From 5e59c95be403b26e59e8b914e2b43fe31441dd9f Mon Sep 17 00:00:00 2001
From: Trung Nguyen <ndactrung@gmail.com>
Date: Sat, 10 Sep 2022 02:45:06 -0500
Subject: [PATCH] Moved temp variables inside loops

---
 lib/gpu/lal_amoeba.cu | 187 ++++++++++++++++++------------------------
 1 file changed, 82 insertions(+), 105 deletions(-)

diff --git a/lib/gpu/lal_amoeba.cu b/lib/gpu/lal_amoeba.cu
index fb66158d06..105f18cfa8 100644
--- a/lib/gpu/lal_amoeba.cu
+++ b/lib/gpu/lal_amoeba.cu
@@ -1643,75 +1643,52 @@ __kernel void k_fphi_uind(const __global numtyp4 *restrict x_,
   int ii=tid+BLOCK_ID_X*BLOCK_SIZE_X;
 
   if (ii<inum) {
-    numtyp4 ix; fetch4(ix,ii,pos_tex); //x_[i];
+    //numtyp4 ix; fetch4(ix,ii,pos_tex); //x_[i];
     acctyp fdip_buf[32];
 
-    int j,k,m;
-    numtyp v0,v1,v2,v3;
-    numtyp u0,u1,u2,u3;
-    numtyp t0,t1,t2,t3;
-    numtyp t0_1,t0_2,t1_1,t1_2;
-    numtyp t2_1,t2_2,tq_1,tq_2;
-    numtyp tu00,tu10,tu01,tu20,tu11;
-    numtyp tu02,tu30,tu21,tu12,tu03;
-    numtyp tu00_1,tu01_1,tu10_1;
-    numtyp tu00_2,tu01_2,tu10_2;
-    numtyp tu20_1,tu11_1,tu02_1;
-    numtyp tu20_2,tu11_2,tu02_2;
-    numtyp tuv100_1,tuv010_1,tuv001_1;
-    numtyp tuv100_2,tuv010_2,tuv001_2;
-    numtyp tuv200_1,tuv020_1,tuv002_1;
-    numtyp tuv110_1,tuv101_1,tuv011_1;
-    numtyp tuv200_2,tuv020_2,tuv002_2;
-    numtyp tuv110_2,tuv101_2,tuv011_2;
-    numtyp tuv000,tuv100,tuv010,tuv001;
-    numtyp tuv200,tuv020,tuv002,tuv110;
-    numtyp tuv101,tuv011,tuv300,tuv030;
-    numtyp tuv003,tuv210,tuv201,tuv120;
-    numtyp tuv021,tuv102,tuv012,tuv111;
-
+    int j,k,m;    
     int nlpts = (bsorder-1) / 2;
 
     // extract the permanent multipole field at each site
 
-    tuv100_1 = (numtyp)0.0;
-    tuv010_1 = (numtyp)0.0;
-    tuv001_1 = (numtyp)0.0;
-    tuv200_1 = (numtyp)0.0;
-    tuv020_1 = (numtyp)0.0;
-    tuv002_1 = (numtyp)0.0;
-    tuv110_1 = (numtyp)0.0;
-    tuv101_1 = (numtyp)0.0;
-    tuv011_1 = (numtyp)0.0;
-    tuv100_2 = (numtyp)0.0;
-    tuv010_2 = (numtyp)0.0;
-    tuv001_2 = (numtyp)0.0;
-    tuv200_2 = (numtyp)0.0;
-    tuv020_2 = (numtyp)0.0;
-    tuv002_2 = (numtyp)0.0;
-    tuv110_2 = (numtyp)0.0;
-    tuv101_2 = (numtyp)0.0;
-    tuv011_2 = (numtyp)0.0;
-    tuv000 = (numtyp)0.0;
-    tuv001 = (numtyp)0.0;
-    tuv010 = (numtyp)0.0;
-    tuv100 = (numtyp)0.0;
-    tuv200 = (numtyp)0.0;
-    tuv020 = (numtyp)0.0;
-    tuv002 = (numtyp)0.0;
-    tuv110 = (numtyp)0.0;
-    tuv101 = (numtyp)0.0;
-    tuv011 = (numtyp)0.0;
-    tuv300 = (numtyp)0.0;
-    tuv030 = (numtyp)0.0;
-    tuv003 = (numtyp)0.0;
-    tuv210 = (numtyp)0.0;
-    tuv201 = (numtyp)0.0;
-    tuv120 = (numtyp)0.0;
-    tuv021 = (numtyp)0.0;
-    tuv102 = (numtyp)0.0;
-    tuv012 = (numtyp)0.0;
-    tuv111 = (numtyp)0.0;
+    numtyp tuv100_1 = (numtyp)0.0;
+    numtyp tuv010_1 = (numtyp)0.0;
+    numtyp tuv001_1 = (numtyp)0.0;
+    numtyp tuv200_1 = (numtyp)0.0;
+    numtyp tuv020_1 = (numtyp)0.0;
+    numtyp tuv002_1 = (numtyp)0.0;
+    numtyp tuv110_1 = (numtyp)0.0;
+    numtyp tuv101_1 = (numtyp)0.0;
+    numtyp tuv011_1 = (numtyp)0.0;
+    numtyp tuv100_2 = (numtyp)0.0;
+    numtyp tuv010_2 = (numtyp)0.0;
+    numtyp tuv001_2 = (numtyp)0.0;
+    numtyp tuv200_2 = (numtyp)0.0;
+    numtyp tuv020_2 = (numtyp)0.0;
+    numtyp tuv002_2 = (numtyp)0.0;
+    numtyp tuv110_2 = (numtyp)0.0;
+    numtyp tuv101_2 = (numtyp)0.0;
+    numtyp tuv011_2 = (numtyp)0.0;
+    numtyp tuv000 = (numtyp)0.0;
+    numtyp tuv001 = (numtyp)0.0;
+    numtyp tuv010 = (numtyp)0.0;
+    numtyp tuv100 = (numtyp)0.0;
+    numtyp tuv200 = (numtyp)0.0;
+    numtyp tuv020 = (numtyp)0.0;
+    numtyp tuv002 = (numtyp)0.0;
+    numtyp tuv110 = (numtyp)0.0;
+    numtyp tuv101 = (numtyp)0.0;
+    numtyp tuv011 = (numtyp)0.0;
+    numtyp tuv300 = (numtyp)0.0;
+    numtyp tuv030 = (numtyp)0.0;
+    numtyp tuv003 = (numtyp)0.0;
+    numtyp tuv210 = (numtyp)0.0;
+    numtyp tuv201 = (numtyp)0.0;
+    numtyp tuv120 = (numtyp)0.0;
+    numtyp tuv021 = (numtyp)0.0;
+    numtyp tuv102 = (numtyp)0.0;
+    numtyp tuv012 = (numtyp)0.0;
+    numtyp tuv111 = (numtyp)0.0;
 
     k = igrid[4*ii+2] - nzlo_out - nlpts;
     for (int kb = 0; kb < bsorder; kb++) {
@@ -1722,32 +1699,32 @@ __kernel void k_fphi_uind(const __global numtyp4 *restrict x_,
       v3 = thetai3[m][kb][3];
       */
       int i3 = ii*4*bsorder + 4*kb;
-      v0 = thetai3[i3];
-      v1 = thetai3[i3+1];
-      v2 = thetai3[i3+2];
-      v3 = thetai3[i3+3];
-      tu00_1 = (numtyp)0.0;
-      tu01_1 = (numtyp)0.0;
-      tu10_1 = (numtyp)0.0;
-      tu20_1 = (numtyp)0.0;
-      tu11_1 = (numtyp)0.0;
-      tu02_1 = (numtyp)0.0;
-      tu00_2 = (numtyp)0.0;
-      tu01_2 = (numtyp)0.0;
-      tu10_2 = (numtyp)0.0;
-      tu20_2 = (numtyp)0.0;
-      tu11_2 = (numtyp)0.0;
-      tu02_2 = (numtyp)0.0;
-      tu00 = (numtyp)0.0;
-      tu10 = (numtyp)0.0;
-      tu01 = (numtyp)0.0;
-      tu20 = (numtyp)0.0;
-      tu11 = (numtyp)0.0;
-      tu02 = (numtyp)0.0;
-      tu30 = (numtyp)0.0;
-      tu21 = (numtyp)0.0;
-      tu12 = (numtyp)0.0;
-      tu03 = (numtyp)0.0;
+      numtyp v0 = thetai3[i3];
+      numtyp v1 = thetai3[i3+1];
+      numtyp v2 = thetai3[i3+2];
+      numtyp v3 = thetai3[i3+3];
+      numtyp tu00_1 = (numtyp)0.0;
+      numtyp tu01_1 = (numtyp)0.0;
+      numtyp tu10_1 = (numtyp)0.0;
+      numtyp tu20_1 = (numtyp)0.0;
+      numtyp tu11_1 = (numtyp)0.0;
+      numtyp tu02_1 = (numtyp)0.0;
+      numtyp tu00_2 = (numtyp)0.0;
+      numtyp tu01_2 = (numtyp)0.0;
+      numtyp tu10_2 = (numtyp)0.0;
+      numtyp tu20_2 = (numtyp)0.0;
+      numtyp tu11_2 = (numtyp)0.0;
+      numtyp tu02_2 = (numtyp)0.0;
+      numtyp tu00 = (numtyp)0.0;
+      numtyp tu10 = (numtyp)0.0;
+      numtyp tu01 = (numtyp)0.0;
+      numtyp tu20 = (numtyp)0.0;
+      numtyp tu11 = (numtyp)0.0;
+      numtyp tu02 = (numtyp)0.0;
+      numtyp tu30 = (numtyp)0.0;
+      numtyp tu21 = (numtyp)0.0;
+      numtyp tu12 = (numtyp)0.0;
+      numtyp tu03 = (numtyp)0.0;
 
       j = igrid[4*ii+1] - nylo_out - nlpts;
       for (int jb = 0; jb < bsorder; jb++) {
@@ -1758,17 +1735,17 @@ __kernel void k_fphi_uind(const __global numtyp4 *restrict x_,
         u3 = thetai2[m][jb][3];
         */
         int i2 = ii*4*bsorder+4*jb;
-        u0 = thetai2[i2];
-        u1 = thetai2[i2+1];
-        u2 = thetai2[i2+2];
-        u3 = thetai2[i2+3];
-        t0_1 = (numtyp)0.0;
-        t1_1 = (numtyp)0.0;
-        t2_1 = (numtyp)0.0;
-        t0_2 = (numtyp)0.0;
-        t1_2 = (numtyp)0.0;
-        t2_2 = (numtyp)0.0;
-        t3 = (numtyp)0.0;
+        numtyp u0 = thetai2[i2];
+        numtyp u1 = thetai2[i2+1];
+        numtyp u2 = thetai2[i2+2];
+        numtyp u3 = thetai2[i2+3];
+        numtyp t0_1 = (numtyp)0.0;
+        numtyp t1_1 = (numtyp)0.0;
+        numtyp t2_1 = (numtyp)0.0;
+        numtyp t0_2 = (numtyp)0.0;
+        numtyp t1_2 = (numtyp)0.0;
+        numtyp t2_2 = (numtyp)0.0;
+        numtyp t3 = (numtyp)0.0;
 
         int i = igrid[4*ii] - nxlo_out - nlpts;
         for (int ib = 0; ib < bsorder; ib++) {
@@ -1789,8 +1766,8 @@ __kernel void k_fphi_uind(const __global numtyp4 *restrict x_,
           numtyp w2 = thetai1[i1+2];
           numtyp w3 = thetai1[i1+3];
           int gidx = 2*(k*ngridxy + j*ngridx + i);
-          tq_1 = grid[gidx];
-          tq_2 = grid[gidx+1];
+          numtyp tq_1 = grid[gidx];
+          numtyp tq_2 = grid[gidx+1];
           t0_1 += tq_1*w0;
           t1_1 += tq_1*w1;
           t2_1 += tq_1*w2;
@@ -1813,9 +1790,9 @@ __kernel void k_fphi_uind(const __global numtyp4 *restrict x_,
         tu20_2 += t2_2*u0;
         tu11_2 += t1_2*u1;
         tu02_2 += t0_2*u2;
-        t0 = t0_1 + t0_2;
-        t1 = t1_1 + t1_2;
-        t2 = t2_1 + t2_2;
+        numtyp t0 = t0_1 + t0_2;
+        numtyp t1 = t1_1 + t1_2;
+        numtyp t2 = t2_1 + t2_2;
         tu00 += t0*u0;
         tu10 += t1*u0;
         tu01 += t0*u1;