Fix portable library build on Windows (pytorch#13700)

**Note: This is a cherry-pick / identical copy of the existing approved
PR (pytorch#13260). I inadvertently
broke the ghstack when attempting to re-order the stack, leading to the
old PR being merged by ghstack into the ghstack-created branch (not
main).**

Enable build of the portable operator library (and executor_runner, now
that we have portable) on the Windows preset. There are a couple of
patterns that weren't building on Windows. Both std::nan and
std::signbit complain about ambiguous overload resolution when given an
integer argument.

To solve the std::isnan issue, I introduced isnan_override in
math_util.h, following the naming pattern from max_override /
min_override / etc. in the same file. It adds a specialized branch for
integers to avoid calling std::isnan. For std::signbit, I just locally
specialized.

I also re-enabled optimized and quantized kernel libraries in the
Windows preset, as they now build with this change.

Author: GregoryComer

kernels/portable/cpu/op_amax.cpp

@@ -9,6 +9,7 @@
 #include <c10/util/irange.h>
 #include <cmath>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -55,7 +56,7 @@ Tensor& amax_out(
           for (const auto out_ix : c10::irange(begin, end)) {
             out_data[out_ix] = plan.execute<CTYPE>(
                 [](CTYPE v, CTYPE max_v) {
-                  return std::isnan(v) || v > max_v ? v : max_v;
+                  return utils::isnan_override(v) || v > max_v ? v : max_v;
                 },
                 out_ix);
           }

kernels/portable/cpu/op_amin.cpp

@@ -8,6 +8,7 @@
 #include <c10/util/irange.h>
 #include <cmath>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -54,7 +55,7 @@ Tensor& amin_out(
           for (const auto out_ix : c10::irange(begin, end)) {
             out_data[out_ix] = plan.execute<CTYPE>(
                 [](CTYPE v, CTYPE min_v) {
-                  return std::isnan(v) || v < min_v ? v : min_v;
+                  return utils::isnan_override(v) || v < min_v ? v : min_v;
                 },
                 out_ix);
           }

kernels/portable/cpu/op_argmax.cpp

@@ -10,6 +10,7 @@
 #include <cmath>
 #include <tuple>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -58,7 +59,7 @@ Tensor& argmax_out(
                   // the below condition as written is equivalent to
                   // !isnan(accval) && (isnan(v) || v > acc_val). See
                   // argument in op_argmin.cpp.
-                  if (!std::isnan(acc_val) && !(v <= acc_val)) {
+                  if (!utils::isnan_override(acc_val) && !(v <= acc_val)) {
                     acc_val = v;
                     acc_ix = ix;
                   }

kernels/portable/cpu/op_argmin.cpp

@@ -10,6 +10,7 @@
 #include <cmath>
 #include <tuple>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -65,7 +66,7 @@ Tensor& argmin_out(
                   // - false, so the result is true. The result is trivially
                   // - true for the above condition that uses isnan(v) as
                   // - well.
-                  if (!std::isnan(acc_val) && !(v >= acc_val)) {
+                  if (!utils::isnan_override(acc_val) && !(v >= acc_val)) {
                     acc_val = v;
                     acc_ix = ix;
                   }

kernels/portable/cpu/op_max.cpp

@@ -10,6 +10,7 @@
 #include <cmath>
 #include <tuple>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -88,8 +89,8 @@ std::tuple<Tensor&, Tensor&> max_out(
               for (const auto out_ix : c10::irange(begin, end)) {
                 std::tuple<CTYPE, long> acc = reduce_over_dim<CTYPE>(
                     [](CTYPE v, long ix, CTYPE acc_val, long acc_ix) {
-                      if (!std::isnan(acc_val) &&
-                          (std::isnan(v) || v > acc_val)) {
+                      if (!utils::isnan_override(acc_val) &&
+                          (utils::isnan_override(v) || v > acc_val)) {
                         acc_val = v;
                         acc_ix = ix;
                       }
@@ -132,7 +133,7 @@ max_unary_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
       data_out[0] = lower_bound<CTYPE_OUT>();
       for (const auto i : c10::irange(in.numel())) {
         CTYPE_OUT val = static_cast<CTYPE_OUT>(data_in[i]);
-        if (std::isnan(val)) {
+        if (utils::isnan_override(val)) {
           data_out[0] = val;
           break;
         }

kernels/portable/cpu/op_min.cpp

@@ -10,6 +10,7 @@
 #include <cmath>
 #include <tuple>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/kernels/portable/cpu/util/reduce_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
@@ -88,8 +89,8 @@ std::tuple<Tensor&, Tensor&> min_out(
               for (const auto out_ix : c10::irange(begin, end)) {
                 std::tuple<CTYPE, long> acc = reduce_over_dim<CTYPE>(
                     [](CTYPE v, long ix, CTYPE acc_val, long acc_ix) {
-                      if (!std::isnan(acc_val) &&
-                          (std::isnan(v) || v < acc_val)) {
+                      if (!utils::isnan_override(acc_val) &&
+                          (utils::isnan_override(v) || v < acc_val)) {
                         acc_val = v;
                         acc_ix = ix;
                       }
@@ -132,7 +133,7 @@ min_unary_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
       data_out[0] = upper_bound<CTYPE_OUT>();
       for (const auto i : c10::irange(in.numel())) {
         CTYPE_OUT val = static_cast<CTYPE_OUT>(data_in[i]);
-        if (std::isnan(val)) {
+        if (utils::isnan_override(val)) {
           data_out[0] = val;
           break;
         }

kernels/portable/cpu/op_relu.cpp

@@ -9,6 +9,7 @@
 #include <cmath>
 
 #include <executorch/kernels/portable/cpu/util/functional_util.h>
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
@@ -45,7 +46,9 @@ Tensor& relu_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
   ET_SWITCH_REALHBF16_TYPES(in.scalar_type(), ctx, "relu.out", CTYPE, [&]() {
     apply_unary_map_fn(
         [](const CTYPE val_in) {
-          return (std::isnan(val_in) || val_in >= CTYPE(0)) ? val_in : CTYPE(0);
+          return (utils::isnan_override(val_in) || val_in >= CTYPE(0))
+              ? val_in
+              : CTYPE(0);
         },
         in.const_data_ptr<CTYPE>(),
         out.mutable_data_ptr<CTYPE>(),

kernels/portable/cpu/op_sign.cpp

@@ -10,6 +10,7 @@
 #include <cstring>
 
 #include <executorch/kernels/portable/cpu/util/functional_util.h>
+#include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
@@ -42,7 +43,7 @@ Tensor& sign_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
     ET_SWITCH_REALHBF16_TYPES(in.scalar_type(), ctx, "sign.out", CTYPE, [&] {
       apply_unary_map_fn(
           [](const CTYPE val_in) {
-            if (std::isnan(val_in)) {
+            if (utils::isnan_override(val_in)) {
               return val_in;
             } else {
               return static_cast<CTYPE>((val_in > 0) - (val_in < 0));

kernels/portable/cpu/op_topk.cpp

@@ -10,6 +10,8 @@
 #include <cmath>
 #include <tuple>
 
+#include <executorch/kernels/portable/cpu/util/math_util.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
 namespace torch {
@@ -62,7 +64,7 @@ bool float_less_than(T x, T y) {
   if constexpr (std::is_integral_v<T>) {
     return x < y;
   }
-  return (!std::isnan(x) && std::isnan(y)) || x < y;
+  return (!utils::isnan_override(x) && utils::isnan_override(y)) || x < y;
 }
 
 template <typename CTYPE, typename elem_t = std::pair<CTYPE, int64_t>>

kernels/portable/cpu/util/math_util.h

@@ -8,10 +8,14 @@
 
 #pragma once
 
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
+
 #if defined(ET_USE_PYTORCH_HEADERS) && ET_USE_PYTORCH_HEADERS
 #include <ATen/cpu/vec/vec.h>
 #endif
 
+#include <type_traits>
+
 namespace torch {
 namespace executor {
 namespace native {
@@ -29,7 +33,8 @@ template <
     typename std::enable_if<std::is_integral<INT_T>::value, bool>::type = true>
 INT_T floor_divide(INT_T a, INT_T b) {
   const auto quot = a / b;
-  if (std::signbit(a) == std::signbit(b)) {
+  // MSVC does not like signbit on integral types.
+  if ((a < 0) == (b < 0)) {
     return quot;
   }
   const auto rem = a % b;
@@ -52,6 +57,20 @@ FLOAT_T floor_divide(FLOAT_T a, FLOAT_T b) {
   return div;
 }
 
+/**
+ * A wrapper around std::isnan that works with MSVC. When building with MSVC,
+ * std::isnan calls with integer inputs fail to compile due to ambiguous
+ * overload resolution.
+ */
+template <typename T>
+bool isnan_override(T a) {
+  if constexpr (!std::is_integral_v<T>) {
+    return std::isnan(a);
+  } else {
+    return false;
+  }
+}
+
 /**
  * Override min/max so we can emulate PyTorch's behavior with NaN entries.
  */