[MXNET-33] Enhance mkldnn pooling to support full convention (#11047)

* fix mkldnn pooling to support full convention

* backward with full convention

* fix

* add pooling test for full convention

* add function for computing padding size

* fix unit test

* only support max-pooling

* fix pooling bwd

* address review comment

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

@@ -113,20 +113,12 @@ inline bool SupportMKLDNNPooling(const PoolingParam &param,
   if (!ret)
     return false;
 
-  if (param.pooling_convention == pool_enum::kValid)
+  if (param.pooling_convention == pool_enum::kValid) {
     return true;
-  else
-    return false;
-
-// need to support pooling convention full
-// https://issues.apache.org/jira/browse/MXNET-33
-#if 0
-  if (((dshape[2] + 2 * param.pad[0] - param.kernel[0]) % param.stride[0] == 0) &&
-      ((dshape[3] + 2 * param.pad[1] - param.kernel[1]) % param.stride[1] == 0))
-    return true;
-  else
-    return false;
-#endif
+  } else {
+    // currently, only max-pooling is supported for full convention
+    return param.pool_type == pool_enum::kMaxPooling;
+  }
 }
 
 inline bool MKLDNNRequireWorkspace(const PoolingParam &param) {

src/operator/nn/mkldnn/mkldnn_pooling.cc

@@ -134,6 +134,14 @@ mkldnn::algorithm GetMKLDNNPoolAlgo(const PoolingParam &param) {
   }
 }
 
+static inline int GetPaddingSizeFull(int x, int padl, int padr, int k, int s) {
+  if ((x + padl + padr - k) % s != 0) {
+    return (padr + s - ((x + padl + padr - k) % s));
+  } else {
+    return padr;
+  }
+}
+
 mkldnn::pooling_forward::primitive_desc GetPoolingFwdPdesc(
     const PoolingParam &param, const bool is_train, const memory::desc &data_md,
     const memory::desc &out_md) {
@@ -154,11 +162,17 @@ mkldnn::pooling_forward::primitive_desc GetPoolingFwdPdesc(
   int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
   int stride_h_ = param.stride[0], stride_w_ = param.stride[1];
 
+  if (param.pooling_convention == pool_enum::kFull) {
+    pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
+    pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
+  }
+
   const mkldnn::engine engine = CpuEngine::Get()->get_engine();
   if (param.global_pool) {
     pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
     stride_h_ = stride_w_ = 1;
   }
+
   if (pad_t_ != 0 || pad_l_ != 0) {
     CHECK(param.pool_type == pool_enum::kAvgPooling ||
           param.pool_type == pool_enum::kMaxPooling)
@@ -167,7 +181,6 @@ mkldnn::pooling_forward::primitive_desc GetPoolingFwdPdesc(
     CHECK_LT(pad_t_, kernel_h_);
   }
 
-
   const mkldnn::algorithm alg = GetMKLDNNPoolAlgo(param);
   mkldnn::prop_kind kind = mkldnn::prop_kind::forward_scoring;
   if (is_train && alg != algorithm::pooling_avg) {
@@ -227,17 +240,22 @@ MKLDNNPoolingFwd &GetPoolingFwd(const PoolingParam &param,
     int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
     int stride_h_ = param.stride[0], stride_w_ = param.stride[1];
 
+    if (param.pooling_convention == pool_enum::kFull) {
+      pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
+      pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
+    }
+
     if (param.global_pool) {
-        pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
-        stride_h_ = stride_w_ = 1;
+      pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
+      stride_h_ = stride_w_ = 1;
     }
 
     if (pad_t_ != 0 || pad_l_ != 0) {
-        CHECK(param.pool_type == pool_enum::kAvgPooling ||
-              param.pool_type == pool_enum::kMaxPooling)
-              << "Padding implemented only for average and max pooling.";
-        CHECK_LT(pad_l_, kernel_w_);
-        CHECK_LT(pad_t_, kernel_h_);
+      CHECK(param.pool_type == pool_enum::kAvgPooling ||
+            param.pool_type == pool_enum::kMaxPooling)
+            << "Padding implemented only for average and max pooling.";
+      CHECK_LT(pad_l_, kernel_w_);
+      CHECK_LT(pad_t_, kernel_h_);
     }
 
     const mkldnn::algorithm alg = GetMKLDNNPoolAlgo(param);
@@ -353,6 +371,12 @@ MKLDNNPoolingBwd &GetPoolingBwd(const PoolingParam &param,
     int pad_t_ = param.pad[0], pad_b_ = param.pad[0];
     int pad_l_ = param.pad[1], pad_r_ = param.pad[1];
     int stride_h_ = param.stride[0], stride_w_ = param.stride[1];
+
+    if (param.pooling_convention == pool_enum::kFull) {
+      pad_b_ = GetPaddingSizeFull(data_md.data.dims[2], pad_t_, pad_b_, kernel_h_, stride_h_);
+      pad_r_ = GetPaddingSizeFull(data_md.data.dims[3], pad_l_, pad_r_, kernel_w_, stride_w_);
+    }
+
     if (param.global_pool) {
       pad_t_ = pad_b_ = pad_l_ = pad_r_ = 0;
       stride_h_ = stride_w_ = 1;

tests/python/gpu/test_operator_gpu.py

@@ -996,6 +996,35 @@ def test_3d_pooling(pool_type, p_value=2, count_include_pad=True):
     test_3d_pooling('lp', p_value=3)
 
 
+@with_seed()
+def test_pooling_full_2d():
+    def test_pooling_full_2d_type(pool_type):
+        data = (2, 2, 10, 10)
+        kernel = (4, 5)
+        pad = (1, 2)
+        stride = (3, 4)
+
+        convention = 'full'
+        ctx_list = []
+        sym_list = []
+
+        # o_h = ceil((10 + 1 + 1 - 4) / 3) + 1 = 4
+        # o_w = ceil((10 + 2 + 2 - 5) / 4) + 1 = 4
+        ctx_list.append({'ctx': mx.cpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+        sym_list.append(mx.sym.Pooling(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type,
+                                       pooling_convention=convention, global_pool=False, name='pool'))
+
+        ctx_list.append({'ctx': mx.gpu(0), 'pool_data': data, 'type_dict': {'pool_data': np.float32}})
+        sym_list.append(mx.sym.Pooling(kernel=kernel, pad=pad, stride=stride, pool_type=pool_type,
+                                       pooling_convention=convention, global_pool=False, name='pool'))
+
+        check_consistency(sym_list, ctx_list)
+
+    test_pooling_full_2d_type('max')
+    test_pooling_full_2d_type('avg')
+    test_pooling_full_2d_type('sum')
+
+
 @with_seed()
 def test_global_pooling():
     def test_1d_pooling(pool_type, p_value=2):