Add experimental Windows wheel build support

include/medium.h

@@ -69,7 +69,9 @@ namespace batoid {
         const Medium* getDevPtr() const override;
 
     private:
-        const double _B1, _B2, _B3, _C1, _C2, _C3;
+        // ``_C1`` / ``_C2`` / ``_C3`` collide with macros from the
+        // Windows SDK (ctype.h), so use a Sellmeier-prefixed name.
+        const double _B1, _B2, _B3, _sellC1, _sellC2, _sellC3;
     };
 
 

setup.py

@@ -36,12 +36,32 @@ def build_extension(self, ext):
             "-DPYTHON_EXECUTABLE=" + sys.executable,
         ]
 
+        # Make pybind11 discoverable in PEP 517 isolated builds, where
+        # CMake's normal find_package would not see the temporary site
+        # packages.  ``python -m pybind11 --cmakedir`` returns the
+        # directory shipped with the pybind11 wheel.
+        try:
+            import pybind11
+            cmake_args.append("-Dpybind11_DIR=" + pybind11.get_cmake_dir())
+        except (ImportError, AttributeError):
+            pass
+
         cfg = "Debug" if self.debug else "RelWithDebInfo"
         build_args = ["--config", cfg]
 
         cmake_args += ["-DCMAKE_BUILD_TYPE=" + cfg]
+        # Hand parallelism to ``cmake --build --parallel`` instead of
+        # forwarding ``-j8`` to the underlying tool: MSBuild rejects
+        # ``-j8`` outright (it expects ``/m`` or ``-maxCpuCount``), while
+        # Make/Ninja are happy with the CMake-level flag.  ``--parallel``
+        # picks up CMAKE_BUILD_PARALLEL_LEVEL when set, or defaults to a
+        # generator-appropriate count.
         if "TF_BUILD" not in env:
-            build_args += ["--", "-j8"]
+            jobs = os.environ.get("CMAKE_BUILD_PARALLEL_LEVEL")
+            if jobs:
+                build_args += ["--parallel", str(jobs)]
+            else:
+                build_args += ["--parallel"]
 
         if not os.path.exists(self.build_temp):
             os.makedirs(self.build_temp)

src/batoid.cpp

@@ -1,4 +1,5 @@
 #include "batoid.h"
+#include <cstddef>  // ptrdiff_t for OpenMP-safe signed loop counters on MSVC.
 #include <limits>
 
 
@@ -10,8 +11,12 @@ namespace batoid {
         double* x, double* y, double* z,
         size_t n, int max_threads
     ) {
+        // MSVC's OpenMP requires a signed integral loop counter; reuse a
+        // ptrdiff_t mirror of ``n`` so the parallel-for index can be signed
+        // without clipping.
+        ptrdiff_t nloop = static_cast<ptrdiff_t>(n);
         #pragma omp parallel for num_threads(max_threads)
-        for(size_t i=0; i<n; i++) {
+        for(ptrdiff_t i=0; i<nloop; i++) {
             double dx = x[i]-dr[0];
             double dy = y[i]-dr[1];
             double dz = z[i]-dr[2];
@@ -26,8 +31,9 @@ namespace batoid {
         double* x, double* y, double* z,
         size_t n, int max_threads
     ) {
+        ptrdiff_t nloop = static_cast<ptrdiff_t>(n);
         #pragma omp parallel for num_threads(max_threads)
-        for(size_t i=0; i<n; i++) {
+        for(ptrdiff_t i=0; i<nloop; i++) {
             double xx = x[i]*drot[0] + y[i]*drot[1] + z[i]*drot[2] + dr[0];
             double yy = x[i]*drot[3] + y[i]*drot[4] + z[i]*drot[5] + dr[1];
             double zz = x[i]*drot[6] + y[i]*drot[7] + z[i]*drot[8] + dr[2];

src/medium.cpp

@@ -137,15 +137,16 @@ namespace batoid {
             double B1, double B2, double B3,
             double C1, double C2, double C3
         ) :
-            Medium(), _B1(B1), _B2(B2), _B3(B3), _C1(C1), _C2(C2), _C3(C3)
+            Medium(), _B1(B1), _B2(B2), _B3(B3),
+            _sellC1(C1), _sellC2(C2), _sellC3(C3)
         {}
 
         SellmeierMedium::~SellmeierMedium() {}
 
         double SellmeierMedium::getN(double wavelength) const {
             // Sellmeier coefficients assume wavelength is in microns, so we have to multiply (1e6)**2
             double x = wavelength*wavelength*1e12;
-            return std::sqrt(1.0 + _B1*x/(x-_C1) + _B2*x/(x-_C2) + _B3*x/(x-_C3));
+            return std::sqrt(1.0 + _B1*x/(x-_sellC1) + _B2*x/(x-_sellC2) + _B3*x/(x-_sellC3));
         }
 
     #if defined(BATOID_GPU)
@@ -160,7 +161,7 @@ namespace batoid {
             Medium* ptr;
             #pragma omp target map(from:ptr)
             {
-                ptr = new SellmeierMedium(_B1, _B2, _B3, _C1, _C2, _C3);
+                ptr = new SellmeierMedium(_B1, _B2, _B3, _sellC1, _sellC2, _sellC3);
             }
             _devPtr = ptr;
             return ptr;

src/obscuration.cpp

@@ -212,7 +212,7 @@ namespace batoid {
                             if (y1 != y2) {
                                 xinters = (y-y1)*(x2-x1)/(y2-y1)+x1;
                             }
-                            if (x1 == x2 or x <= xinters) {
+                            if (x1 == x2 || x <= xinters) {
                                 inside = !inside;
                             }
                         }
@@ -251,7 +251,7 @@ namespace batoid {
         xinters.reserve(16);  // 2 is probably most common, but it's cheap to allocate 16
         for (int j=0; j<ny; j++) {
             double y = ygrid[j];
-            if ((y < ymin) or (y > ymax)) {
+            if ((y < ymin) || (y > ymax)) {
                 for (int i=0; i<nx; i++) {
                     out[j*nx+i] = false;
                 }

src/table.cpp

@@ -58,7 +58,7 @@ namespace batoid {
     double Table::eval(double x, double y) const {
         int ix = int(std::floor((x-_x0)/_dx));
         int iy = int(std::floor((y-_y0)/_dy));
-        if ((ix >= (_nx-1)) or (ix < 0) or (iy >= (_ny-1)) or (iy < 0)) {
+        if ((ix >= (_nx-1)) || (ix < 0) || (iy >= (_ny-1)) || (iy < 0)) {
             return _use_nan ? NAN : 0.0;
         }
         double xgrid = _x0 + ix*_dx;
@@ -95,7 +95,7 @@ namespace batoid {
     ) const {
         int ix = int(std::floor((x-_x0)/_dx));
         int iy = int(std::floor((y-_y0)/_dy));
-        if ((ix >= (_nx-1)) or (ix < 0) or (iy >= (_ny-1)) or (iy < 0)) {
+        if ((ix >= (_nx-1)) || (ix < 0) || (iy >= (_ny-1)) || (iy < 0)) {
             if (_use_nan) {
                 dzdx = NAN;
                 dzdy = NAN;