ocl-misc: adjusted some defaults and updated documentation (#445)

* Operate on a unique list of triplets (tune_multiply.sh). Adjusted documentation. * Made ACC_OPENCL_DEVSPLIT=1 the default (disable with ACC_OPENCL_DEVSPLIT=0). * Code/format cleanup, updated LIBXSMM (Daint-CI).
cp2k · Apr 26, 2021 · 8a30155 · 8a30155
1 parent fb3c5de
commit 8a30155
Show file tree

Hide file tree

Showing 5 changed files with 7 additions and 4 deletions.
diff --git a/.ci/daint.cscs.ch/ocl.build.sh b/.ci/daint.cscs.ch/ocl.build.sh
@@ -25,7 +25,7 @@ if [ ! -d "${HOME}/libxsmm" ]; then
 fi
 cd "${HOME}/libxsmm"
 git fetch
-git checkout 66bab63ee6c7565b725e1fb70f66cdb2fa507250
+git checkout af0dcb9018ed1f92173d60bc04dcf00a4e1854f1
 make -j
 cd ..
 

diff --git a/src/acc/opencl/acc_opencl.c b/src/acc/opencl/acc_opencl.c
@@ -193,7 +193,7 @@ int c_dbcsr_acc_init(void)
             cl_uint j = 0, n;
             for (; j < ndevices; ++j) {
 #if defined(CL_VERSION_1_2)
-              if ( (NULL == env_device_split || '0' == *env_device_split)
+              if ( (NULL != env_device_split && '0' == *env_device_split)
                 || (c_dbcsr_acc_opencl_ndevices + 1) == ACC_OPENCL_DEVICES_MAXCOUNT
                 || (CL_SUCCESS != clCreateSubDevices(devices[j], properties, 0, NULL, &n)))
 #endif

diff --git a/src/acc/opencl/smm/README.md b/src/acc/opencl/smm/README.md
@@ -130,4 +130,4 @@ cd src/acc/opencl/smm
 ./tune_multiply.sh 300  8 1  4 10 15, 6 7 8, 23
 ```
 
-The script `tune_multiply.sh` is tuning 1444 kernels by default (`./acc_bench_smm 300 8 1` taking approximately 15 hours per part). If the process is interrupted earlier (per SIGINT or Ctrl-C), the execution terminates for all requested kernels (triplet specification) unless an environment variable `CONTINUE=1` is set (proceeds to the next kernel).
+The script `tune_multiply.sh` is tuning 1266 kernels by default (`./tune_multiply.sh 300 8 1` taking approximately 13 hours per part). If the process is interrupted earlier (per SIGINT or Ctrl-C), the execution terminates for all requested kernels (triplet specification) unless an environment variable `CONTINUE=1` is set (proceeds to the next kernel).
diff --git a/src/acc/opencl/smm/opencl_libsmm.c b/src/acc/opencl/smm/opencl_libsmm.c
@@ -973,7 +973,7 @@ int libsmm_acc_process(const int* host_param_stack, const int* dev_param_stack,
           if (EXIT_SUCCESS == result) {
             const double gflops = (2.0 * m_max * n_max * k_max * stack_size) / duration;
 #   if LIBXSMM_VERSION3(1, 16, 1) <= LIBXSMM_VERSION3(LIBXSMM_VERSION_MAJOR, \
-         LIBXSMM_VERSION_MINOR, LIBXSMM_VERSION_UPDATE) && 1159 <= LIBXSMM_VERSION_PATCH
+       LIBXSMM_VERSION_MINOR, LIBXSMM_VERSION_UPDATE) && 1159 <= LIBXSMM_VERSION_PATCH
             const size_t size = sizeof(config->size) / sizeof(*config->size); assert(2 <= size);
             libxsmm_kahan_sum(log(gflops), &config->gflops_sumlog, &config->gflops_comp);
             if (size <= config->nexec) {

diff --git a/src/acc/opencl/smm/tune_multiply.sh b/src/acc/opencl/smm/tune_multiply.sh
@@ -73,6 +73,9 @@ if [ "${SED}" ] && [ "${LS}" ] && [ "${RM}" ] && [ "${WC}" ]; then
     SPEC=$(echo "${SPECS}" | ${SED} -e "s/^x//g" -e "s/x$//g" -e "s/x/,/g")
     MNKS="${MNKS} $(eval printf "%s" "{${SPEC}}x{${SPEC}}x{${SPEC}}\" \"" | ${SED} -e "s/{//g" -e "s/}//g")"
   done
+  if [ "$(command -v sort)" ] && [ "$(command -v xargs)" ]; then
+    MNKS=$(echo "${MNKS}" | xargs -n1 | sort -u | xargs)
+  fi
   NTRIPLETS=$(echo "${MNKS}" | wc -w)
   PARTSIZE=$(((NTRIPLETS+NPARTS-1)/NPARTS))
   PARTOFFS=$(((PART-1)*PARTSIZE))