|
 |
33fa5a |
From ad278554860b0da7d5848262a7bf35e058266cb1 Mon Sep 17 00:00:00 2001
|
|
|
33fa5a |
From: Andreas Arnez <arnez@linux.ibm.com>
|
|
|
33fa5a |
Date: Wed, 12 Dec 2018 20:06:27 +0100
|
|
|
33fa5a |
Subject: [PATCH 5/8] Optimizations for IBM z13
|
|
|
33fa5a |
|
|
|
33fa5a |
Perform some optimizations for IBM z13:
|
|
|
33fa5a |
- Compile with -O2 instead of -O.
|
|
|
33fa5a |
- Streamline vector loads/stores.
|
|
|
33fa5a |
- Define the vvrsum2 macro.
|
|
|
33fa5a |
|
|
|
33fa5a |
Also, use the compile option -march=z13 instead of -march=native.
|
|
|
33fa5a |
---
|
|
|
33fa5a |
CONFIG/src/atlcomp.txt | 8 +++-----
|
|
|
33fa5a |
include/atlas_simd.h | 11 +++++------
|
|
|
33fa5a |
2 files changed, 8 insertions(+), 11 deletions(-)
|
|
|
33fa5a |
|
|
|
33fa5a |
diff --git a/CONFIG/src/atlcomp.txt b/CONFIG/src/atlcomp.txt
|
|
|
33fa5a |
index aa31604..2ac71cf 100644
|
|
|
33fa5a |
--- a/CONFIG/src/atlcomp.txt
|
|
|
33fa5a |
+++ b/CONFIG/src/atlcomp.txt
|
|
|
33fa5a |
@@ -246,12 +246,10 @@ MACH=IBMz9,IBMz10,IBMz196 OS=ALL LVL=500 COMPS=f77
|
|
|
33fa5a |
'gfortran' '-O3 -funroll-loops'
|
|
|
33fa5a |
MACH=IBMz9,IBMz10,IBMz196,IBMz12 OS=ALL LVL=500 COMPS=smc,dmc,skc,dkc,icc,xcc,gcc
|
|
|
33fa5a |
'gcc' '-O3 -funroll-loops'
|
|
|
33fa5a |
-MACH=IBMz13 OS=ALL LVL=1000 COMPS=dmc,skc,dkc,icc,xcc,gcc
|
|
|
33fa5a |
- 'gcc' '-march=native -O -mvx -mzvector'
|
|
|
33fa5a |
-MACH=IBMz13 OS=ALL LVL=1000 COMPS=smc
|
|
|
33fa5a |
- 'gcc' '-march=native -O -mvx -mzvector -fno-peephole -fno-peephole2'
|
|
|
33fa5a |
+MACH=IBMz13 OS=ALL LVL=1000 COMPS=smc,dmc,skc,dkc,icc,xcc,gcc
|
|
|
33fa5a |
+ 'gcc' '-march=z13 -mtune=z13 -O2'
|
|
|
33fa5a |
MACH=IBMz13 OS=ALL LVL=1000 COMPS=f77
|
|
|
33fa5a |
- 'gfortran' '-march=native -O -mvx -mzvector'
|
|
|
33fa5a |
+ 'gfortran' '-march=z13 -mtune=z13 -O2'
|
|
|
33fa5a |
#
|
|
|
33fa5a |
# Windows defaults ; need to make SSE/SSE2 arch dep.
|
|
|
33fa5a |
#
|
|
|
33fa5a |
diff --git a/include/atlas_simd.h b/include/atlas_simd.h
|
|
|
33fa5a |
index 68daf79..f171933 100644
|
|
|
33fa5a |
--- a/include/atlas_simd.h
|
|
|
33fa5a |
+++ b/include/atlas_simd.h
|
|
|
33fa5a |
@@ -384,8 +384,8 @@
|
|
|
33fa5a |
#endif
|
|
|
33fa5a |
#define ATL_VTYPE vector double
|
|
|
33fa5a |
#if (defined(DREAL) || defined(DCPLX))
|
|
|
33fa5a |
- #define ATL_vld(v_, p_) {v_[0] = *(p_); v_[1] = (p_)[1]; }
|
|
|
33fa5a |
- #define ATL_vst(p_, v_) {*(p_) = v_[0]; (p_)[1] = v_[1];}
|
|
|
33fa5a |
+ #define ATL_vld(v_, p_) v_ = *(ATL_VTYPE *)(p_)
|
|
|
33fa5a |
+ #define ATL_vst(p_, v_) *(ATL_VTYPE *)(p_) = v_
|
|
|
33fa5a |
#else
|
|
|
33fa5a |
#define ATL_vld(v_, p_) v_ = vec_ld2f(p_);
|
|
|
33fa5a |
#define ATL_vst(p_, v_) vec_st2f(v_, p_);
|
|
|
33fa5a |
@@ -400,10 +400,9 @@
|
|
|
33fa5a |
#define ATL_vmul(d_, s1_, s2_) d_ = s1_ * s2_
|
|
|
33fa5a |
#define ATL_vmac(d_, s1_, s2_) d_ = __builtin_s390_vec_madd(s1_, s2_, d_)
|
|
|
33fa5a |
#define ATL_vvrsum1(s0_) \
|
|
|
33fa5a |
- { ATL_VTYPE t_;\
|
|
|
33fa5a |
- t_ = vec_splat(s0_, 1); \
|
|
|
33fa5a |
- s0_ += t_; \
|
|
|
33fa5a |
- }
|
|
|
33fa5a |
+ { s0_ = vec_mergeh(s0_, s0_) + vec_mergel(s0_, s0_); }
|
|
|
33fa5a |
+ #define ATL_vvrsum2(s0_, s1_) \
|
|
|
33fa5a |
+ { s0_ = vec_mergeh(s0_, s1_) + vec_mergel(s0_, s1_); }
|
|
|
33fa5a |
#define ATL_vsplat0(d_, s_) d_ = vec_splat(s_, 0)
|
|
|
33fa5a |
#define ATL_vsplat1(d_, s_) d_ = vec_splat(s_, 1)
|
|
|
33fa5a |
#elif defined(ATL_NEON) && (defined(SREAL) || defined(SCPLX))
|
|
|
33fa5a |
--
|
|
|
33fa5a |
2.23.0
|
|
|
33fa5a |
|