summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: b76e6b5)
raw | patch | inline | side by side (parent: b76e6b5)
author | Jianzhong Xu <xuj@ti.com> | |
Thu, 14 Apr 2016 14:24:36 +0000 (14:24 +0000) | ||
committer | Jianzhong Xu <xuj@ti.com> | |
Thu, 14 Apr 2016 14:24:36 +0000 (14:24 +0000) |
src/ti/linalg/tuning/common/tune_com.h | patch | blob | history | |
src/ti/linalg/tuning/dgemm_tune/dgemm_tune.c | patch | blob | history | |
src/ti/linalg/tuning/make.inc | patch | blob | history |
index 851a47b8c90b322345c03612614f58e6dc89e82e..fcf71a2d31ab50bd0e42c9f4dfe06947756fc7bf 100644 (file)
#include <stdio.h>
#define TUNING_START_SIZE_SQUARE_MATRIX 16
-#define TUNING_START_SIZE_RECTAN_MATRIX 8
-#define NUM_MATRIX_SIZE_TO_BENCHMARK 16
+#define TUNING_START_SIZE_RECTAN_MATRIX 64
+#define NUM_MATRIX_SIZE_TO_BENCHMARK 4 //16
#define HAS_MEMORY 1
#define NO_MEMORY 0
#define OFFLOAD 1
diff --git a/src/ti/linalg/tuning/dgemm_tune/dgemm_tune.c b/src/ti/linalg/tuning/dgemm_tune/dgemm_tune.c
index 5f88529ad662f58d14b5b5ff556b3ce1b44fc187..d4b043fdfa1115b85a6cd5866221c6bd8c9b4532 100644 (file)
@@ -223,8 +223,7 @@ int run_dgemm_dsp_and_arm(int M, int N, int K, float *time_dsp, float *time_arm,
/*------------------------------------------------------------------------
* Time DSP dgemm
*-----------------------------------------------------------------------*/
- //ti_cblas_offload_config("001"); /* force offloading level 3 to DSP */
- //printf("Running on DSP.\n");
+ printf("Running on DSP.\n");
TI_CBLAS_L3_OFFLOAD = 1;
tick();
@@ -249,8 +248,7 @@ int run_dgemm_dsp_and_arm(int M, int N, int K, float *time_dsp, float *time_arm,
/*-------------------------------------------------------------------------
* Time ARM dgemm
*------------------------------------------------------------------------*/
- //ti_cblas_offload_config("000"); /* force no offloading */
- //printf("Running on ARM.\n");
+ printf("Running on ARM.\n");
TI_CBLAS_L3_OFFLOAD = 0;
tick();
index ceb605e53de506c8fed95a2c7c3ec5c58614beec..ea6f25acd974acec1a43aa1db90942009fe132a4 100644 (file)
+UNAME_M :=$(shell uname -m)
+ifneq (,$(findstring 86, $(UNAME_M)))
+ # In a cross compile environment we are assuming that the EVM file system
+ # is located on the build host and necessary ARM libraries are installed
+ # on that file system.
+ ifneq ($(MAKECMDGOALS),clean)
+ ifeq ($(TARGET_ROOTDIR),)
+ $(error Environment variable TARGET_ROOTDIR must be defined. Set it to point at the EVM root file system)
+ endif
+ endif
+ # gcc ARM cross compiler will not, by default, search the host's
+ # /usr/include. Explicitly specify here to find dependent vendor headers
+ CC = arm-linux-gnueabihf-gcc
+else
+ CC = gcc
+endif
-CC = gcc
-CFLAGS = -g -O2 -I/usr/include
+CFLAGS = -g -O2 -I$(TARGET_ROOTDIR)/usr/include -I$(LINALG_DIR)/packages/ti/linalg
-BLAS_LIB_DIR = /usr/lib/
-BLASLIB = $(BLAS_LIB_DIR)libcblas_armplusdsp.a $(BLAS_LIB_DIR)libblis.a -lOpenCL -locl_util -lstdc++ -lrt -lm -lgomp
+LIB_DIR = $(LINALG_DIR)/packages/ti/linalg/lib/
+LD_FLAGS=-L$(TARGET_ROOTDIR)/lib -L$(TARGET_ROOTDIR)/usr/lib -Wl,-rpath-link,$(TARGET_ROOTDIR)/lib -Wl,-rpath-link,$(TARGET_ROOTDIR)/usr/lib
+BLASLIB = $(LIB_DIR)libcblas_armplusdsp.a -lOpenCL -locl_util -lstdc++ -lrt -lm -lgomp -lpthread
+LAPACKLIB = $(LIB_DIR)libcblaswr.a $(LIB_DIR)liblapack.a $(LIB_DIR)libf2c.a $(LIB_DIR)libcblas_armplusdsp.a -lOpenCL -locl_util -lstdc++ -lrt -lm -lgomp -lpthread
TUNE_UTILS = ../common/print_header.o
TUNE_UTILS_OBJ = print_header.o