Improve FC perf when no_bias=False (apache#15033)

* Improve FC perf when no_bias=False

* Add Issue number in comment

* Correct req

Author: anirudh2290

src/operator/nn/fully_connected-inl.h

@@ -36,6 +36,7 @@
 #include "../elemwise_op_common.h"
 #include "../linalg.h"
 #include "../../common/utils.h"
+#include "../tensor/broadcast_reduce_op.h"
 
 namespace mxnet {
 namespace op {
@@ -169,7 +170,18 @@ void FCBackward(const OpContext &ctx, const FullyConnectedParam &param,
   // gradient of bias
   if (!param.no_bias) {
     Tensor<xpu, 1, DType> gbias = in_grad[fullc::kBias].get<xpu, 1, DType>(s);
-    Assign(gbias, req[fullc::kBias], sum_rows(grad));
+    TBlob grad_blob = TBlob(grad);
+    TBlob gbias_blob = TBlob(gbias);
+    mxnet::TShape x(1, 0);
+    mxnet::TShape small;
+    if (shape_assign(&gbias_blob.shape_, Shape2(param.num_hidden, 1))) {
+      small = gbias_blob.shape_;
+    } else {
+      small = ReduceAxesShapeImpl(grad_blob.shape_, dmlc::optional<mxnet::TShape>(x), true, false);
+    }
+    ReduceAxesComputeImpl<xpu, mshadow::red::sum, false, false,
+                          mshadow_op::identity>(ctx, {grad_blob}, {req[fullc::kBias]},
+                                                {in_grad[fullc::kBias]}, small);
   }
   // gradient of data
   // Legacy approach shown here for comparison:

src/operator/nn/fully_connected.cc

@@ -316,11 +316,9 @@ NNVM_REGISTER_OP(_backward_FullyConnected)
   const FullyConnectedParam& params = nnvm::get<FullyConnectedParam>(attrs.parsed);
   return params.no_bias ? 2 : 3;
 })
-#if MXNET_USE_MKLDNN == 1
 .set_attr<FResourceRequest>("FResourceRequest", [](const NodeAttrs& n) {
   return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
 })
-#endif
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
 .set_attr<nnvm::FInplaceOption>("FInplaceOption", [](const NodeAttrs& attrs){
   return std::vector<std::pair<int, int> >{{1, 0}};

tests/python/unittest/test_operator.py

@@ -696,6 +696,27 @@ def test_symbol_pow():
     check_symbolic_backward(test, [data_tmp, exp_tmp], [np.ones(shape)], [data_dir, exp_dir])
 
 
+@with_seed()
+def test_fully_connected():
+    data = mx.sym.var("data")
+    fc_weight = mx.sym.var("weight")
+    fc_bias = mx.sym.var("bias")
+    fc = mx.sym.FullyConnected(data=data, weight=fc_weight, bias=fc_bias, num_hidden=10, no_bias=False, name='fc')
+    data = mx.nd.random.uniform(shape=(5, 5, 5, 13), dtype=np.float32)
+    fc_weight = mx.nd.random.uniform(shape=(10, 325), dtype=np.float32)
+    fc_bias = mx.nd.random.uniform(shape=(10), dtype=np.float32)
+    fc_bias2 = mx.nd.random.uniform(shape=(10, 1), dtype=np.float32)
+    data_np = data.asnumpy().reshape(5, 325)
+    fc_weight_np = np.transpose(fc_weight.asnumpy())
+    fc_bias_np = fc_bias.asnumpy()
+    res = np.dot(data_np, fc_weight_np) + fc_bias.asnumpy()
+    check_symbolic_forward(fc, {'data': data_np, 'weight': fc_weight.asnumpy(), 'bias': fc_bias_np}, {'fc_output': res})
+    check_numeric_gradient(fc, {'data': data_np, 'weight': fc_weight.asnumpy(), 'bias': fc_bias_np},
+                           numeric_eps=1e-2, rtol=1e-4, atol=1e-2)
+    # TODO: Fix Bug #15032 when bias has ndim > 1
+    #check_symbolic_forward(fc, {'data': data_np, 'weight': fc_weight.asnumpy(), 'bias': fc_bias2.asnumpy()}, {'fc_output': res})
+
+
 @with_seed()
 def test_pow_fn():
     shape = (3, 4)