cp2k · hfp · Feb 8, 2021 · Feb 4, 2021 · Feb 4, 2021 · Feb 4, 2021
diff --git a/src/acc/acc_bench_smm.c b/src/acc/acc_bench_smm.c
@@ -106,15 +106,18 @@ int main(int argc, char* argv[])
   printf("%s%s%i %i %i %i %i %i %i %i\n", 0 < argc ? argv[0] : "", 0 < argc ? " " : "",
     nrepeat, stack_size, m, n, k, nc, na, nb);
   CHECK(c_dbcsr_acc_init(), &result);
+  /* note: libsmm_acc_init() may imply acc_init() */
+  CHECK(libsmm_acc_init(), &result);
   CHECK(c_dbcsr_acc_get_ndevices(&ndevices), &result);
   if (0 < ndevices) {
 #if defined(_DEBUG)
     fprintf(stderr, "number of devices found: %i\n", ndevices);
 #endif
   }
   else {
-#if defined(_DEBUG)
-    fprintf(stderr, "Error: no device found!\n");
+    fprintf(stderr, "No ACC-device found!\n");
+#if !defined(__CUDA)
+    CHECK(libsmm_acc_finalize(), NULL);
 #endif
     CHECK(c_dbcsr_acc_finalize(), NULL);
     return result;
@@ -165,14 +168,14 @@ int main(int argc, char* argv[])
     CHECK(libsmm_acc_transpose(trans_dev, 0/*offset*/, nb, bmat_dev,
       DBCSR_TYPE(ELEM_TYPE), n, k, MAX_KERNEL_DIM, stream), &result);
   }
-#if defined(USE_LIBXSMM)
+# if defined(USE_LIBXSMM)
   CHECK(c_dbcsr_acc_stream_sync(stream), &result);
   start = libxsmm_timer_tick();
-#endif
+# endif
   /* to perform NN-SMMs on the device, all B-matrices are transposed upfront (SMM-kernel is limited to NT) */
   CHECK(libsmm_acc_transpose(trans_dev, 0/*offset*/, nb, bmat_dev,
     DBCSR_TYPE(ELEM_TYPE), k, n, MAX_KERNEL_DIM, stream), &result);
-#if defined(USE_LIBXSMM)
+# if defined(USE_LIBXSMM)
   CHECK(c_dbcsr_acc_stream_sync(stream), &result);
   transpose = libxsmm_timer_duration(start, libxsmm_timer_tick());
 # endif
@@ -282,6 +285,9 @@ int main(int argc, char* argv[])
   CHECK(c_dbcsr_acc_dev_mem_deallocate(bmat_dev), NULL);
   CHECK(c_dbcsr_acc_dev_mem_deallocate(cmat_dev), NULL);
   CHECK(c_dbcsr_acc_stream_destroy(stream), NULL);
+#if !defined(__CUDA)
+  CHECK(libsmm_acc_finalize(), NULL);
+#endif
   CHECK(c_dbcsr_acc_finalize(), NULL);
   if (EXIT_SUCCESS != result) {
     fprintf(stderr, "FAILED\n");

diff --git a/src/acc/acc_bench_trans.c b/src/acc/acc_bench_trans.c
@@ -91,15 +91,18 @@ int main(int argc, char* argv[])
   assert(m <= (mn / n) && 0 == (mn % n));
   printf("%s%s%i %i %i %i\n", 0 < argc ? argv[0] : "", 0 < argc ? " " : "", nrepeat, stack_size, m, n);
   CHECK(c_dbcsr_acc_init(), &result);
+  /* note: libsmm_acc_init() may imply acc_init() */
+  CHECK(libsmm_acc_init(), &result);
   CHECK(c_dbcsr_acc_get_ndevices(&ndevices), &result);
   if (0 < ndevices) {
 #if defined(_DEBUG)
     fprintf(stderr, "number of devices found: %i\n", ndevices);
 #endif
   }
   else {
-#if defined(_DEBUG)
-    fprintf(stderr, "Error: no device found!\n");
+    fprintf(stderr, "No ACC-device found!\n");
+#if !defined(__CUDA)
+    CHECK(libsmm_acc_finalize(), NULL);
 #endif
     CHECK(c_dbcsr_acc_finalize(), NULL);
     return result;
@@ -210,6 +213,9 @@ int main(int argc, char* argv[])
   CHECK(c_dbcsr_acc_dev_mem_deallocate(stack_dev), NULL);
   CHECK(c_dbcsr_acc_dev_mem_deallocate(mat_dev), NULL);
   CHECK(c_dbcsr_acc_stream_destroy(stream), NULL);
+#if !defined(__CUDA)
+  CHECK(libsmm_acc_finalize(), NULL);
+#endif
   CHECK(c_dbcsr_acc_finalize(), NULL);
   if (EXIT_SUCCESS != result) {
     fprintf(stderr, "FAILED\n");

diff --git a/src/acc/cuda/Makefile b/src/acc/cuda/Makefile
@@ -1,6 +1,6 @@
 INCACC := $(wildcard *.h*) ../acc.h
 SRCACC := $(wildcard *.cpp)
-OBJACC := $(SRCACC:.cpp=.o) acc_cublas.o
+OBJACC := $(SRCACC:.cpp=.o)
 
 GPUSMM := $(wildcard ../libsmm_acc/kernels/*.h*)
 INCSMM := $(wildcard ../libsmm_acc/*.h*) ../acc_libsmm.h \
@@ -130,10 +130,7 @@ test: ../dbcsr_acc_test
 ../libsmm_acc/smm_acc_kernels.h: $(GPUSMM) Makefile ../libsmm_acc/generate_kernels.py ../libsmm_acc/parameters/parameters_$(WITH_GPU).json
 	@cd ../libsmm_acc && $(PYTHON) ../libsmm_acc/generate_kernels.py ../libsmm_acc/kernels
 
-acc_cublas.o: acc_cublas.cu Makefile
-	$(NVCC) $(addprefix -Xcompiler $(NULL),$(CXXFLAGS)) -c $< -o $@
-
-../dbcsr_acc.a: $(OBJACC) acc_cublas.o ../libsmm_acc/libsmm_acc_init.o
+../dbcsr_acc.a: $(OBJACC) ../libsmm_acc/libsmm_acc_init.o
 	$(AR) -rs $@ $^
 
 ../dbcsr_acc_smm.a: $(OBJSMM)
@@ -153,7 +150,7 @@ acc_bench_trans.o: ../acc_bench_trans.c Makefile
 	$(CXX) $^ $(LDFLAGS) -o $@
 
 dbcsr_acc_test.o: ../../../tests/dbcsr_acc_test.c Makefile
-	$(CC) $(CFLAGS) -c $< -o $@
+	$(CC) $(CFLAGS) -I../.. -c $< -o $@
 ../dbcsr_acc_test: dbcsr_acc_test.o ../dbcsr_acc_smm.a ../dbcsr_acc.a
 	$(CXX) $^ $(LDFLAGS) -o $@
 

diff --git a/src/acc/opencl/README.md b/src/acc/opencl/README.md
@@ -8,7 +8,7 @@ The OpenCL backend implements the [ACC interface](https://github.com/cp2k/dbcsr/
 
 ### Compile-time Settings
 
-Compile-time settings are (implicitly) documented and can be adjusted by editing [acc_opencl.h](https://github.com/cp2k/dbcsr/blob/develop/src/acc/opencl/acc_opencl.h) (adjusting the build-line as per `-D` is possible as well but less convenient). For example, `ACC_OPENCL_STREAM_PRIORITIES` is enabled by default (and further confirmed at runtime/build-time) but can be disabled, or `ACC_OPENCL_VERBOSE` (which is disabled by default) can be enabled for debug purpose. More sensitive/private compile-time settings may be available within particular translation units like in `acc_opencl_mem.c`.
+Compile-time settings are (implicitly) documented and can be adjusted by editing [acc_opencl.h](https://github.com/cp2k/dbcsr/blob/develop/src/acc/opencl/acc_opencl.h) (adjusting the build-line as per `-D` is possible as well but less convenient). For example, `ACC_OPENCL_STREAM_PRIORITIES` is enabled by default (and further confirmed at runtime/build-time) but can be disabled, or `ACC_OPENCL_DEBUG` (which is disabled by default) can be enabled for debug purpose. More sensitive/private compile-time settings may be available within particular translation units like in `acc_opencl_mem.c`.
 
 An application of compile-time settings (and perhaps a valuable contribution) might be to call a GPU library in OpenCL-based LIBSMM. In such case, Shared Virtual Memory support (SVM) in OpenCL comes handy and can be enabled per `ACC_OPENCL_SVM`. The latter allows then to simply take the raw pointer out of an `cl_mem` object, and pass it into such library/function (which in turn can work across language borders, etc.).
 
@@ -19,6 +19,9 @@ Runtime settings are made by the means of environment variables (implemented in
 * `ACC_OPENCL_VENDOR`: character string matching the vendor of the OpenCL device in an case-insensitive fashion, e.g., "intel".
 * `ACC_OPENCL_DEVTYPE`: character string matching the device-kind like "cpu", "gpu", or another kind if neither CPU or GPU.
 * `ACC_OPENCL_DEVICE`: non-negative integer number to select a device from the (internally enumerated) list of devices.
+* `ACC_OPENCL_VERBOSE`: verbosity level (integer).
+    * `ACC_OPENCL_VERBOSE=1`: outputs (stderr) the number of devices found and the name of the selected device.
+    * `ACC_OPENCL_VERBOSE=2`: outputs (stderr) the duration needed to generate a requested kernel.
 
 The OpenCL backend enumerates and orders devices primarily by device-kind (GPU, CPU, and others in that order) and by memory capacity (secondary criterion). Device IDs are zero-based as per ACC interface (and less than what is permitted/returned by `acc_get_ndevices`).
 

diff --git a/src/acc/opencl/acc_opencl.c b/src/acc/opencl/acc_opencl.c
@@ -131,7 +131,11 @@ int c_dbcsr_acc_init(void)
 {
 #if defined(_OPENMP)
   /* initialization/finalization is not meant to be thread-safe */
-  int result = (0 == omp_in_parallel() ? EXIT_SUCCESS : EXIT_FAILURE);
+  int result = ((0 == omp_in_parallel()
+# if /*WORKAROUND*/defined(__DBCSR_ACC)
+    || 0/*master*/ == omp_get_thread_num()
+# endif
+    ) ? EXIT_SUCCESS : EXIT_FAILURE);
 #else
   int result = EXIT_SUCCESS;
 #endif
@@ -177,7 +181,6 @@ int c_dbcsr_acc_init(void)
       if (device_id < acc_opencl_ndevices) {
         if (NULL != env_device_vendor && '\0' != *env_device_vendor) {
           for (i = 0; i < (cl_uint)acc_opencl_ndevices;) {
-            buffer[0] = '\0';
             if (CL_SUCCESS == clGetDeviceInfo(acc_opencl_devices[i],
               CL_DEVICE_VENDOR, ACC_OPENCL_BUFFERSIZE, buffer, NULL))
             {
@@ -216,7 +219,9 @@ int c_dbcsr_acc_init(void)
           }
         }
         if (EXIT_SUCCESS == result) {
+          const char *const env_verbose = getenv("ACC_OPENCL_VERBOSE");
           cl_device_id active_device;
+          acc_opencl_options.verbosity = (NULL == env_verbose ? 0 : atoi(env_verbose));
           result = c_dbcsr_acc_opencl_set_active_device(device_id, &active_device);
 #if defined(_OPENMP) && defined(ACC_OPENCL_THREADLOCAL_CONTEXT)
           if (EXIT_SUCCESS == result) {
@@ -284,7 +289,11 @@ int c_dbcsr_acc_finalize(void)
 {
 #if defined(_OPENMP)
   /* initialization/finalization is not meant to be thread-safe */
-  int result = (0 == omp_in_parallel() ? EXIT_SUCCESS : EXIT_FAILURE);
+  int result = ((0 == omp_in_parallel()
+# if /*WORKAROUND*/defined(__DBCSR_ACC)
+    || 0/*master*/ == omp_get_thread_num()
+# endif
+    ) ? EXIT_SUCCESS : EXIT_FAILURE);
 #else
   int result = EXIT_SUCCESS;
 #endif
@@ -325,7 +334,6 @@ void c_dbcsr_acc_clear_errors(void)
 int c_dbcsr_acc_get_ndevices(int* ndevices)
 {
   int result;
-
 #if defined(__DBCSR_ACC)
   /* DBCSR calls acc_get_ndevices before calling acc_init(). */
   result = c_dbcsr_acc_init();
@@ -375,7 +383,6 @@ int c_dbcsr_acc_opencl_device_vendor(cl_device_id device, const char* vendor)
   char buffer[ACC_OPENCL_BUFFERSIZE];
   int result = EXIT_SUCCESS;
   assert(NULL != device && NULL != vendor);
-  buffer[0] = '\0';
   ACC_OPENCL_CHECK(clGetDeviceInfo(device,
     CL_DEVICE_VENDOR, ACC_OPENCL_BUFFERSIZE, buffer, NULL),
     "retrieve device vendor", result);
@@ -477,8 +484,20 @@ int c_dbcsr_acc_opencl_set_active_device(int device_id, cl_device_id* device)
         ACC_OPENCL_CHECK(result, "create context", result);
       }
     }
-    if (NULL != device) {
-      *device = (EXIT_SUCCESS == result ? active_id : NULL);
+    if (EXIT_SUCCESS == result) {
+      if (NULL != device) *device = active_id;
+      if (0 != acc_opencl_options.verbosity) {
+        char buffer[ACC_OPENCL_BUFFERSIZE];
+        if (CL_SUCCESS == clGetDeviceInfo(active_id,
+          CL_DEVICE_NAME, ACC_OPENCL_BUFFERSIZE, buffer, NULL))
+        {
+          fprintf(stderr, "INFO ACC/OpenCL: ndevices=%i device%i=\"%s\"\n",
+            acc_opencl_ndevices, device_id, buffer);
+        }
+      }
+    }
+    else {
+      if (NULL != device) *device = NULL;
     }
   }
   ACC_OPENCL_RETURN(result);
@@ -546,7 +565,7 @@ int c_dbcsr_acc_opencl_wgsize(cl_device_id device, cl_kernel kernel,
 int c_dbcsr_acc_opencl_kernel(const char* source, const char* build_options,
   const char* kernel_name, cl_kernel* kernel)
 {
-  char buffer[ACC_OPENCL_BUFFERSIZE] = "\0";
+  char buffer[ACC_OPENCL_BUFFERSIZE] = "";
   cl_int result;
   assert(NULL != kernel);
   if (NULL != acc_opencl_context) {

diff --git a/src/acc/opencl/acc_opencl.h b/src/acc/opencl/acc_opencl.h
@@ -86,8 +86,8 @@
 #if !defined(ACC_OPENCL_MEM_ASYNC) && 1
 # define ACC_OPENCL_MEM_ASYNC
 #endif
-#if !defined(ACC_OPENCL_VERBOSE) && 0
-# define ACC_OPENCL_VERBOSE
+#if !defined(ACC_OPENCL_DEBUG) && 0
+# define ACC_OPENCL_DEBUG
 #endif
 #if !defined(ACC_OPENCL_SVM) && 0
 # if defined(CL_VERSION_2_0)
@@ -189,9 +189,12 @@ extern "C" {
 
 /** Settings depending on OpenCL vendor or standard level (discovered/setup in acc_init). */
 typedef struct acc_opencl_options_t {
-  /** Asynchronous memory operations may crash for some OpenCL implementations. */
+  /** Asynchronous memory operations (may crash for some OpenCL implementations). */
   cl_bool async_memops;
+  /** Runtime SVM support (needs ACC_OPENCL_SVM at compile-time). */
   cl_bool svm_interop;
+  /** Runtime verbosity (output on stderr). */
+  cl_int verbosity;
 } acc_opencl_options_t;
 
 extern acc_opencl_options_t acc_opencl_options;

diff --git a/src/acc/opencl/acc_opencl_event.c b/src/acc/opencl/acc_opencl_event.c
@@ -107,7 +107,7 @@ int c_dbcsr_acc_event_query(void* event, acc_bool_t* has_occurred)
   }
   assert(NULL != has_occurred);
   *has_occurred = (CL_COMPLETE == status || 0 > status);
-#if defined(ACC_OPENCL_VERBOSE) && defined(_DEBUG)
+#if defined(ACC_OPENCL_DEBUG) && defined(_DEBUG)
   fprintf(stderr, "c_dbcsr_acc_event_query(%p, %i)\n", event, *has_occurred);
 #endif
   ACC_OPENCL_RETURN(result);
@@ -118,7 +118,7 @@ int c_dbcsr_acc_event_synchronize(void* event)
 { /* Waits on the host-side. */
   int result = EXIT_SUCCESS;
   assert(NULL != event);
-#if defined(ACC_OPENCL_VERBOSE) && defined(_DEBUG)
+#if defined(ACC_OPENCL_DEBUG) && defined(_DEBUG)
   fprintf(stderr, "c_dbcsr_acc_event_synchronize(%p)\n", event);
 #endif
   ACC_OPENCL_CHECK(clWaitForEvents(1, ACC_OPENCL_EVENT(event)),

diff --git a/src/acc/opencl/acc_opencl_stream.c b/src/acc/opencl/acc_opencl_stream.c
@@ -165,7 +165,7 @@ int c_dbcsr_acc_stream_sync(void* stream)
 { /* Blocks the host-thread. */
   int result = EXIT_SUCCESS;
   assert(NULL != stream);
-#if defined(ACC_OPENCL_VERBOSE) && defined(_DEBUG)
+#if defined(ACC_OPENCL_DEBUG) && defined(_DEBUG)
   fprintf(stderr, "c_dbcsr_acc_stream_sync(%p)\n", stream);
 #endif
   ACC_OPENCL_CHECK(clFinish(*ACC_OPENCL_STREAM(stream)),
@@ -178,7 +178,7 @@ int c_dbcsr_acc_stream_wait_event(void* stream, void* event)
 { /* Wait for an event (device-side). */
   int result = EXIT_SUCCESS;
   assert(NULL != stream && NULL != event);
-#if defined(ACC_OPENCL_VERBOSE) && defined(_DEBUG)
+#if defined(ACC_OPENCL_DEBUG) && defined(_DEBUG)
   fprintf(stderr, "c_dbcsr_acc_stream_wait_event(%p, %p)\n", stream, event);
 #endif
 #if defined(ACC_OPENCL_STREAM_SYNCFLUSH)

diff --git a/src/acc/opencl/smm/README.md b/src/acc/opencl/smm/README.md
@@ -14,7 +14,16 @@ The `OPENCL_LIBSMM_DEBUG` compile-time setting enables side-by-side validation o
 
 ### Runtime Settings
 
-Runtime settings are made by the means of environment variables (implemented in `opencl_libsmm.c`). There are two categories (for the two major functions) like matrix transpose (`OPENCL_LIBSMM_TRANS_*`) and matrix multiplication (`OPENCL_LIBSMM_SMM_*`). For tranposing matrices:
+Runtime settings are made by the means of environment variables (implemented in `opencl_libsmm.c`). There are two categories (for the two major functions) like matrix transpose (`OPENCL_LIBSMM_TRANS_*`) and matrix multiplication (`OPENCL_LIBSMM_SMM_*`). Common settings are (see OpenCL backend documentation for more details):
+
+* `ACC_OPENCL_VENDOR`: character string matching the vendor of the OpenCL device in an case-insensitive fashion, e.g., "intel".
+* `ACC_OPENCL_DEVTYPE`: character string matching the device-kind like "cpu", "gpu", or another kind if neither CPU or GPU.
+* `ACC_OPENCL_DEVICE`: non-negative integer number to select a device from the (internally enumerated) list of devices.
+* `ACC_OPENCL_VERBOSE`: verbosity level (integer).
+    * `ACC_OPENCL_VERBOSE=1`: outputs (stderr) the number of devices found and the name of the selected device.
+    * `ACC_OPENCL_VERBOSE=2`: outputs (stderr) the duration needed to generate a requested kernel.
+
+For tranposing matrices:
 
 * `OPENCL_LIBSMM_TRANS_BUILDOPTS`: character string with build options (compile and link) supplied to the OpenCL runtime compiler.
 * `OPENCL_LIBSMM_TRANS_INPLACE`: Boolean value (zero or non-zero integer) for inplace matrix transpose not relying on local memory.
@@ -28,13 +37,13 @@ For multiplying matrices:
 * `OPENCL_LIBSMM_SMM_BLOCK_M`: non-negative integer number (less/equal than the M-extent) denoting the blocksize in M-direction.
 * `OPENCL_LIBSMM_SMM_BLOCK_N`: non-negative integer number (less/equal than the N-extent) denoting the blocksize in N-direction.
 
-**NOTE**: above runtime settings may be non-smooth in the sense of enabling a distinct code-path depending on a specific value, e.g., `OPENCL_LIBSMM_SMM_BATCHSIZE=1`.
+**NOTE**: LIBSMM's tunable runtime settings may be non-smooth in the sense of enabling a distinct code-path depending on a specific value, e.g., `OPENCL_LIBSMM_SMM_BATCHSIZE=1` vs. `OPENCL_LIBSMM_SMM_BATCHSIZE=2`.
 
 ## Auto Tuning
 
 Auto tuning code for performance is a practical way to find the "best" setting for parameterized code (e.g., GPU kernels). Introducing effective parameters is a prerequisite, and exploring the (potentially) high-dimensional parameter space in an efficient way is an art. It is desirable to have reasonable defaults even without auto-tuning the parameters. It would be even better to avoid auto-tuning if best performance was possible right away, i.e., if auto-tuning is not able to find better settings.
 
-For the OpenCL based LIBSMM, `OPENCL_LIBSMM_SMM_BATCHSIZE`, `OPENCL_LIBSMM_SMM_BLOCK_M`, and `OPENCL_LIBSMM_SMM_BLOCK_N` are explored using [OpenTuner](http://opentuner.org/). The script [tune_multiply.py](https://github.com/cp2k/dbcsr/blob/develop/src/acc/opencl/smm/tune_multiply.py) leverages for instance the [acc_bench_smm](index.html) benchmark by parsing console output (timing, data type, etc.). This way, the tuning is implemented without being intermingled with subject being tuned. To build the benchmarks:
+For the OpenCL based LIBSMM, `OPENCL_LIBSMM_SMM_BATCHSIZE`, `OPENCL_LIBSMM_SMM_BLOCK_M`, and `OPENCL_LIBSMM_SMM_BLOCK_N` are explored using [OpenTuner](http://opentuner.org/). The script [tune_multiply.py](https://github.com/cp2k/dbcsr/blob/develop/src/acc/opencl/smm/tune_multiply.py) leverages for instance the `acc_bench_smm` benchmark by parsing console output (timing, data type, etc.). This way, the tuning is implemented without being intermingled with subject being tuned. To build the benchmarks:
 
 ```bash
 cd src/acc/opencl
@@ -66,7 +75,7 @@ The OpenTuner script implements multiple objectives ("cost"), primarily "accurac
 [    67s]    INFO opentuner.search.plugin.DisplayPlugin: tests=53, best {'BS': 48, 'BM': 8, 'BN': 1}, cost accuracy=32.20000000, size=1.0, found by UniformGreedyMutation
 ```
 
-The script finally writes a JSON-file with a filename like `tune_multiply-float-12x12x12-60gflops.json` which is encoding the benchmark (multiply), the precision (float), the kernel (12x12x12), and the achieved performance (60gflops). Tuninig starts from an internal default that is supposed to match LIBSMM's internal default parameters. However, tuning can be (re-)started with specific parameters (e.g., `-bs 64`, `-bm 13`, `-bn 1` for `OPENCL_LIBSMM_SMM_BATCHSIZE`, `OPENCL_LIBSMM_SMM_BLOCK_M`, and `OPENCL_LIBSMM_SMM_BLOCK_N` respectively).
+The script finally writes a JSON-file with a filename like `tune_multiply-float-12x12x12-60gflops.json` which is encoding the benchmark (multiply), the precision (float), the kernel (12x12x12), and the achieved performance (60gflops). The script handles SIGINT (like Ctrl-C), and output is still written despite of not terminating normally (can abused to tune interactively). Tuninig starts from an internal default that is supposed to match LIBSMM's internal default parameters. However, tuning can be (re-)started with specific parameters (e.g., `-bs 64`, `-bm 13`, `-bn 1` for `OPENCL_LIBSMM_SMM_BATCHSIZE`, `OPENCL_LIBSMM_SMM_BLOCK_M`, and `OPENCL_LIBSMM_SMM_BLOCK_N` respectively).
 
 ## Optimized Kernels
 
@@ -114,4 +123,4 @@ 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).
+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).