Redefine FBGEMM targets with split_embeddings_utils [1/N] (pytorch#3049)

Summary:
Pull Request resolved: pytorch#3049

X-link: facebookresearch/FBGEMM#146

- Redefine `split_embeddings_utils`, `embedding_ops_common`, and `index_select_ops` targets with `gpu_cpp_library()`

Reviewed By: duduyi2013

Differential Revision: D61889098

fbshipit-source-id: 7c6fae12a5e2bccf382d7746bf510b6b812a3ba5

fbgemm_gpu/codegen/training/backward/embedding_backward_split_template.cu

@@ -22,8 +22,8 @@
 
 #include "fbgemm_gpu/embedding_backward_template_helpers.cuh"
 #include "fbgemm_gpu/utils/tensor_accessor.h"
-#include "fbgemm_gpu/split_embeddings_utils.cuh"
 #include "fbgemm_gpu/sparse_ops.h"
+#include "fbgemm_gpu/split_embeddings_utils.cuh"
 #include "fbgemm_gpu/utils/ops_utils.h"
 
 using Tensor = at::Tensor;

fbgemm_gpu/codegen/training/forward/embedding_forward_split_kernel_template.cu

@@ -34,7 +34,6 @@
 {%- set locs_or_addrs_idx = "row_idx" if ssd else "cache_idx" %}
 
 #include "fbgemm_gpu/embedding_forward_template_helpers.cuh"
-#include "fbgemm_gpu/split_embeddings_cache_cuda.cuh"
 
 using Tensor = at::Tensor;
 using namespace fbgemm_gpu;

fbgemm_gpu/codegen/training/forward/embedding_forward_split_template.cu

@@ -34,14 +34,15 @@
 {%- if not is_index_select %}
 ////////////////////////////////////////////////////////////////////////////////
 // Required for op registrations
-#include "fbgemm_gpu/utils/ops_utils.h"
 ////////////////////////////////////////////////////////////////////////////////
+#include "fbgemm_gpu/utils/ops_utils.h"
 {%- endif %}
 #include "fbgemm_gpu/embedding_forward_template_helpers.cuh"
 #include "fbgemm_gpu/split_embeddings_cache_cuda.cuh"
 
 using Tensor = at::Tensor;
 using namespace fbgemm_gpu;
+
 ////////////////////////////////////////////////////////////////////////////////
 // External Function Declarations
 ////////////////////////////////////////////////////////////////////////////////

fbgemm_gpu/fbgemm_gpu/sparse_ops.py

@@ -26,9 +26,7 @@
         torch.ops.load_library(
             "//deeplearning/fbgemm/fbgemm_gpu/codegen:embedding_ops_hip"
         )
-        torch.ops.load_library(
-            "//deeplearning/fbgemm/fbgemm_gpu/codegen:index_select_ops_hip"
-        )
+
     else:
         torch.ops.load_library("//deeplearning/fbgemm/fbgemm_gpu:sparse_ops")
         torch.ops.load_library(