Add vectorization in elementwise_util (#9432)

This is a first cut at #9241 . In this PR I've vectorized a small
initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul,
pow, and sigmoid. In addition, the following ops should have gotten
vectorized automatically because they already used generic lambdas: add,
div, rsub, sub. I've left covering ops that use the `unary_ufunc_*`
utilities in
[pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h)
for a follow-up push, because pattern.h and elementwise_util need some
work before we can migrate pattern.h's utilities to be backed by
elementwise_util.

This PR adds an interesting testing problem: in theory, *all* operators
might need test cases long enough to tickle vectorization, because we
might accidentally vectorize ops unexpectedly and break their lambdas
due to anticipated differences in semantics. I address this issue by
using Vectorized for the scalar prologue/epilogue in debug mode (we run
tests in both debug and release) so that we can detect broken lambdas. I
additionally intentionally introduced a bug in the vectorized path in
elementwise_util and manually verified that we saw test failures for
each vectorized op called out above.

Author: swolchok

.lintrunner.toml

@@ -272,6 +272,8 @@ exclude_patterns = [
     'exir/verification/bindings.cpp',
     'extension/**',
     # Uses properly-gated (ET_USE_PYTORCH_HEADERS) ATen include.
+    'kernels/portable/cpu/util/elementwise_util.h',
+    'kernels/portable/cpu/util/math_util.h',
     'kernels/portable/cpu/util/vectorized_math.h',
     'kernels/optimized/**',
     'runtime/core/exec_aten/**',

kernels/portable/cpu/op_add.cpp

@@ -102,14 +102,18 @@ Tensor& add_scalar_out(
   static constexpr const char op_name[] = "add.Scalar_out";
 
   ET_SWITCH_REALB_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
+    CTYPE_COMPUTE val_b = utils::scalar_to<CTYPE_COMPUTE>(b);
+    CTYPE_COMPUTE val_alpha = utils::scalar_to<CTYPE_COMPUTE>(alpha);
+    auto val_alpha_times_b = val_alpha * val_b;
     utils::apply_unitensor_elementwise_fn<
         CTYPE_COMPUTE,
         op_name,
         utils::SupportedTensorDtypes::SAME_AS_COMMON>(
-        [b, alpha](const auto val_a) {
-          CTYPE_COMPUTE val_b = utils::scalar_to<CTYPE_COMPUTE>(b);
-          CTYPE_COMPUTE val_alpha = utils::scalar_to<CTYPE_COMPUTE>(alpha);
-          return val_a + val_alpha * val_b;
+        [val_alpha_times_b](const auto val_a) {
+          // Cast here supports vectorization; either it does nothing
+          // or it casts from CTYPE_COMPUTE to
+          // Vectorized<CTYPE_COMPUTE>.
+          return val_a + decltype(val_a)(val_alpha_times_b);
         },
         ctx,
         a,

kernels/portable/cpu/op_atan2.cpp

@@ -60,7 +60,7 @@ Tensor& atan2_out(
         op_name,
         utils::SupportedTensorDtypes::FLOATHBF16>(
         [](const auto val_a, const auto val_b) {
-          return std::atan2(val_a, val_b);
+          return executorch::math::atan2(val_a, val_b);
         },
         ctx,
         a,

kernels/portable/cpu/op_clamp.cpp

@@ -138,9 +138,8 @@ Tensor& clamp_out(
         CTYPE_COMPUTE,
         op_name,
         utils::SupportedTensorDtypes::SAME_AS_COMMON>(
-        [has_min, min_opt, has_max, max_opt](const CTYPE_COMPUTE val_in) {
-          // TODO: rewrite this to be vectorization-capable.
-          CTYPE_COMPUTE val_out = val_in;
+        [has_min, min_opt, has_max, max_opt](const auto val_in) {
+          auto val_out = val_in;
           if (has_min) {
             val_out = utils::max_override(
                 val_out, utils::scalar_to<CTYPE_COMPUTE>(min_opt.value()));

kernels/portable/cpu/op_elu.cpp

@@ -48,8 +48,7 @@ Tensor& elu_out(
         CTYPE,
         op_name,
         utils::SupportedTensorDtypes::SAME_AS_COMMON>(
-        [negcoef, math_scale, math_input_scale](const auto x) {
-          // TODO: rewrite this to be vectorization-capable.
+        [negcoef, math_scale, math_input_scale](const CTYPE x) {
           return MathT(x) <= MathT(0)
               ? std::expm1(MathT(x) * math_input_scale) * negcoef
               : MathT(x) * math_scale;

kernels/portable/cpu/op_fmod.cpp

@@ -61,7 +61,7 @@ Tensor& fmod_Tensor_out(
         utils::SupportedTensorDtypes::REALHBF16>(
         [&div_by_zero_error](
             const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
-          // TODO: rewrite this to be vectorization-capable.
+          // TODO: rewrite this to be vectorization-capable?
           CTYPE_COMPUTE value = 0;
           if (is_integral_type<CTYPE_COMPUTE, /*includeBool=*/true>::value) {
             if (val_b == 0) {
@@ -138,10 +138,8 @@ Tensor& fmod_Scalar_out(
         CTYPE_COMPUTE,
         op_name,
         utils::SupportedTensorDtypes::REALHBF16>(
-        [val_b](const CTYPE_COMPUTE val_a) {
-          // TODO: rewrite this to be vectorization-capable.
-          CTYPE_COMPUTE value = std::fmod(val_a, val_b);
-          return value;
+        [val_b](const auto val_a) {
+          return executorch::math::fmod(val_a, (decltype(val_a))val_b);
         },
         ctx,
         a,

kernels/portable/cpu/op_maximum.cpp

@@ -49,7 +49,7 @@ Tensor& maximum_out(
         CTYPE_COMPUTE,
         op_name,
         utils::SupportedTensorDtypes::REALHBBF16>(
-        [](const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
+        [](const auto val_a, const auto val_b) {
           return utils::max_override(val_a, val_b);
         },
         ctx,

kernels/portable/cpu/op_minimum.cpp

@@ -49,8 +49,7 @@ Tensor& minimum_out(
         CTYPE_COMPUTE,
         op_name,
         utils::SupportedTensorDtypes::REALHBBF16>(
-        [](const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
-          // TODO: rewrite this to be vectorization-capable.
+        [](const auto val_a, const auto val_b) {
           return utils::min_override(val_a, val_b);
         },
         ctx,

kernels/portable/cpu/op_mul.cpp

@@ -72,9 +72,7 @@ Tensor& mul_out(
           CTYPE_COMPUTE,
           op_name,
           utils::SupportedTensorDtypes::REALHBBF16>(
-          [](const CTYPE_COMPUTE val_a, const CTYPE_COMPUTE val_b) {
-            return val_a * val_b;
-          },
+          [](const auto val_a, const auto val_b) { return val_a * val_b; },
           ctx,
           a,
           utils::SupportedTensorDtypes::REALHBBF16,

kernels/portable/cpu/op_native_dropout.cpp

@@ -57,8 +57,11 @@ std::tuple<Tensor&, Tensor&> native_dropout_out(
     }
     ET_SWITCH_FLOATHBF16_TYPES(
         input.scalar_type(), ctx, op_name, CTYPE_COMPUTE, [&]() {
-          utils::apply_bitensor_elementwise_fn<CTYPE_COMPUTE, op_name>(
-              [](const auto val, const auto mask_val) {
+          utils::apply_bitensor_elementwise_fn<
+              CTYPE_COMPUTE,
+              op_name,
+              utils::SupportedTensorDtypes::SAME_AS_COMMON>(
+              [](const CTYPE_COMPUTE val, const CTYPE_COMPUTE mask_val) {
                 if (!mask_val) {
                   return static_cast<decltype(val)>(0);
                 }
@@ -70,8 +73,7 @@ std::tuple<Tensor&, Tensor&> native_dropout_out(
               mask,
               // TODO: should really be just BOOL
               utils::SupportedTensorDtypes::BOOL_OR_BYTE,
-              out,
-              utils::SupportedTensorDtypes::SAME_AS_COMMON);
+              out);
         });
   } else if (input.numel() > 0) {
     std::memcpy(out.mutable_data_ptr(), input.data_ptr(), input.nbytes());