Add DXIL 1.8 op code cap and move WaveMatrix intrisics to SM 6.9

[hekota]

Adds max opcode value for DXIL 1.8 and moves WaveMatrix intrinsics into future shader model 6.9.

Contributes to #6133
Related to #6125

[pow2clk]

Ironically, the failures reveal success. The shaders that were targeting earlier shader models are failing. It would be nice to have a simple test that verifies those failures.

[pow2clk]

Looks good! Thanks for the quick work!

[pow2clk]

I changed the description from "fixes" to "contributes to" in order to keep the bug open since we'd still like a test for the original failure of allowing these in earlier shader models.

[tex3d]

Looks good!

[MarijnS95]

Hi! Is there any possible way to test the new cs_6_9 profile now? Judging from the commit Remove lit.local.config - this directory is not scanned by lit it is not intended to be reachable at all, and I get error: invalid profile cs_6_9 out of CGHLSLMS.cpp, presumably because hlsl::ShaderModel is invalid as there is no parser variant for the number '9' yet, not even on main as of today:

DirectXShaderCompiler/lib/DXIL/DxilShaderModel.cpp

Lines 300 to 314 in ef043e9

	case '8':
	if (Major == 6) {
	Minor = 8;
	break;
	} else
	return GetInvalid();
	// VALRULE-TEXT:END
	case 'x':
	if (kind == Kind::Library && Major == 6) {
	Minor = kOfflineMinor;
	break;
	} else
	return GetInvalid();
	default:
	return GetInvalid();

Is this intended to be reenabled for testing at some point? It is blocking us from upgrading DXC as we have experimental shaders with WaveMatrix intrinsics.

EDIT

The bump apparently already happened in 15fe220 (and #6465 probably relies on it?) for the mesh-node preview branch.

That release indeed accepts the cs_6_9 profile!

Unfortunately it doesn't help us much because one particular reason to upgrade the compiler is to revert our workaround that was addressed with the closing of #6173 / #6087.

[hekota]

Sorry @MarijnS95, WaveMatrix support is getting removed in #6807.

[llvm-beanz]

Hi! Is there any possible way to test the new cs_6_9 profile now? Judging from the commit Remove lit.local.config - this directory is not scanned by lit

Just to clarify something here. The HLSLFileCheck test directory has never been scanned by LIT. Those tests have a separate runner, and are scanned by the unit test implemented here. It's an odd legacy of the way DXC's Windows-only testing strategy was done.

[MarijnS95]

Sorry @MarijnS95, WaveMatrix support is getting removed in #6807.

@hekota thanks for the update, but is there anywhere I can read up on what this is about? Are the ops deprecated and will they be replaced by a different spec in the future, or just out of scope for DXC at the time?

---

Just to clarify something here. The HLSLFileCheck test directory has never been scanned by LIT. Those tests have a separate runner, and are scanned by the unit test implemented here. It's an odd legacy of the way DXC's Windows-only testing strategy was done.

@llvm-beanz thanks for poinging that out, looks like it was a red herring because the file was only temporarily introduced in this PR.

The thing making sure that the tests don't run at all with the unsupported _6_9 profile must be #6807 then.

[hekota]

@hekota thanks for the update, but is there anywhere I can read up on what this is about? Are the ops deprecated and will they be replaced by a different spec in the future, or just out of scope for DXC at the time?

AFAIK the ops are marked reserved. I am not sure if they would be reused in the future or stay reserved forever. @damyanp?

[damyanp]

@hekota thanks for the update, but is there anywhere I can read up on what this is about? Are the ops deprecated and will they be replaced by a different spec in the future, or just out of scope for DXC at the time?

AFAIK the ops are marked reserved. I am not sure if they would be reused in the future or stay reserved forever. @damyanp?

It's unclear what the exact future here is at the moment. However, I would strongly advise against making any assumptions about the meaning of these opcodes being stable.

[MarijnS95]

@damyanp thanks, I'm not specifically looking at these opcodes but more about the WaveMatrix feature in general. I haven't written this shader code and am completely out of the loop about this experimental feature, and am merely seeking for advice on what to tell my coworkers about this future of their shader and the impact on updating our shader compiler. If this is something that is expected to continue working on the driver end, we'll likely hold off DXC upgrades or include an older DXC to keep using it.

[damyanp]

I can't speak for the GPU vendors and their plans for supporting this experimental feature that was never released.

We are continuing to look into possible future features that support the scenarios and hardware that WaveMatrix was intended for, but I would be very surprised if whatever we end up with allows these experimental shaders to continue working without modification.