aether-dsc 0.7.2__tar.gz → 0.7.3__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (85) hide show
  1. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/CMakeLists.txt +23 -5
  2. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/PKG-INFO +4 -2
  3. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/README.md +1 -1
  4. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cuda/DriverRuntime.inl +10 -0
  5. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/HostVectorMath.h +5 -4
  6. aether_dsc-0.7.3/aether_dsc/_payload/0f27028f2c1b4ff1e23e90a4a69c665549eb7811c737b6ff74ba55b2dfb44767.bin +0 -0
  7. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/c_core.py +0 -5
  8. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/payload_headers.txt +2 -0
  9. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/pyproject.toml +3 -2
  10. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/conftest.py +1 -1
  11. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_aether_dsc.py +13 -0
  12. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_arrays.py +6 -1
  13. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_openmp.py +33 -0
  14. aether_dsc-0.7.2/aether_dsc/_payload/5dc495598ffcb1b6760e0e0b34d90b0dfbd8fac172880d9cc001d9d45e5076e6.bin +0 -0
  15. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/LICENSE +0 -0
  16. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/NOTICE +0 -0
  17. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_build_backend.py +0 -0
  18. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketArcTrig.h +0 -0
  19. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketExpLog.h +0 -0
  20. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketExpLogTables.h +0 -0
  21. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketHyperbolic.h +0 -0
  22. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketMathCore.h +0 -0
  23. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketMisc.h +0 -0
  24. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketRounding.h +0 -0
  25. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/backend/cpu/simd/math/PacketSinCos.h +0 -0
  26. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/chunk/Chunk.h +0 -0
  27. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/chunk/Copy.h +0 -0
  28. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/chunk/Fill.h +0 -0
  29. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/chunk/StreamOrdered.h +0 -0
  30. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/device/Device.h +0 -0
  31. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/dtype/DType.h +0 -0
  32. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/dtype/WorkingType.h +0 -0
  33. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/dtype/dlpack.h +0 -0
  34. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/err/Error.h +0 -0
  35. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/expr/Expression.h +0 -0
  36. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/expr/Reduce.h +0 -0
  37. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/fft/Fft.h +0 -0
  38. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/fft/pocketfft_hdronly.h +0 -0
  39. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/index/BundleIndex.h +0 -0
  40. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/index/Offset.h +0 -0
  41. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/index/SampleIndex.h +0 -0
  42. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/interop/DLPack.h +0 -0
  43. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/interop/DLPackLegacy.h +0 -0
  44. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/layout/Extents.h +0 -0
  45. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/layout/Layout.h +0 -0
  46. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/layout/detail/Carray.h +0 -0
  47. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/macros.h +0 -0
  48. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/BandedFwd.h +0 -0
  49. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/BoundedTrig.h +0 -0
  50. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/HostVectorMathRoute.h +0 -0
  51. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/MathDispatch.h +0 -0
  52. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/detail/MathDispatchUndef.h +0 -0
  53. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/explog.h +0 -0
  54. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/hyperbolic.h +0 -0
  55. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/invtrig.h +0 -0
  56. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/math.h +0 -0
  57. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/root.h +0 -0
  58. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/round.h +0 -0
  59. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/special.h +0 -0
  60. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/math/trig.h +0 -0
  61. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/residency/PeerAccess.h +0 -0
  62. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/residency/Transport.h +0 -0
  63. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/typedefs.h +0 -0
  64. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/view/MakeRuntimeView.h +0 -0
  65. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/view/RuntimeView.h +0 -0
  66. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/_sdist_include/aether/view/View.h +0 -0
  67. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether/__init__.py +0 -0
  68. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether/_dtype.py +0 -0
  69. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether/fft.py +0 -0
  70. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/__init__.py +0 -0
  71. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/_core.py +0 -0
  72. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/_openmp.py +0 -0
  73. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/_rtc_check.py +0 -0
  74. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/_seal_impl.py +0 -0
  75. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/payload_headers_host.txt +0 -0
  76. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/aether_dsc/seal.py +0 -0
  77. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/aether_core.cpp +0 -0
  78. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/cufft_plan.h +0 -0
  79. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/driver_runtime.cpp +0 -0
  80. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/plugin.cpp +0 -0
  81. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/plugin.map +0 -0
  82. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/src/seam.h +0 -0
  83. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_c_core.py +0 -0
  84. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_fft.py +0 -0
  85. {aether_dsc-0.7.2 → aether_dsc-0.7.3}/tests/test_ft_payload.py +0 -0
@@ -88,6 +88,22 @@ if(UNIX)
88
88
  else()
89
89
  set_target_properties(_core PROPERTIES INSTALL_RPATH "$ORIGIN/_omp")
90
90
  endif()
91
+ # The runtime must be the one this compiler's code calls: libgomp for GCC, LLVM's libomp for Clang. Shipping the other one
92
+ # loads fine and then fails or runs on a second runtime. In a conda environment `libgomp.so.1` can be a symlink to LLVM's
93
+ # libomp, so the check is on the resolved file, not on the name a linker was given.
94
+ get_filename_component(_omp_file "${_omp_real}" NAME)
95
+ if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
96
+ set(_omp_want "libgomp")
97
+ elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
98
+ set(_omp_want "lib(omp|iomp)")
99
+ else()
100
+ set(_omp_want "")
101
+ endif()
102
+ if(_omp_want AND NOT _omp_file MATCHES "^${_omp_want}")
103
+ message(FATAL_ERROR "the OpenMP runtime ${_omp_real} does not match the ${CMAKE_CXX_COMPILER_ID} compiler ${CMAKE_CXX_COMPILER} "
104
+ "(wanted ${_omp_want}*): a libgomp.so.1 that links to LLVM's libomp, as in a conda environment, is the usual cause. "
105
+ "Build with the compiler's own runtime visible (a toolchain outside the conda prefix, or OpenMP_CXX_LIBRARIES set).")
106
+ endif()
91
107
  # Refuse by name anything that is not a binary (a text linker script copied here fails at import: "invalid ELF header").
92
108
  file(READ "${_omp_real}" _omp_magic LIMIT 4 HEX)
93
109
  if(NOT (_omp_magic STREQUAL "7f454c46" OR _omp_magic STREQUAL "cffaedfe" OR _omp_magic STREQUAL "cafebabe"))
@@ -100,6 +116,9 @@ if(UNIX)
100
116
  message(FATAL_ERROR "the OpenMP runtime ${_omp_real} has no SONAME (readelf -d found none)")
101
117
  endif()
102
118
  set(_omp_name "${CMAKE_MATCH_1}")
119
+ if(_omp_want AND NOT _omp_name MATCHES "^${_omp_want}")
120
+ message(FATAL_ERROR "the OpenMP runtime ${_omp_real} has SONAME ${_omp_name}, not the ${CMAKE_CXX_COMPILER_ID} compiler's (wanted ${_omp_want}*)")
121
+ endif()
103
122
  endif()
104
123
  message(STATUS "OpenMP runtime shipped as aether/_omp/${_omp_name}: ${_omp_real}")
105
124
  install(FILES "${_omp_real}" DESTINATION aether/_omp RENAME ${_omp_name})
@@ -118,10 +137,9 @@ if(UNIX)
118
137
  if(NOT _cores)
119
138
  message(FATAL_ERROR "no aether/_core*.so under ${_dest}: cannot point it at the shipped OpenMP runtime")
120
139
  endif()
121
- # The shipped libomp carries Homebrew's signature, which install_name_tool would invalidate (with a warning): drop it
122
- # first. `_core` is linker-signed and must keep its signature space (install_name_tool refuses an unsigned one).
123
- # Both are re-signed (ad hoc) and verified below: Apple silicon kills a process that maps an invalid signature.
124
- execute_process(COMMAND codesign --remove-signature "${_dest}/_omp/libomp.dylib" COMMAND_ERROR_IS_FATAL ANY)
140
+ # install_name_tool invalidates the signatures (it warns) and needs them in place: a stripped libomp leaves a
141
+ # __LINKEDIT newer Xcode refuses to edit, and `_core` must keep its signature space. Both are re-signed (ad hoc)
142
+ # and verified below: Apple silicon kills a process that maps an invalid signature.
125
143
  execute_process(COMMAND install_name_tool -id @rpath/libomp.dylib "${_dest}/_omp/libomp.dylib" COMMAND_ERROR_IS_FATAL ANY)
126
144
  # Rewrite from what `_core` actually records (the linker wrote the runtime's own install name: an absolute Homebrew
127
145
  # path), then refuse a build that still has an OpenMP reference outside @rpath.
@@ -163,7 +181,7 @@ endif()
163
181
  # (the CUDA runtime functions aether's headers call, defined over the driver API), and nvcc links them. It needs
164
182
  # libcuda.so.1 at run time and no CUDA runtime library. No CMake CUDA language: `_core` stays CUDA-free.
165
183
  function(_aether_plugin)
166
- set(_hints "$ENV{AETHER_CUDA_ROOT}")
184
+ set(_hints "$ENV{AETHER_CUDA_ROOT}" "$ENV{CUDA_PATH}") # an explicit root first, then the toolkit the environment names
167
185
  file(GLOB _cands "/usr/local/cuda-12.*" "/usr/local/cuda")
168
186
  list(SORT _cands COMPARE NATURAL ORDER DESCENDING)
169
187
  find_program(_aether_nvcc nvcc HINTS ${_hints} ${_cands} PATH_SUFFIXES bin)
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: aether-dsc
3
- Version: 0.7.2
3
+ Version: 0.7.3
4
4
  Summary: aether-dsc: the sealed AETHER header payload (aether_dsc) and CPU/GPU arrays for Python (aether)
5
5
  Author: Alessandro Masat
6
6
  License: Apache-2.0
@@ -12,6 +12,8 @@ Project-URL: Repository, https://github.com/amasat01/aether
12
12
  Requires-Python: >=3.9
13
13
  Provides-Extra: test
14
14
  Requires-Dist: pytest; extra == "test"
15
+ Requires-Dist: numpy; extra == "test"
16
+ Requires-Dist: raptor-core<0.8,>=0.7.3; extra == "test"
15
17
  Provides-Extra: cufft12
16
18
  Requires-Dist: nvidia-cufft-cu12<12,>=11; extra == "cufft12"
17
19
  Provides-Extra: cufft13
@@ -74,4 +76,4 @@ everything they depend on as one block of C text. Compile it with
74
76
 
75
77
  ## Threading
76
78
 
77
- Free-threaded CPython (3.13t/3.14t) is supported: the compiled modules declare GIL-free operation and the GIL stays disabled after import. Any number of threads may call module-level functions, build, compile, plan and cache concurrently. Distinct objects may be used from distinct threads without synchronisation. One stateful object (a stream, capture, launcher, graph, composer, plan, pipeline, active set, host kernel or arg block) shared by several threads is memory-safe — its calls serialise and a consumed object raises — but the ORDER of those calls is the caller's responsibility, exactly as for a NumPy array or a CuPy stream. CUDA adds two rules of its own: a stream capture is begun, filled and ended by one thread, and while any capture is open no thread may synchronise the whole device (stream-level synchronisation is fine). GPU routes are supported on free-threaded 3.14; on 3.13t the CPU route runs.
79
+ Free-threaded CPython (3.14t/3.15t) is supported: the compiled modules declare GIL-free operation and the GIL stays disabled after import. Any number of threads may call module-level functions, build, compile, plan and cache concurrently. Distinct objects may be used from distinct threads without synchronisation. One stateful object (a stream, capture, launcher, graph, composer, plan, pipeline, active set, host kernel or arg block) shared by several threads is memory-safe — its calls serialise and a consumed object raises — but the ORDER of those calls is the caller's responsibility, exactly as for a NumPy array or a CuPy stream. CUDA adds two rules of its own: a stream capture is begun, filled and ended by one thread, and while any capture is open no thread may synchronise the whole device (stream-level synchronisation is fine). GPU routes are supported on free-threaded 3.14; the CPU route runs elsewhere.
@@ -54,4 +54,4 @@ everything they depend on as one block of C text. Compile it with
54
54
 
55
55
  ## Threading
56
56
 
57
- Free-threaded CPython (3.13t/3.14t) is supported: the compiled modules declare GIL-free operation and the GIL stays disabled after import. Any number of threads may call module-level functions, build, compile, plan and cache concurrently. Distinct objects may be used from distinct threads without synchronisation. One stateful object (a stream, capture, launcher, graph, composer, plan, pipeline, active set, host kernel or arg block) shared by several threads is memory-safe — its calls serialise and a consumed object raises — but the ORDER of those calls is the caller's responsibility, exactly as for a NumPy array or a CuPy stream. CUDA adds two rules of its own: a stream capture is begun, filled and ended by one thread, and while any capture is open no thread may synchronise the whole device (stream-level synchronisation is fine). GPU routes are supported on free-threaded 3.14; on 3.13t the CPU route runs.
57
+ Free-threaded CPython (3.14t/3.15t) is supported: the compiled modules declare GIL-free operation and the GIL stays disabled after import. Any number of threads may call module-level functions, build, compile, plan and cache concurrently. Distinct objects may be used from distinct threads without synchronisation. One stateful object (a stream, capture, launcher, graph, composer, plan, pipeline, active set, host kernel or arg block) shared by several threads is memory-safe — its calls serialise and a consumed object raises — but the ORDER of those calls is the caller's responsibility, exactly as for a NumPy array or a CuPy stream. CUDA adds two rules of its own: a stream capture is begun, filled and ended by one thread, and while any capture is open no thread may synchronise the whole device (stream-level synchronisation is fine). GPU routes are supported on free-threaded 3.14; the CPU route runs elsewhere.
@@ -287,6 +287,15 @@ cudaError_t cudaLaunchKernel(const void* func, dim3 grid, dim3 block, void** arg
287
287
  AETHER_DRV(launch, cuLaunchKernel);
288
288
  return fromDriver(launch(f, grid.x, grid.y, grid.z, block.x, block.y, block.z, static_cast<unsigned>(shared), stream(s), args, nullptr));
289
289
  }
290
+ // CUDA 13's `<<<>>>` stub resolves a handle once (__cudaGetKernel) and launches through it; CUDA 12's nvcc defines
291
+ // __NV_LEGACY_LAUNCH and calls cudaLaunchKernel. The handle is the host stub itself, so both paths meet above.
292
+ cudaError_t __cudaGetKernel(cudaKernel_t* kernel, const void* func) {
293
+ *kernel = reinterpret_cast<cudaKernel_t>(const_cast<void*>(func));
294
+ return cudaSuccess;
295
+ }
296
+ cudaError_t __cudaLaunchKernel(cudaKernel_t k, dim3 grid, dim3 block, void** args, std::size_t shared, cudaStream_t s) {
297
+ return cudaLaunchKernel(reinterpret_cast<const void*>(k), grid, block, args, shared, s);
298
+ }
290
299
 
291
300
  // ---- errors and versions -----------------------------------------------------------------------------------
292
301
  cudaError_t cudaGetLastError(void) {
@@ -490,6 +499,7 @@ cudaError_t cudaThreadExchangeStreamCaptureMode(cudaStreamCaptureMode* mode) {
490
499
  // ---- the per-thread default stream spellings (`--default-stream per-thread` builds) -----------------------------
491
500
  inline cudaStream_t perThreadStream(cudaStream_t s) { return s == nullptr ? reinterpret_cast<cudaStream_t>(CU_STREAM_PER_THREAD) : s; }
492
501
  cudaError_t cudaLaunchKernel_ptsz(const void* f, dim3 g, dim3 b, void** a, std::size_t sh, cudaStream_t s) { return cudaLaunchKernel(f, g, b, a, sh, perThreadStream(s)); }
502
+ cudaError_t __cudaLaunchKernel_ptsz(cudaKernel_t k, dim3 g, dim3 b, void** a, std::size_t sh, cudaStream_t s) { return __cudaLaunchKernel(k, g, b, a, sh, perThreadStream(s)); }
493
503
  cudaError_t cudaMallocAsync_ptsz(void** p, size_t n, cudaStream_t s) { return cudaMallocAsync(p, n, perThreadStream(s)); }
494
504
  cudaError_t cudaFreeAsync_ptsz(void* p, cudaStream_t s) { return cudaFreeAsync(p, perThreadStream(s)); }
495
505
  cudaError_t cudaMemcpyAsync_ptsz(void* d, const void* src, size_t n, cudaMemcpyKind k, cudaStream_t s) { return cudaMemcpyAsync(d, src, n, k, perThreadStream(s)); }
@@ -91,8 +91,8 @@
91
91
  *
92
92
  * Routed: exp, exp2, exp10, expm1, log, log2, log10, log1p, pow, sin,
93
93
  * cos, tan, tanh, sinh, cosh, asinh, acosh, atanh, atan, atan2, asin,
94
- * acos, cbrt, hypot, sqrt, rsqrt, floor, ceil, trunc, round, rint, fdim
95
- * (vector ABI); abs, copysign, fmax, fmin (inline, see `hostvec::fmax`).
94
+ * acos, cbrt, hypot, rsqrt, floor, ceil, trunc, round, rint, fdim
95
+ * (vector ABI); sqrt, abs, copysign, fmax, fmin (inline, see `hostvec::fmax`).
96
96
  * Not routed (no packet; stay on `std::`): fmod, fma, erf, erfc; the
97
97
  * numpy-semantics `remainder` is built on `fmod`. `sincos` reaches the hook
98
98
  * through its `sin` and `cos` legs: GCC does not vectorise a call that
@@ -285,7 +285,6 @@ AETHER_HVM_DEFINE_UNARY(asin, vasin)
285
285
  AETHER_HVM_DEFINE_UNARY(acos, vacos)
286
286
  AETHER_HVM_DEFINE_UNARY(cbrt, vcbrt)
287
287
  AETHER_HVM_DEFINE_BINARY(hypot, vhypot)
288
- AETHER_HVM_DEFINE_UNARY(sqrt, sqrt)
289
288
  AETHER_HVM_DEFINE_UNARY(rsqrt, vrsqrt)
290
289
  AETHER_HVM_DEFINE_UNARY(floor, vfloor)
291
290
  AETHER_HVM_DEFINE_UNARY(expm1, vexpm1)
@@ -363,7 +362,9 @@ AETHER_HVM_ROUTE1(asin, std::asin(x))
363
362
  AETHER_HVM_ROUTE1(acos, std::acos(x))
364
363
  AETHER_HVM_ROUTE1(cbrt, std::cbrt(x))
365
364
  AETHER_HVM_ROUTE2(hypot)
366
- AETHER_HVM_ROUTE1(sqrt, std::sqrt(x))
365
+ // sqrt: the compiler widens it natively (vsqrtpd), so no vector-ABI symbol.
366
+ inline double sqrt(double x) { return std::sqrt(x); }
367
+ inline float sqrt(float x) { return std::sqrt(x); }
367
368
  AETHER_HVM_ROUTE1(rsqrt, 1.0f / std::sqrt(x))
368
369
  AETHER_HVM_ROUTE1(floor, std::floor(x))
369
370
  AETHER_HVM_ROUTE1(expm1, std::expm1(x))
@@ -56,11 +56,6 @@ def _memo(item: CFunction, rank: int) -> CFunction:
56
56
  return _REGISTRY.setdefault(item.name, item)
57
57
 
58
58
 
59
- def real_type(width: str) -> str:
60
- _check_width(width)
61
- return _WIDTH[width][0]
62
-
63
-
64
59
  def _t(width: str) -> str:
65
60
  return f"aether_c_{width}_t"
66
61
 
@@ -19,6 +19,7 @@
19
19
  # `aether/aether.h` (the umbrella) is deliberately not listed: it
20
20
  # transitively includes every one of the ~28 excluded headers, so it is
21
21
  # not yet part of the payload either.
22
+ aether/accum/TwoSum.h
22
23
  aether/accum/run.h
23
24
  aether/device.h
24
25
  aether/device/Device.h
@@ -51,6 +52,7 @@ aether/linalg/Cholesky.h
51
52
  aether/linalg/Eigh.h
52
53
  aether/linalg/Lu.h
53
54
  aether/linalg/Triangular.h
55
+ aether/linalg/WarpEigh.h
54
56
  aether/linalg/detail/Common.h
55
57
  aether/linalg/linalg.h
56
58
  aether/linalg/team.h
@@ -35,7 +35,8 @@ classifiers = [
35
35
  dependencies = []
36
36
 
37
37
  [project.optional-dependencies]
38
- test = ["pytest"]
38
+ # test_arrays compares with numpy; test_ft_payload uses raptor's free-threading rows
39
+ test = ["pytest", "numpy", "raptor-core>=0.7.3,<0.8"]
39
40
  # GPU FFTs: cuFFT, opened on the first device transform (not bundled: NVIDIA licence; not imported otherwise)
40
41
  cufft12 = ["nvidia-cufft-cu12>=11,<12"]
41
42
  cufft13 = ["nvidia-cufft>=12,<13"]
@@ -67,7 +68,7 @@ provider = "scikit_build_core.metadata.setuptools_scm"
67
68
  [tool.pytest.ini_options]
68
69
  testpaths = ["tests"]
69
70
  markers = [
70
- "ft: free-threading rows (raptor.conformance.freethreading) -- skip on a GIL build; select with -m ft on 3.13t/3.14t.",
71
+ "ft: free-threading rows (raptor.conformance.freethreading) -- skip on a GIL build; select with -m ft on 3.14t/3.15t.",
71
72
  ]
72
73
  # Makes `import aether_dsc` work against the checked-out tree without an
73
74
  # install — this repo root, prepended to sys.path before collection (mirrors
@@ -6,4 +6,4 @@ so the marker cannot live in pyproject.toml."""
6
6
 
7
7
 
8
8
  def pytest_configure(config):
9
- config.addinivalue_line("markers", "ft: the free-threaded smoke rows (run on 3.13t/3.14t with -m ft)")
9
+ config.addinivalue_line("markers", "ft: the free-threaded smoke rows (run on 3.14t/3.15t with -m ft)")
@@ -393,3 +393,16 @@ def test_the_target_deprecation_notice_does_not_count_as_a_diagnostic():
393
393
  "nvrtc: warning: Architectures prior to '<compute/sm>_75' are deprecated and may be "
394
394
  "removed in a future release")
395
395
  assert not _is_target_deprecation_notice("error: identifier \"foo\" is undefined")
396
+
397
+
398
+ def test_release_wheel_jobs_build_for_the_portable_isa():
399
+ """A source build targets this machine (-march=native); every wheel job must say x86-64-v3 instead."""
400
+ release = Path(__file__).resolve().parents[2] / ".github" / "workflows" / "release.yml"
401
+ if not release.is_file():
402
+ pytest.skip("release.yml is not next to the tests (installed-wheel run)")
403
+ jobs = re.split(r"(?m)^ (?=[A-Za-z0-9_-]+:\s*$)", release.read_text())
404
+ wheel_jobs = [j for j in jobs if "cibuildwheel" in j and "runs-on" in j]
405
+ assert wheel_jobs, "no cibuildwheel job found: the pattern needs updating"
406
+ for job in wheel_jobs:
407
+ envs = len(re.findall(r"CIBW_ENVIRONMENT(?:_LINUX|_MACOS)?:", job))
408
+ assert envs and job.count("AETHER_MARCH=x86-64-v3") >= envs, f"wheel job {job.split(':', 1)[0]}: every CIBW_ENVIRONMENT needs AETHER_MARCH=x86-64-v3"
@@ -5,6 +5,7 @@ numpy, cupy and torch rows skip when that library is absent (aether itself needs
5
5
 
6
6
  import ctypes
7
7
  import gc
8
+ import importlib.util
8
9
  import pathlib
9
10
  import struct
10
11
  import subprocess
@@ -30,6 +31,7 @@ def _gpu():
30
31
 
31
32
 
32
33
  gpu = pytest.mark.skipif(not _gpu(), reason="needs a CUDA plugin that loads and a device")
34
+ needs_cupy = pytest.mark.skipif(importlib.util.find_spec("cupy") is None, reason="needs cupy (the subprocess rows import it)")
33
35
 
34
36
 
35
37
  def _run(code, **env):
@@ -280,6 +282,7 @@ def test_zero_fill_capture_two_replays():
280
282
 
281
283
 
282
284
  @gpu
285
+ @needs_cupy
283
286
  def test_staged_upload_then_dropping_its_stream_does_not_break_the_next_copy():
284
287
  """A pageable host to device copy on a user stream leaves the staging buffer busy; the next staged copy must
285
288
  not wait on that stream, which the caller has since destroyed (it used to segfault)."""
@@ -300,6 +303,7 @@ def test_staged_upload_then_dropping_its_stream_does_not_break_the_next_copy():
300
303
  @gpu
301
304
  @pytest.mark.parametrize("blocking", [False, True])
302
305
  @pytest.mark.parametrize("mode", ["legacy", "stream"])
306
+ @needs_cupy
303
307
  def test_gc_free_inside_global_capture(mode, blocking):
304
308
  out = _run(f"""
305
309
  import gc, aether, cupy as cp
@@ -348,6 +352,7 @@ def test_early_drop_after_launch_is_stream_ordered(mode):
348
352
 
349
353
  # ---- row 6: lifetimes at exit
350
354
  @gpu
355
+ @needs_cupy
351
356
  def test_arrays_alive_at_exit_and_stream_object_held():
352
357
  out = _run("""
353
358
  import aether, cupy as cp
@@ -363,7 +368,7 @@ def test_arrays_alive_at_exit_and_stream_object_held():
363
368
  assert out.returncode == 0 and out.stdout.strip() == "ok", (out.returncode, out.stderr[-400:])
364
369
 
365
370
 
366
- # ---- row 7: threads (the free-threaded smoke when run on 3.13t/3.14t with -m ft)
371
+ # ---- row 7: threads (the free-threaded smoke when run on 3.14t/3.15t with -m ft)
367
372
  @pytest.mark.ft
368
373
  def test_threads_allocate_free_export():
369
374
  errors, before = [], _core._exports_live()
@@ -2,8 +2,13 @@
2
2
  # SPDX-License-Identifier: Apache-2.0
3
3
 
4
4
  """`aether_dsc.openmp()`: where the one shipped OpenMP runtime is, for the packages that link it."""
5
+ import ctypes
6
+ import os
7
+ import re
5
8
  import sys
6
9
 
10
+ import pytest
11
+
7
12
  import aether_dsc
8
13
 
9
14
 
@@ -28,3 +33,31 @@ def test_a_package_with_the_runtime_reports_its_file_and_flags(tmp_path, monkeyp
28
33
  assert found.rpath() == str(omp)
29
34
  assert f"-Wl,-rpath,{omp}" in found.link_flags and str(omp / name) in found.link_flags
30
35
  assert "-fopenmp" in found.compile_flags
36
+
37
+
38
+ _RUNTIME = re.compile(r"lib(gomp|omp|iomp5)[^/]*\.(so(\.\d+)*|dylib)$")
39
+
40
+
41
+ def _loaded_runtimes():
42
+ """The distinct OpenMP runtime files mapped into this process."""
43
+ if sys.platform == "darwin":
44
+ dyld = ctypes.CDLL(None)
45
+ dyld._dyld_get_image_name.restype = ctypes.c_char_p
46
+ names = [dyld._dyld_get_image_name(i).decode() for i in range(dyld._dyld_image_count())]
47
+ else:
48
+ names = [line.split()[-1] for line in open("/proc/self/maps") if "/" in line.split()[-1]]
49
+ return {os.path.realpath(n) for n in names if _RUNTIME.search(n)}
50
+
51
+
52
+ def test_one_openmp_runtime_is_loaded_and_it_is_the_shipped_one():
53
+ shipped = aether_dsc.openmp()
54
+ if shipped is None:
55
+ pytest.skip("this aether install ships no OpenMP runtime (a source checkout)")
56
+ import aether
57
+ from aether import fft
58
+
59
+ fft.fft(aether.zeros((64, 4096, 2), "float64")) # the first OpenMP region runs here
60
+ runtimes = _loaded_runtimes()
61
+ assert runtimes == {os.path.realpath(str(shipped.library))}, runtimes
62
+ if (os.cpu_count() or 1) > 1 and "OMP_NUM_THREADS" not in os.environ:
63
+ assert ctypes.CDLL(str(shipped.library)).omp_get_max_threads() > 1
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes