From eb16cf21e96a9c6907f6092e113f5231c09f9f97 Mon Sep 17 00:00:00 2001
From: Incardona Pietro <incardon@mpi-cbg.de>
Date: Sun, 12 Dec 2021 22:31:07 +0100
Subject: [PATCH] Fixing performance
---
example/Performance/miniBUDE/main.cu | 15 ++++++++++++---
.../3_gray_scott_3d_sparse_gpu_cs/Makefile | 5 +++--
openfpm_data | 2 +-
openfpm_devices | 2 +-
4 files changed, 17 insertions(+), 7 deletions(-)
diff --git a/example/Performance/miniBUDE/main.cu b/example/Performance/miniBUDE/main.cu
index cebbeabc4..d845ef21f 100644
--- a/example/Performance/miniBUDE/main.cu
+++ b/example/Performance/miniBUDE/main.cu
@@ -64,10 +64,15 @@ Params params;
typedef struct
{
- openfpm::vector_gpu<aggregate<float[3],int>> d_protein;
- openfpm::vector_gpu<aggregate<float[3],int>> d_ligand;
- openfpm::vector_gpu<aggregate<int,float,float,float>> d_forcefield;
+ // _lin = AOS
+ openfpm::vector_gpu_lin<aggregate<float[3],int>> d_protein;
+ // AOS
+ openfpm::vector_gpu_lin<aggregate<float[3],int>> d_ligand;
+ // AOS
+ openfpm::vector_gpu_lin<aggregate<int,float,float,float>> d_forcefield;
+ // SOA
openfpm::vector_gpu<aggregate<float>> d_results;
+ // SOA
openfpm::vector_gpu<aggregate<float,float,float,float,float,float>> d_poses;
openfpm::vector<double> gflops_data;
@@ -601,14 +606,18 @@ void loadParameters(int argc, char *argv[], OpenFPM & _openfpm)
fclose(file);
}
+#ifndef __APPLE__
#include <fenv.h>
#include <xmmintrin.h>
#include <pmmintrin.h>
+#endif
int main(int argc, char *argv[])
{
+#ifndef __APPLE__
_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
_MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);
+#endif
init_wrappers();
OpenFPM _openfpm;
diff --git a/example/SparseGrid/3_gray_scott_3d_sparse_gpu_cs/Makefile b/example/SparseGrid/3_gray_scott_3d_sparse_gpu_cs/Makefile
index 52d3e5efa..6d5330383 100644
--- a/example/SparseGrid/3_gray_scott_3d_sparse_gpu_cs/Makefile
+++ b/example/SparseGrid/3_gray_scott_3d_sparse_gpu_cs/Makefile
@@ -5,8 +5,9 @@ CUDA_CC=
CUDA_CC_LINK=
ifdef CUDA_ON_CPU
CUDA_CC=mpic++ -x c++ $(INCLUDE_PATH)
- INCLUDE_PATH_NVCC=
- CUDA_CC_LINK=mpic++
+ INCLUDE_PATH_NVCC=
+ CUDA_CC_LINK=mpic++
+ LIBS_SELECT=$(LIBS)
else
ifeq (, $(shell which nvcc))
CUDA_CC=mpic++ -x c++ $(INCLUDE_PATH)
diff --git a/openfpm_data b/openfpm_data
index f84f0a370..df7c8617d 160000
--- a/openfpm_data
+++ b/openfpm_data
@@ -1 +1 @@
-Subproject commit f84f0a370f906dcf2f35ef2747d7adbbe2a397d4
+Subproject commit df7c8617df88632b0de820a9f9d258f3d6b86c44
diff --git a/openfpm_devices b/openfpm_devices
index 4e23dd287..33dfcd4c1 160000
--- a/openfpm_devices
+++ b/openfpm_devices
@@ -1 +1 @@
-Subproject commit 4e23dd287dd78e69b0072d0b1c3fe832e0c45966
+Subproject commit 33dfcd4c1486af8ff33dcb856facd941e044e254
--
GitLab