optimize fc jit (#21878)

test=develop

paddle/fluid/operators/math/fc.cc

@@ -61,19 +61,13 @@ class FCFunctor<platform::CPUDeviceContext, T> {
                             "When bias is NULL, relu can not be true."));
       return;
     }
-    if (relu) {
     auto compute =
-          jit::KernelFuncs<jit::VAddReluTuple<T>, platform::CPUPlace>::Cache()
+        relu
+            ? jit::KernelFuncs<jit::VAddReluTuple<T>,
+                               platform::CPUPlace>::Cache()
+                  .At(N)
+            : jit::KernelFuncs<jit::VAddTuple<T>, platform::CPUPlace>::Cache()
                   .At(N);
-      for (int i = 0; i < M; i++) {
-        T* dst = Y + i * N;
-        T* src = (padding_weights) ? Y1_data + i * (N + 4) : dst;
-        compute(B, src, dst, N);
-      }
-    } else {
-      auto compute =
-          jit::KernelFuncs<jit::VAddTuple<T>, platform::CPUPlace>::Cache().At(
-              N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
@@ -83,7 +77,6 @@ class FCFunctor<platform::CPUDeviceContext, T> {
       compute(B, src, dst, N);
     }
   }
-  }
 };
 
 template class FCFunctor<platform::CPUDeviceContext, float>;