Repository navigation

madspace: fix SIMD mode - #222

Open

theoheimel wants to merge 26 commits into

mainfrom

fix-madspace-simd

Open

madspace: fix SIMD mode#222
theoheimel wants to merge 26 commits into
mainfrom
fix-madspace-simd

Commits on Oct 7, 2026
- SIMD bugfixes, NEON-specific and general
  theoheimel
  committed
  9d07621
- more SIMD bugfixes, including the one proposed in #209
  theoheimel
  committed
  6f50fb9
- make madspace simd mode selection consistent with madmatrix
  theoheimel
  committed
  12b37ac
- add avx512y mode, similar to madmatrix
  theoheimel
  committed
  a36ef82
- fix avx512y build with GCC: enable AVX2 and FMA explicitly
  theoheimel
  committed
  b21a448
- test madspace and mg7 runs for every supported SIMD mode
  theoheimel
  committed
  f95a83c
- test gridpacks trained in scalar mode across SIMD modes
  theoheimel
  committed
  9215112
- fix NaN in cut-bounded t-channel inverse, test inverse with a pt cut
  theoheimel
  committed
  7095b41
- madspace: fix cut_all/cut_any bound rank, contiguous_dims prefix, broadcast tensor_add, discrete probs gradient, generator differentiable flag
  
  theoheimel
  and
  claude
  committed
  0874881
- madspace: fix sub(0, x) folding, fold stacks of scalar literals
  
  theoheimel
  and
  claude
  committed
  c5c120b
- madspace: avoid constant gathers and full(1) products with a single flavor
  
  theoheimel
  and
  claude
  committed
  dadfe3a
- madspace: contiguous SIMD loads/stores, single inputs as a separate kernel type
  
  theoheimel
  and
  claude
  committed
  bef49aa
- madspace: store literals padded to the SIMD vector size
  
  theoheimel
  and
  claude
  committed
  a5f4b5d
- madspace: write unused gradients of literal and integer inputs to scratch memory
  
  theoheimel
  and
  claude
  committed
  734a34d
- madspace: make external CPU runtime arguments batch-contiguous
  
  theoheimel
  and
  claude
  committed
  1f89dcd
- madspace: test layout independence, single inputs, folding and broadcast gradients
  
  theoheimel
  and
  claude
  committed
  f7113de
- madspace: fix use-after-free of matrix element inputs with the dummy matrix element
  
  theoheimel
  and
  claude
  committed
  9a08ca0
Commits on Oct 12, 2026
- Merge branch 'main' into fix-madspace-simd
  theoheimel
  committed
  7344421
- Merge branch 'feat-madspace-contig-simd' into fix-madspace-simd
  
  theoheimel
  and
  claude
  committed
  4686a48
Commits on Oct 13, 2026
- madspace: make the event field type an enum
  
  theoheimel
  and
  claude
  committed
  974d3a5
- madspace: fix reading and appending event files
  
  theoheimel
  and
  claude
  committed
  326a010
- madspace: make the max-weight freeze independent of thread timing
  
  theoheimel
  and
  claude
  committed
  3f4efe4
Commits on Oct 17, 2026

Navigation Menu

Repository navigation

madspace: fix SIMD mode - #222

madspace: fix SIMD mode#222
theoheimel wants to merge 26 commits into
mainfrom
fix-madspace-simd

Commits on Oct 7, 2026

SIMD bugfixes, NEON-specific and general

more SIMD bugfixes, including the one proposed in #209

make madspace simd mode selection consistent with madmatrix

add avx512y mode, similar to madmatrix

fix avx512y build with GCC: enable AVX2 and FMA explicitly

test madspace and mg7 runs for every supported SIMD mode

test gridpacks trained in scalar mode across SIMD modes

fix NaN in cut-bounded t-channel inverse, test inverse with a pt cut

madspace: fix cut_all/cut_any bound rank, contiguous_dims prefix, broadcast tensor_add, discrete probs gradient, generator differentiable flag

madspace: fix sub(0, x) folding, fold stacks of scalar literals

madspace: avoid constant gathers and full(1) products with a single flavor

madspace: contiguous SIMD loads/stores, single inputs as a separate kernel type

madspace: store literals padded to the SIMD vector size

madspace: write unused gradients of literal and integer inputs to scratch memory

madspace: make external CPU runtime arguments batch-contiguous

madspace: test layout independence, single inputs, folding and broadcast gradients

madspace: fix use-after-free of matrix element inputs with the dummy matrix element

Commits on Oct 12, 2026

Merge branch 'main' into fix-madspace-simd

Merge branch 'feat-madspace-contig-simd' into fix-madspace-simd

Commits on Oct 13, 2026

madspace: make the event field type an enum

madspace: fix reading and appending event files

madspace: make the max-weight freeze independent of thread timing

Commits on Oct 17, 2026

Merge remote-tracking branch 'origin/main' into fix-madspace-simd

madspace: default to auto SIMD mode, scalar on ARM and simd_512 > simd_256 on x86

madspace: auto SIMD mode as C++ default, refuse late mode switches, unpadded literal outputs

madspace tests: SIMD layout tests cpu_only, gradient tests on the test device