[SM6.9] Enable trivial native vector Dxil Operations plus a few #7324

llvm-beanz · 2025-04-08T16:56:16Z

This enables the generation of native vector DXIL Operations
that are "trivial", meaning they take only a single DXOp Call
instruction to implement as well as a few others that either only took
such a call and some llvm operations or were of particular interest for
other reasons.

This involves allowing the overloads by adding the vector indication in
hctdb, altering the lowering to maintain the vectors instead of
scalarizing them, and a few sundry changes to fix issues along the way.

The "trivial" dxil operations that return a different value from the
overload type had to be moved out of the way and given their own
lowering function so that the main function could generate vectors
conditional on the version and vector type. These will be added in a
later change.

While the long vector supporting intrinsics that weren't given this
treatment will continue to generate scalarized operations, some of them
needed some work as well. The dot product for float vectors longer than
4 had to take the integer fallback path, which required some small
modificaitons and a rename.
Additionally, a heuristic for pow that malfunctioned with too many
elements had to have a limit placed on it.

Since the or()/and()/select() intrinsics translate directly to LLVM ops,
they can have their lowering scalarization removed and what future
scalarization might be needed by the current version can be done by
later passes as with other LLVM operators.

An issue with a special value used to represent unassined dimensions had
to be addressed since new dimensions can exceed that value. It's now
MAX_INT.

Contributes to #7120, but I'd prefer to leave it open until all
intrinsics are covered

Primary work by @pow2clk

Fixes #7297 & #7120

This enables the generation of native vector DXIL Operations that are "trivial", meaning they take only a single DXOp Call instruction to implement as well as a few others that either only took such a call and some llvm operations or were of particular interest for other reasons. This involves allowing the overloads by adding the vector indication in hctdb, altering the lowering to maintain the vectors instead of scalarizing them, and a few sundry changes to fix issues along the way. The "trivial" dxil operations that return a different value from the overload type had to be moved out of the way and given their own lowering function so that the main function could generate vectors conditional on the version and vector type. These will be added in a later change. While the long vector supporting intrinsics that weren't given this treatment will continue to generate scalarized operations, some of them needed some work as well. The dot product for float vectors longer than 4 had to take the integer fallback path, which required some small modificaitons and a rename. Additionally, a heuristic for pow that malfunctioned with too many elements had to have a limit placed on it. Since the or()/and()/select() intrinsics translate directly to LLVM ops, they can have their lowering scalarization removed and what future scalarization might be needed by the current version can be done by later passes as with other LLVM operators. An issue with a special value used to represent unassined dimensions had to be addressed since new dimensions can exceed that value. It's now MAX_INT. Contributes to microsoft#7120, but I'd prefer to leave it open until all intrinsics are covered

Was using int dot for the float operands as it was originally an int-only lowering function.

github-actions · 2025-04-08T16:57:14Z

✅ With the latest revision this PR passed the C/C++ code formatter.

lib/HLSL/HLOperationLower.cpp

tools/clang/lib/Sema/SemaHLSL.cpp

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-intrinsics.hlsl

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-scalarized-intrinsics.hlsl

damyanp · 2025-04-08T18:32:38Z

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-trivial-scalarized-intrinsics.hlsl

+StructuredBuffer< vector<TYPE, 8> > buf;
+ByteAddressBuffer rbuf;
+
+float4 main(uint i : SV_PrimitiveID, uint4 m : M) : SV_Target {


@alsepkow - the approach used in this file might be interesting to think about for the HLKs you're working on

lib/HLSL/HLOperationLower.cpp

../tools/clang/test/CodeGenDXIL/hlsl/types/longvec-intrinsics.hlsl

tex3d

A few things:

Some intrinsic expansions aren't using vector DXIL ops when they are available - I think they should.
I think whether vectors are supported by an operation can be determined deeper in common TrivialDxilOperation code based on the operation, rather than by callers of each of the TrivialDxil*Operation functions. This will also likely fix the prior issue with expansions, while reducing the number of changes.
This reminded me that we don't have tests for literal vector DXIL op evaluation. Did I forget about this testing in an earlier change, or am I missing something about how those could be handled automatically?
I think we had a static method on hlsl::OP which queried specifically for vector overload support for a particular DXIL operation (and probably specific overload dimension index, defaulting to 0). Right now, you'd have to rely on IsOverloadLegal and supply the full (multi-)overload type.
All the tests for the shader model version to determine whether vector dxil is supported makes me wish there was a specific accessor for this property on DxilModule and HLModule instead (plus one on hlsl::OP would help).

tex3d · 2025-04-08T19:13:24Z

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-scalarized-intrinsics.hlsl

+  // CHECK: call float @dx.op.unary.f32(i32 17, float %{{.*}}) ; Atan(value)
+  // CHECK: call float @dx.op.unary.f32(i32 17, float %{{.*}}) ; Atan(value)
+  // CHECK: call float @dx.op.unary.f32(i32 17, float %{{.*}}) ; Atan(value)
+  // CHECK: call float @dx.op.unary.f32(i32 17, float %{{.*}}) ; Atan(value)


Why don't these use the supported vector overloads for expansions?
I think it would have made sense for each of these operations with corresponding vector DXIL ops:

atan2 (Atan)

fmod (FAbs, Frc)

ldexp (Exp)

pow (Log, Exp)

modf (Round_z)

tex3d · 2025-04-08T19:35:10Z

lib/HLSL/HLOperationLower.cpp

+
+  // If supported and the overload type is a vector with more than 1 element,
+  // create a native vector operation.
+  if (SupportsVectors && Ty->isVectorTy() && Ty->getVectorNumElements() > 1) {


Instead of the caller supplying SupportsVectors, couldn't this be looked up by the operation and the module?

Suggested change

if (SupportsVectors && Ty->isVectorTy() && Ty->getVectorNumElements() > 1) {

if (Ty->isVectorTy() && Ty->getVectorNumElements() > 1 &&

hlslOP->GetModule()->GetHLModule()->GetShaderModel()->IsSM69Plus() &&

OP::IsOverloadLegal(opcode, Ty)) {

The only iffy one here is using OP::IsOverloadLegal as opposed to a more specific accessor to see if the operation supports native vector types. I did that because that accessor doesn't currently exist, but it would make sense to add it for this purpose.

There's only one oddity in using OP::IsOverloadLegal instead of a more specific check for vector support that I can think of at the moment:
The weird corner case where we still rely on illegal double overloads for literal evaluation. It would reject the vector path for an unsupported literal (double) vector and scalarize the DXIL operation, which would be evaluated and folded later - so scalarization in that case wouldn't do any harm. In fact, I don't think we have any testing for literal DXIL op evaluation with native vector overloads (do we even handle that?!).

Great catch @tex3d!

Fixing this actually corrects at least some of the missing intrinsics you cited in your comment above. I'll roll this change through and do a pass at removing spurious diffs too. I think this is a fantastic simplification of the change.

This also results in correctly disabling scalarization for a number of intrinsics too. ../tools/clang/test/CodeGenDXIL/hlsl/types/longvec-scalarized-intrinsics .hlsl ../tools/clang/test/CodeGenDXIL/hlsl/types/longvec-trivial-scalarized-in trinsics.hlsl

llvm-beanz · 2025-04-08T23:27:13Z

After adapting to @tex3d's feedback this really simplified and cleaned up a lot. Please let me know if there are other changes I should make, but I think I got everything that has been suggested.

damyanp

FWIW, LGTM.

llvm-beanz · 2025-04-09T00:00:35Z

I'm looking at the test failure now. Looks like one of my commits from today is causing a failure in int16 operation lowering. Will fix.

tex3d

Looks good!

tex3d · 2025-04-09T20:00:48Z

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-scalarized-intrinsics.hlsl

Nit: Given that only WaveMatch is truly "scalarized", the name of the test isn't great. Perhaps "expanded" instead. But I don't think this justifies an update on its own. This can be cleaned up later.

Greg Roth and others added 7 commits April 8, 2025 11:40

Fix wrong mul type and tighted up dot() testing

7e3137f

Was using int dot for the float operands as it was originally an int-only lowering function.

Add IR test for dxilgen pass

48cac3d

clang-format

318d3b1

Fix code mis-merged

ad1ac0e

Generated source updates

463dade

A few fixups to conform to coding standards in new code

d39001b

llvm-beanz requested a review from tex3d April 8, 2025 16:56

github-project-automation bot added this to HLSL Roadmap Apr 8, 2025

github-project-automation bot moved this to New in HLSL Roadmap Apr 8, 2025

llvm-beanz changed the title ~~Enable trivial native vector Dxil Operations plus a few~~ [SM6.9] Enable trivial native vector Dxil Operations plus a few Apr 8, 2025

llvm-beanz commented Apr 8, 2025

View reviewed changes

lib/HLSL/HLOperationLower.cpp Outdated Show resolved Hide resolved

llvm-beanz added 3 commits April 8, 2025 11:58

Update lib/HLSL/HLOperationLower.cpp

ede9b02

A few other style cleanups on modified lines

ff8835e

clang-format

317f1d9

damyanp added this to HLSL Support Apr 8, 2025

damyanp assigned tex3d and llvm-beanz Apr 8, 2025

damyanp moved this to Needs Review in HLSL Support Apr 8, 2025

damyanp reviewed Apr 8, 2025

View reviewed changes

tools/clang/lib/Sema/SemaHLSL.cpp Outdated Show resolved Hide resolved

damyanp reviewed Apr 8, 2025

View reviewed changes

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-intrinsics.hlsl Show resolved Hide resolved

damyanp reviewed Apr 8, 2025

View reviewed changes

tools/clang/test/CodeGenDXIL/hlsl/types/longvec-scalarized-intrinsics.hlsl Outdated Show resolved Hide resolved

damyanp reviewed Apr 8, 2025

View reviewed changes

lib/HLSL/HLOperationLower.cpp Outdated Show resolved Hide resolved

llvm-beanz added 2 commits April 8, 2025 15:00

A few minor fixups to use C++ rather than C and Windows-isms

5e9a096

Add coverage for additional vector sizes

1b6b934

../tools/clang/test/CodeGenDXIL/hlsl/types/longvec-intrinsics.hlsl

tex3d reviewed Apr 8, 2025

View reviewed changes

llvm-beanz added 2 commits April 8, 2025 17:54

Reducing diff size

fe19f0d

Revert change

cb6cfb1

damyanp previously approved these changes Apr 8, 2025

View reviewed changes

Fixing things I broke...

2cb4fb7

llvm-beanz dismissed damyanp’s stale review via 2cb4fb7 April 9, 2025 00:07

damyanp approved these changes Apr 9, 2025

View reviewed changes

damyanp assigned alsepkow Apr 9, 2025

damyanp mentioned this pull request Apr 9, 2025

[0030][SM6.9] Add list of ops with vector overload microsoft/hlsl-specs#481

Merged

tex3d approved these changes Apr 9, 2025

View reviewed changes

llvm-beanz merged commit 5d2fa92 into microsoft:main Apr 9, 2025
12 checks passed

llvm-beanz deleted the cbieneman/longvec branch April 9, 2025 21:41

github-project-automation bot moved this from New to Done in HLSL Roadmap Apr 9, 2025

damyanp moved this from Needs Review to Closed in HLSL Support Apr 22, 2025

damyanp removed this from HLSL Support Jun 25, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[SM6.9] Enable trivial native vector Dxil Operations plus a few #7324

[SM6.9] Enable trivial native vector Dxil Operations plus a few #7324

Uh oh!

llvm-beanz commented Apr 8, 2025 •

edited

Loading

Uh oh!

github-actions bot commented Apr 8, 2025 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

damyanp Apr 8, 2025

Uh oh!

Uh oh!

tex3d left a comment

Uh oh!

tex3d Apr 8, 2025

Uh oh!

tex3d Apr 8, 2025

Uh oh!

llvm-beanz Apr 8, 2025

Uh oh!

llvm-beanz commented Apr 8, 2025

Uh oh!

damyanp left a comment

Uh oh!

llvm-beanz commented Apr 9, 2025

Uh oh!

tex3d left a comment

Uh oh!

tex3d Apr 9, 2025

Uh oh!

Uh oh!

Uh oh!

-  if (SupportsVectors && Ty->isVectorTy() && Ty->getVectorNumElements() > 1) {
+  if (Ty->isVectorTy() && Ty->getVectorNumElements() > 1 &&
+      hlslOP->GetModule()->GetHLModule()->GetShaderModel()->IsSM69Plus() &&
+      OP::IsOverloadLegal(opcode, Ty)) {

[SM6.9] Enable trivial native vector Dxil Operations plus a few #7324

[SM6.9] Enable trivial native vector Dxil Operations plus a few #7324

Uh oh!

Conversation

llvm-beanz commented Apr 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Apr 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

damyanp Apr 8, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

tex3d left a comment

Choose a reason for hiding this comment

Uh oh!

tex3d Apr 8, 2025

Choose a reason for hiding this comment

Uh oh!

tex3d Apr 8, 2025

Choose a reason for hiding this comment

Uh oh!

llvm-beanz Apr 8, 2025

Choose a reason for hiding this comment

Uh oh!

llvm-beanz commented Apr 8, 2025

Uh oh!

damyanp left a comment

Choose a reason for hiding this comment

Uh oh!

llvm-beanz commented Apr 9, 2025

Uh oh!

tex3d left a comment

Choose a reason for hiding this comment

Uh oh!

tex3d Apr 9, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

llvm-beanz commented Apr 8, 2025 •

edited

Loading

github-actions bot commented Apr 8, 2025 •

edited

Loading