Implement GetGroupWaveIndex and GetGroupWaveCount #7959
Conversation
|
✅ With the latest revision this PR passed the Python code formatter. |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
|
Could you add something to ReleaseNotes.md please? |
tex3d
left a comment
tex3d
left a comment
There was a problem hiding this comment.
We have several SM 6.10 features staged in PRs now. In order to reduce conflicts, I'd like to suggest we make a separate PR reserving the HLSL Intrinsics and the experimental DXIL ops ahead of the rest of the changes.
This means a PR with just the additions to gen_intrin_main.txt, hctdb.py, nullptr , generated changes (hlsl_intrinsic_opcodes.json, DxilConstants.h, DXIL.rst, DxilInstructions.h, DxilOperations.cpp), and initial EmptyLower table entries for gLowerTable in HLOperationLower.cpp.
Of course, this will also depend on getting #7947 in first.
tex3d
left a comment
tex3d
left a comment
There was a problem hiding this comment.
I've put a PR up for reserving these operations here: #7995.
Once that's merged, this should be rebased onto main, but you won't want some of your changes here.
You don't need the changes in the generated files, or the addition of the HLSL instrinsics in gen_intrin_main.txt or DXIL ops in hctdb.py, or the HL op reservations in hlsl_intrinsic_opcodes.json.
You'll want to take what's in main for HLOperationLower.cpp's gLowerTable, then update the lowering function to TranslateWaveToVal.
| elif i.name.startswith("GetGroupWave"): | ||
| i.category = "Group Wave Ops" | ||
| i.shader_model = 6, 10 | ||
| i.shader_stages = ("compute", "mesh", "amplification", "library") |
There was a problem hiding this comment.
You can back these changes out when you rebase on top of main with merged #7995.
| counters=("floats",), | ||
| ) | ||
|
|
||
| # Group Wave Operations |
There was a problem hiding this comment.
You can back these changes out when you rebase on top of main with merged #7995.
| assert op_count == 312, ( | ||
| "312 is expected next operation index but encountered %d and thus opcodes are broken" | ||
| assert op_count == 314, ( | ||
| "314 is expected next operation index but encountered %d and thus opcodes are broken" |
There was a problem hiding this comment.
We definitely do not want this change.
| "IOP_isnormal": 394 | ||
| "IOP_isnormal": 394, | ||
| "IOP_GetGroupWaveCount": 395, | ||
| "IOP_GetGroupWaveIndex": 396 |
There was a problem hiding this comment.
You can back these changes out when you rebase on top of main with merged #7995.
| bool [[]] QuadAny(in bool cond); | ||
| bool [[]] QuadAll(in bool cond); | ||
| uint [[rn]] GetGroupWaveIndex(); | ||
| uint [[rn]] GetGroupWaveCount(); |
There was a problem hiding this comment.
You can back these changes out when you rebase on top of main with merged #7995.
|
I did a rebase and a test update, while catching a fault with the ordering of the lowering table entries in the opcode reservation change. Since it's not based on this branch (it's a rebase), I put the branch up here: You can reset your branch to this and iterate from there to save a bunch of work. |
3 participants
Implement Group Wave Index and Group Wave Count as proposed by:
https://github.com/microsoft/hlsl-specs/blob/main/proposals/0048-group-wave-index.md
Added two new intrinsics:
Limited to Shader Model 6.10 and Compute, Mesh, Node and Amp. shaders.
Added basic test.