Optimzed backward pass for ROCm devices (pt 2) (pytorch#3511)

Summary:

X-link: facebookresearch/FBGEMM#594

- Break up D66310520 (pytorch#3367) into backend and frontend diffs.  This is the frontend diff, and followup to D66986498

Differential Revision: D67407935

Author: q10

fbgemm_gpu/codegen/training/backward/embedding_backward_dense_host_cpu.cpp

@@ -172,7 +172,7 @@ Tensor split_embedding_codegen_lookup_dense_function(
     c10::SymInt /* max_B = -1 */,
     c10::SymInt /* max_B_feature_rank = -1 */,
     c10::SymInt /* vbe_output_size = -1 */,
-    bool /* mixed_D = true */) {
+    bool /* mixed_D = false */) {
   return SplitLookupFunction_Dense_Op::apply(
       host_weights,
       weights_offsets,
@@ -191,15 +191,15 @@ Tensor split_embedding_codegen_lookup_dense_function(
 // Deprecated for fb namespace! Please use fbgemm namespace instead!
 TORCH_LIBRARY_FRAGMENT(fb, m) {
   m.def(
-      "dense_embedding_codegen_lookup_function(Tensor dev_weights, Tensor weights_offsets, Tensor D_offsets, SymInt total_D, SymInt max_D, Tensor hash_size_cumsum, int total_hash_size_bits, Tensor indices, Tensor offsets, int pooling_mode, Tensor? indice_weights, Tensor? feature_requires_grad, int output_dtype=0, Tensor? B_offsets=None, Tensor? vbe_output_offsets_feature_rank=None, Tensor? vbe_B_offsets_rank_per_feature=None, SymInt max_B=-1, SymInt max_B_feature_rank=-1, SymInt vbe_output_size=-1, bool mixed_D=True) -> Tensor");
+      "dense_embedding_codegen_lookup_function(Tensor dev_weights, Tensor weights_offsets, Tensor D_offsets, SymInt total_D, SymInt max_D, Tensor hash_size_cumsum, int total_hash_size_bits, Tensor indices, Tensor offsets, int pooling_mode, Tensor? indice_weights, Tensor? feature_requires_grad, int output_dtype=0, Tensor? B_offsets=None, Tensor? vbe_output_offsets_feature_rank=None, Tensor? vbe_B_offsets_rank_per_feature=None, SymInt max_B=-1, SymInt max_B_feature_rank=-1, SymInt vbe_output_size=-1, bool mixed_D=False) -> Tensor");
   DISPATCH_TO_CPU(
       "dense_embedding_codegen_lookup_function",
       split_embedding_codegen_lookup_dense_function);
 }
 
 TORCH_LIBRARY_FRAGMENT(fbgemm, m) {
   m.def(
-      "dense_embedding_codegen_lookup_function(Tensor dev_weights, Tensor weights_offsets, Tensor D_offsets, SymInt total_D, SymInt max_D, Tensor hash_size_cumsum, int total_hash_size_bits, Tensor indices, Tensor offsets, int pooling_mode, Tensor? indice_weights, Tensor? feature_requires_grad, int output_dtype=0, Tensor? B_offsets=None, Tensor? vbe_output_offsets_feature_rank=None, Tensor? vbe_B_offsets_rank_per_feature=None, SymInt max_B=-1, SymInt max_B_feature_rank=-1, SymInt vbe_output_size=-1, bool mixed_D=True) -> Tensor");
+      "dense_embedding_codegen_lookup_function(Tensor dev_weights, Tensor weights_offsets, Tensor D_offsets, SymInt total_D, SymInt max_D, Tensor hash_size_cumsum, int total_hash_size_bits, Tensor indices, Tensor offsets, int pooling_mode, Tensor? indice_weights, Tensor? feature_requires_grad, int output_dtype=0, Tensor? B_offsets=None, Tensor? vbe_output_offsets_feature_rank=None, Tensor? vbe_B_offsets_rank_per_feature=None, SymInt max_B=-1, SymInt max_B_feature_rank=-1, SymInt vbe_output_size=-1, bool mixed_D=False) -> Tensor");
   DISPATCH_TO_CPU(
       "dense_embedding_codegen_lookup_function",
       split_embedding_codegen_lookup_dense_function);

fbgemm_gpu/codegen/training/backward/embedding_backward_split_host_template.cpp

@@ -1083,7 +1083,7 @@ Tensor {{ bwd_mdesc }}_embedding_codegen_lookup_{{ optimizer }}_function(
     {%- else %}
     const c10::SymInt vbe_output_size = -1,
     {%- endif %}
-    const bool mixed_D = true
+    const bool mixed_D = false
 ) {
   // TODO: refactor into macro
   {%- if has_gpu_support %}
@@ -1200,7 +1200,7 @@ TORCH_LIBRARY_FRAGMENT({{ lib_name }}, m) {
           "    Tensor[]? ssd_tensors=None,"
           {%- endif %}
           "   float gwd_lower_bound=0, "
-          "   bool mixed_D=True"
+          "   bool mixed_D=False"
           ") -> Tensor",
           {PT2_COMPLIANT_TAG});
 

fbgemm_gpu/codegen/training/python/lookup_args.template

@@ -49,6 +49,7 @@ class CommonArgs(NamedTuple):
     {%- if ssd %}
     ssd_tensors: Dict[str, torch.Tensor]
     {%- endif %}
+    mixed_D: bool
 
 
 class OptimizerArgs(NamedTuple):

fbgemm_gpu/codegen/training/python/split_embedding_codegen_lookup_invoker.template

@@ -409,5 +409,6 @@ def invoke(
         use_homogeneous_placements=common_args.use_homogeneous_placements,
         apply_global_weight_decay=apply_global_weight_decay,
         gwd_lower_bound=gwd_lower_bound,
+        mixed_D=common_args.mixed_D,
     )
     {%- endif %}

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -744,6 +744,7 @@ def __init__(  # noqa C901
                 not mixed_D
             ), "OptimType.NONE does not support mixed embedding dimension"
 
+        self.mixed_D: bool = mixed_D
         if device is None:
             self.current_device: torch.device = (
                 torch.device("cpu")
@@ -1806,6 +1807,7 @@ def forward(  # noqa: C901
             is_experimental=self.is_experimental,
             use_uniq_cache_locations_bwd=self.use_uniq_cache_locations_bwd,
             use_homogeneous_placements=self.use_homogeneous_placements,
+            mixed_D=self.mixed_D,
         )
 
         if self.optimizer == OptimType.NONE:
@@ -3581,6 +3583,14 @@ def __init__(
         )
         assert self.D_offsets.numel() == T + 1
 
+        mixed_D = False
+        D = dims[0]
+        for d in dims:
+            if d != D:
+                mixed_D = True
+                break
+        self.mixed_D: bool = mixed_D
+
         # Required for VBE
         self.register_buffer(
             "feature_dims",
@@ -3694,6 +3704,7 @@ def forward(
             max_B=vbe_metadata.max_B,
             max_B_feature_rank=vbe_metadata.max_B_feature_rank,
             vbe_output_size=vbe_metadata.output_size,
+            mixed_D=self.mixed_D,
         )
 
     @torch.jit.export

fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py

@@ -1615,6 +1615,7 @@ def forward(
             },
             # pyre-fixme[6]: Expected `lookup_args_ssd.VBEMetadata` but got `lookup_args.VBEMetadata`
             vbe_metadata=vbe_metadata,
+            mixed_D=False,
         )
 
         self.timesteps_prefetched.pop(0)