Issues

• Detect max-context activation workspaces retained by one-row inference paths

  jxsl13/perfscan#888 · Filter by author jxsl13 opened on Aug 24, 2026
  1

• Detect fused-weight optimizations that duplicate the unfused resident weights

  jxsl13/perfscan#883 · Filter by author jxsl13 opened on Aug 24, 2026

• Detect execution-policy drift between sibling accelerator types

  jxsl13/perfscan#882 · Filter by author jxsl13 opened on Aug 24, 2026

• Detect repeated objective gradient materialization followed by immediate optimizer consumption

  jxsl13/perfscan#880 · Filter by author jxsl13 opened on Aug 24, 2026

• Port the remaining 53 GoAI-only checks before retiring the internal scanner

  jxsl13/perfscan#877 · Filter by author jxsl13 opened on Aug 24, 2026
  2

• Detect larger graph candidates across adjacent fused differentiable boundaries

  jxsl13/perfscan#872 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Recognize trainable pre-norm attention chains as graph-fusion candidates

  jxsl13/perfscan#871 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect synchronized accelerator composite boundaries eligible for graph fusion

  jxsl13/perfscan#870 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect synchronization-only FFI calls that can piggyback on the next native operation

  jxsl13/perfscan#869 · Filter by author jxsl13 opened on Aug 23, 2026

• Preflight privacy-protected inputs before xctrace launches

  jxsl13/perfscan#867 · Filter by author jxsl13 opened on Aug 23, 2026
  2

• Interleave GPU tile variants: grouped count runs fabricated a Q4_K rows-per-SIMD win

  jxsl13/perfscan#865 · Filter by author jxsl13 opened on Aug 23, 2026

• M2 Metal: reject Q8 activation staging without an integer-dot primitive

  jxsl13/perfscan#864 · Filter by author jxsl13 opened on Aug 23, 2026

• Analyzer idea: flag runtime shader source compilation when an offline IR artifact is possible

  jxsl13/perfscan#863 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Analyzer idea: identify output-invariant reductions nested inside quantized matvec row loops

  jxsl13/perfscan#862 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect false GPU threadgroup-width optimizations with shape-matrix gates

  jxsl13/perfscan#861 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Model GPU register-retention wins and reduction reassociation risk

  jxsl13/perfscan#860 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Detect threadgroup-width tuning that serializes logical reduction lanes

  jxsl13/perfscan#859 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect two-pass GPU reductions eligible for threadgroup-local fusion

  jxsl13/perfscan#858 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Detect lane-invariant SIMD work hidden behind single-lane output stores

  jxsl13/perfscan#857 · Filter by author jxsl13 opened on Aug 23, 2026

• Metal: forced full unrolling is not a shape-stable optimization on Apple M2

  jxsl13/perfscan#852 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• Account for Amdahl share before promoting high-speedup fused leaves

  jxsl13/perfscan#851 · Filter by author jxsl13 opened on Aug 23, 2026

• M2: model width-sensitive cost before recommending quant projection fusion

  jxsl13/perfscan#850 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect command-replay work whose encoding cost is already hidden

  jxsl13/perfscan#849 · Filter by author jxsl13 opened on Aug 23, 2026

• Detect cross-package benchmark contention before accepting regression guards

  jxsl13/perfscan#848 · Filter by author jxsl13 opened on Aug 23, 2026
  1

• arm64: model vector reuse before recommending split load/FMLA schedules

  jxsl13/perfscan#846 · Filter by author jxsl13 opened on Aug 23, 2026
  2