Back out "Add EmbeddingSpMDM8Bit_Sve" (#4961)

Summary:
Pull Request resolved: #4961

X-link: https://github.com/facebookresearch/FBGEMM/pull/1980

Original commit changeset: fa17a82bbeea

Original Phabricator Diff: D72112120 introduced an issue that caused prediction results off for some IG models. Specifically, the scale (L2-norm) of the returned tensor is off by 10%, see S567079 for details.

Backing it out to mitigate S556160.

Reviewed By: q10

Differential Revision: D83710388

fbshipit-source-id: 67fc03282cefa81f2a195732e5bda5d285c99f7a

Author: joshuuuasu

BUILD.bazel

@@ -152,7 +152,7 @@ cc_library(
         ],
         ":linux-aarch64": [
             "-fopenmp",
-            "-march=armv9-a+sve2+fp16+bf16",
+            "-march=armv9-a+sve2+fp16",
         ],
         "//conditions:default": [],
     }),

include/fbgemm/FbgemmEmbedding.h

@@ -362,31 +362,6 @@ void compressed_indices_remap_avx512(
     float* out_weights);
 #endif
 
-// Specialization for uint8_t* input on aarch64 called by GenerateEmbeddingSpMDM
-template <
-    typename IndexType,
-    typename OffsetType,
-    typename OutType,
-    bool NoBag,
-    bool EnablePrefetching>
-FBGEMM_API bool EmbeddingSpMDM8Bit_Sve(
-    const int64_t block_size,
-    const int64_t output_size,
-    const int64_t index_size,
-    const int64_t data_size,
-    const uint8_t* input,
-    const IndexType* indices,
-    const OffsetType* offsets_or_lengths,
-    const float* weights, // optional, can be null for non-weighted sum
-    const bool normalize_by_lengths,
-    OutType* out,
-    const bool is_weight_positional,
-    const bool use_offsets,
-    const int64_t output_stride,
-    const int64_t input_stride,
-    const bool scale_bias_last,
-    const bool is_bf16_out);
-
 } // namespace internal
 
 template <typename IndexType>

include/fbgemm/Utils.h

@@ -23,8 +23,6 @@
 #ifndef HAVE_SVE
 #if defined(__aarch64__) && __ARM_FEATURE_SVE
 #define HAVE_SVE 1
-#include <arm_neon_sve_bridge.h> // @manual
-#include <arm_sve.h>
 #else
 #define HAVE_SVE 0
 #endif

src/EmbeddingSpMDM.cc

@@ -18,7 +18,6 @@
 #include <tuple>
 #include "./CodeCache.h" // @manual
 #include "./EmbeddingSpMDMAutovec.h" // @manual
-#include "./EmbeddingSpMDMSve.h"
 #include "./MaskAvx2.h" // @manual
 #include "./RefImplementations.h" // @manual
 #include "fbgemm/FbgemmEmbedding.h"
@@ -1127,76 +1126,6 @@ typename EmbeddingSpMDMKernelSignature<inType, indxType, offsetType, outType>::
   }
 #endif // CPUINFO_ARCH_X86 || CPUINFO_ARCH_X86_64
 
-#if HAVE_SVE
-  if constexpr (std::is_same<inType, uint8_t>::value) {
-    if (!is_asmjit_disabled()) {
-      if (no_bag) {
-        return [=](int64_t output_size,
-                   int64_t index_size,
-                   int64_t data_size,
-                   const uint8_t* input_u8,
-                   const indxType* indices,
-                   const offsetType* offsets_or_lengths,
-                   const float*
-                       weights, // optional, can be null for non-weighted sum
-                   outType* out) {
-          return internal::
-              EmbeddingSpMDM8Bit_Sve<indxType, offsetType, outType, true, true>(
-                  block_size,
-                  output_size,
-                  index_size,
-                  data_size,
-                  input_u8,
-                  indices,
-                  offsets_or_lengths,
-                  weights,
-                  normalize_by_lengths,
-                  out,
-                  is_weight_positional,
-                  use_offsets,
-                  output_stride,
-                  input_stride,
-                  scale_bias_last,
-                  is_bf16_out);
-        };
-      } else {
-        return [=](int64_t output_size,
-                   int64_t index_size,
-                   int64_t data_size,
-                   const uint8_t* input_u8,
-                   const indxType* indices,
-                   const offsetType* offsets_or_lengths,
-                   const float* weights, // optional, can be null for
-                                         // non-weighted sum
-                   outType* out) {
-          return internal::EmbeddingSpMDM8Bit_Sve<
-              indxType,
-              offsetType,
-              outType,
-              false,
-              true>(
-              block_size,
-              output_size,
-              index_size,
-              data_size,
-              input_u8,
-              indices,
-              offsets_or_lengths,
-              weights,
-              normalize_by_lengths,
-              out,
-              is_weight_positional,
-              use_offsets,
-              output_stride,
-              input_stride,
-              scale_bias_last,
-              is_bf16_out);
-        };
-      };
-    }
-  }
-#endif
-
 #ifdef FBGEMM_AUTOVEC_AVAILABLE
   if (!cpuinfo_initialize()) {
     throw std::runtime_error("Failed to initialize cpuinfo!");