Emit SPIR-V from WSL compiler (Part 1)

• Component: WebGPU
• Status: RESOLVED FIXED

[Myles C. Maxfield]

Emit SPIR-V from WSL compiler

[Myles C. Maxfield]

Created attachment 322993 [details]
WIP

[Myles C. Maxfield]

Created attachment 323038 [details]
WIP

[Myles C. Maxfield]

Created attachment 323064 [details]
WIP

[Myles C. Maxfield]

Created attachment 323065 [details]
WIP

[Myles C. Maxfield]

Created attachment 323116 [details]
Patch

[Filip Pizlo]

Comment on attachment 323116 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=323116&action=review

> Tools/WebGPUShadingLanguageRI/SPIRVCodegen.js:104
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +

So much space!

> Tools/WebGPUShadingLanguageRI/SPIRVCodegen.js:120
> +                    switch (type[0]) {

It seems like you really want the TypeRef, not the instantiated type.  You should try to see how you can structure this code so that you get that.  Then, you can case on types more easily.

FWIW the inliner already leaves behind the uninstantiated types in other places where we already needed them.

I don't think that this concern should block landing, since the plan here is to land a WIP that we can iterate on.

> Tools/WebGPUShadingLanguageRI/SPIRVCodegen.js:190
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +

So much space!

> Tools/WebGPUShadingLanguageRI/SPIRVCodegen.js:244
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +
> +

So space.

> Tools/WebGPUShadingLanguageRI/SPIRVTypeAnalyzer.js:102
> +    // FIXME: Using toString() in these functions is a hack. Instead, we should implement a proper type deduper.

Yeah, it's a hack.  But also, can you explain why you need a map at all?

> Tools/WebGPUShadingLanguageRI/SPIRVTypeAnalyzer.js:138
> +            this.typeMap.set(node.toString(), {id: id, fieldTypes: fieldTypes});

ES6 lets you say just {id, fieldTypes} in this case.

> Tools/WebGPUShadingLanguageRI/SPIRVVariableAnalyzer.js:141
> +        let builtInStructs = [
> +            "struct vec2<> { int32 x; int32 y }",
> +            "struct vec2<> { uint32 x; uint32 y; }",
> +            "struct vec2<> { float32 x; float32 y; }",
> +            "struct vec2<> { float64 x; float64 y; }",
> +            "struct vec3<> { int32 x; int32 y; int32 z; }",
> +            "struct vec3<> { uint32 x; uint32 y; uint32 z; }",
> +            "struct vec3<> { float32 x; float32 y; float32 z; }",
> +            "struct vec3<> { float64 x; float64 y; float64 z; }",
> +            "struct vec4<> { int32 x; int32 y; int32 z; int32 w; }",
> +            "struct vec4<> { uint32 x; uint32 y; uint32 z; uint32 w; }",
> +            "struct vec4<> { float32 x; float32 y; float32 z; float32 w; }",
> +            "struct vec4<> { float64 x; float64 y; float64 z; float64 w; }"
> +        ];
> +        for (let builtInStruct of builtInStructs) {
> +            if (node.toString() == builtInStruct) {
> +                this.result.push({ name: this.nameComponents.join(""), id: this._currentId++, type: this.typeMap.get(node.toString()).id, location: this._currentLocation++  });
> +                return;
> +            }
> +        }

I think you really want Intrinsics to stash fields in those types to describe their spirv behavior.  I think that the vec types should be native types so that this becomes easier.

[WebKit Commit Bot]

Comment on attachment 323116 [details]
Patch

Clearing flags on attachment: 323116

Committed r223246: <https://trac.webkit.org/changeset/223246>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.

[Radar WebKit Bug Importer]

<rdar://problem/34959874>

[Myles C. Maxfield]

Migrated to https://github.com/gpuweb/WHLSL/issues/164