Remove Ampere-Turing matmul scheduler

[naoyam]

As part of #5722, I dropped the matmul scheduler for Ampere/Turing. I tried to just disable Swizzle2D but wasn't quite sure how to correctly update the scheduler. Instead of figuring it out, I decided just dropping it.

The matmul tests now only run on Hopper or upper.

[github-actions]

Review updated until commit 060c34c

Description

• Remove Ampere-Turing matmul scheduler implementation completely

• Update architecture requirements from Ampere (8.0) to Hopper (9.0) across all matmul tests

• Replace scheduler calls with error messages for unsupported architectures

• Delete matmul_ampere-.cpp and matmul_ampere-.h files entirely

Changes walkthrough

	Relevant files
Enhancement	5 files matmul.cpp Update architecture guard from Ampere to Hopper                    +1/-1      matmul.cpp Remove Ampere-Turing scheduler include and add error handling +3/-2      matmul_ampere-.cpp Delete entire Ampere-Turing scheduler implementation file +0/-1338 matmul.h Update comment to reflect Hopper+ only support                      +1/-1      matmul_ampere-.h Delete entire Ampere-Turing scheduler header file                +0/-206
Tests	4 files test_matmul.cpp Update matmul tests to require Hopper+ architecture            +10/-4    test_matmul_scheduler.cpp Update scheduler tests to require Hopper+ architecture      +9/-2      test_multidevice_matmul.cpp Update distributed matmul tests to require Hopper+              +2/-2      test_translate_mma.cpp Update MMA translation tests to require Hopper+                    +6/-6
Configuration changes	1 files CMakeLists.txt Remove Ampere-Turing scheduler source file from build        +0/-1

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Breaking Change Documentation The PR throws an NVF_THROW error when Turing or Ampere macros are detected, which is a breaking change for existing users. Ensure this change is properly documented in release notes and migration guides, as users on these architectures will no longer be able to use the matmul scheduler functionality. if (isTuring(macro) || isAmpere(macro)) { NVF_THROW( "Support for Turing and Ampere was dropped. Scheduling functionality " "should still be largely available except for swizzling"); } else if (isHopper(macro)) { Benchmark Architecture Restriction The SMEM guard was updated from (8,0) to (9,0), which means benchmarks will now only run on Hopper+ architectures. Verify that this is the intended behavior and that benchmark results documentation reflects this change. 9, 0, getSmemSize(cta_tile, number_of_stage), benchmark_state);

[naoyam]

!test

[greptile-apps]

Greptile Summary

Removes the Ampere/Turing matmul scheduler as part of the effort to eliminate Swizzle2D support. The PR deletes ~1,544 lines of scheduler implementation code and updates all matmul tests to require Hopper (SM 9.0) or newer GPUs.

Key Changes:

• Deleted csrc/scheduler/matmul_ampere-.cpp and csrc/scheduler/matmul_ampere-.h
• Replaced Ampere/Turing scheduler invocation with an error throw explaining support was dropped
• Updated all test files to skip on pre-Hopper architectures using NVFUSER_TEST_CUDA_ARCH_GUARD(9, 0)
• Removed matmul_ampere-.cpp from CMake build configuration

Minor Issues:

• Dead code in benchmarks/cpp/matmul.cpp:404-409 - the architecture range check for [8.0, 9.0) is unreachable since the earlier guard already skips anything below 9.0

Confidence Score: 5/5

• Safe to merge - well-executed code removal with comprehensive test updates
• The PR cleanly removes deprecated functionality with no remaining references to the deleted code. All affected tests have been properly updated with architecture guards. The only issue is minor dead code that doesn't affect functionality.
• No files require special attention - the dead code in benchmarks/cpp/matmul.cpp is a cosmetic issue

Important Files Changed

Filename	Overview
csrc/scheduler/matmul.cpp	Removed include and replaced Ampere/Turing scheduler call with error throw
csrc/scheduler/matmul_ampere-.cpp	File deleted - removed 1338 lines of Ampere/Turing scheduler implementation
csrc/scheduler/matmul_ampere-.h	File deleted - removed 206 lines of Ampere/Turing scheduler declarations
tests/cpp/test_matmul.cpp	Added Hopper guard (9.0+) to MatmulTest base class and MLPBenchmarkTest
benchmarks/cpp/matmul.cpp	Changed arch guard from 8.0 to 9.0, but left dead code checking for 8.0-9.0 range

Sequence Diagram

sequenceDiagram
    participant Client
    participant MatmulScheduler
    participant ArchCheck
    participant HopperScheduler
    participant AmpereMinus as AmpereMinus (REMOVED)
    
    Client->>MatmulScheduler: schedule(fusion, params)
    MatmulScheduler->>MatmulScheduler: Get MMA macro
    
    alt isTuring(macro) || isAmpere(macro)
        MatmulScheduler->>AmpereMinus: BEFORE: schedule_matmul::AmpereMinus(fusion, mparams).run()
        Note over AmpereMinus: Removed 1544 lines<br/>(matmul_ampere-.cpp/h)
        MatmulScheduler->>MatmulScheduler: AFTER: throw NVF_THROW error
        Note right of MatmulScheduler: "Support for Turing and<br/>Ampere was dropped"
    else isHopper(macro)
        MatmulScheduler->>HopperScheduler: schedule_matmul::Hopper(fusion, mparams).run()
        HopperScheduler-->>MatmulScheduler: Success
    else isBlackwell(macro)
        MatmulScheduler->>MatmulScheduler: schedule_matmul::Blackwell(fusion, mparams).run()
    end
    
    Note over ArchCheck: All tests now require<br/>Hopper (arch 9.0+)
    ArchCheck->>ArchCheck: BEFORE: CUDA arch >= 7.5/8.0
    ArchCheck->>ArchCheck: AFTER: CUDA arch >= 9.0
    
    MatmulScheduler-->>Client: Scheduled fusion or error

Loading

[greptile-apps]

Additional Comments (4)

1. tests/cpp/test_matmul_scheduler.cpp, line 142 (link)

   logic: Test will fail on Turing/Ampere since scheduler now throws error for these architectures. Should be NVFUSER_TEST_CUDA_ARCH_RANGE_GUARD(9, 0, 10, 0);

2. tests/cpp/test_matmul_scheduler.cpp, line 168 (link)

   logic: Ampere (8.0) no longer supported. Update to NVFUSER_TEST_CUDA_ARCH_GUARD(9, 0);

3. tests/cpp/test_matmul_scheduler.cpp, line 142-3245 (link)

   logic: 48 tests in this file still have CUDA arch guards for Turing (7.5) and/or Ampere (8.0). Since the scheduler now throws for these architectures (see csrc/scheduler/matmul.cpp:80), all these tests will fail. All guards with 7, 5 should start at 9, 0 and all guards with upper bound 9, 0 should have upper bound 10, 0. Are these tests intentionally meant to verify error handling for unsupported architectures, or should they be updated to only run on Hopper+?

4. csrc/scheduler/matmul.h, line 49 (link)

   syntax: Stale reference to AmpereMinus in comment - this class was removed

_{7 files reviewed, 4 comments}

_{Edit Code Review Agent Settings | Greptile}

[naoyam]

!test

[greptile-apps]

Additional Comments (1)

1. benchmarks/cpp/matmul.cpp, line 404-409 (link)

   style: dead code - the NVFUSER_BENCHMARK_ARCH_SMEM_GUARD(9, 0, ...) on line 401 already skips anything below arch 9.0, so this check for architectures in range [8.0, 9.0) will never execute

_{10 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}