[Backport to 16] Add multiple missing Cooperative Matrix commits by fda0 · Pull Request #3184 · KhronosGroup/SPIRV-LLVM-Translator

fda0 · 2025-05-22T18:38:13Z

12 commits that were needed to cleanly apply 9b00b88 which fixes Cooperative Matrix prefetch.

KhronosGroup#1835) It specifies how to interpret 'Component Type' when components of a joint matrix are storages for values of different types, for example float for TF32, unsigned short for bfloat16. At this point only tf32 type interpretation is added during SPIR-V generation. Adding it to bf16 is a breaking change and requires adaptation across drivers. Spec update: intel/llvm#8175 Signed-off-by: Sidorov, Dmitry dmitry.sidorov@intel.com (cherry picked from commit b7c5218)

…onosGroup#1913) Extension name will be preserved for a while for binary compatibility. Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 68855f6)

…operative matrices (KhronosGroup#2213) Previously added scalar/vector ConvertFToBF16INTEL, ConvertBF16ToFINTEL and RoundFToTF32INTEL conversions are now enabled for cooperative matrix type under SPV_INTEL_joint_matrix extension following the spec: https://github.com/intel/llvm/blob/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_joint_matrix.asciidoc Note, joint matrices are not allowed as input/output for these conversions as it is being deprecated. Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 1010efc)

KhronosGroup#2214) This PR aims to introduce entities related to OpCooperativeMatrixApplyFunctionINTEL in llvm-spirv translator, according to https://github.com/intel/llvm/blob/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_joint_matrix.asciidoc. Co-authored-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 467edf9)

…n type (KhronosGroup#1995) This target extension type is created here: https://github.com/intel/vc-intrinsics/blob/master/GenXIntrinsics/lib/GenXIntrinsics/GenXSPIRVWriterAdaptor.cpp#L245 As with other target extension types, reverse translation is not yet supported. Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com> Co-authored-by: Victor Mustya <victor.mustya@intel.com> (cherry picked from commit 60746d5) (cherry picked from commit ac6aa17dc92885d9ecbe65fb7d0fb6f75ace2b9e)

…Group#2088) This patch adds joint_matrix reverse translation to target extension type and starts rewriting all of the tests. Some tests are being removed as outdated Remaining tests to add after the patch: 1. tf32 test 2. element wise operations test Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 465eb3c)

…onosGroup#1935) The patch adds TypeJointMatrixINTELv2 which maps to new type OpCode 6184. Under new OpCode matrix type no longer has Layout parameter. The patch also moved 'scope' to optional matrix muladd instruction. The changes are done only in the consumer part to prepare the switch and make E2E switch backward compatible by preparing consumers ahead of time. Unfortunately there is no way to add a test foe this unless it's binary test, but it seems to be a bit unsafe to add this, so the patch was tested locally. Spec change: intel/llvm#8175 Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit a6fcade)

…2143) visit method of the sort relies on getNonLiteralOperands method of the SPIRVType which is being inserted in the module. Without it dependent types can be inserted in the module in incorrect order. For example: TypeCooperativeMatrixKHR %ID% TypeStruct ... %ID% is the correct order, but without the patch in some cases the translator could generate the opposite order. Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 436c497)

…up#2137) (cherry picked from commit 5d4ebd1)

…ope parameters (KhronosGroup#2223) Use should be: MatrixA, MatrixB or Accumulator. Scope must be at max Invocation (others are not supported by the translator). Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit f18e64d)

…2234) - change Scope argument to one of two available options: ScopeWorkgroup/ScopeWorkgroup - fix arguments order in calls to OpCooperativeMatrixLoadKHR() (cherry picked from commit 4e1d3e0)

…trixPrefetchINTEL" (KhronosGroup#2293)" (KhronosGroup#2560) This reverts commit 93f2783. (cherry picked from commit 9b00b88)

fda0 · 2025-05-22T18:43:53Z

It would be best to avoid squashing for this PR.

MrSidims

It won't harm to merge the PR as is, yet may be it's better to discuss few things in advance.

MrSidims · 2025-05-23T09:52:44Z

lib/SPIRV/SPIRVReader.cpp

                                       (unsigned)S};
    if (auto *Use = MT->getUse())
      Params.push_back(static_cast<SPIRVConstant *>(Use)->getZExtIntValue());
+    auto *CTI = MT->getComponentTypeInterpretation();


This patch contradicts current SPV_INTEL_joint_matrix specification, where type interpretation is a part of MulAdd. Note, IGC also expects type interpretation be a part of MulAdd and not a part of the type. Feel free to IM me to discuss this.

@mateuszchudyk please follow up with @vmaksimo about this comment (I'm OOO for the next 2 weeks), she will point out on spec changes and IGC patch, that adds type interpretation support.
Basically I'm fine to merge it as is, but please make sure, that we know, what we are doing with matrix special types. Please also note, that there is no such thing as Int4 interpretation (it is now a proper TypeInt 4 - see SPV_INTEL_int4 (this is to be backported by us soon)) and Int2 interpretation (and there won't be any counterpart).

MrSidims · 2025-05-23T09:53:55Z

lib/SPIRV/SPIRVInternal.h

 const static char ConstantPipeStorage[] = "ConstantPipeStorage";
 const static char VmeImageINTEL[] = "VmeImageINTEL";
 const static char JointMatrixINTEL[] = "JointMatrixINTEL";
+const static char BufferSurfaceINTEL[] = "BufferSurfaceINTEL";


@sarnex @vmustya please correct me if I'm wrong - this is being deprecated, right?

@sarnex @vmustya any updates on this?

Sorry, yes as far as I know it's being deprecated but @vmustya should confirm.

AFAIK, the BufferSurfaceINTEL stateful raw buffers are still used for some workloads.

MrSidims · 2025-05-23T09:57:08Z

lib/SPIRV/libSPIRV/SPIRVType.cpp

  Decoder >> Id >> CompType >> Args;
 }

+void SPIRVTypeCooperativeMatrixKHR::validate() const {


Not sure if it's needed for IGC, but guess it's added just to avoid merge conflicts, right?

mateuszchudyk · 2025-05-26T17:17:03Z

@fda0 is out of office until 23rd of June. I'll be the owner of this PR during that time.

EDIT: Due to other high priority tasks I have, this PR will wait until @fda0 returns.

CLAassistant · 2026-03-03T06:54:52Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 3 committers have signed the CLA.

✅ sarnex
✅ VyacheslavLevytskyy
❌ fda0
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

Dmitry Sidorov and others added 12 commits May 22, 2025 16:48

[Backport to 16] Rename ConvertFToTF32INTEL to RoundFToTF32INTEL (Khr…

7a457b6

…onosGroup#1913) Extension name will be preserved for a while for binary compatibility. Signed-off-by: Sidorov, Dmitry <dmitry.sidorov@intel.com> (cherry picked from commit 68855f6)

Test conversion instructions for TypeCooperativeMatrixKHR (KhronosGro…

5c09881

…up#2137) (cherry picked from commit 5d4ebd1)

Fix cooperative matrix prefetch test (scope parameter) (KhronosGroup#…

9d8262d

…2234) - change Scope argument to one of two available options: ScopeWorkgroup/ScopeWorkgroup - fix arguments order in calls to OpCooperativeMatrixLoadKHR() (cherry picked from commit 4e1d3e0)

Revert "Revert "Remove CoordX and CoordY arguments of OpCooperativeMa…

8a3917e

…trixPrefetchINTEL" (KhronosGroup#2293)" (KhronosGroup#2560) This reverts commit 93f2783. (cherry picked from commit 9b00b88)

fda0 changed the title ~~[Backport to 16] Add ComponentTypeInterpretation for joint matrix type (#1835)~~ [Backport to 16] Add multiple missing Cooperative Matrix commits May 22, 2025

michalpaszkowski requested a review from MrSidims May 22, 2025 18:59

MrSidims approved these changes May 23, 2025

View reviewed changes

MrSidims requested a review from vmaksimo May 26, 2025 16:06

vmaksimo mentioned this pull request Jun 4, 2025

[Backport to LLVM 16] Reland "Remove CoordX and CoordY arguments of OpCooperativeMatrixPrefetchINTEL" (#2560) #3205

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Backport to 16] Add multiple missing Cooperative Matrix commits#3184

[Backport to 16] Add multiple missing Cooperative Matrix commits#3184
fda0 wants to merge 12 commits intoKhronosGroup:llvm_release_160from
fda0:llv16_backmerges

fda0 commented May 22, 2025 •

edited

Loading

Uh oh!

fda0 commented May 22, 2025

Uh oh!

MrSidims left a comment

Uh oh!

MrSidims May 23, 2025

Uh oh!

MrSidims May 27, 2025 •

edited

Loading

Uh oh!

MrSidims May 23, 2025

Uh oh!

bokrzesi Jun 4, 2025

Uh oh!

sarnex Jun 4, 2025

Uh oh!

vmustya Jun 4, 2025

Uh oh!

MrSidims May 23, 2025

Uh oh!

mateuszchudyk commented May 26, 2025 •

edited

Loading

Uh oh!

CLAassistant commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

Conversation

fda0 commented May 22, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fda0 commented May 22, 2025

Uh oh!

MrSidims left a comment

Choose a reason for hiding this comment

Uh oh!

MrSidims May 23, 2025

Choose a reason for hiding this comment

Uh oh!

MrSidims May 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

MrSidims May 23, 2025

Choose a reason for hiding this comment

Uh oh!

bokrzesi Jun 4, 2025

Choose a reason for hiding this comment

Uh oh!

sarnex Jun 4, 2025

Choose a reason for hiding this comment

Uh oh!

vmustya Jun 4, 2025

Choose a reason for hiding this comment

Uh oh!

MrSidims May 23, 2025

Choose a reason for hiding this comment

Uh oh!

mateuszchudyk commented May 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

CLAassistant commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

fda0 commented May 22, 2025 •

edited

Loading

MrSidims May 27, 2025 •

edited

Loading

mateuszchudyk commented May 26, 2025 •

edited

Loading