[BACKPORT][BUGFIX][FEATURE] Add oneDNN 1D and 3D deconvolution support and fix bias (#20292)

* [v1.x][BUGFIX] Implement oneDNN deconvolution primitives to deconvolution 2D (#20107)

* Use mkldnn deconvolution primitive in deconvolution

* Apply clang-format

* Refactor deconvolution version 1

* Refactor deconvolution version 2 and use permute_axes in IOLogicalSwapDesc

* Refactor deconvolution version 3

* Enable Deconvolution2D test

* Fix sanity

* Fix windows builds

* Fix deconvolution with bias test

* [v1.x][FEATURE] Add MKLDNN Deconvolution 1D and 3D support (#20137)

* Use MXNET_USE_ONEDNN

* Fix test

* Apply formatter

* Add native support for 3D deconvolution

* Remove outdated check

* Replace math.prod with np.prod

* Check convolution layout only when it has value

* Remove outdated check

* Change tests

* Increase default workspace size to mach convolution

* Fix deconv workspace size

* Increase default deconv workspace size in python API

* Disable 3D tests for GPU

* Add deconv arguments checks

* Remove next_impl calls until it is fixed

* Share workspace

* Fix documentation

* Add test_deconv_dilation

* Fix check

* Fix include order

Author: PawelGlomski-Intel

Diff for: python/mxnet/ndarray/numpy_extension/_op.py

@@ -617,9 +617,9 @@ def convolution(data=None, weight=None, bias=None, kernel=None, stride=None, dil
 @set_module('mxnet.ndarray.numpy_extension')
 def deconvolution(data=None, weight=None, bias=None, kernel=None, stride=None, dilate=None,
                   pad=None, adj=None, target_shape=None, num_filter=1, num_group=1,
-                  workspace=512, no_bias=False, cudnn_tune=None,
+                  workspace=1024, no_bias=False, cudnn_tune=None,
                   cudnn_off=False, layout=None):
-    r"""Computes 1D or 2D transposed convolution (aka fractionally strided convolution) of
+    r"""Computes 1D, 2D or 3D transposed convolution (aka fractionally strided convolution) of
     the input tensor. This operation can be seen as the gradient of Convolution operation
     with respect to its input. Convolution usually reduces the size of the input.
     Transposed convolution works the other way, going from a smaller input

Diff for: python/mxnet/numpy_extension/_op.py

@@ -586,9 +586,9 @@ def convolution(data=None, weight=None, bias=None, kernel=None, stride=None, dil
 @set_module('mxnet.numpy_extension')
 def deconvolution(data=None, weight=None, bias=None, kernel=None, stride=None, dilate=None,
                   pad=None, adj=None, target_shape=None, num_filter=1, num_group=1,
-                  workspace=512, no_bias=False, cudnn_tune=None,
+                  workspace=1024, no_bias=False, cudnn_tune=None,
                   cudnn_off=False, layout=None):
-    r"""Computes 1D or 2D transposed convolution (aka fractionally strided convolution) of
+    r"""Computes 1D, 2D or 3D transposed convolution (aka fractionally strided convolution) of
     the input tensor. This operation can be seen as the gradient of Convolution operation
     with respect to its input. Convolution usually reduces the size of the input.
     Transposed convolution works the other way, going from a smaller input

Diff for: src/operator/deformable_convolution-inl.h

@@ -328,7 +328,7 @@ class DeformableConvolutionOp : public Operator {
   index_t num_kernels_col2im_;
   bool bias_term_;  // has bias term?
   bool is_1x1_;
-};  // class ConvolutionOp
+};  // class DeformableConvolutionOp
 
 template <typename xpu>
 Operator* CreateOp(DeformableConvolutionParam param,

Diff for: src/operator/modulated_deformable_convolution-inl.h

@@ -389,7 +389,7 @@ class ModulatedDeformableConvolutionOp : public Operator {
   index_t im2col_step_;
   bool bias_term_;  // has bias term?
   bool is_1x1_;
-};  // class ConvolutionOp
+};  // class ModulatedDeformableConvolutionOp
 
 template <typename xpu>
 Operator* CreateOp(ModulatedDeformableConvolutionParam param,

Diff for: src/operator/nn/convolution-inl.h

@@ -97,12 +97,6 @@ static bool DeconvolutionShape(const nnvm::NodeAttrs& attrs,
                                mxnet::ShapeVector* in_shape,
                                mxnet::ShapeVector* out_shape) {
   const DeconvolutionParam& param_ = nnvm::get<DeconvolutionParam>(attrs.parsed);
-#if MXNET_USE_CUDNN == 0
-  if (param_.kernel.ndim() > 2) {
-    LOG(FATAL) << "If not using CUDNN, only 1D or 2D Deconvolution is supported";
-    return false;
-  }
-#endif  // CUDNN
 
   using namespace mshadow;
   if (!param_.no_bias) {
@@ -412,9 +406,9 @@ DMLC_REGISTER_PARAMETER(DeconvolutionParam);
 NNVM_REGISTER_OP(Deconvolution)
     .add_alias("_npx_deconvolution")
     .describe(
-        "Computes 1D or 2D transposed convolution (aka fractionally strided convolution) of the "
-        "input tensor. This operation can be seen as the gradient of Convolution operation with "
-        "respect to its input. Convolution usually reduces the size of the input. Transposed "
+        "Computes 1D, 2D or 3D transposed convolution (aka fractionally strided convolution) of "
+        "the input tensor. This operation can be seen as the gradient of Convolution operation "
+        "with respect to its input. Convolution usually reduces the size of the input. Transposed "
         "convolution works the other way, going from a smaller input to a larger output while "
         "preserving the connectivity pattern.")
     .set_num_inputs([](const NodeAttrs& attrs) {