Fix build with older gcc versions (pytorch#525)

Summary:
Pull Request resolved: pytorch#525

circleci build fails because reduce_add is not available with gcc 5.4. In fact, it's not available with gcc < 7.

The code with reduce_add was introduced in D26584575 (pytorch@953e671)

```
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc: In function 'void fbgemm::internal::SparseDenseInt8MVAvx512(const std::unique_ptr<fbgemm::BCSRMatrix<> >&, const uint8_t*, int, int32_t*, uint8_t*, fbgemm::trRequantizationParams_t&, bool, int, int)':
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:177:70: error: cannot convert 'const uint8_t* {aka const unsigned char*}' to 'const int*' for argument '2' to '__m512i _mm512_i32gather_epi32(__m512i, const int*, int)'
         __m512i b_v = _mm512_i32gather_epi32(b_idx, cur_B, block_size);
                                                                      ^
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:189:70: error: cannot convert 'const uint8_t* {aka const unsigned char*}' to 'const int*' for argument '2' to '__m512i _mm512_i32gather_epi32(__m512i, const int*, int)'
         __m512i b_v = _mm512_i32gather_epi32(b_idx, cur_B, block_size);
                                                                      ^
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:52: error: there are no arguments to '_mm512_reduce_add_epi32' that depend on a template parameter, so a declaration of '_mm512_reduce_add_epi32' must be available [-fpermissive]
       int32_t res_i32 = _mm512_reduce_add_epi32(res);
                                                    ^
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:52: note: (if you use '-fpermissive', G++ will accept your code, but allowing the use of an undeclared name is deprecated)
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc: In instantiation of 'void fbgemm::internal::SparseDenseInt8MVAvx512(const std::unique_ptr<fbgemm::BCSRMatrix<> >&, const uint8_t*, int, int32_t*, uint8_t*, fbgemm::trRequantizationParams_t&, bool, int, int) [with bool FUSE_RELU = true; fbgemm::QuantizationGranularity Q_GRAN = (fbgemm::QuantizationGranularity)0; uint8_t = unsigned char; int32_t = int]':
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:231:1:   required from here
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:48: error: '_mm512_reduce_add_epi32' was not declared in this scope
       int32_t res_i32 = _mm512_reduce_add_epi32(res);
                                                ^
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc: In instantiation of 'void fbgemm::internal::SparseDenseInt8MVAvx512(const std::unique_ptr<fbgemm::BCSRMatrix<> >&, const uint8_t*, int, int32_t*, uint8_t*, fbgemm::trRequantizationParams_t&, bool, int, int) [with bool FUSE_RELU = true; fbgemm::QuantizationGranularity Q_GRAN = (fbgemm::QuantizationGranularity)2; uint8_t = unsigned char; int32_t = int]':
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:232:1:   required from here
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:48: error: '_mm512_reduce_add_epi32' was not declared in this scope
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc: In instantiation of 'void fbgemm::internal::SparseDenseInt8MVAvx512(const std::unique_ptr<fbgemm::BCSRMatrix<> >&, const uint8_t*, int, int32_t*, uint8_t*, fbgemm::trRequantizationParams_t&, bool, int, int) [with bool FUSE_RELU = false; fbgemm::QuantizationGranularity Q_GRAN = (fbgemm::QuantizationGranularity)0; uint8_t = unsigned char; int32_t = int]':
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:233:1:   required from here
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:48: error: '_mm512_reduce_add_epi32' was not declared in this scope
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc: In instantiation of 'void fbgemm::internal::SparseDenseInt8MVAvx512(const std::unique_ptr<fbgemm::BCSRMatrix<> >&, const uint8_t*, int, int32_t*, uint8_t*, fbgemm::trRequantizationParams_t&, bool, int, int) [with bool FUSE_RELU = false; fbgemm::QuantizationGranularity Q_GRAN = (fbgemm::QuantizationGranularity)2; uint8_t = unsigned char; int32_t = int]':
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:234:1:   required from here
/root/project/src/FbgemmSparseDenseVectorInt8Avx512.cc:195:48: error: '_mm512_reduce_add_epi32' was not declared in this scope
CMakeFiles/fbgemm_avx512.dir/build.make:158: recipe for target 'CMakeFiles/fbgemm_avx512.dir/src/FbgemmSparseDenseVectorInt8Avx512.cc.o' failed
```

Reviewed By: jianyuh

Differential Revision: D26715238

fbshipit-source-id: dfc8e04c5ca5f02aff4d3404fa1822f138be1ba6

src/FbgemmSparseDenseVectorInt8Avx512.cc

@@ -16,6 +16,14 @@ namespace fbgemm {
 
 namespace internal {
 
+static inline int32_t horizontal_add(__m256i a) {
+  __m256i t1 = _mm256_hadd_epi32(a, a);
+  __m256i t2 = _mm256_hadd_epi32(t1, t1);
+  __m128i t3 = _mm256_extracti128_si256(t2, 1);
+  __m128i t4 = _mm_add_epi32(_mm256_castsi256_si128(t2), t3);
+  return _mm_cvtsi128_si32(t4);
+}
+
 template <
     bool FUSE_RELU,
     bool ACT_ZP_0, // is activation zero point 0?
@@ -174,7 +182,8 @@ void SparseDenseInt8MVAvx512(
       for (; r < r_end_aligned; r += VLEN_INT32) {
         __m512i a_v = _mm512_loadu_si512(values + r * block_size);
         __m512i b_idx = _mm512_loadu_si512(col_idx + r);
-        __m512i b_v = _mm512_i32gather_epi32(b_idx, cur_B, block_size);
+        __m512i b_v = _mm512_i32gather_epi32(
+            b_idx, reinterpret_cast<const int32_t*>(cur_B), block_size);
         __m512i c_i16_v = _mm512_maddubs_epi16(b_v, a_v);
         __m512i c_i32_v = _mm512_madd_epi16(one_16bit_v, c_i16_v);
         res = _mm512_add_epi32(res, c_i32_v);
@@ -186,13 +195,21 @@ void SparseDenseInt8MVAvx512(
         __m512i a_v =
             _mm512_maskz_loadu_epi32(mask_int32_v, values + r * block_size);
         __m512i b_idx = _mm512_maskz_loadu_epi32(mask_int32_v, col_idx + r);
-        __m512i b_v = _mm512_i32gather_epi32(b_idx, cur_B, block_size);
+        __m512i b_v = _mm512_i32gather_epi32(
+            b_idx, reinterpret_cast<const int32_t*>(cur_B), block_size);
         __m512i c_i16_v = _mm512_maddubs_epi16(b_v, a_v);
         __m512i c_i32_v = _mm512_madd_epi16(one_16bit_v, c_i16_v);
         res = _mm512_add_epi32(res, c_i32_v);
       }
       // Horizontal reduce
+      // _mm512_reduce_add_epi32 is only available for gcc version > 7
+#if __GNUC__ >= 7
       int32_t res_i32 = _mm512_reduce_add_epi32(res);
+#else
+      __m256i low = _mm512_castsi512_si256(res);
+      __m256i high = _mm512_extracti64x4_epi64(res, 1);
+      int32_t res_i32 = horizontal_add(_mm256_add_epi32(low, high));
+#endif
 
       // store the results
       if (accum || kt > 0) {