| Age | Commit message (Collapse) | Author |
|---|
|
(#1318)
TL;DR: This is a tweak the rules for layout that only affects a corner case for people who actually use `interface`-type shader parameters (which for now is just our own test cases). The tweaked rules seem like they make it easier to write the application code for interfacing with Slang, but even if we change our minds later the risk here should be low (again: nobody is using this stuff right now).
Slang already has a rule that a constant buffer that contains no ordinary/uniform data doesn't actually allocate a constant buffer `binding`/`register`:
struct A { float4 x; Texture2D y; } // has uniform/ordinary data
struct B { Texture2D u; SamplerState v; } // has none
ConstantBuffer<A> gA; // gets a constant buffer register/binding
ConstantBuffer<B> gB; // does not
There is similar logic for `ParameterBlock`, where the feature makes more sense. A user would be somewhat surprised if they declared a parmaeter block with a texture and a sampler in it, but then the generating code reserved Vulkan `binding=0` for a constant buffer they never asked for. The behavior in the case of a plain `ConstantBuffer` is chosen to be consistent with the parameter block case.
(Aside: all of this is a non-issue for targets with direct support for pointers, like CUDA and CPU. On those platforms a constant buffer or parameter block always translates to a pointer to the contained data.)
Now, suppose the user declares a constant buffer with an interface type in it:
interface IFoo { ... }
ConstantBuffer<IFoo> gBuffer;
When the layout logic sees the declaration of `gBuffer` it doesn't yet know what type will be plugged in as `IFoo` there. Will it contain uniform/ordinary data, such that a constant buffer is needed?
The existing logic in the type layout step implemented a complicated rule that amounted to:
* A `ConstantBuffer` or `cbuffer` that only contains `interface`/existential-type data will *not* be allocated a constant buffer `register`/`binding` during the initial layout process (on unspecialized code). That means that any resources declared after it will take the next consecutive `register`/`binding` without leaving any "gap" for the `ConstantBuffer` variable.
* After specialization (e.g., when we know that `Thing` should be plugged in for `IFoo`), if we discover that there is uniform/ordinary data in `Thing` then we will allocate a constant buffer `register`/`binding` for the `ConstantBuffer`, but that register/binding will necessarily come *after* any `register`s/`binding`s that were allocated to parameters during the first pass.
* Parameter blocks were intended to work the same when when it comes to whether or not they allocate a default `space`/`set`, but that logic appears to not have worked as intended.
These rules make some logical sense: a `ConstantBuffer` declaration only pays for what the element type actually needs, and if that changes due to specialization then the new resource allocation comes after the unspecialized resources (so that the locations of unspecialized parameters are stable across specializations).
The problem is that in practice it is almost impossible to write client application code that uses the Slang reflection API and makes reasonable choices in the presence of these rules. A general-purpose `ShaderObject` abstraction in application code ends up having to deal with multiple possible states that an object could be in:
1. An object where the element type `E` contains no uniform/ordinary data, and no interface/existential fields, so a constant buffer doesn't need to be allocated or bound.
2. An object where the element type `E` contains no uniform/ordinary data, but has interace/existential fields, with two sub-cases:
a. When no values bound to interface/existential fields use uniform/ordinary dat, then the parent object must not bind a buffer
b. When the type of value bound to an interface/existential field uses uniform/ordinary data, then the parent object needs to have a buffer allocated, and bind it.
3. When the element type `E` contains uniform/ordinary data, then a buffer should be allocated and bound (although its size/contents may change as interface/existential fields get re-bound)
Needing to deal with a possible shift between cases (2a) and (2b) based on what gets bound at runtime is a mess, and it is important to note that even though both (2a) and (3) require a buffer to be bound, the rules about *where* the buffer gets bound aren't consistent (so that the application needs to undrestand the distinction between "primary" and "pending" data in a type layout).
This change introduces a different rule, which seems to be more complicated to explain, but actually seems to simplify things for the application:
* A `ConstantBuffer` or `cbuffer` that only contains `interface`/existential-type data always has a constant buffer `register`/`binding` allocated for it "just in case."
* If after specialization there is any uniform/ordinary data, then that will use the buffer `register`/`binding` that was already allocated (that's easy enough).
* If after speciazliation there *isn't* any uniform/ordinary data, then the generated HLSL/GLSL shader code won't declare a buffer, but the `register`/`binding` is still claimed.
* A `ParameterBlock` behaves equivalently, so that if it contains any `interface`/existential fields, then it will always allocate a `space`/`set` "just in case"
The effect of these rules is to streamline the cases that an application needs to deal with down to two:
1. If the element type `E` of a shader object contains no uniform/ordinary or interface/existential fields, then no buffer needs to be allocated or bound
2. If the element type `E` contains *any* uniform/ordinary or interface/existential fields, then it is always safe to allocate and bind a buffer (even in the cases where it might be ignored).
Furthermore, the reflection data for the constant buffer `register`/`binding` becomes consistent in case (2), so that the application can always expect to find it in the same way.
|
|
The `TEST_INPUT` facility allows textual Slang test cases to provide two kinds of information to the `render-test` tool:
1. Information on what shader inputs exist
2. Information on what values/objects to bind into those shader inputs
Under the first category of information, there exists supporting for attaching a `dxbinding(...)` annotation to a `TEST_INPUT` which seemingly indicates what HLSL `register` the input uses. There is a similar `glbinding(...)` annotation, used for OpenGL and Vulkan.
It turns out that these annotations were, in practice, completely ignored and had no bearing on how `render-test` allocates or bindings graphics API objects. There was some amount of code attempting to validate that explicit registers/bindings were being set appropriately, but the actual values were being ignored.
The visible consequence of the `dxbinding` and `glbinding` annotations being ignored is issue #1036: the order of `TEST_INPUT` lines was *de facto* determining the registers/bindings that were being used by `render-test`.
This change simply removes the placebo features and strips things down to what is implemented in practice: the `TEST_INPUT` lines do not need target-API-specific binding/register numbers, because their order in the file implicitly defines them.
I added logic to the parsing of `TEST_INPUT` lines to make sure I got an error message on any leftover annotations, and went ahead and systematicaly deleted all of the placebo annotations from our test cases.
If we decide to make `TEST_INPUT` lines *not* depend on order of declaration in the future, we can build it up as a new and better considered feature.
The main alternative I considered was to keep the annotations in place, and change `render-test` and the `gfx` abstraction layer to properly respect them, but that path actually creates much more opportunity for breakage (since every single test case would suddenly be specifying its root signature / pipeline layout via a different path using data that has never been tested). The approach in this change has the benefit of giving me high confidence that all the test cases continue to work just as they had before.
|
|
* Start exposing a new COM-lite API
This change is mostly about exposing a new API to the Slang compiler that allows more fine-grained control over the compilation flow. The basic concepts in the new API are:
* An `IGlobalSession` is the granularity at which we load/parse the Slang stdlib, and therefore gives applications a way to amortize startup cost for the library across multiple compiles. This is a concept that might be able to go away in a future version of Slang.
* An `ISession` owns all the code that gets loaded/compiled/generated. Any `import`ed modules are shared across everything in a session (we don't re-parse/-check the code when we see another `import` for the same module). Any generic- or interface-based code in the session can be specialized using types from the same session (but not necessarily across sessions).
* An `IModule` is the unit of code loading and scoping. It doesn't expose any API in this change, but would be the right scope for looking up types or entry points by name.
* An `IProgram` is a "linked" combination of modules and entry points from which code can be generated and reflection information queried.
This change re-uses the existing reflection API types, rather than introduce a new API that duplicates that functionality. That will probably change in a future revision.
There are two major pieces of functionality added here that aren't related to the new API:
* We now have an API concept of "entry point groups" which are one or more entry points that are intended to be used together so that they need to have non-overlapping parameters. For now this is being used to handle "hit groups" and local root signatures for ray tracing, but I'm not sure this is a concept we will keep in the long run.
* We have a very special-case (client-application-specific) flag that ascribes special meaning to the `shared` keyword, so that it can be attached to global parameters to indicate that they are actually to be part of the local root signature rather than the global one for DXR.
None of the API design (including naming) here is finalized; the only reason to check in the changes at this point to avoid having a long-running branch that leads to merge pain. Clients should *not* try to depend on the new API just yet, since it is still a work in progress.
* fixup: clang warning
* fixup: try to detect clang C++11 support
* fixup
* fixup
* fixup
* fixup
* fixup: review feedback
|
|
Previously, interface types were allowed to be used directly as function parameters, local variables, and global shader parameters.
Using an interface type as a field of a `struct` type or a `cbuffer` declaration was not implemented.
This change adds that support, and fixes several unrelated issues that caused problems in doing so.
* The most important work here was adding a case for `IRStructType` to `maybeSpecializeBindExistentialsType` that creates a specialized variant of a `struct` type on-demand based on specialization operands. This logic loops over the fields of the original struct, and creates new fields by binding the existentials/interfaces in the type of each field. Caching is used to ensure that the same `struct` type specialized to the same operands should yield the same result.
* To allow subsequent specialization to occur when a `struct` with interface-type fields is used, it was also necessary to specialize field-address and field-extract instructions in cases where the value that the field is being extracted from is a `wrapExistential`.
* Similarly, we neede to make sure that the logic for specializing called functions based on the concrete types for interfaces in the argument list would also take into account `struct` types with existential-type fields inside of them.
* Doing the above changes revealed some serious flaws in how the `ir-specialize.cpp` logic was tracking which instructions still needed to be processed. It had previously been assuming that it could assume any relevant instructions were on its work list, and when the work list went empty it could exit. This runs into two problems: (1) sometimes we create new instructions when specializing, and it may be impossible to ensure that all the new instructions (e.g., those created by utility routines in other files) get added to the work list, and (2) sometimes the instruction(s) that need to be re-visited when we specialize something aren't its direct users, but instead somethign that transitively depends on the instruction.
These issues were fixed by two changes to the pass: (1) we now maintain a list of known "clean" instructions instead of implicitly using the work-list as a list of "dirty" instructions (so that implicitly any new instruction is dirty), and periodically iterating over all instructions to add the non-clean ones to the work list for processing, and (2) when an instruction is specialized/replaced we mark everything that transitively depends on it "dirty" (by removing it from the "clean" list).
* Added some logic to "fix up" the type of an IR function after changes that might modify its parameter list. Failing to have this logic meant that certain types were still live (because they were referenced by a function type) that couldn't actually be emitted as legal HLSL/GLSL.
* Added some special cases to IR instruction creation for `wrapExistential` and `BindExistentialsType` so that they act as no-ops when there are no "slots" providing specialization information. This helps avoid some special cases when specializing structure fields (since some fields specialization and others don't, so in general there are zero or more operands specific to each field).
* Added a test case that uses an interface type in a `cbuffer`, as well as an interface type in a `struct` passed as an entry-point `uniform` parameter.
* Fixed up some parts of the `.natvis` files to reflect naming changes from a previous PR and thus restore some of the useful Visual Studio debugging experience for Slang.
|
