Added adaptive_max_poolNd batch rule (#263)

* Added adaptive_max_poolNd fw/bw batch rules Description: - Added adaptive_max_poolNd fw/bw batch rules - Updated tests Related to #240 Notes: I created two additional macros to handle adaptive_max_pool2d and adaptive_max_pool3d_backward. Not sure if we could make a generic rule to handle max_pool2d_with_indices_backward_batch_rule and adaptive_max_pool3d_backward, as max_pool2d_with_indices_backward_batch_rule requires some args in the middle between gradOutput, input and indices. * Replaced EXISTING_BDIM_MULTIOUT by EXISTING_BDIM_ALL_BOXED * Removed specific implementations with indices.contiguous() for - max_pool2d_with_indices_backward - adaptive_max_pool2d_backward - adaptive_max_pool3d_backward and added ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED_CONTIG1 to handle that
pytorch · Nov 23, 2021 · da27398 · da27398
1 parent 10df6ca
commit da27398
Show file tree

Hide file tree

Showing 5 changed files with 26 additions and 42 deletions.
diff --git a/functorch/csrc/BatchRulesDecompositions.cpp b/functorch/csrc/BatchRulesDecompositions.cpp
@@ -19,6 +19,7 @@ namespace at { namespace functorch {
 TORCH_LIBRARY_IMPL(aten, FT_BATCHED_KEY, m) {
   OP_DECOMPOSE(absolute);
   OP_DECOMPOSE(avg_pool1d);
+  OP_DECOMPOSE(adaptive_max_pool1d);
   OP_DECOMPOSE(adaptive_avg_pool1d);
   OP_DECOMPOSE(adaptive_avg_pool2d);
   OP_DECOMPOSE(adaptive_avg_pool3d);

diff --git a/functorch/csrc/BatchRulesHelper.h b/functorch/csrc/BatchRulesHelper.h
@@ -265,7 +265,7 @@ inline void boxed_existing_bdim_all_batch_rule(
 #define EXISTING_BDIM_ALL_BOXED(op) \
   m.impl(#op, torch::CppFunction::makeFromBoxedFunction<boxed_existing_bdim_all_batch_rule>());
 
-template <int64_t feature_rank>
+template <int64_t feature_rank, int64_t contig_tensor_index=-1>
 inline void boxed_all_tensors_have_optional_bdim(
     const c10::OperatorHandle& op, torch::jit::Stack* stack) {
   const auto& schema = op.schema();
@@ -302,11 +302,19 @@ inline void boxed_all_tensors_have_optional_bdim(
     }
     if (*is_no_batch_dim_case) {
       TORCH_INTERNAL_ASSERT(logical_rank == feature_rank);
-      (*stack)[args_begin + tensor_pos[tensor_idx]] = moveBatchDimToFront(value_, bdim);
+      value_ = moveBatchDimToFront(value_, bdim);
+      if (tensor_idx == contig_tensor_index) {
+        value_ = value_.contiguous();
+      }
+      (*stack)[args_begin + tensor_pos[tensor_idx]] = value_;
       continue;
     }
     TORCH_INTERNAL_ASSERT(logical_rank == feature_rank + 1);
-    (*stack)[args_begin + tensor_pos[tensor_idx]] = reshape_dim_into(*bdim, 0, value_);
+    value_ = reshape_dim_into(*bdim, 0, value_);
+    if (tensor_idx == contig_tensor_index) {
+      value_ = value_.contiguous();
+    }
+    (*stack)[args_begin + tensor_pos[tensor_idx]] = value_;
   }
 
   op.callBoxed(stack);
@@ -330,6 +338,14 @@ inline void boxed_all_tensors_have_optional_bdim(
 #define ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED(feature_rank, op) \
   m.impl(#op, torch::CppFunction::makeFromBoxedFunction<boxed_all_tensors_have_optional_bdim<feature_rank>>());
 
+#define ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED_CONTIG1(feature_rank, op, contig_tensor_index) \
+  m.impl(#op, \
+         torch::CppFunction::makeFromBoxedFunction<\
+             boxed_all_tensors_have_optional_bdim<\
+                 feature_rank, \
+                 contig_tensor_index>\
+             >());
+
 template <typename A, A a, typename C>
 struct ExistingBdimBatchRuleHelper;
 

diff --git a/functorch/csrc/BatchRulesPooling.cpp b/functorch/csrc/BatchRulesPooling.cpp
@@ -26,38 +26,6 @@ static Tensor reshape_bdim_into_front(
   return reshape_dim_into(*bdim, 0, value_);
 }
 
-// We can't use ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED because the CUDA
-// kernel rightfully assumes that indices is contiguous.
-std::tuple<Tensor,optional<int64_t>> max_pool2d_with_indices_backward_batch_rule(
-    const Tensor& gradOutput, optional<int64_t> gradOutput_bdim,
-    const Tensor& input, optional<int64_t> input_bdim,
-    IntArrayRef kernel_size,
-    IntArrayRef stride,
-    IntArrayRef padding,
-    IntArrayRef dilation,
-    bool ceil_mode,
-    const Tensor& indices, optional<int64_t> indices_bdim) {
-  TORCH_INTERNAL_ASSERT(input_bdim.has_value() ^ !indices_bdim.has_value());
-  const auto bdim_size = get_bdim_size2(gradOutput, gradOutput_bdim, input, input_bdim);
-  const auto input_logical_rank = rankWithoutBatchDim(input, input_bdim);
-  bool chw_case = input_logical_rank == 3;
-
-  const auto gradOutput_ = reshape_bdim_into_front(gradOutput, gradOutput_bdim, bdim_size, chw_case);
-  const auto input_ = reshape_bdim_into_front(input, input_bdim, bdim_size, chw_case);
-  const auto indices_ = reshape_bdim_into_front(indices, indices_bdim, bdim_size, chw_case);
-
-  const auto result = at::max_pool2d_with_indices_backward(
-      gradOutput_, input_, kernel_size, stride, padding, dilation, ceil_mode,
-      // max_pool2d_with_indices rightfully assumes that indices is contiguous
-      indices_.contiguous());
-
-  if (chw_case) {
-    return std::make_tuple(std::move(result), 0);
-  } else {
-    return std::make_tuple(reshape_dim_outof(0, bdim_size, result), 0);
-  }
-}
-
 std::tuple<Tensor,optional<int64_t>,Tensor,optional<int64_t>>
 max_pool2d_with_indices_batch_rule(
     const Tensor& self, optional<int64_t> self_bdim,
@@ -91,8 +59,13 @@ TORCH_LIBRARY_IMPL(aten, FT_BATCHED_KEY, m) {
   EXISTING_BDIM(avg_pool3d);
   EXISTING_BDIM_ALL_BOXED(avg_pool2d_backward);
   EXISTING_BDIM_ALL_BOXED(avg_pool3d_backward);
+  EXISTING_BDIM_ALL_BOXED(adaptive_max_pool2d);
+  EXISTING_BDIM_ALL_BOXED(adaptive_max_pool3d);
+  ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED_CONTIG1(3, adaptive_max_pool2d_backward, 2);
+  ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED_CONTIG1(4, adaptive_max_pool3d_backward, 2);
+
   VMAP_SUPPORT("max_pool2d_with_indices", max_pool2d_with_indices_batch_rule);
-  VMAP_SUPPORT("max_pool2d_with_indices_backward", max_pool2d_with_indices_backward_batch_rule);
+  ALL_TENSORS_HAVE_OPTIONAL_BDIM_BOXED_CONTIG1(3, max_pool2d_with_indices_backward, 2);
 }
 
 }}
diff --git a/test/test_ops.py b/test/test_ops.py
@@ -492,9 +492,6 @@ def test_vmapvjp(self, device, dtype, op):
         xfail('diagonal_scatter'),
         xfail('double', 'channels_last'),
         xfail('linalg.cross'),
-        xfail('nn.functional.adaptive_max_pool1d'),
-        xfail('nn.functional.adaptive_max_pool2d'),
-        xfail('nn.functional.adaptive_max_pool3d'),
         xfail('nn.functional.conv1d'),
         xfail('nn.functional.gaussian_nll_loss'),
         xfail('nn.functional.hardsigmoid'),

diff --git a/test/test_vmap.py b/test/test_vmap.py
@@ -3201,9 +3201,6 @@ def test_vmap_exhaustive(self, device, dtype, op):
         xfail('slice_scatter'),
         xfail('unique_consecutive'),
         xfail('unique'),
-        xfail('nn.functional.adaptive_max_pool1d'),
-        xfail('nn.functional.adaptive_max_pool2d'),
-        xfail('nn.functional.adaptive_max_pool3d'),
         xfail('nn.functional.conv1d'),
         xfail('nn.functional.cosine_embedding_loss'),
         # xfail('nn.functional.cross_entropy'),