Fixes Apple build (#174)

* fix apple

* fix env

* another try

* sse3 skip

* fix windows build

* fix build issue

* fix windows build

_doc/api/validation_cpu.rst

@@ -16,4 +16,12 @@ _validation
 
 .. autofunction:: onnx_extended.validation.cpu._validation.benchmark_cache_tree
 
+.. autofunction:: onnx_extended.validation.cpu._validation.double2float_rn
+
 .. autofunction:: onnx_extended.validation.cpu._validation.murmurhash3_bytes_s32
+
+.. autofunction:: onnx_extended.validation.cpu._validation.float2half_rn
+
+.. autofunction:: onnx_extended.validation.cpu._validation.half2float
+
+.. autofunction:: onnx_extended.validation.cpu._validation.has_sse3

_unittests/ut_validation/test_cpu_fpemu.cpp

@@ -5,13 +5,19 @@
 using namespace cpu_fpemu;
 
 void test_cast() {
+
+#if defined(__SSSE3__)
+
   float f = 1.f;
   double d = 1.f;
   float ff = __double2float_rn(d);
   ASSERT_THROW(f == ff);
   unsigned short u = __float2half_rn(f);
   float bu = __half2float(u);
   ASSERT_THROW(f == bu);
+
+#endif
+
 }
 
 int main(int, char**) {

_unittests/ut_validation/test_cpu_fpemu.py

@@ -1,14 +1,17 @@
 import unittest
 from onnx_extended.ext_test_case import ExtTestCase
-from onnx_extended.validation.cpu._validation import (
-    double2float_rn,
-    float2half_rn,
-    half2float,
-)
+from onnx_extended.validation.cpu._validation import has_sse3
 
 
 class TestCpuFpEmu(ExtTestCase):
+    @unittest.skipIf(not has_sse3(), "SSE3 not available")
     def test_cast(self):
+        from onnx_extended.validation.cpu._validation import (
+            double2float_rn,
+            float2half_rn,
+            half2float,
+        )
+
         self.assertEqual(double2float_rn(1), 1)
         self.assertEqual(half2float(float2half_rn(1)), 1)
 

onnx_extended/validation/cpu/_validation.cpp

@@ -87,6 +87,13 @@ The code is `benchmark_cache_tree
 :return: hash
 )pbdoc");
 
+#if defined(__SSSE3__)
+
+  m.def(
+      "has_sse3", []() -> bool { return true; },
+      R"pbdoc(Tells if SSE3 instructions are available. 
+They are needed to convert floart to half and half to float.)pbdoc");
+
   m.def("double2float_rn", &cpu_fpemu::__double2float_rn, py::arg("d"),
         R"pbdoc(Converts a double into float.)pbdoc");
 
@@ -96,6 +103,15 @@ The code is `benchmark_cache_tree
   m.def("half2float", &cpu_fpemu::__half2float, py::arg("d"),
         R"pbdoc(Converts a half represented as an unsigned short into float.)pbdoc");
 
+#else
+
+  m.def(
+      "has_sse3", []() -> bool { return false; },
+      R"pbdoc(Tells if SSE3 instructions are available. 
+They are needed to convert floart to half and half to float.)pbdoc");
+
+#endif
+
   m.def("sparse_struct_to_dense", &sparse_struct_to_dense, py::arg("v"),
         R"pbdoc(Converts a sparse structure stored in a float tensor
 into a dense vector.)pbdoc");

onnx_extended/validation/cpu/cpu_fpemu.hpp

@@ -13,10 +13,16 @@
 
 #pragma once
 
+#if defined(__SSSE3__)
+
 #include <immintrin.h>
 
+#endif
+
 namespace cpu_fpemu {
 
+#if defined(__SSSE3__)
+
 inline float __double2float_rn(double inval) {
   float out[4] = {0};
   __m128 vout = _mm_cvtpd_ps(_mm_set1_pd(inval));
@@ -47,4 +53,6 @@ inline float __half2float(unsigned short h_val) { return _cvtsh_ss(h_val); }
 
 #endif
 
+#endif
+
 } // namespace cpu_fpemu

pyproject.toml

@@ -113,15 +113,17 @@ archs = ["x86_64"]
 build = "cp*"
 skip = "pypy* *musllinux* cp36-* cp37-* cp38-* cp39-* cp312-* cp313-* cp314-*"
 manylinux-x86_64-image = "manylinux_2_28"
-before-build = "pip install auditwheel-symbols abi3audit"
+before-build = "pip install auditwheel-symbols abi3audit;python -c 'import sysconfig;print(sysconfig.get_platform())'"
 build-verbosity = 1
 repair-wheel-command = "auditwheel-symbols --manylinux 2_28 {wheel} ; abi3audit {wheel} ; auditwheel repair -w {dest_dir} {wheel} || exit 0"
 test-command = "yum list installed&&python -m pip install onnxruntime&&python -c \"import onnx_extended;onnx_extended.check_installation(val=True,ortops=True)\"&&python -c \"import onnx_extended;onnx_extended.check_installation(val=True,ortcy=True,verbose=True)\""
 
 [tool.cibuildwheel.macos]
 archs = ["x86_64"]
 build = "cp*"
+before-build = "brew install llvm libomp"
 skip = "pypy* pp* cp36-* cp37-* cp38-* cp39-* cp312-* cp313-* cp314-*"
+# environment="LDFLAGS='-L/opt/homebrew/opt/llvm/lib -L/opt/homebrew/opt/libomp/lib' CPPFLAGS='-I/opt/homebrew/opt/llvm/include -I/opt/homebrew/opt/libomp/include'"
 
 [tool.cibuildwheel.windows]
 archs = ["AMD64"]