90d3fb
From ad278554860b0da7d5848262a7bf35e058266cb1 Mon Sep 17 00:00:00 2001
90d3fb
From: Andreas Arnez <arnez@linux.ibm.com>
90d3fb
Date: Wed, 12 Dec 2018 20:06:27 +0100
90d3fb
Subject: [PATCH 5/8] Optimizations for IBM z13
90d3fb
90d3fb
Perform some optimizations for IBM z13:
90d3fb
- Compile with -O2 instead of -O.
90d3fb
- Streamline vector loads/stores.
90d3fb
- Define the vvrsum2 macro.
90d3fb
90d3fb
Also, use the compile option -march=z13 instead of -march=native.
90d3fb
---
90d3fb
 CONFIG/src/atlcomp.txt |  8 +++-----
90d3fb
 include/atlas_simd.h   | 11 +++++------
90d3fb
 2 files changed, 8 insertions(+), 11 deletions(-)
90d3fb
90d3fb
diff --git a/CONFIG/src/atlcomp.txt b/CONFIG/src/atlcomp.txt
90d3fb
index aa31604..2ac71cf 100644
90d3fb
--- a/CONFIG/src/atlcomp.txt
90d3fb
+++ b/CONFIG/src/atlcomp.txt
90d3fb
@@ -246,12 +246,10 @@ MACH=IBMz9,IBMz10,IBMz196 OS=ALL LVL=500 COMPS=f77
90d3fb
    'gfortran' '-O3 -funroll-loops'
90d3fb
 MACH=IBMz9,IBMz10,IBMz196,IBMz12 OS=ALL LVL=500 COMPS=smc,dmc,skc,dkc,icc,xcc,gcc
90d3fb
    'gcc' '-O3 -funroll-loops'
90d3fb
-MACH=IBMz13 OS=ALL LVL=1000 COMPS=dmc,skc,dkc,icc,xcc,gcc
90d3fb
-   'gcc' '-march=native -O -mvx -mzvector'
90d3fb
-MACH=IBMz13 OS=ALL LVL=1000 COMPS=smc
90d3fb
-   'gcc' '-march=native -O -mvx -mzvector -fno-peephole -fno-peephole2'
90d3fb
+MACH=IBMz13 OS=ALL LVL=1000 COMPS=smc,dmc,skc,dkc,icc,xcc,gcc
90d3fb
+   'gcc' '-march=z13 -mtune=z13 -O2'
90d3fb
 MACH=IBMz13 OS=ALL LVL=1000 COMPS=f77
90d3fb
-   'gfortran' '-march=native -O -mvx -mzvector'
90d3fb
+   'gfortran' '-march=z13 -mtune=z13 -O2'
90d3fb
 #
90d3fb
 # Windows defaults ; need to make SSE/SSE2 arch dep.
90d3fb
 #
90d3fb
diff --git a/include/atlas_simd.h b/include/atlas_simd.h
90d3fb
index 68daf79..f171933 100644
90d3fb
--- a/include/atlas_simd.h
90d3fb
+++ b/include/atlas_simd.h
90d3fb
@@ -384,8 +384,8 @@
90d3fb
    #endif
90d3fb
    #define ATL_VTYPE vector double
90d3fb
    #if (defined(DREAL) || defined(DCPLX))
90d3fb
-      #define ATL_vld(v_, p_) {v_[0] = *(p_); v_[1] = (p_)[1]; }
90d3fb
-      #define ATL_vst(p_, v_)  {*(p_) = v_[0]; (p_)[1] = v_[1];}
90d3fb
+      #define ATL_vld(v_, p_) v_ = *(ATL_VTYPE *)(p_)
90d3fb
+      #define ATL_vst(p_, v_) *(ATL_VTYPE *)(p_) = v_
90d3fb
    #else
90d3fb
       #define ATL_vld(v_, p_) v_ = vec_ld2f(p_);
90d3fb
       #define ATL_vst(p_, v_) vec_st2f(v_, p_);
90d3fb
@@ -400,10 +400,9 @@
90d3fb
    #define ATL_vmul(d_, s1_, s2_) d_ =  s1_ * s2_
90d3fb
    #define ATL_vmac(d_, s1_, s2_) d_ = __builtin_s390_vec_madd(s1_, s2_, d_)
90d3fb
    #define ATL_vvrsum1(s0_) \
90d3fb
-   {  ATL_VTYPE t_;\
90d3fb
-      t_ = vec_splat(s0_, 1); \
90d3fb
-      s0_ += t_; \
90d3fb
-   }
90d3fb
+   {  s0_ = vec_mergeh(s0_, s0_) + vec_mergel(s0_, s0_); }
90d3fb
+   #define ATL_vvrsum2(s0_, s1_) \
90d3fb
+   {  s0_ = vec_mergeh(s0_, s1_) + vec_mergel(s0_, s1_); }
90d3fb
    #define ATL_vsplat0(d_, s_) d_ = vec_splat(s_, 0)
90d3fb
    #define ATL_vsplat1(d_, s_) d_ = vec_splat(s_, 1)
90d3fb
 #elif defined(ATL_NEON) && (defined(SREAL) || defined(SCPLX))
90d3fb
-- 
90d3fb
2.23.0
90d3fb