[v1.7.x] backport Invoke mkldnn and cudnn BatchNorm when axis != 1 t…

…o v1.7.x (apache#18676)

* [Improvement] Invoke mkldnn and cudnn BatchNorm when axis != 1 (apache#18504)

* fix batch norm when fix_gamma is True

* support gradient accumulation for batch norm

* mkldnn batchnorm support grad add

* unittest for bn

* fix bn arg

* fix lint

* fix mkldnn

* fix mkldnn bn

* fix grad when fixing gamma

* fix naive gpu bn

* fix lint

* invoke mkldnn and cudnn batchnorm when axis != 1

* backport 18500

* change condition

* fix

* fix

* add mkldnn_off for bn

* remove mkldnn_off

* recover save_000800.json

* cast

* remove  and fix flaky test

Co-authored-by: JackieWu <[email protected]>

src/operator/nn/batch_norm.cc

@@ -422,10 +422,14 @@ static bool BatchNormType(const nnvm::NodeAttrs& attrs,
 
 #if MXNET_USE_MKLDNN == 1
 static inline bool SupportMKLDNNBN(const NDArray &input, const BatchNormParam &param) {
-  mxnet::TShape shape = input.shape();
-  return SupportMKLDNN(input) && shape.ndim() == 4
-      && param.axis == mxnet::op::batchnorm::DEFAULT_AXIS
-      && !mxnet::op::batchnorm::disable_mkl;
+  if (mxnet::op::batchnorm::disable_mkl) return false;
+  const mxnet::TShape shape = input.shape();
+  const int ndim = shape.ndim();
+  if (ndim == 0 || shape.Size() == 0) return false;
+  const int dtype = input.dtype();
+  return (dtype == mshadow::kFloat32 ||
+          dtype == mshadow::kBfloat16) &&
+          SupportStorageMKLDNN(input.storage_type());
 }
 
 void BatchNormComputeExCPU(const nnvm::NodeAttrs &attrs,

src/operator/nn/batch_norm.cu

@@ -704,8 +704,7 @@ void BatchNormCompute<gpu>(const nnvm::NodeAttrs& attrs,
 
   param.axis = mxnet::op::batchnorm::GetRealAxis(shape, param.axis);
 #if MXNET_USE_CUDNN == 1
-  if (!param.use_global_stats && !param.cudnn_off
-      && param.axis == mxnet::op::batchnorm::DEFAULT_AXIS) {
+  if (!param.use_global_stats && !param.cudnn_off) {
     MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
       GetCuDNNOp<DType>(param).Forward(ctx, in_data, req, outputs, aux_states);
     })
@@ -733,8 +732,7 @@ void BatchNormGradCompute<gpu>(const nnvm::NodeAttrs& attrs,
 
   param.axis = mxnet::op::batchnorm::GetRealAxis(shape, param.axis);
 #if MXNET_USE_CUDNN == 1
-  if (!param.use_global_stats && !param.cudnn_off
-      && param.axis == mxnet::op::batchnorm::DEFAULT_AXIS) {
+  if (!param.use_global_stats && !param.cudnn_off) {
     MSHADOW_REAL_TYPE_SWITCH(dtype, DType, {
       GetCuDNNOp<DType>(param).Backward(ctx, inputs, req, outputs);
     })

src/operator/nn/cudnn/cudnn_batch_norm-inl.h

@@ -260,15 +260,27 @@ class CuDNNBatchNormOp {
 
  private:
   void Init(const TBlob &in_data) {
-    if (in_data.ndim() == 4) {
-      for (int i = 0; i < 4; ++i)
-        shape_[i] = in_data.shape_[i];
+    CHECK_GE(param_.axis, 0);
+    CHECK_LT(param_.axis, in_data.ndim());
+    if (param_.axis == 1) {
+      if (in_data.ndim() == 4) {
+        for (int i = 0; i < 4; ++i)
+          shape_[i] = in_data.shape_[i];
+      } else {
+        // when in_data.ndim() != 4
+        shape_[0] = in_data.shape_[0];
+        shape_[1] = in_data.ndim() > 1 ? in_data.shape_[1] : 1;
+        shape_[2] = 1;
+        shape_[3] = static_cast<dim_t>(in_data.shape_.ProdShape(2,
+              in_data.ndim()));
+      }
     } else {
-      // when in_data.ndim() != 4
-      shape_[0] = in_data.shape_[0];
-      shape_[1] = in_data.ndim() > 1 ? in_data.shape_[1] : 1;
+      // reshape to (N, C, 1, D), C is the `param_.axis` dimension
+      shape_[0] = static_cast<dim_t>(in_data.shape_.ProdShape(0, param_.axis));
+      shape_[1] = in_data.shape_[param_.axis];
       shape_[2] = 1;
-      shape_[3] = in_data.shape_.ProdShape(2, in_data.ndim());
+      shape_[3] = static_cast<dim_t>(in_data.shape_.ProdShape(param_.axis + 1,
+            in_data.ndim()));
     }
 
     CUDNN_CALL(cudnnSetTensor4dDescriptor(io_desc_,

src/operator/nn/mkldnn/mkldnn_batch_norm-inl.h

@@ -157,7 +157,25 @@ void MKLDNNBatchNormForward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
                             const std::vector<NDArray> &inputs, const std::vector<OpReqType> &req,
                             const std::vector<NDArray> &outputs, bool fuse_relu) {
   const BatchNormParam &param = nnvm::get<BatchNormParam>(attrs.parsed);
-  const std::vector<NDArray> in_data(inputs.begin(), inputs.begin() + batchnorm::kInMovingMean);
+  std::vector<NDArray> in_data(inputs.begin(), inputs.begin() + batchnorm::kInMovingMean);
+
+  mxnet::TShape shape = inputs[batchnorm::kData].shape();
+  const int real_axis = mxnet::op::batchnorm::GetRealAxis(shape, param.axis);
+  CHECK_LT(real_axis, shape.ndim());
+  NDArray out = outputs[batchnorm::kOut];
+  if (param.axis != 1 || shape.ndim() != 4) {
+    // reshape to (N, C, 1, D)
+    mxnet::TShape new_shape{
+      static_cast<dim_t>(shape.ProdShape(0, real_axis)),
+      shape[real_axis],
+      1,
+      static_cast<dim_t>(shape.ProdShape(real_axis + 1,
+            static_cast<int>(shape.ndim())))
+    };
+    in_data[batchnorm::kData] = in_data[batchnorm::kData].Reshape(new_shape);
+    out = out.Reshape(new_shape);
+  }
+
   const std::vector<NDArray> aux_states(inputs.begin() + batchnorm::kInMovingMean, inputs.end());
   TmpMemMgr::Get()->Init(ctx.requested[batchnorm::kTempSpace]);
   mkldnn::normalization_flags flags = _GetFlags(in_data,
@@ -166,7 +184,6 @@ void MKLDNNBatchNormForward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
                                                 fuse_relu);
   const NDArray &data = in_data[batchnorm::kData];
   auto &fwd = GetBNForward<DType>(param, ctx, data, flags);
-  const NDArray &out = outputs[batchnorm::kOut];
 
   // for output memory
   auto out_mem = const_cast<NDArray &>(out).CreateMKLDNNData(fwd.GetPd().dst_desc());
@@ -325,9 +342,9 @@ void MKLDNNBatchNormBackward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
                                                 ctx.is_train && !param.use_global_stats,
                                                 fuse_relu);
 
-  const NDArray &data         = in_data[batchnorm::kData];
-  const NDArray &diff         = out_grad[batchnorm::kOut];
-  const NDArray &gradIn       = in_grad[batchnorm::kData];
+  NDArray data         = in_data[batchnorm::kData];
+  NDArray diff         = out_grad[batchnorm::kOut];
+  NDArray gradIn       = in_grad[batchnorm::kData];
   const NDArray &moving_mean  = aux_states[batchnorm::kMovingMean];
   const NDArray &moving_var   = aux_states[batchnorm::kMovingVar];
   const NDArray &out_mean     = out_data[batchnorm::kMean];
@@ -338,6 +355,23 @@ void MKLDNNBatchNormBackward(const nnvm::NodeAttrs &attrs, const OpContext &ctx,
   CHECK(moving_mean.IsDefaultData());
   CHECK(moving_var.IsDefaultData());
 
+  mxnet::TShape shape = data.shape();
+  const int real_axis = mxnet::op::batchnorm::GetRealAxis(shape, param.axis);
+  CHECK_LT(real_axis, shape.ndim());
+  if (param.axis != 1 || shape.ndim() != 4) {
+    // reshape to (N, C, 1, D)
+    mxnet::TShape new_shape{
+      static_cast<dim_t>(shape.ProdShape(0, real_axis)),
+      shape[real_axis],
+      1,
+      static_cast<dim_t>(shape.ProdShape(real_axis + 1,
+            static_cast<int>(shape.ndim())))
+    };
+    data = data.Reshape(new_shape);
+    diff = diff.Reshape(new_shape);
+    gradIn = gradIn.Reshape(new_shape);
+  }
+
   auto data_mem  = data.GetMKLDNNData();
   auto diff_mem  = diff.GetMKLDNNData();
   // MKLDNN batchnorm should run on special layouts. If one of them isn't, we

tests/python/unittest/test_numpy_op.py

@@ -1541,7 +1541,7 @@ def _test_batchnorm_impl(shape, fix_gamma, cudnn_off, output_mean_var,
                 assert_almost_equal(
                     bn_beta.grad.asnumpy(), adb.asnumpy(), atol=atol, rtol=rtol)
 
-    shapes = [(24, 2), (24, 3, 4), (24, 8, 4, 5), (24, 5, 6, 4, 5)]
+    shapes = [(4, 2), (4, 3, 4), (4, 6, 4, 5), (4, 5, 6, 4, 5)]
     bools = [False, True]
     for shape, fix_gamma, cudnn_off, output_mean_var in itertools.product(
             shapes, bools, bools, bools):

tests/python/unittest/test_operator.py

@@ -1964,7 +1964,7 @@ def _test_batchnorm_impl(op_name, shape, fix_gamma, cudnn_off, output_mean_var,
                     bn_beta.grad.asnumpy(), adb.asnumpy(), atol=atol, rtol=rtol)
 
     op_names = ['BatchNorm', 'SyncBatchNorm']
-    shapes = [(24, 2), (24, 3, 4), (24, 8, 4, 5), (24, 5, 6, 4, 5)]
+    shapes = [(4, 2), (4, 3, 4), (4, 6, 4, 5), (4, 5, 6, 4, 5)]
     bools = [False, True]
     for op_name, shape, fix_gamma, cudnn_off, output_mean_var in itertools.product(
             op_names, shapes, bools, bools, bools):

tests/python/unittest/test_symbol.py

@@ -272,36 +272,6 @@ def check_symbol_consistency(sym1, sym2, ctx, skip_grad=False, equal_nan=False):
                                     grad_req='null' if skip_grad else 'write',
                                     equal_nan=equal_nan)
 
-def test_load_000800():
-    with mx.AttrScope(ctx_group='stage1'):
-        data = mx.symbol.Variable('data', lr_mult=0.2)
-        weight = mx.sym.Variable(name='fc1_weight', lr_mult=1.2)
-        fc1  = mx.symbol.FullyConnected(data = data, weight=weight, name='fc1', num_hidden=128, wd_mult=0.3)
-        act1 = mx.symbol.Activation(data = fc1, name='relu1', act_type="relu")
-
-    set_stage1 = set(act1.list_arguments())
-    with mx.AttrScope(ctx_group='stage2'):
-        fc2  = mx.symbol.FullyConnected(data = act1, name = 'fc2', num_hidden = 64, lr_mult=0.01)
-        act2 = mx.symbol.Activation(data = fc2, name='relu2', act_type="relu")
-        fc3  = mx.symbol.FullyConnected(data = act2, name='fc3', num_hidden=10)
-        fc3 = mx.symbol.BatchNorm(fc3, name='batchnorm0')
-        sym1  = mx.symbol.SoftmaxOutput(data = fc3, name = 'softmax')
-
-    curr_path = os.path.dirname(os.path.abspath(os.path.expanduser(__file__)))
-    sym2 = mx.sym.load(os.path.join(curr_path, 'save_000800.json'))
-
-    attr1 = sym1.attr_dict()
-    attr2 = sym2.attr_dict()
-    for k, v1 in attr1.items():
-        assert k in attr2, k
-        v2 = attr2[k]
-        for kk, vv1 in v1.items():
-            if kk.startswith('__') and kk.endswith('__'):
-                assert kk in v2 and v2[kk] == vv1, k + str(v1) + str(v2)
-
-    check_symbol_consistency(sym1, sym2,
-        {'ctx': mx.cpu(0), 'group2ctx': {'stage1' : mx.cpu(1), 'stage2' : mx.cpu(2)}, 'data': (1,200)})
-
 
 def test_blockgrad():
     a = mx.sym.Variable('a')