Re: Binary vs Text

Thanks for the issues! I took a glance and it seems like useful feedback.

For what it’s worth - we have an actual implementation of WHLSL on top of Metal, so it seems implementable on at least all the GPUs supported by Metal.

It’s possible that either we’ve missed some important subtleties, or the set of GPUs you know is disjoint from the set of GPUs supported by Metal

Regards,
Maciej


> On Nov 12, 2018, at 4:18 PM, David Neto <dneto@google.com> wrote:
> 
> Hi.  I'm very late to this thread (sorry).  I have not caught up on it (double-sorry).
> 
> At the moment WHLSL is unimplementable on the GPUs I know, due to oversimplification in WHLSL's memory model.  I have filed 7 issues against the WHLSL spec. (#247 through #253).  Some are more fundamental than others.
> 
> thanks
> david
> 
> 
> 
> On Mon, Nov 12, 2018 at 6:54 PM Maciej Stachowiak <mjs@apple.com <mailto:mjs@apple.com>> wrote:
> 
> 
>> On Nov 12, 2018, at 2:41 PM, James Darpinian <jdarpinian@google.com <mailto:jdarpinian@google.com>> wrote:
>> 
>> > Too much complexity is bad for both humans writing the language and software consuming it.
>> 
>> You are conflating two different kinds of complexity. Features that make reading or writing the language less complex for humans may make the implementation more complex and vice versa.
> 
> Sure, this may sometimes be true. Other times these two notions of complexity are aligned. 
> 
>> Unfortunately the evidence that you request can only really be gathered by implementation experience. If we continue down the path of implementing both SPIR-V and WHLSL ingestion, I doubt that will help us come to agreement. The evidence we gather is still subject to interpretation which we will likely still disagree on, and the more we build the more we will have to throw away, which we will naturally be reluctant to do.
> 
> I share your concern. I’m not sure how we get out of this impasse. One thing we could  try is to lay out up front what sort of evidence, if presented, would lead each side to change their position.
> 
>> However, we have a lot of evidence already available to us from previous implementers of modern graphics APIs, who have universally chosen to provide ingestion formats that are different from their shading languages.
> 
> To the extent that this is true, I don’t place a lot of weight on it. The web is a different environment, and often the right choice for the web is different. If you choose web formats as the reference class rather than shader formats for modern graphics APIs, the evidence is strongly in favor of text based formats. I think “web-based languages" is a better choice of reference, because almost every modern technology has required significant rethinking and adaptation for the web, and many of the lessons learned are universal.
> 
> 
> It’s also worth noting that DX12 and Metal both have not chosen to make an exclusive ingestion format that’s different from the shading language. It’s possible to use the actual shading language at runtime. In the case of Metal, the binary format is not even a compile target for third parties; the only official input point is Metal.   It’s just that apps are allowed to bundle a precompiled binary shader. Failing to directly handle a human-authorable format at all would be the more unusual choice.
> 
> 
> Regards,
> Maciej
> 
>> 
>> On Thu, Nov 8, 2018 at 11:59 PM Maciej Stachowiak <mjs@apple.com <mailto:mjs@apple.com>> wrote:
>> 
>> 
>>> On Nov 8, 2018, at 2:51 PM, James Darpinian <jdarpinian@google.com <mailto:jdarpinian@google.com>> wrote:
>>> 
>>> > Specifically, I don’t agree that the ingestion format can or should be “non-evolving”
>>> 
>>> Let's put that question aside for now. I'd like to find some things we can all agree on.
>> 
>> It’s good to find things we can agree on. It’s also important to be clear about what we don’t yet agree on. I’ll try   to do both.
>> 
>>> Can we agree that the ingestion format and the shading language have different requirements that sometimes conflict,
>> 
>> Depends on what you mean by “sometimes". I think I was pretty explicit about my position, but to state it again:
>> 
>> - I agree that it’s possible in theory that we could find a such a conflict.
>> - I don’t agree that we have already found one.
>> - I agree that if we find a conflict, this may push us to use different languages for these things, if the best available compromise between the requirements is more harmful on net than the harm of having two separate languages.
>> - I note that even for a single purpose of a language, there may be conflicting requirements that call for tradeoffs to be made.
>> 
>> 
>>> and in particular HLSL compatibility vs. simplicity is one of those conflicts?
>> 
>> I don’t fully agree with this. To elaborate:
>> 
>> * A good level of simplicity is a goal for both a human-writable shader language and an ingestion format. There’s a minimum level of complexity is set by the requirements of the domain (i.e. a shader language/format has to have the expressiveness and capabilities needed for shaders). Too much complexity is bad for both humans writing the language and software consuming it. 
>> 
>> * Perfect HLSL compatibility is likely not achievable for a human-writable shader language for the web, because regular HLSL doesn’t have the right safety properties. The question is how far to go in that direction. Being at least superficially similar is helpful for shader authors. Being real-world compatible with at least some HLSL shaders is even nicer, if it’s practical. 
>> 
>> * There is indeed some tradeoff between more HLSL compatibility and more complexity. But more complexity is a downside for humans too. So this tradeoff exists before you even consider the needs of software consuming the language. I suspect the best range in this tradeoff space is also a good spot for software ingestion needs. But I could be convinced otherwise by evidence.
>> 
>> 
>> I guess there are some factual questions that could shed light on the matter:
>> - Does WHLSL have good enough HLSL compatibility to allow any useful shaders at all to be brought over, or only  enough for vague familiarity?
>> - Can more compatibility be added without:
>>  - Violating web safety requirements?
>>  - Adding a level of complexity that’s bad for authors?
>>  - Making the language too hard to process safely and robustly?
>> - If more compatibility is added, will that actually allow more real existing shaders to run, or would it just add a bit more familiarity?
>> 
>> I don’t know enough about HLSL or the world of HLSL shaders out there to answer these questions myself.
>> 
>> Regards,
>> Maciej
>> 
>> 
>>> 
>>> On Thu, Nov 8, 2018 at 1:52 PM Maciej Stachowiak <mjs@apple.com <mailto:mjs@apple.com>> wrote:
>>> 
>>> 
>>>> On Nov 8, 2018, at 1:09 PM, James Darpinian <jdarpinian@google.com <mailto:jdarpinian@google.com>> wrote:
>>>> 
>>>> > > Would you be interested in a non-evolving AST-level ingestion format?
>>>> > Yes, if that format is text on the wire, since that is the most efficient and simple way to express an AST format.
>>>> 
>>>> Perhaps there's something we can agree on here then. Can we agree that the ingestion format and the shading language have different requirements that sometimes conflict, e.g.  compatibility with existing HLSL vs. simplicity,
>>> 
>>> I agree that it may be true, but not that it has been shown to be true on this thread so far. Specifically, I don’t agree that the ingestion format can or should be “non-evolving”. It should probably evolve more slowly than other web languages, and likely will regardless, due to the nature of the domain. But that’s about it.
>>> 
>>> 
>>>> and we should, as a group, investigate making the ingestion format different from the shading language to better satisfy both sets of requirements?
>>> 
>>> I think we are already investigating it in that we’re considering a web dialect of SPIR-V as one of the ingestion formats, and no one thinks it’s a human-writable shader language.
>>> 
>>> Whether we ultimately decide that the ingestion format is different from the human-writable format remains to be seen. In my mind, it depends on if we find that they actually have conflicting requirements, and that the compromises necessary to satisfy both are a higher cost than having two formats.
>>> 
>>> I tend to think a single text-based language can both be an adequate compiler target for other languages, still nice to write directly, and secure and robust enough to use as a wire format on the web, so I’m not yet convinced we need two formats.
>>> 
>>> Regards,
>>> Maciej
>>> 
>>>> 
>>>> On Thu, Nov 8, 2018 at 8:45 AM Filip Pizlo <fpizlo@apple.com <mailto:fpizlo@apple.com>> wrote:
>>>> 
>>>> 
>>>> On Nov 7, 2018, at 10:57 PM, Kai Ninomiya <kainino@google.com <mailto:kainino@google.com>> wrote:
>>>> 
>>>>> Maciej: You're right that comparing WHLSL with JavaScript is not a fair analogy. I mistook your statement "The evidence from WebAssembly vs JavaScript suggests this probably won’t be true"  to be trying to make that analogy, but I see now that it was about a more specific point. I apologize for digging at this rathole.
>>>>> 
>>>>> Filip: WebAssembly is a little hard to compare with SPIR-V since it's not SSA as you pointed out. WHLSL may be comparable to WebAssembly in that it is, in essence, an AST-level language. However, WHLSL is most definitely not at the level of WebAssembly when it comes to actual language complexity, if we are going to support existing HLSL code,
>>>> 
>>>> I’m not sure that is true. Like WebAssembly, WHLSL just contains the low level features you need to build other things out of. 
>>>> 
>>>> The only manner in which WHLSL feels more complex to me is the addition of:
>>>> 
>>>> - GPU style concurrency, which has more quirks than CPU style. 
>>>> 
>>>> - API for doing graphics things. WebAssembly is only concerned with the language and it has basically no api exposed to the wasm program. WHLSL has lots of spec-mandated functions exposed to the WHLSL program. 
>>>> 
>>>> So, I don’t think that WHLSL is more complex except where it absolutely has to be to do graphics. SPIR-V also has these additional complexities. 
>>>> 
>>>>> and especially if we are going to add additional features (e.g. templates/generics or operator overloading) to the language.
>>>> 
>>>> We aren’t proposing to add templates to WHLSL at this time. I think that when debating about WHLSL versus other languages, we should focus on what is being proposed rather than what might be proposed. I’m not a fan of critiquing something that might be proposed but hasn’t been proposed, since such a critique has no limiting principle - you could make up whatever you think WHLSL might have and point out that you don’t like it. 
>>>> 
>>>>> WebAssembly does not need updates when C++ gains new language features,
>>>> 
>>>> That’s not really true!  WebAssembly has to evolve to support some new features like threads and maybe simd. 
>>>> 
>>>>> and I think this is a strength of both WebAssembly and SPIR-V.
>>>> 
>>>> Both of them have been revved with new stuff in the past. Both of them will probably be revved with new stuff in the future. 
>>>> 
>>>>> 
>>>>> Would you be interested in a non-evolving AST-level ingestion format?
>>>> 
>>>> Yes, if that format is text on the wire, since that is the most efficient and simple way to express an AST format. One of the lessons I learned from wasm is that binary serialization of ASTs is really hard, and considering the time it took to reach consensus on the technique wasm ended up using, I think that it’s just simpler to use a text format. 
>>>> 
>>>> Specifically:
>>>> 
>>>> - text formats basically mean using delimiters (like { and }) around blocks of code. If you go binary you either have to invent some other delimiter or use block headers that tell the length. From a parsing standpoint, binary is just not any better than text. 
>>>> 
>>>> - text formats are trivial to introspect. There is no need for a separate text encoding used for View Source. 
>>>> 
>>>> I think that any argument in favor of binary has to be strong enough to counterbalance text’s benefits for view source.
>>>> 
>>>>> Maybe we should discuss it. (Although, IMO, existing HLSL is already too complex to use as a WASM-level AST-style format; Inventing a new format or repurposing WASM would be painful because it gets us neither an existing tool ecosystem nor an existing application ecosystem.)
>>>> 
>>>> WHLSL (i.e. WSL at the time) started out as more of the thing you want, since it didn’t initially have all the stuff necessary to support all of HLSL. We removed generics to make the language even simpler. 
>>>> 
>>>> In the last call, we talked about going for full HLSL compatibility. That’s making WHLSL less like the thing that you want. For example, WHLSL currently avoids some complexity by having less of the lvalue magic that C has and by having a more restrictive parser. WHLSL also uses operator overloading to make many primitive operations (like +) exist outside the language itself - the language just views + as a function call. 
>>>> 
>>>> Personally, I’d be happy with a text shader format that goes for extreme simplicity. You could imagine making some additional simplifications, like requiring that all variables are declared at the top of function. Maybe there is even more that can be done to reduce complexity. My position is that these are the good things we want in a web shader format:
>>>> 
>>>> 1) text
>>>> 2) security
>>>> 3) simplicity
>>>> 4) compiler target
>>>> 5) similar level of abstraction to SPIR-V
>>>> 
>>>> WHLSL currently satisfies 1, 2, 4, and 5 but may be diverging from 3 because of the desire for full HLSL compat. 
>>>> 
>>>> You could even imagine this:
>>>> 
>>>> - WHLSL is like a kernel language (not in the sense of numerical kernel but in the sense of just having the core functionality) and doesn’t evolve much. 
>>>> - some other HLSL flavor has All The Features. 
>>>> - programmers can use WHLSL directly or they can use it as a compiler target. 
>>>> 
>>>>> 
>>>>> > Before SSA, folks used IRs with numbered temporaries like 3AC.
>>>>> 
>>>>> IMO, 3AC is more like SSA than like AST when it comes to most issues, such as applying code transformations.
>>>> 
>>>> I agree. 
>>>> 
>>>>> Regardless, I agree that coming up with new variable names is not particularly problematic.
>>>>> 
>>>>> On Wed, Nov 7, 2018 at 2:42 PM Filip Pizlo <fpizlo@apple.com <mailto:fpizlo@apple.com>> wrote:
>>>>> 
>>>>> 
>>>>> On Nov 7, 2018, at 5:15 PM, Kai Ninomiya <kainino@google.com <mailto:kainino@google.com>> wrote:
>>>>> 
>>>>>> > OpLifetimeStart and OpLifetimeEnd are instructions in the SPIR-V language, which presumably means that lifetimes are not clearly expressed with those instructions. Even with the addition of those instructions, they can’t be trusted because they have to be validated, which means they could lie.
>>>>>> 
>>>>>> According to your links, OpLifetimeStart/OpLifetimeEnd are only valid with Kernel capability (i.e. OpenCL). I would guess this is related to physical pointers.
>>>>>> 
>>>>>> > Things like plumbing bounds around with other objects would require rewriting functions and variables and operations on those variables. It would require generating new SSA IDs or possibly regenerating / reassigning them
>>>>>> 
>>>>>> Generating and reassigning SSA IDs is extremely simple compared with non-SSA IDs. This is why SSA is used in modern compilers to begin with.
>>>>> 
>>>>> Before SSA, folks used IRs with numbered temporaries like 3AC. The thing that SSA brings to the table is that it makes it easy to find the definition of a variable given its use. That’s why compilers use it. If all they wanted was an easy way to generate IDs then it’s just as easy to do without SSA as with SSA.
>>>>> 
>>>>> That said, I think both of you guys have a point:
>>>>> 
>>>>> - It’s true that editing SPIR-V to insert checks will mean that you’re not simply passing a SPIR-V blob through. You’re going to have to decode it to an SSA object graph and then encode that graph back to a blob. 
>>>>> 
>>>>> - It’s true that SPIR-V’s use of 32-bit variable IDs makes generating new ones straightforward.
>>>>> 
>>>>> But I should note that since WebHLSL is not a higher order language, generating new variable names is pretty easy. Any name not already used is appropriate, which isn’t significantly different from finding a spare 32-but variable ID. 
>>>>> 
>>>>>> 
>>>>>> > The evidence from WebAssembly vs JavaScript suggests this probably won’t be true (if by “easier” you mean either “faster” or “simpler to code correctly”).
>>>>>> 
>>>>>> It sounds like you are claiming that the JavaScript parser/code generator is not more complex than the WASM parser/code generator. Is this correct? Can you provide evidence for this claim?
>>>>> 
>>>>> Depends on what you mean by complexity. And it depends on a lot of things that are not really inherent to the languages. And it depends on whether you account for the handicap in JS due to JS being a more complex language in ways that have nothing to do with binary versus text. 
>>>>> 
>>>>> Without a doubt, parsing JavaScript is de facto more code than parsing WebAssembly. This happens mostly because those parsers have been hyper optimized over a long time (decade or more in some cases, like the one in JSC). Maybe it’s also more code to parse JS even if you didn’t do those optimizations, but I’m not sure we have an easy way of knowing just by looking at an existing JS parser or wasm parser. 
>>>>> 
>>>>> What is sure is that JavaScript has better startup time than WebAssembly. See: https://pspdfkit.com/blog/2018/a-real-world-webassembly-benchmark/ <https://pspdfkit.com/blog/2018/a-real-world-webassembly-benchmark/>
>>>>> 
>>>>> So if “complexity” is about time then I don’t think that WebAssembly wins. 
>>>>> 
>>>>> Looks like this varies by browser and it also looks like cases where one language is faster to start than the other have more to do with the compiler backend than parsing.
>>>>> 
>>>>> If by complexity you mean bugs, then WebAssembly parsing has bugs as does JS. JS parsing has less bugs for us, but that may have to do more with JS being very mature. It may also be because parsing text is easier to get right.
>>>>> 
>>>>> If by complexity you mean amount of code or difficulty of code after the parser but before the backend, then it’s unclear. WebAssembly and JavaScript both have some quirks that implementations have to deal with before emitting code to the backend. JSC does weird stuff to JS before emitting bytecode and it has significant complexity in how it interprets wasm to produce B3 IR. Also, WebAssembly opted against SSA - it’s more of an AST serialization disguised as a stack-based bytecode than SSA. I think wasm opted for that because dealing with something AST-like as an input was thought to be easier than dealing with SSA as an input. 
>>>>> 
>>>>> -Filip
>>>>> 
>>>>>> 
>>>>>> On Wed, Nov 7, 2018 at 2:11 PM Myles C. Maxfield <mmaxfield@apple.com <mailto:mmaxfield@apple.com>> wrote:
>>>>>>> On Nov 6, 2018, at 3:55 PM, Jeff Gilbert <jgilbert@mozilla.com <mailto:jgilbert@mozilla.com>> wrote:
>>>>>>> 
>>>>>>> I don't think it's necessarily helpful to think of this discussion as
>>>>>>> predominately binary vs text.
>>>>>>> 
>>>>>>> I think there is a lot of value in a constrained, targeted ingestion
>>>>>>> format, *and separately* I think SPIR-V is a natural choice for this
>>>>>>> ingestion format.
>>>>>>> 
>>>>>>> SPIR-V's core format is very, very easy to parse,
>>>>>> 
>>>>>> 
>>>>>> SPIR-V is a sequence of 32-bit words, so you’re right that it’s easy to read a sequence of 32-bit words.
>>>>>> 
>>>>>> However, a Web browser’s job is to understand any possible sequence of inputs. What should a browser do when it encounters two OpEntryPoint <https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpEntryPoint> instructions that happen to have the same name but different execution models? What happens when an ArrayStride <https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#Decoration> decoration is set to 17 bytes? What happens when both SpecId and BuiltIn decorations are applied to the same value <https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#_a_id_shadervalidation_a_validation_rules_for_shader_a_href_capability_capabilities_a>? SPIR-V today is clearly not a dream for ingestion. It is more difficult for a browser to understand a SPIR-V program than a WHLSL program.
>>>>>> 
>>>>>>> and lends itself
>>>>>>> well to simple but robust parsing. Lifetimes are clearly expressed,
>>>>>>> instruction invocations are very explicit, and ecosystem support is
>>>>>>> already good. It's a dream format for ingestion.
>>>>>>> 
>>>>>>> Binning it with other (particularly older) binary formats is just
>>>>>>> inaccurate. Doing the initial parse gives you the structures
>>>>>>> (functions, types, bindings) you want pretty immediately. By
>>>>>>> construction, most unsafe constructs are impossible or trivially
>>>>>>> validatable. (SSA, instruction requirements, unsafe types, pointers)
>>>>>>> 
>>>>>>> For what it's worth, text formats are technically binary formats
>>>>>>> with a charset. I would rather consume a constrained,
>>>>>>> rigidly-structured (SSA-like? s-expressions?) text-based assembly
>>>>>>> than some binary formats I've worked with. (DER, ugh!)
>>>>>>> 
>>>>>>> Disentangling our ingestion format from the pressures of both
>>>>>>> redundancies and elisions that are desirable in directly-authored
>>>>>>> languages, simplifies things and actually prevents ambiguity. It
>>>>>>> immediately frees the authoring language to change and evolve at a
>>>>>>> faster rate, and tolerates more experimentation.
>>>>>>> 
>>>>>>> I would rather solve the compilation tool distribution use-case
>>>>>>> without sacrificing simplicity and robustness in ingestion. A
>>>>>>> authoring-to-ingestion language compiler in a JS library would let us
>>>>>>> trivially share everything above the web-IR->host-IR translation,
>>>>>>> including optimization passes.
>>>>>>> On Tue, Nov 6, 2018 at 3:16 PM Ken Russell <kbr@google.com <mailto:kbr@google.com>> wrote:
>>>>>>>> 
>>>>>>>> Hi Myles,
>>>>>>>> 
>>>>>>>> Our viewpoint is based on the experience of using GLSL as WebGL's input language, and dealing with hundreds of bugs associated with parsing, validating, and passing a textual shading language through to underlying drivers.
>>>>>>>> 
>>>>>>>> Kai wrote this up at the beginning of the year in this Github issue: https://github.com/gpuweb/gpuweb/issues/44 <https://github.com/gpuweb/gpuweb/issues/44> , and there is a detailed bug list (which is still only a sampling of the associated bugs we fixed over the years) in this spreadsheet:
>>>>>>>> https://docs.google.com/spreadsheets/d/1bjfZJcvGPI4M6Df5HC8BPQXbl847RpfsFKw6SI6_T30/edit#gid=0 <https://docs.google.com/spreadsheets/d/1bjfZJcvGPI4M6Df5HC8BPQXbl847RpfsFKw6SI6_T30/edit#gid=0>
>>>>>>>> 
>>>>>>>> Unlike what I said on the call, the main issues aren't really around the parsing of the input language or string handling. Both the preprocessor's and compiler's parsers in ANGLE's shader translator are autogenerated from grammars. Of more concern were situations where we had to semi-arbitrarily restrict the source language so that we wouldn't pass shaders through to the graphics driver which would crash its own shader compiler. Examples included having to restrict the "complexity" or "depth" of expression trees to avoid stack overflows in some drivers (this was added as an implementation-specific security workaround rather than to the spec), working around bugs in variable scoping and shadowing, defeating incorrect compiler optimizations, and more. Please take the time to read Kai's writeup and go through the spreadsheet.
>>>>>>>> 
>>>>>>>> The question will come up: would using a lower-level representation like SPIR-V for WebGPU's shaders really address these problems? I think it would. SPIR-V uses  SSA form and simple numbers for variables, which will eliminate entire classes of bugs in mishandling of language-level identifiers, variables, and scopes. SPIR-V's primitives are lower level than those in a textual shader language, and if it turns out restrictions on shaders are still needed in WebGPU's environment spec in order to work around driver bugs, they'll be easier to define more precisely against SPIR-V than source text. Using SPIR-V as WebGPU's shader ingestion format would bring other advantages, including that it's based on years of experience developing a portable binary shader representation, and has been designed in conjunction with GPU vendors across the industry.
>>>>>>>> 
>>>>>>>> On the conference call I didn't mean to over-generalize the topic to "binary formats vs. text formats in the browser", so apologies if I misspoke.
>>>>>>>> 
>>>>>>>> -Ken
>>>>>>>> 
>>>>>>>> 
>>>>>>>> 
>>>>>>>> On Mon, Nov 5, 2018 at 10:58 PM Myles C. Maxfield <mmaxfield@apple.com <mailto:mmaxfield@apple.com>> wrote:
>>>>>>>>> 
>>>>>>>>> Hi!
>>>>>>>>> 
>>>>>>>>> When we were discussing WebGPU today, the issue of binary vs text was raised. We are confused at the viewpoint that binary languages on the Web are inherently safer and more portable than text ones. All of our browsers accept HTML, CSS, JavaScript, binary image formats, binary font files, GLSL, and WebAssembly, and so we don’t understand how our teams came to opposite conclusions given similar circumstances.
>>>>>>>>> 
>>>>>>>>> Can you describe the reasons for this viewpoint (as specifically as possible, preferably)? We’d like to better understand the reasoning.
>>>>>>>>> 
>>>>>>>>> Thanks,
>>>>>>>>> Myles
>>>>>>> 
>>> 
>> 
>