Revert D28836794: SumKernel (BFloat16): use float as accumulation type

Test Plan: revert-hammer

Differential Revision:
D28836794 (4f5c688)

Original commit changeset: 46ed3a862c2b

fbshipit-source-id: 3b586eeb752b7cdee909fa97a4c78876a6014770

Author: VitalyFedyunin

aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h

@@ -742,26 +742,4 @@ inline Vectorized<BFloat16> convert_float_bfloat16(const Vectorized<float>& a, c
 
 #endif
 
-struct Vec2f {
-  Vectorized<float> val0, val1;
-  Vec2f() {}
-  Vec2f(float v) : val0(v), val1(v) {}
-  Vec2f(Vectorized<float> v0, Vectorized<float> v1) : val0(v0), val1(v1) {}
-  operator Vectorized<BFloat16>() const {
-    return convert_float_bfloat16(val0, val1);
-  }
-};
-inline Vec2f& operator+= (Vec2f& a, const Vec2f& b) {
-  a.val0 += b.val0;
-  a.val1 += b.val1;
-  return a;
-}
-inline Vec2f& operator+= (Vec2f& a, const Vectorized<BFloat16>& b) {
-  Vectorized<float> b0, b1;
-  std::tie(b0, b1) = convert_bfloat16_float(b);
-  a.val0 += b0;
-  a.val1 += b1;
-  return a;
-}
-
 }}}

aten/src/ATen/native/cpu/SumKernel.cpp

@@ -11,16 +11,6 @@ namespace at {
 namespace native {
 namespace {
 
-// use float as accumulation type for BFloat16
-template <typename scalar_t> struct AccType { using type = scalar_t; };
-template <> struct AccType<BFloat16> { using type = float; };
-
-template <typename scalar_t> struct AccType<Vectorized<scalar_t>> { using type = Vectorized<scalar_t>; };
-template <> struct AccType<Vectorized<BFloat16>> { using type = Vec2f; };
-
-template <typename scalar_t>
-using acc_type = typename AccType<scalar_t>::type;
-
 template <typename scalar_t>
 struct LoadPolicy {
   static scalar_t load(const char * C10_RESTRICT data, int64_t stride, int64_t index) {
@@ -217,9 +207,8 @@ std::array<scalar_t, nrows> multi_row_sum(
   const int64_t level_mask = level_step - 1;
 
   // NOLINTNEXTLINE(modernize-avoid-c-arrays,cppcoreguidelines-avoid-c-arrays)
-  using accscalar_t = acc_type<scalar_t>;
-  accscalar_t acc[num_levels][nrows];
-  std::fill_n(&acc[0][0], num_levels * nrows, accscalar_t(0));
+  scalar_t acc[num_levels][nrows];
+  std::fill_n(&acc[0][0], num_levels * nrows, scalar_t(0));
 
   int64_t i = 0;
   for (; i + level_step <= size;) {
@@ -239,7 +228,7 @@ std::array<scalar_t, nrows> multi_row_sum(
       #endif
       for (int64_t k = 0; k < nrows; ++k) {
         acc[j][k] += acc[j-1][k];
-        acc[j-1][k] = accscalar_t(0);
+        acc[j-1][k] = scalar_t(0);
       }
 
       const auto mask = (level_mask << (j * level_power));
@@ -271,7 +260,7 @@ std::array<scalar_t, nrows> multi_row_sum(
   // NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
   std::array<scalar_t, nrows> ret;
   for (int64_t k = 0; k < nrows; ++k) {
-    ret[k] = scalar_t(acc[0][k]);
+    ret[k] = acc[0][k];
   }
   return ret;
 }