[llvm-branch-commits] [llvm] [AMDGPU] Fix the packed F32 unpack window in SIPreEmitPeephole (PR #233721)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Nov 2 00:43:02 PST 2026
================
@@ -371,6 +371,24 @@ inline CoExecInfo CoExecInfo::build(unsigned UnitOccupancy,
/// Get co-execution info for a gfx950 MFMA instruction.
CoExecInfo getMFMACoExecInfo(unsigned Opcode);
+/// Get the first cycle, counted from \p MFMA issuing at cycle 0, at which a
+/// VALU can issue on gfx942 with the peak TOPS limiter enabled, or on gfx950.
----------------
lijinpei-amd wrote:
**Default:** PTL is off by default on MI308X: `amdgpu.ptl` defaults to `-1`, which the driver treats as disabled ([driver](https://github.com/ROCm/ROCK-Kernel-Driver/blob/820212794d184710298efcecce42c0215bfc9c6a/drivers/gpu/drm/amd/amdgpu/gfx_v9_4_3.c#L2486-L2509)).
**Profiling:** rocprofiler-sdk still disables PTL during counter collection ([controller.cpp](https://github.com/ROCm/rocm-systems/blob/d2e0de357464cd34cabaea859b71c10da88e7ae9/projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/controller.cpp#L224-L235)), and KFD re-enables it when the profiling process exits ([kfd_process.c](https://github.com/ROCm/ROCK-Kernel-Driver/blob/820212794d184710298efcecce42c0215bfc9c6a/drivers/gpu/drm/amd/amdkfd/kfd_process.c#L1257-L1258)). Thread trace and PC sampling leave it alone. The ATT numbers in this PR were taken with PTL set explicitly, so profiling doesn't affect them.
@shiltian @arsenm, should I add a target feature (or a gfx942 sub-target distinction) for PTL state? The compiler can't tell MI300X from MI308X, or PTL on from off. A feature would let this window pick the right VALU cycle.
https://github.com/llvm/llvm-project/pull/233721
More information about the llvm-branch-commits
mailing list