From d4dda8628e40d7604f30b396916f1e6802a090a1 Mon Sep 17 00:00:00 2001
From: GaoWei8 <53294385+GaoWei8@users.noreply.github.com>
Date: Mon, 23 Dec 2019 12:09:56 +0800
Subject: [PATCH] optimize fc jit (#21878)

test=develop
---
 paddle/fluid/operators/math/fc.cc | 29 +++++++++++------------------
 1 file changed, 11 insertions(+), 18 deletions(-)

diff --git a/paddle/fluid/operators/math/fc.cc b/paddle/fluid/operators/math/fc.cc
index 27a75f7631a..9309a992f73 100644
--- a/paddle/fluid/operators/math/fc.cc
+++ b/paddle/fluid/operators/math/fc.cc
@@ -61,27 +61,20 @@ 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()
-              .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);
+    auto compute =
+        relu
+            ? jit::KernelFuncs<jit::VAddReluTuple<T>,
+                               platform::CPUPlace>::Cache()
+                  .At(N)
+            : jit::KernelFuncs<jit::VAddTuple<T>, platform::CPUPlace>::Cache()
+                  .At(N);
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
-      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);
-      }
+    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);
     }
   }
 };
-- 
GitLab