Persuading more efficient memory format to be preferred (#44078)

* - blind shot fix

* - workaround

* - compilation fix

* - Hack

paddle/fluid/platform/mkldnn_helper.h

@@ -233,15 +233,21 @@ inline dnnl::memory::format_tag GetMKLDNNFormat(dnnl::memory::desc mem_desc) {
     if (inner_nblks == 0) {
       if (strides[0] >= strides[1] && strides[1] >= strides[2] &&
           strides[2] >= strides[3]) {
-        return dnnl::memory::format_tag::nchw;
+        return dnnl::memory::format_tag::abcd;
+      } else if (strides[2] >= strides[3] && strides[3] >= strides[1] &&
+                 strides[1] >= strides[0]) {
+        return dnnl::memory::format_tag::cdba;
+      } else if (strides[0] >= strides[2] && strides[2] >= strides[3] &&
+                 strides[3] >= strides[1]) {
+        return dnnl::memory::format_tag::acdb;
+      } else if (strides[0] >= strides[1] && strides[1] >= strides[3] &&
+                 strides[3] >= strides[2]) {
+        return dnnl::memory::format_tag::abdc;
       } else if (strides[2] >= strides[3] && strides[3] >= strides[1] &&
                  strides[1] >= strides[0]) {
         return dnnl::memory::format_tag::cdba;
-      } else if (strides[3] >= strides[2] && strides[2] >= strides[0] &&
-                 strides[0] >= strides[1]) {
-        return dnnl::memory::format_tag::dcab;
       } else {
-        return dnnl::memory::format_tag::nhwc;
+        return dnnl::memory::format_tag::dcab;
       }
     } else if (inner_nblks == 1) {
       if (inner_blks[0] == 16 && inner_idxs[0] == 1) {

paddle/fluid/platform/mkldnn_utils.h

@@ -51,15 +51,21 @@ inline dnnl::memory::format_tag GetMKLDNNFormat(dnnl::memory::desc mem_desc) {
     if (inner_nblks == 0) {
       if (strides[0] >= strides[1] && strides[1] >= strides[2] &&
           strides[2] >= strides[3]) {
-        return dnnl::memory::format_tag::nchw;
+        return dnnl::memory::format_tag::abcd;
+      } else if (strides[2] >= strides[3] && strides[3] >= strides[1] &&
+                 strides[1] >= strides[0]) {
+        return dnnl::memory::format_tag::cdba;
+      } else if (strides[0] >= strides[2] && strides[2] >= strides[3] &&
+                 strides[3] >= strides[1]) {
+        return dnnl::memory::format_tag::acdb;
+      } else if (strides[0] >= strides[1] && strides[1] >= strides[3] &&
+                 strides[3] >= strides[2]) {
+        return dnnl::memory::format_tag::abdc;
       } else if (strides[2] >= strides[3] && strides[3] >= strides[1] &&
                  strides[1] >= strides[0]) {
         return dnnl::memory::format_tag::cdba;
-      } else if (strides[3] >= strides[2] && strides[2] >= strides[0] &&
-                 strides[0] >= strides[1]) {
-        return dnnl::memory::format_tag::dcab;
       } else {
-        return dnnl::memory::format_tag::nhwc;
+        return dnnl::memory::format_tag::dcab;
       }
     } else if (inner_nblks == 1) {
       if (inner_blks[0] == 16 && inner_idxs[0] == 1) {