for for creating a vector 128 from 264 bit vectors or for creating a 2 v, vector 256 from 2 vector 128, we should have a create that takes lower and upper and then for the case, where we're 0 extending we should have a a 2 or either an extend 2 or a 2, vector 128 method on vector, 64, I. Think.

D

That makes sense, because 4 will create to create vector 256. For example, you could have four vector 64, it's yeah. Well, it doesn't make sense for that, to be instance, method or something right, I think it makes sense for that. Okay and then extending makes sense for the other two. So.

C

Then the question we do, we want to call it to or extend to I think.

G

The semantics are, it's always the lower one is what's at right, right, yes, okay and then the by the dual nature of. Do you want the upper zeroed or uninitialized right? Okay,.

E

Probably they extend to because you'd probably want suffixes after that 128 256 512 great.

C

So extend to vector 128 in extent any vector, 128 unsafe and then extend to you back to and extend two vectors. Okay, when.

A

You can't think of them as simple type, conversions and so like to type name, might not be too bad. Cuz I mean I could save it. You know by two in would work why the initialize the lower version of the end and make the rest is zeros right and.

E

If, for whatever reason, it doesn't trust me in too many characters, we should also really just say what the extender 128 extended to produce things if it ends up yeah.

A

I can go either way with just two or.

A

The xxx would be the signs right, yeah, the thing vector doesn't add much right. Well, I guess, would you convert or anything but a vector ever like? Would you ever convert to a an actual literal s or an actual scalar problemo.

C

So we have a scalar method, I'll be if you scroll down to the actual generic type, it's an instance method, a scalar, but that's different, because it's not a conversion. It's a it's just pulling the bottom element out. Okay,.

A

That's a good plan. We should probably call it to them. We should probably call it as because Kim was always this expensive operation kind of thing and answers. Always this interpret the already existing bits. Image in you would expect those to be cheap operation. I. Don't expect this to be super expensive.

B

E

Ever need a like I get vector, 128 dogs and Road. We ever need like dots where it every bit is set for whatever reason, dot.

C

One yeah all right, I guess not one true I, think it's I think it falls into the realm. Oh yes, it would be theoretically nice if we had it but I'm, not sure if it's worth exposing as part of public API, unless we get feedback requesting it.

C

Oh have all the arranged with the council yeah.

E

C

Reinterpret cast zero as scalar, which just gets the lowest element T, so e zero again can set element, helpers, get lower and get upper, but no set lower set opera as of T as a generic, so that was they they and a few other people have brought this up multiple times that some people want in as T you last time, I asked this I think just MO and Jeremy and maybe Christoph were left in the room and they said we shouldn't expose.

C

It just tell people to use unsafe, as but I said that on the PR there was a few people that came back, push back and so I said, I'd bring it up again. Yeah.

B

But there is a problem that if we use unsafe and the cogent is rely on the optimization of JIT compiler, especially the inlining, but we have defeated. That's all the heroin trees and have no cogent difference between here, zero or higher up here. So that will get some job, always okay, your program from zero, zero that the unsafe as well generates a function call and some redundant.

E

So what we actually have scenarios right now and within the runtime, where we do as or vector of T and.

G

E

So say that you're calling index hub, for instance- and we know that the incoming T is int char byte. As might something like that right. What we do is under the covers. We turn it all into bytes. We turn it all into a bachelor by hand, and then we do our our search operation and then we need to take that vector a bite and then convert it back to the original vector of T.

E

But now, if these are all the api's we have upside from using unsafe now we would have to have a giant switch statement. That says what was the actual we're trying to converter.

C

Right so so, when writing generic code not having a generic ads, does make code more difficult to write. Yes,.

E

It forces us to fall down to understate, which, again we can do in the run time, but we don't necessarily want to force our customers to have to write upset code. What.

D

Is the operation I would say.

C

It's you, it's not. You have to fall down to C sharp unsafe code. It's you have to fall back to component.

A

No, but next, but is the operation like giving like other any combinations of T's and use where it is actually unsafe? If.

C

T is not a primitive type, but.

A

C

Fail down at that point, yeah.

A

The operation would just blow up, you will just get so it's always safe. You might just get an exception at runtime. Is that.

C

Yeah yeah getting exception. If you did like vector 256 approval, yeah.

A

And their kids I think having a generic conversion seems fine, I think the last time. Let me talked about this. The question was just: do we want to allow people to write generic code and I think the answer was generally no, but but you just an artist, not right. Engineering Co, ideal code is to just say: I started with a vector of char and I want to go to a vector of wine. I'm gonna go back to a vector of char, but I.

A

Think it's you're, not really writing generic code that operates over arbitrary T's right, because if you need to do that, like everything fights against you, why do you have to like yeah.

E

Yeah yeah we have, we have. We have generic methods because we expect to work with. You know the tenor, chloride.

G

For primitives.

E

But we expect to go down the single helper method and then eventually turn it back into what the user expects care.

E

Policy I think.

A

It was like also some arguments in Europe like floating bytes like you know, there might not be great, but then you know again, if you we should their own up there. No.

E

People- it's not it's not unsafe to do that. It's just the data gets corrupted election mark.

A

What even that might be depending what the algorithm is? Actually what you want man again.

A

So I think yeah I mean given it easy money. If you guys, they expect. You know more to use I think that's probably also okay,.

A

All right so then we covered that part to.

C

Just general questions of our epi shapes correct. Oh.

B

One more question about the data element, so here the settlement return type is void. That means that is a mutable operation, but uh yeah.

C

B

But yeah we this data, all the interesting, should be pure functional and why don't we have winter and X? You know for that water method. Oh, why.

C

Didn't you index because it if, if you don't have hardware acceleration, it is not a cheap operation.

A

Well, what should be used two times then.

C

It's it's a case of.

A

Like I mean to find cheap, I mean like your your computing, a memory slot and you put the video in there and I mean hooks, but shouldn't it be with it's. Not it.

C

Compared using the hardware, so the hardware intrinsic is like one cycle and otherwise you.

G

C

Store it, you have to store it to the sack, read the address out of that sack and then return it right.

A

But I mean you.

C

A

That the vast majority of people, the news vector of 128 I'm gonna- have hardware acceleration right. That is a cheap operation right yeah, so I mean I. Think it just makes code really ugly to use. If you have don't even index all right. So.

C

Just I wouldn't say: I mean get into.

A

Element of typical things that we'd use indexes for now won't even go a step further and say even if you're not accelerated the amount of row, but usually do reineck, sir, it's probably all in the mind to more expense to them.

A

When you guys have here all right, like it's really like really well in excess air power properties they only about if it's a no one operation right, you don't want to have an Owen or other crazy stuff, but like expensive is you know the outer beholder, but I mean computing, a memory address and writing into it. I mean that should be in the realm of what we consider cheap right. Okay, something.

E

That we may want to do as documentation here is like looking at house name 32, for instance, because this is a reinterpret cast rather than an actual conversion like if you have a vector 128 of floats, and you call as into 32 like we should document this isn't going around like it's actually going to give you. The mains gets that worker float.

C

Basically, that does the same thing: expand.

E

Yeah I think the API is okay. There, it's just a dogfish, yeah.

A

So you guys don't want some other api's for aesthetic methods like one or something you think they had them on the other ones.

C

A

I'm just asking, because we I think we had two in the other type and I remember: somebody was very mentally fighting for I did for heading it or something III.

C

Think, having a once, we had the jet support for optimizing constructors that take all constant values. Then having a public static, read-only vector, you know, 256 true equals static cast from you know, create you int 128 or whatever it's going to be super cheap. The JIT will optimize at all and it won't be a concern. I, don't think we should be exposing values we think might be common other than zero and less people explicitly. After all, that's.

A

A

I think I made that argument last time. I lost it. That's one.

A

All right, so, okay, that's the only other side variation. This is sorry.

E

You go back up to oh.

A

Yeah thank the only thing. We should add a thing. We want to add the s. What is it as T coma? You yeah that the shape you want, though,.

C

That's the same shape as we have burden what.

A

Do we need to generous, where isn't it just one? If.

E

It's a static method. Oh it's.

A

E

A

We may decide we're through.

C

Yeah, we could just do it as you yeah, because.

A

I mean if you have two of them is always ugly because you never get one of beginning. You cannot partially infer until what it does best by both yeah and given that you're passing in an instant seems dumb. It's like there's only one correct answer.

E

For 256, you have hopper api's to get the lower and upper components as vector 128. Would you ever need the vector of 64, or did you just call like get lower, get lower, I? Think.

C

Just yeah, otherwise you get into a confusing state of what do you call me mid made upper yeah.

A

Bigler middle, less middle middle one little.

A

All the way, don't know anything all right. So then we talked about the third one. That was your quick one right.

B

Actually, last review something about the black original. We have. One really knows that to expose all the lab values secret.

F

And all to massive yeah, yeah but I, know yeah.

B

But there are some overlap there that skin exposure that can have different somatic in different environments, so we exploded them to three types yeah and we named this values to as close as possible to the the past pacified. Without luck here is a example just now yeah, so I think for the flat insider yeah. We already found butter for the intrinsic in name yeah. There are some feedback yeah, for example, the the state +1 slider. We directly use some flux name in the intrinsic name.

B

However, here the flux, for example, the state flag, Oh blog as flock that in normal instructions, for example, to say, means --carrie and all means overflowed, but in this enchanting and it is instruction, it means total different things. So we decided to include a somatic referring to that right. Yeah functioning, for example, that yeah here we yeah for example. Here we rename the state flock to compare pass match, has an edge yeah and the knots a flood name, the to compare no match yeah.

B

But there is one special thing is not stay flock and another thief flock. Doesn't we rename is to compare no match and the rights not terminated right, Carol suggested that's okay, yeah that last one, but that may be too verbose. But, okay, we did so far. We didn't get a better name. This.

C

Also impacts sse4, one test test, see we current. We currently have one SC.

B

You can see.

C

B

But I think that's the lines. Okay, but in fantasy and that's that's means yeah same thing: yeah.

D

Asking for flat yeah.

B

Yeah yeah yeah, the testator song and the thing means the arrow. Yeah technically also means.

C

I think I think my point there is that, if we're one, if we're having to change them off of see in the processing for one, we shouldn't have them as seeing Z for okay.

B

C

Verb sse for one, oh yeah, that makes sense to move so so if we can come up with something that makes sense, we should update both, because it has not see and not see, doesn't can very much unless you I.

A

Mean like the more specialized instructions are right. Then you pass in like a bunch of arguments that do some crazy permutation on the arguments when, basically, you have to read the docs to know how to use the API effectively.

A

I think like at some point the questions like how else we want to go to be right, like I mean if it's already like like a thing, I agree with for creators here is that I mean I like API calls where you can read the code, make sense of it, but at the same time the question is: does it add value right like if I still cannot reason about it? But now I have to write like a hundred ex of what I'm?

A

Normally writing at that point, you're, probably better off or like you know, it's good off, I. Guess then, just using a shorter name and just say no, you have to read the docs I, don't know.

E

A

I mean hopefully, our own dog people when we document the heart for intrinsics will do a good enough job that I may just get away with hitting f1 like. If that's not enough, then that's sad but then again like if yeah. If you have to read the introduction- and the other question is like do I have enough like. Are these the right keywords that have be very likely to find something in your dog's? The answer is: no, then we're actually doing a disservice by making it like.

E

So one of the problems we have a lot in recent weeks was I have a very particular instruction. I was trying to call and I'm, like I, don't even know which public API corresponds to this instruction right like if you could somehow make that better, then the might not be all that important. Well.

A

My suggestion was earlier: it just should have the actual instruction names in the docs and then, ideally, if our dogs are indexed in some same fashion, you would to be able to just enter the actual instruction name in the search, and then it should find the public API that doesn't and.

C

We tried to do that in the so light for all of our existing API surface area. We are current dot. Comments are basically here's the actual x86 instruction in the operand encoding that supports and here's the native intrinsic it's.

B

A good point: if.

C

We had something like the Utah intrinsic site that you, you can also do a cross reference search thing yeah. That would be infinitely better yeah.

A

I'm not advocating for I'm not going to school I am on the 28th right. I would say at some point like it's okay to say, though, that is different naming conventions too long, as it's somewhat mechanical translation to the other, but like once, you start running out of english word to describe operation like it I think it's like wow. It's.

E

A

E

Specialized operations at this point, yeah.

A

Like you're you're.

E

Getting into offers operations that perform very specific to training there are sites right. Is there scenarios? That's fine. Could.

C

We presented things 'flag see fine if we want to keep similar names, but still have them more dotnet II. Could we just call it like.

C

Like carry zero sign over low so like taxi, you would be test. Carry you.

E

Know like that's, if I'm reading, if I'm reading the documentation, I'm, probably more used to the CFS hostile.

G

A

E

Which means that I'm, looking for specifically the words very easily by myself, yep I, mean.

A

I found that I think this isn't the realm of what I said earlier about tried to make the amiko mentioned fit, but you try to preserve. You know the spirit where it's a mechanical translation and you're likely to find this stuff right. We.

E

Also, don't want to have something: that's so radical and.net from what c++, like you know, someone ever the most awesome yeah.

E

A

You know I mean I, mean I, think our feedback was the game. He looked at the API for like ten seconds and they said. Oh, this looks complicated. It can be simplified, that's right, but then, if the entries are this more complicated them? Yes, it's like going the wrong direction right, but yeah. We have some ideas like this. That's like yeah I'm,.

G

Not a hundred percent sure I'm on board I guess with this plan like it, it just feels like it's such an implementation, detail of the processor. Then it uses the carry flag in the zero flag and the sign flag right and then to keep propagating that implementation detail all the way up. Just it doesn't seem I guess that's. It doesn't.

A

G

A good API right, I mean.

A

I mean I would say, I'm not sure, I'm a hundred percent disagreeing of you. At the same time, though I'm not sure I agree hundred percent, because if you think of hardw intrinsics I mean they're not in the abstraction over the processor right, they're, literally giving you the instruction like access to the instructions, but still preserving the fact that you pass pages, you don't have to write the register locator in your head right so that that's fundamentally, what is api's are doing right. Yeah.

G

Trying to give you like 20 minutes, we just spent on the other section, was like explicitly going away from how the C API a lot right. If I'm looking for rosette scaler, it's like well, no I got a go, call create on the thing right, which is a completely different verb and think I have to make that mental mapping anyways and a bunch of these. Those.

A

Are so like those ones, I would say are different because they just help us but they're, not actually in Twinings but they're, not giving you anything from the processor. It's just or whatever data hole that you have I mean. That's the API servers there right. So that's why I think of them more thing. I'm, just listening.

G

That is like an example, there's a there's, a bunch of other ones where we've gone away from like the C API. Did it this way yeah and we're like no we're explicitly doing it. The more natural c-sharp API way clear.

A

I'm not advocating for this mirroring blindly the the I wanted kind of, like you know, underscore no school, 1:28 writer like anything missing.

G

Carrie flake and zero flake to our test zero or test Carrie. It's like that doesn't make any sense whatsoever you from an API perspective, I.

C

Think it would depend on so maybe for SSE for one test. It does make sense, because it's actually testing for Carrie and stuff, but this one where Carrie doesn't actually mean Carrie.

G

C

I guess the question we when Levi goes and updates our span parsers to use these.

H

C

How how thing of a headache is equal in to get for writing this code and trying to figure out what means what yeah.

H

A

Mean to be clear like a month and.

G

Another person coming in trying to read it also write the next person who wants to come in and read it and figure out what does carry flag mean so yeah, maybe I missed the part about two for two or the for one test stuff like if that's really what it's doing then calling it to carry plague or the see flag or whatever totally makes sense, but in this case with the strings, it's like. We are using this flag to mean this completely other thing: yeah. That's.

E

The other name is I.

A

Mean maybe begin you said before the names a bit better like, if you mean, if your suggestions for how to rename them, if they think the bigger problem here is not so much. The naming themselves is just the fact that it's very verbose and it does not make it easier to understand. Maybe it's just a function of what getting better names here. Most.

C

Of the names are also pretty short, it's just the Nazi Nazi flag, one.

G

That's the hardest, one for sure I mean I'm, the original PR. Oh, we had, we had suggested I think at least three or four different ones. All right this this list does seem like the best of the suggestions, I guess, but we can go over that last one again.

A

All right, then, I'll record that and then I think we get proof the first one, the last one, the middle one. We just need to review again. I guess all right! Thank you.