Msvc ops changes (pytorch#15226)

JacobSzwejbka · web-flow · commit 8be08b988b73 · 2025-10-21T20:48:17.000-07:00
### Summary
MSVC really doesn't like the constexpr variables that I've removed in
this PR. It was complaining about them not actually being static at
compile time. Searching around suggests that its the act of putting them
in a variable that breaks it and if I just place the value directly in
all the used locations it should be fine. I tried that and it works.

The other changes are fixing up some compiler flags, adding a missing
include, fixing an interaction between lambdas the ternary operator and
auto, and removing unneeded statement expressions

### Test plan
current ci to verify no regression on linux/mac. Manual testing of msvc
will add ci soon
diff --git a/kernels/optimized/CMakeLists.txt b/kernels/optimized/CMakeLists.txt
@@ -21,7 +21,10 @@ if(NOT EXECUTORCH_ROOT)
   set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../..)
 endif()
 
-set(_common_compile_options -Wno-deprecated-declarations)
+set(_common_compile_options
+    $<$<CXX_COMPILER_ID:MSVC>:/wd4996>
+    $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-Wno-deprecated-declarations>
+)
 
 # Note for apple platform we can rely on Accelerate framework Will come back to
 # this
diff --git a/kernels/optimized/cpu/op_bmm.cpp b/kernels/optimized/cpu/op_bmm.cpp
@@ -150,7 +150,8 @@ Tensor& opt_bmm_out(
   ET_KERNEL_CHECK(
       ctx, check_bmm_out_args(self, mat2, out), InvalidArgument, out);
 
-  constexpr auto name = "bmm.out";
+  static constexpr auto name = "bmm.out";
+
   auto self_type = self.scalar_type();
 
   if (executorch::runtime::isComplexType(self_type)) {
diff --git a/kernels/portable/cpu/op_convolution_backward.cpp b/kernels/portable/cpu/op_convolution_backward.cpp
@@ -305,7 +305,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> convolution_backward_out(
         ret_val);
   }
 
-  constexpr auto name = "convolution_backward.out";
+  static constexpr auto name = "convolution_backward.out";
 
   ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
     conv2d_backward_impl<CTYPE>(
diff --git a/kernels/portable/cpu/op_gather.cpp b/kernels/portable/cpu/op_gather.cpp
@@ -85,7 +85,7 @@ Tensor& gather_out(
       InvalidArgument,
       out);
 
-  constexpr auto name = "gather.out";
+  static constexpr auto name = "gather.out";
 
   ET_SWITCH_REALHBBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&]() {
     gather_helper<CTYPE>(in, index, out, dim);
diff --git a/kernels/portable/cpu/op_masked_scatter.cpp b/kernels/portable/cpu/op_masked_scatter.cpp
@@ -41,13 +41,13 @@ Tensor& masked_scatter_out(
       InvalidArgument,
       out);
 
-  constexpr auto op_name = "masked_scatter.out";
-
   int64_t idx = 0;
   int64_t src_numel = src.numel();
   bool src_numel_check = true;
 
-  ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, op_name, CTYPE, [&]() {
+  static constexpr auto name = "masked_scatter.out";
+
+  ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, name, CTYPE, [&]() {
     const CTYPE* const src_data = src.const_data_ptr<CTYPE>();
     apply_binary_elementwise_fn<CTYPE, bool, CTYPE>(
         [src_data, &idx, &src_numel, &src_numel_check](
diff --git a/kernels/portable/cpu/op_max.cpp b/kernels/portable/cpu/op_max.cpp
@@ -124,7 +124,7 @@ max_unary_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
 
   ET_KERNEL_CHECK(ctx, canCast(in_type, out_type), InvalidArgument, out);
 
-  constexpr auto name = "max.unary_out";
+  static constexpr auto name = "max.unary_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, name, CTYPE_IN, [&] {
     ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE_OUT, [&] {
diff --git a/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp b/kernels/portable/cpu/op_max_pool2d_with_indices_backward.cpp
@@ -169,7 +169,7 @@ Tensor& max_pool2d_with_indices_backward_out(
       InvalidArgument,
       grad_input);
 
-  constexpr auto name = "max_pool2d_with_indices_backward.grad_input";
+  static constexpr auto name = "max_pool2d_with_indices_backward.grad_input";
 
   ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
     max_pool_backward_impl<CTYPE, false>(grad_input, grad_output, indices);
diff --git a/kernels/portable/cpu/op_min.cpp b/kernels/portable/cpu/op_min.cpp
@@ -124,7 +124,7 @@ min_unary_out(KernelRuntimeContext& ctx, const Tensor& in, Tensor& out) {
 
   ET_KERNEL_CHECK(ctx, canCast(in_type, out_type), InvalidArgument, out);
 
-  constexpr auto name = "min.unary_out";
+  static constexpr auto name = "min.unary_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, name, CTYPE_IN, [&] {
     ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE_OUT, [&] {
diff --git a/kernels/portable/cpu/op_native_batch_norm.cpp b/kernels/portable/cpu/op_native_batch_norm.cpp
@@ -102,7 +102,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> _native_batch_norm_legit_no_training_out(
   size_t outer = getLeadingDims(in, C_dim);
   size_t inner = getTrailingDims(in, C_dim);
 
-  constexpr auto name = "native_batch_norm_legit_no_training.out";
+  static constexpr auto name = "native_batch_norm_legit_no_training.out";
 
   ET_SWITCH_FLOATHBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&] {
     const CTYPE* in_data = in.const_data_ptr<CTYPE>();
@@ -259,7 +259,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> _native_batch_norm_legit_no_stats_out(
       InvalidArgument,
       ret_val);
 
-  constexpr auto name = "_native_batch_norm_legit.no_stats_out";
+  static constexpr auto name = "_native_batch_norm_legit.no_stats_out";
 
   ET_SWITCH_FLOATHBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&] {
     const CTYPE* in_data = in.const_data_ptr<CTYPE>();
diff --git a/kernels/portable/cpu/op_native_group_norm.cpp b/kernels/portable/cpu/op_native_group_norm.cpp
@@ -190,7 +190,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> native_group_norm_out(
         ret_val);
   }
 
-  constexpr auto name = "native_group_norm.out";
+  static constexpr auto name = "native_group_norm.out";
 
   ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {
     group_norm<CTYPE>(
diff --git a/kernels/portable/cpu/op_pdist_forward.cpp b/kernels/portable/cpu/op_pdist_forward.cpp
@@ -40,7 +40,7 @@ Tensor& _pdist_forward_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "_pdist_forward.out";
+  static constexpr auto name = "_pdist_forward.out";
 
   ET_SWITCH_FLOATHBF16_TYPES(
       in_type, ctx, name, CTYPE, [&] { pdist<CTYPE>(in, out, p); });
diff --git a/kernels/portable/cpu/op_prod.cpp b/kernels/portable/cpu/op_prod.cpp
@@ -32,7 +32,7 @@ Tensor& prod_out(
 
   ScalarType in_type = in.scalar_type();
   ScalarType out_type = out.scalar_type();
-  constexpr auto name = "prod.int_out";
+  static constexpr auto name = "prod.int_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, name, CTYPE_IN, [&] {
     ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE_OUT, [&] {
@@ -72,7 +72,7 @@ Tensor& prod_int_out(
 
   ScalarType in_type = in.scalar_type();
   ScalarType out_type = out.scalar_type();
-  constexpr auto name = "prod.int_out";
+  static constexpr auto name = "prod.int_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in_type, ctx, name, CTYPE_IN, [&] {
     ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE_OUT, [&] {
diff --git a/kernels/portable/cpu/op_reflection_pad1d.cpp b/kernels/portable/cpu/op_reflection_pad1d.cpp
@@ -44,7 +44,7 @@ Tensor& reflection_pad1d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "reflection_pad1d.out";
+  static constexpr auto name = "reflection_pad1d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad1d<CTYPE>(reflection_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_reflection_pad2d.cpp b/kernels/portable/cpu/op_reflection_pad2d.cpp
@@ -44,7 +44,7 @@ Tensor& reflection_pad2d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "reflection_pad2d.out";
+  static constexpr auto name = "reflection_pad2d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad2d<CTYPE>(reflection_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_reflection_pad3d.cpp b/kernels/portable/cpu/op_reflection_pad3d.cpp
@@ -44,7 +44,7 @@ Tensor& reflection_pad3d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "reflection_pad3d.out";
+  static constexpr auto name = "reflection_pad3d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad3d<CTYPE>(reflection_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_repeat_interleave.cpp b/kernels/portable/cpu/op_repeat_interleave.cpp
@@ -72,7 +72,7 @@ Tensor& repeat_interleave_Tensor_out(
 
   int64_t repeats_sum = 0;
 
-  constexpr auto name = "repeat_interleave.Tensor_out";
+  static constexpr auto name = "repeat_interleave.Tensor_out";
 
   ET_SWITCH_TWO_TYPES(Int, Long, repeats.scalar_type(), ctx, name, CTYPE, [&] {
     const CTYPE* repeats_data = repeats.const_data_ptr<CTYPE>();
diff --git a/kernels/portable/cpu/op_replication_pad1d.cpp b/kernels/portable/cpu/op_replication_pad1d.cpp
@@ -36,7 +36,7 @@ Tensor& replication_pad1d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "replication_pad1d.out";
+  static constexpr auto name = "replication_pad1d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad1d<CTYPE>(replication_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_replication_pad2d.cpp b/kernels/portable/cpu/op_replication_pad2d.cpp
@@ -36,7 +36,7 @@ Tensor& replication_pad2d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "replication_pad2d.out";
+  static constexpr auto name = "replication_pad2d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad2d<CTYPE>(replication_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_replication_pad3d.cpp b/kernels/portable/cpu/op_replication_pad3d.cpp
@@ -36,7 +36,7 @@ Tensor& replication_pad3d_out(
       out);
 
   ScalarType in_type = in.scalar_type();
-  constexpr auto name = "replication_pad3d.out";
+  static constexpr auto name = "replication_pad3d.out";
 
   ET_SWITCH_ALL_TYPES(in_type, ctx, name, CTYPE, [&] {
     pad3d<CTYPE>(replication_ix, in, out, padding);
diff --git a/kernels/portable/cpu/op_roll.cpp b/kernels/portable/cpu/op_roll.cpp
@@ -80,7 +80,7 @@ Tensor& roll_out(
   size_t dim_shift_array_length = static_cast<size_t>(in.dim()); // NOLINT
   IntArrayRef dim_shifts(dim_shift_array, dim_shift_array_length);
 
-  constexpr auto name = "roll.out";
+  static constexpr auto name = "roll.out";
 
   ET_SWITCH_REALHBBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&] {
     const CTYPE* in_data = in.const_data_ptr<CTYPE>();
diff --git a/kernels/portable/cpu/op_scalar_tensor.cpp b/kernels/portable/cpu/op_scalar_tensor.cpp
@@ -22,7 +22,7 @@ scalar_tensor_out(KernelRuntimeContext& ctx, const Scalar& s, Tensor& out) {
 
   ScalarType out_type = out.scalar_type();
 
-  constexpr auto name = "scalar_tensor.out";
+  static constexpr auto name = "scalar_tensor.out";
 
   ET_SWITCH_REALHBBF16_TYPES(out_type, ctx, name, CTYPE, [&]() {
     auto opt_val_casted = utils::internal::check_overflow_scalar_cast<CTYPE>(s);
diff --git a/kernels/portable/cpu/op_scatter.cpp b/kernels/portable/cpu/op_scatter.cpp
@@ -119,7 +119,7 @@ Tensor& scatter_src_out(
   ET_KERNEL_CHECK(
       ctx, resize_tensor(out, in.sizes()) == Error::Ok, InvalidArgument, out);
 
-  constexpr auto name = "scatter.src_out";
+  static constexpr auto name = "scatter.src_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&]() {
     scatter_src_helper<CTYPE>(in, dim, index, src, out);
@@ -146,7 +146,7 @@ Tensor& scatter_value_out(
   ET_KERNEL_CHECK(
       ctx, resize_tensor(out, in.sizes()) == Error::Ok, InvalidArgument, out);
 
-  constexpr auto name = "scatter.value_out";
+  static constexpr auto name = "scatter.value_out";
 
   ET_SWITCH_REALHBBF16_TYPES(in.scalar_type(), ctx, name, CTYPE, [&]() {
     auto opt_val = utils::internal::check_overflow_scalar_cast<CTYPE>(value);
diff --git a/kernels/portable/cpu/op_topk.cpp b/kernels/portable/cpu/op_topk.cpp
@@ -6,14 +6,14 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <c10/util/irange.h>
-#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>
 
+#include <c10/util/irange.h>
+#include <cmath>
+#include <tuple>
+
 namespace torch {
 namespace executor {
 namespace native {
@@ -117,14 +117,18 @@ void perform_topk(
         queue[i].second = i;
       }
 
-      // Perform topk on the queue
-      const auto elem_greater = [](const elem_t& x, const elem_t& y) -> bool {
+      // Perform topk on the queue, explict typing for the lambda to satisfy
+      // msvc compiler.
+      bool (*elem_greater)(const elem_t&, const elem_t&) =
+          [](const elem_t& x, const elem_t& y) -> bool {
         return float_less_than(y.first, x.first);
       };
-      const auto elem_less = [](const elem_t& x, const elem_t& y) -> bool {
+      bool (*elem_less)(const elem_t&, const elem_t&) =
+          [](const elem_t& x, const elem_t& y) -> bool {
         return float_less_than(x.first, y.first);
       };
-      const auto cmp = largest ? elem_greater : elem_less;
+      bool (*cmp)(const elem_t&, const elem_t&) =
+          largest ? elem_greater : elem_less;
       if (use_partial_sort) {
         std::partial_sort(queue, queue + k, queue + dim_size, cmp);
       } else {
@@ -187,7 +191,7 @@ std::tuple<Tensor&, Tensor&> topk_values(
       InvalidArgument,
       out);
 
-  constexpr auto name = "topk.values";
+  static constexpr auto name = "topk.values";
 
   if (in.numel() == 0 || (k == 0 && in.dim() > 0)) {
     return out;
diff --git a/kernels/portable/cpu/op_var.cpp b/kernels/portable/cpu/op_var.cpp
@@ -91,7 +91,7 @@ Tensor& var_out(
   const size_t num = get_reduced_dim_product(in, dim_list);
   const size_t denom = unbiased ? num - 1 : num;
 
-  constexpr auto name = "var.out";
+  static constexpr auto name = "var.out";
 
   ET_SWITCH_FLOATHBF16_TYPES(in.scalar_type(), ctx, name, CTYPE_IN, [&] {
     ET_SWITCH_FLOATHBF16_TYPES(out.scalar_type(), ctx, name, CTYPE_OUT, [&] {
@@ -123,7 +123,7 @@ Tensor& var_correction_out(
       InvalidArgument,
       out);
 
-  constexpr auto name = "var.correction_out";
+  static constexpr auto name = "var.correction_out";
 
   double correction_val = 1;
   if (correction.has_value()) {
diff --git a/kernels/portable/cpu/op_view_as_real_copy.cpp b/kernels/portable/cpu/op_view_as_real_copy.cpp
@@ -64,7 +64,7 @@ Tensor& view_as_real_copy_out(
   ET_KERNEL_CHECK(
       ctx, tensors_have_same_dim_order(self, out), InvalidArgument, out);
 
-  constexpr auto op_name = "view_as_real_copy.out";
+  static constexpr auto op_name = "view_as_real_copy.out";
 
   ET_SWITCH_COMPLEXH_TYPES(self.scalar_type(), ctx, op_name, CTYPE_IN, [&] {
     ET_SWITCH_FLOATH_TYPES(out.scalar_type(), ctx, op_name, CTYPE_OUT, [&] {
diff --git a/kernels/portable/cpu/util/CMakeLists.txt b/kernels/portable/cpu/util/CMakeLists.txt
@@ -21,7 +21,10 @@ endif()
 
 list(TRANSFORM _kernels_util_all_deps__srcs PREPEND "${EXECUTORCH_ROOT}/")
 
-set(_common_compile_options -Wno-deprecated-declarations)
+set(_common_compile_options
+    $<$<CXX_COMPILER_ID:MSVC>:/wd4996>
+    $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:-Wno-deprecated-declarations>
+)
 
 add_library(kernels_util_all_deps ${_kernels_util_all_deps__srcs})
 target_link_libraries(kernels_util_all_deps PRIVATE executorch_core)
diff --git a/kernels/portable/cpu/util/elementwise_util.h b/kernels/portable/cpu/util/elementwise_util.h
@@ -85,7 +85,7 @@ inline void dtype_specialized_elementwise_fn_impl(
   static_assert(
       (std::is_same_v<Args, std::pair<const Tensor*, SupportedTensorDtypes>> &&
        ...));
-  constexpr auto kNumInputs = sizeof...(inputs);
+  static constexpr auto kNumInputs = sizeof...(inputs);
   // All inputs must be of type CTYPE_COMPUTE.
   ET_DCHECK(
       ((inputs.first->scalar_type() ==
@@ -229,7 +229,7 @@ inline void apply_elementwise_fn_generic_impl(
     const Tensor& out,
     SupportedTensorDtypes out_dtypes,
     Args... inputs) {
-  constexpr auto kNumInputs = sizeof...(inputs);
+  static constexpr auto kNumInputs = sizeof...(inputs);
 
   struct InputInfo {
     load_to_compute_fn<CTYPE_COMPUTE> load_to_compute;
diff --git a/runtime/core/exec_aten/util/tensor_util.h b/runtime/core/exec_aten/util/tensor_util.h

Original file line number	Diff line number	Diff line change
`@@ -305,7 +305,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> convolution_backward_out(`
`305`	`305`	`ret_val);`
`306`	`306`	`}`
`307`	`307`
`308`		`- constexpr auto name = "convolution_backward.out";`
	`308`	`+ static constexpr auto name = "convolution_backward.out";`
`309`	`309`
`310`	`310`	`ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {`
`311`	`311`	`conv2d_backward_impl<CTYPE>(`
Original file line number	Diff line number	Diff line change
`@@ -190,7 +190,7 @@ std::tuple<Tensor&, Tensor&, Tensor&> native_group_norm_out(`
`190`	`190`	`ret_val);`
`191`	`191`	`}`
`192`	`192`
`193`		`- constexpr auto name = "native_group_norm.out";`
	`193`	`+ static constexpr auto name = "native_group_norm.out";`
`194`	`194`
`195`	`195`	`ET_SWITCH_FLOATHBF16_TYPES(input.scalar_type(), ctx, name, CTYPE, [&]() {`
`196`	`196`	`group_norm<CTYPE>(`