Per https://github.com/EfficientComputer/deveng/issues/921, --disable-memory-ordering is not intended for public usage at this time, however 16 apps in this repo currently include it.
Applied unconditionally — 13 apps
Each is a single eff_subtarget_compile_options(<name>_fabric PRIVATE ...) with no enclosing condition. Note that in every case the flag sits mid-line among ordinary tuning options — -O3, -I, --pnr-rounds, -DUSE_PRAGMAS. Nothing marks it out, which is why it reads as boilerplate and gets copied along with the rest of the line.
Gated behind if(EFF_BLD_HAND_OPTIMIZED) — 3 apps
The gate defaults off ([app_examples/CMakeLists.txt#L6]
|
# set(EFF_BLD_HAND_OPTIMIZED ON) |
) is a commented-out
# set(EFF_BLD_HAND_OPTIMIZED ON)), so a default build hits only the 13 above. But it is a
documented opt-in, not a guard — [app_examples/README.md#L54]
https://github.com/EfficientComputer/e1x_examples/blob/7fb3a8600576237a9625c5b3a0aaa6a09c9b74f1/app_examples/README.md)
tells the reader to uncomment that line to build the hand-optimized versions, which walks them into all 16.
Per https://github.com/EfficientComputer/deveng/issues/921, --disable-memory-ordering is not intended for public usage at this time, however 16 apps in this repo currently include it.
Applied unconditionally — 13 apps
quickstart-DUSE_PRAGMAS --disable-memory-orderingconv3x3-DUSE_PRAGMAS --disable-memory-ordering -O3conv3x3_dw-DUSE_PRAGMAS --disable-memory-orderingconv3x3xn-DUSE_PRAGMAS --disable-memory-orderingconv5x5-DUSE_PRAGMAS --disable-memory-ordering --pnr-rounds=8 --mof-passes=convert-parallel-for,lift-scf-parallel --disable-mof-option-explorationdmm-I ${CMAKE_SOURCE_DIR}/shared -DUSE_PRAGMAS --disable-memory-orderingdmm_i8-I ${CMAKE_SOURCE_DIR}/shared -DUSE_PRAGMAS --disable-memory-orderingdmv_fp-DUSE_PRAGMAS --disable-memory-orderingfft-DUSE_PRAGMAS --disable-memory-orderingfir-DUSE_PRAGMAS --disable-memory-ordering --pnr-rounds=10knapsack-DUSE_PRAGMAS --disable-memory-ordering --mof-passes=convert-parallel-for,lift-scf-parallel --disable-mof-option-explorationsmv-I ${CMAKE_SOURCE_DIR}/shared -DUSE_PRAGMAS --disable-memory-ordering --mof-passes=convert-parallel-for,lift-scf-parallel --disable-mof-option-explorationspadd-I ${CMAKE_SOURCE_DIR}/shared -DUSE_PRAGMAS --disable-memory-ordering --mof-passes=convert-parallel-for,lift-scf-parallel --disable-mof-option-explorationEach is a single
eff_subtarget_compile_options(<name>_fabric PRIVATE ...)with no enclosing condition. Note that in every case the flag sits mid-line among ordinary tuning options —-O3,-I,--pnr-rounds,-DUSE_PRAGMAS. Nothing marks it out, which is why it reads as boilerplate and gets copied along with the rest of the line.Gated behind
if(EFF_BLD_HAND_OPTIMIZED)— 3 appsbfs--disable-memory-orderingcholesky_decomp--disable-memory-orderingjpeg--disable-memory-ordering --mof-passes=convert-parallel-for,lift-scf-parallel --disable-mof-option-exploration --mark-critical-pathsThe gate defaults off ([app_examples/CMakeLists.txt#L6]
e1x_examples/app_examples/CMakeLists.txt
Line 6 in 7fb3a86
# set(EFF_BLD_HAND_OPTIMIZED ON)), so a default build hits only the 13 above. But it is a documented opt-in, not a guard — [app_examples/README.md#L54] https://github.com/EfficientComputer/e1x_examples/blob/7fb3a8600576237a9625c5b3a0aaa6a09c9b74f1/app_examples/README.md)tells the reader to uncomment that line to build the hand-optimized versions, which walks them into all 16.