From a19b3225a1da8c31fc996bace3ac09e6f5f177ef Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Sat, 17 Nov 2018 14:56:43 +0000
Subject: [PATCH] fix jitcode small size

test=develop
---
 paddle/fluid/operators/math/jit_code.cc        | 12 ++++++++----
 paddle/fluid/operators/math/jit_kernel_test.cc | 10 +++++-----
 2 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/paddle/fluid/operators/math/jit_code.cc b/paddle/fluid/operators/math/jit_code.cc
index a080079a2d..e484e9a3c7 100644
--- a/paddle/fluid/operators/math/jit_code.cc
+++ b/paddle/fluid/operators/math/jit_code.cc
@@ -59,9 +59,10 @@ void VXXJitCode::generate() {
     offset += sizeof(float) * YMM_FLOAT_BLOCK;
   }
   int rest = num_ % YMM_FLOAT_BLOCK;
-  int block = XMM_FLOAT_BLOCK;
   while (rest > 0) {
+    int block = XMM_FLOAT_BLOCK;
     if (rest >= 4) {
+      block = 4;
       if (scalar_index_ != 1) {
         vmovups(xmm_src1, ptr[param1 + offset]);
       }
@@ -69,6 +70,7 @@ void VXXJitCode::generate() {
         vmovups(xmm_src2, ptr[param2 + offset]);
       }
     } else if (rest >= 2) {
+      block = 2;
       if (scalar_index_ != 1) {
         vmovq(xmm_src1, ptr[param1 + offset]);
       }
@@ -76,6 +78,7 @@ void VXXJitCode::generate() {
         vmovq(xmm_src2, ptr[param2 + offset]);
       }
     } else {
+      block = 1;
       if (scalar_index_ != 1) {
         vmovss(xmm_src1, ptr[param1 + offset]);
       }
@@ -105,7 +108,6 @@ void VXXJitCode::generate() {
     }
     offset += sizeof(float) * block;
     rest -= block;
-    block /= 2;
   }
   ret();
 }
@@ -167,13 +169,16 @@ void VActJitCode::generate() {
     offset += sizeof(float) * YMM_FLOAT_BLOCK;
   }
   int rest = num_ % YMM_FLOAT_BLOCK;
-  int block = XMM_FLOAT_BLOCK;
   while (rest > 0) {
+    int block = XMM_FLOAT_BLOCK;
     if (rest >= 4) {
+      block = 4;
       vmovups(xmm_src, ptr[param1 + offset]);
     } else if (rest >= 2) {
+      block = 2;
       vmovq(xmm_src, ptr[param1 + offset]);
     } else {
+      block = 1;
       vmovss(xmm_src, ptr[param1 + offset]);
     }
     switch (type_) {
@@ -201,7 +206,6 @@ void VActJitCode::generate() {
     }
     offset += sizeof(float) * block;
     rest -= block;
-    block /= 2;
   }
   ret();
 }
diff --git a/paddle/fluid/operators/math/jit_kernel_test.cc b/paddle/fluid/operators/math/jit_kernel_test.cc
index 932fa4c000..b6c62a2634 100644
--- a/paddle/fluid/operators/math/jit_kernel_test.cc
+++ b/paddle/fluid/operators/math/jit_kernel_test.cc
@@ -69,7 +69,7 @@ void vrelu_intri8(const int n, const float* x, float* y) {
 
 TEST(JitKernel, vrelu) {
   namespace jit = paddle::operators::math::jitkernel;
-  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
+  for (int d : {3, 7, 8, 15, 16, 30, 256, 512}) {
     std::vector<float> x(d);
     std::vector<float> zref(d), ztgt(d);
     RandomVec<float>(d, x.data(), -10.f, 1.f);
@@ -159,7 +159,7 @@ void vexp_mkl(const int n, const float* x, float* y) {
 
 TEST(JitKernel, vexp) {
   namespace jit = paddle::operators::math::jitkernel;
-  for (int d : {7, 8, 12, 15, 16, 20, 30, 128, 256}) {
+  for (int d : {1, 3, 4, 6, 7, 8, 12, 15, 16, 20, 30, 128, 256}) {
     std::vector<float> x(d);
     std::vector<float> zref(d), ztgt(d);
     RandomVec<float>(d, x.data(), -2.f, 2.f);
@@ -234,7 +234,7 @@ void vsigmoid_better(
 
 TEST(JitKernel, vsigmoid) {
   namespace jit = paddle::operators::math::jitkernel;
-  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+  for (int d : {1, 3, 4, 6, 7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
     std::vector<float> x(d);
     std::vector<float> zref(d), ztgt(d);
     RandomVec<float>(d, x.data(), -2.f, 2.f);
@@ -298,7 +298,7 @@ void vtanh_better(
 
 TEST(JitKernel, vtanh) {
   namespace jit = paddle::operators::math::jitkernel;
-  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
+  for (int d : {1, 2, 3, 4, 5, 6, 7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
     std::vector<float> x(d);
     std::vector<float> zref(d), ztgt(d);
     RandomVec<float>(d, x.data(), -2.f, 2.f);
@@ -389,7 +389,7 @@ void lstm_ctht_better(
 
 TEST(JitKernel, lstm) {
   namespace jit = paddle::operators::math::jitkernel;
-  for (int d : {7, 8, 15, 16, 30, 32, 64, 100}) {
+  for (int d : {1, 2, 3, 4, 5, 6, 7, 8, 15, 16, 30, 32, 64, 100}) {
     int d4 = d * 4;
     int d3 = d * 3;
     std::vector<float> x(d4), xref(d4);
-- 
GitLab