Add creation-time expressions #2668

alan-baker · 2022-03-15T17:28:43Z

Allows vectors and matrices to have abstract components
Renames constexpr to creation-time expression and greatly expands
functionality to include most expressions
- to be creation-time expressions, all identifiers must be
  creation-time constants or creation-time functions
- removes const_expression grammar (replacement for it)
Add new declaration type: creation-time constant
- uses const keyword
- usable at module and function scopes
Restrict let to function scope
Add @const attribute to declare creation-time constants
- only usable for built-in functions currently
Define override expression in terms of creation-time expressions
Update fixed-size array declarations to allow creation-time
expressions
Update texture functions to use creation-time expressions
Update builtin functions to specify which are usable with abstract
types and which are usable with creation-time expressions
Update expressions to specify which are usable with abstract types
Clarify that let and var are always concrete
De-duplicate some builtin function table entries
Update some overload resolution examples

This builds on #2227 and implements the consensus on constexpr from #1272. Take a look at the last commit.

Does not add support for specifying arguments as required to be creation-time expressions. Does not add staticAssert yet.

Some of the examples use a grammar ambiguity that was pre-existing (e.g. vec3(...)).

wgsl/index.bs

dneto0

This is really nice.
I have very minor edits/suggestions.

wgsl/index.bs

dneto0

Thanks for the update.
I think two things remain:

element_count_expression changes
the 's' type parameter issue for matrix arithmetic

wgsl/index.bs

dneto0

Thanks!
LGTM. Now we only need W3C WebGPU group agreement.

wgsl/index.bs

kdashg · 2022-03-30T18:11:08Z

WGSL meeting minutes 2022-03-29

MM: const x = 5; is the type AbstractInt, or in32.
AB: In this PR, it’s i32. When you add ‘const’ as a declarator, then it becomes abstract
MM: Glad there is an answer. No strong opinion about what the answer should be.
MM: Why have @const when authors can’t write it. Why not have an explicit list in the spec.
AB: Thinking forward, we can add that feature later. It’s easy to explain it this way.
AB: Happy to bikeshed the name like ‘constfn’ vs @const fn.
MM: Think user-defined constexpr functions is a big thing, and needs careful consideration about how it’s shaped. So for now want to not add a syntax for it.
AB: I want a single place to look in the spec to know about a function, rather than look in two places. As long as it’s inline it’s important.
MM: Inline is no problem. Could be a column, or a sentence.
KG: Agree with Alan somewhat. When I look at documentation I often just look at a declaration. So it would be cool to see a ‘const’ there, or @const. You could argue it’s just a table with no borders.
MM: I view the whole spec as the thing you’re describing. If someone made a list of everything in the std library that info would be in the list.
DN: The way AB has it is there’s an @const attrib, but users can’t use it, so this would be something we can harmlessly changed later with no user-visibility. Did we add a note that this is just for explaining, and not “real” syntax?
MM: RIght, this is an editorial comment.
MM: Confirm you can still make a let at internal function scope.
AB: Yes.
MM: Not blocking. If I wrote a PR to replace @const with something that is inline next to the function, would that be a harm?
AB: would need to see it.
KG: Currently prefer @const.
MM: Ok, let’s go forward, and I can write a PR.
AB: I’ll rebase, convert to non-draft.
KG: I’ll review
AB: Call attention that some builtin functions now take abstract numeric types, see if you agree with the set. And also abstract types can be for scalars, vectors, and matrices. Not arrays and structs.
MM: I’m most interested in the math functions working on these abstract values, and they don’t operate on arrays and structs, generally.
AB: An odd case is frexp and modf, which return structs. Decided one way, and open to changing to the other.

kdashg · 2022-04-12T07:49:30Z

WGSL meeting minutes 2022-04-05

AB: Waiting on more reviews. Keep on rebasing.
KG:** I promise to review this week.**

* Allows vectors and matrices to have abstract components * Renames constexpr to creation-time expression and greatly expands functionality to include most expressions * to be creation-time expressions, all identifiers must be creation-time constants or creation-time functions * removes `const_expression` grammar (replacement for it) * Add new declaration type: creation-time constant * uses `const` keyword * usable at module and function scopes * Restrict `let` to function scope * Add `@const` attribute to declare creation-time constants * only usable for built-in functions currently * Define override expression in terms of creation-time expressions * Update fixed-size array declarations to allow creation-time expressions * Update texture functions to use creation-time expressions * Update builtin functions to specify which are usable with abstract types and which are usable with creation-time expressions * Update expressions to specify which are usable with abstract types * Clarify that `let` and `var` are always concrete * De-duplicate some builtin function table entries * Update some overload resolution exmaples * Clarify formulae for matrix arithmetic * Change element count expressions for arrays to be shift expresions and bitwise expressions * Add creation-time constants to uniformity analysis * Change texture built-in function parameters to remove references to module-scope let declarations * Clarify shift operations * Add unsigned conversion from abstract integer * Fixes gpuweb#2719 * Add conversion to u32 scalar and vector types from abstract integer scalar and vector respectively

dneto0

Thanks!

dneto0 · 2022-04-12T19:44:24Z

wgsl/index.bs

      <td>Reinterpretation of bits.<br>
-          The result is the unique value in [=u32=] that is equal to (|e| mod 2<sup>32</sup>).
-  <tr algorithm="scalar conversion from binary32 floating point to unsigned integer">
+          The result is the unique value in [=u32=] that has the same bit pattern as |e|.


I guess AbstractInt values have a bit pattern, because we define bitwise operations on them.

This rephrasing would make it an error to try to stuff a too-big AbstractInt value into a u32. It would be overflow. Similarly for a negative abstract-int value. That makes sense to me.

dneto0 · 2022-04-12T19:48:48Z

wgsl/index.bs

@@ -4371,8 +4433,8 @@ For details on conversion to and from floating point types, see [[#floating-poin
  <tr algorithm="scalar reinterpretation from unsigned to signed">
      <td>|e|: u32<td>`i32(`|e|`)`: i32
      <td>Reinterpretation of bits.<br>
-          The result is the unique value in [=i32=] that is equal to (|e| mod 2<sup>32</sup>).
-  <tr algorithm="scalar conversion from binary32 floating point to signed integer">
+          The result is the unique value in [=i32=] that has the same bit pattern as |e|.


Missing AbstractInt here?

Ah, discussed offline: The AbstractInt -> i32 conversion is already the default. So writing i32( ... ) around an AbstractInt already just works as intended.

Cool

@dneto0

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

@dneto0

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

@dneto0

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

kdashg · 2022-04-13T17:04:11Z

WGSL meeting minutes 2022-04-12

Previously:
- KG: I promise to review this week.
KG**: **Think we should land this. Progress
AB to resolve conflicts, then David to land.

* Allows vectors and matrices to have abstract components * Renames constexpr to creation-time expression and greatly expands functionality to include most expressions * to be creation-time expressions, all identifiers must be creation-time constants or creation-time functions * removes `const_expression` grammar (replacement for it) * Add new declaration type: creation-time constant * uses `const` keyword * usable at module and function scopes * Restrict `let` to function scope * Add `@const` attribute to declare creation-time constants * only usable for built-in functions currently * Define override expression in terms of creation-time expressions * Update fixed-size array declarations to allow creation-time expressions * Update texture functions to use creation-time expressions * Update builtin functions to specify which are usable with abstract types and which are usable with creation-time expressions * Update expressions to specify which are usable with abstract types * Clarify that `let` and `var` are always concrete * De-duplicate some builtin function table entries * Update some overload resolution exmaples * Clarify formulae for matrix arithmetic * Change element count expressions for arrays to be shift expresions and bitwise expressions * Add creation-time constants to uniformity analysis * Change texture built-in function parameters to remove references to module-scope let declarations * Clarify shift operations * Add unsigned conversion from abstract integer * Fixes #2719 * Add conversion to u32 scalar and vector types from abstract integer scalar and vector respectively

* Allows vectors and matrices to have abstract components * Renames constexpr to creation-time expression and greatly expands functionality to include most expressions * to be creation-time expressions, all identifiers must be creation-time constants or creation-time functions * removes `const_expression` grammar (replacement for it) * Add new declaration type: creation-time constant * uses `const` keyword * usable at module and function scopes * Restrict `let` to function scope * Add `@const` attribute to declare creation-time constants * only usable for built-in functions currently * Define override expression in terms of creation-time expressions * Update fixed-size array declarations to allow creation-time expressions * Update texture functions to use creation-time expressions * Update builtin functions to specify which are usable with abstract types and which are usable with creation-time expressions * Update expressions to specify which are usable with abstract types * Clarify that `let` and `var` are always concrete * De-duplicate some builtin function table entries * Update some overload resolution exmaples * Clarify formulae for matrix arithmetic * Change element count expressions for arrays to be shift expresions and bitwise expressions * Add creation-time constants to uniformity analysis * Change texture built-in function parameters to remove references to module-scope let declarations * Clarify shift operations * Add unsigned conversion from abstract integer * Fixes gpuweb#2719 * Add conversion to u32 scalar and vector types from abstract integer scalar and vector respectively

alan-baker added the wgsl WebGPU Shading Language Issues label Mar 15, 2022

alan-baker requested review from jimblandy, litherum, kdashg and dneto0 March 15, 2022 17:28

alan-baker marked this pull request as draft March 15, 2022 17:30

alan-baker commented Mar 23, 2022

View reviewed changes

wgsl/index.bs Outdated Show resolved Hide resolved

wgsl/index.bs Outdated Show resolved Hide resolved

wgsl/index.bs Outdated Show resolved Hide resolved

dneto0 requested changes Mar 23, 2022

View reviewed changes

dneto0 requested changes Mar 25, 2022

View reviewed changes

wgsl/index.bs Outdated Show resolved Hide resolved

dneto0 approved these changes Mar 25, 2022

View reviewed changes

dneto0 approved these changes Mar 28, 2022

View reviewed changes

alan-baker force-pushed the constexpr branch from 79fee4d to 92351bb Compare March 29, 2022 19:24

alan-baker marked this pull request as ready for review March 29, 2022 19:24

jrprice reviewed Mar 29, 2022

View reviewed changes

wgsl/index.bs Outdated Show resolved Hide resolved

alan-baker force-pushed the constexpr branch from 92351bb to 32fde7e Compare March 30, 2022 00:40

dneto0 mentioned this pull request Mar 30, 2022

context-aware tokenization: Sometimes >> is better parsed as two copies of > (similar for ]] ) #2092

Closed

alan-baker mentioned this pull request Apr 1, 2022

Add scalar and vector constructors from abstract numeric types #2719

Closed

alan-baker force-pushed the constexpr branch from 6da1c8f to 5dadb06 Compare April 12, 2022 19:00

alan-baker force-pushed the constexpr branch from d629954 to 5b827b8 Compare April 12, 2022 19:18

dneto0 approved these changes Apr 12, 2022

View reviewed changes

dneto0 merged commit 83f4b92 into gpuweb:main Apr 12, 2022

github-actions bot added a commit that referenced this pull request Apr 12, 2022

Add creation-time constants and expressions (#2668)

3b98f63

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

github-actions bot added a commit that referenced this pull request Apr 12, 2022

Add creation-time constants and expressions (#2668)

cc1ee3a

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

github-actions bot added a commit that referenced this pull request Apr 12, 2022

Add creation-time constants and expressions (#2668)

ca0f2b9

SHA: 83f4b92 Reason: push, by @dneto0 Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

alan-baker deleted the constexpr branch April 12, 2022 20:14

This was referenced Apr 12, 2022

wgsl: -0 is not a valid token #2071

Closed

Do we have a plan for non-32bit literals? #1986

Closed

dneto0 mentioned this pull request Apr 12, 2022

[wgsl] Support compile-time-constants (constexpr) #1272

Closed

Gordon-F mentioned this pull request Apr 14, 2022

[wgsl-in] Implement const expressions gfx-rs/naga#1829

Closed

yutannihilation mentioned this pull request Sep 9, 2022

WGSL const-declarations gfx-rs/wgpu#2994

Closed

dneto0 mentioned this pull request Sep 20, 2022

wgsl: somehow we lost arithmetic right shift #3461

Closed

Add creation-time expressions #2668

Add creation-time expressions #2668

Uh oh!

Conversation

alan-baker commented Mar 15, 2022

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dneto0 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dneto0 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

dneto0 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kdashg commented Mar 30, 2022

Uh oh!

kdashg commented Apr 12, 2022

Uh oh!

dneto0 left a comment

Choose a reason for hiding this comment

Uh oh!

dneto0 Apr 12, 2022

Choose a reason for hiding this comment

Uh oh!

dneto0 Apr 12, 2022

Choose a reason for hiding this comment

Uh oh!

dneto0 Apr 12, 2022

Choose a reason for hiding this comment

Uh oh!

kdashg commented Apr 13, 2022

Uh oh!

Uh oh!