Merge branch 'release-2.3.0-rc2'

.github/workflows/release.yml

@@ -27,6 +27,10 @@ jobs:
           -DMPI_EXECUTABLE_SUFFIX=.mpich \
           ..
 
+    - name: Configure git to trust the workspace despite the different owner
+      run:
+        git config --global --add safe.directory "$GITHUB_WORKSPACE"
+
     - name: Build Release Asset
       run: cmake --build build -- dist
 

VERSION

@@ -1,8 +1,8 @@
 MAJOR = 2
 MINOR = 3
-PATCH = 0-rc1
+PATCH = 0-rc2
 # A specific DATE (YYYY-MM-DD) fixes an official release, otherwise
 # it is considered Development version.
-DATE  = 2022-05-03
+DATE  =
 
 

cmake/CompilerConfiguration.cmake

@@ -43,9 +43,9 @@ Please open an issue at https://github.com/cp2k/dbcsr/issues with the reported c
 endif ()
 
 if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
-  set(CMAKE_CXX_FLAGS_RELEASE      "-O3 -g -funroll-loops -Wall -Werror")
-  set(CMAKE_CXX_FLAGS_COVERAGE     "-O0 -g --coverage -Wall -Werror")
-  set(CMAKE_CXX_FLAGS_DEBUG        "-O2 -ggdb -Wall -Werror -fsanitize=undefined -fsanitize=address -fsanitize-recover=all")
+  set(CMAKE_CXX_FLAGS_RELEASE      "-O3 -g -funroll-loops -Wall -Wextra -Werror")
+  set(CMAKE_CXX_FLAGS_COVERAGE     "-O0 -g --coverage -Wall -Wextra -Werror")
+  set(CMAKE_CXX_FLAGS_DEBUG        "-O2 -ggdb -Wall -Wextra -Werror -fsanitize=undefined -fsanitize=address -fsanitize-recover=all")
   if ((NOT (USE_MPI)) OR (NOT ("${MPI_Fortran_LIBRARY_VERSION_STRING}" MATCHES "Open MPI")))
     set(CMAKE_CXX_FLAGS_COVERAGE "${CMAKE_CXX_FLAGS_COVERAGE} -fsanitize=leak")
     set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -fsanitize=leak")

docs/CMakeLists.txt

@@ -14,9 +14,14 @@ add_custom_target(
   COMMENT "Generating API documentation and doc pages"
   COMMAND "${FORD_EXE}" "${FORD_PROJECT_FILE}"
   VERBATIM)
-add_dependencies(doc doc_copy_tests)
+
+if (BUILD_TESTING)
+  add_dependencies(doc doc_copy_tests)
+endif ()
+
 if (WITH_C_API AND WITH_EXAMPLES)
   add_dependencies(doc doc_copy_examples)
 endif ()
+
 add_dependencies(doc fypp) # only depend on the fypp step to avoid building
                            # everything just for the docs

src/acc/acc.h

@@ -16,6 +16,8 @@
 #define DBCSR_CONCATENATE2(A, B) A##B
 #define DBCSR_CONCATENATE(A, B) DBCSR_CONCATENATE2(A, B)
 
+/** used to mark variables used */
+#define DBCSR_MARK_USED(x) (void)(x)
 
 #if defined(__cplusplus)
 extern "C" {

src/acc/acc_bench_smm.c

@@ -14,16 +14,26 @@
 
 #if defined(__LIBXSMM)
 #  include <libxsmm.h>
+#  include <libxsmm_sync.h>
+#  define USE_LIBXSMM
 #  if !defined(LIBXSMM_VERSION_NUMBER)
 #    define LIBXSMM_VERSION_NUMBER \
       LIBXSMM_VERSION4(LIBXSMM_VERSION_MAJOR, LIBXSMM_VERSION_MINOR, LIBXSMM_VERSION_UPDATE, LIBXSMM_VERSION_PATCH)
 #  endif
-#  define USE_LIBXSMM
 #  if defined(_OPENMP)
 #    define ACC_BENCH_USEOMP(FUNC) LIBXSMM_USEOMP(FUNC)
 #  else
 #    define ACC_BENCH_USEOMP(FUNC) (FUNC)
 #  endif
+#  define PRINTF(...) \
+    do { \
+      const size_t print_buffer_size = sizeof(print_buffer) - print_offset; \
+      const int print_buffer_result = LIBXSMM_SNPRINTF(print_buffer + print_offset, print_buffer_size, __VA_ARGS__); \
+      assert(0 <= print_buffer_result && print_buffer_result < (int)print_buffer_size); \
+      print_offset += print_buffer_result; \
+    } while (0)
+#else
+#  define PRINTF(...) printf(__VA_ARGS__)
 #endif
 
 #if !defined(ELEM_TYPE)
@@ -81,17 +91,15 @@ static void parse_params(int argc, char* argv[], FILE** file, const char** snr,
   if (NULL == *file) {
     for (i = 0; i < argc; ++i) args[i] = argv[i + 1];
   }
-  else {
+  else { /* input file is specified */
     argc = 0;
     if (NULL != fgets(buffer, sizeof(buffer), *file)) {
       char* arg = strtok(buffer, DELIMS);
       while (NULL == arg && NULL != fgets(buffer, sizeof(buffer), *file)) {
         arg = strtok(buffer, DELIMS);
       }
       for (; NULL != arg; arg = strtok(NULL, DELIMS)) {
-        if (argc * sizeof(*args) < sizeof(args)) {
-          args[argc++] = arg;
-        }
+        if (argc * sizeof(*args) < sizeof(args)) args[argc++] = arg;
         else { /* malformed command-line */
           fclose(*file);
           *file = NULL;
@@ -237,6 +245,8 @@ int main(int argc, char* argv[]) {
     void* stream = NULL;
 #if defined(USE_LIBXSMM)
     libxsmm_timer_tickint start;
+    int print_offset = 0;
+    char print_buffer[1024];
 #  if defined(__OPENCL)
     const char* const env_smm_repeat = getenv("SMM_NREPEAT");
     const int smm_nrepeat = (NULL == env_smm_repeat ? 1 : MAX(atoi(env_smm_repeat), 1));
@@ -293,8 +303,8 @@ int main(int argc, char* argv[]) {
     assert(m <= (mn / n) && 0 == (mn % n));
     assert(m <= (mk / k) && 0 == (mk % k));
     assert(k <= (kn / n) && 0 == (kn % n));
-    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);
-    printf("typename (id=%i): %s\n", DBCSR_TYPE(ELEM_TYPE), DBCSR_STRINGIFY(ELEM_TYPE));
+    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);
+    PRINTF("typename (id=%i): %s\n", DBCSR_TYPE(ELEM_TYPE), DBCSR_STRINGIFY(ELEM_TYPE));
     if (MAX_KERNEL_DIM < m || MAX_KERNEL_DIM < n || MAX_KERNEL_DIM < k) {
       fprintf(stderr, "Matrix shape exceeds MAX_KERNEL_DIM!\n");
       result = EXIT_FAILURE;
@@ -344,7 +354,7 @@ int main(int argc, char* argv[]) {
     if (NULL != amat_hst && NULL != bmat_hst && NULL != stack_hst) {
       const size_t size = sizeof(ELEM_TYPE) * (mk * na + kn * nb) + sizeof(int) * 3 * stack_size;
       duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
-      printf("copy-in (%i MB): %.2g ms %.1f GB/s\n", (int)((size + (1 << 19)) >> 20), 1000.0 * duration,
+      PRINTF("copy-in (%i MB): %.2g ms %.1f GB/s\n", (int)((size + (1 << 19)) >> 20), 1000.0 * duration,
         size / (duration * (1ULL << 30)));
     }
 #endif
@@ -370,8 +380,8 @@ int main(int argc, char* argv[]) {
 #endif
     /* warmup execution and prebuild SMM-kernel */
     for (r = 0; r < warmup; ++r) {
-      CHECK(libsmm_acc_process(stack_hst, stack_dev, stack_size, 3 /*nparams*/, DBCSR_TYPE(ELEM_TYPE), amat_dev, bmat_dev, cmat_dev,
-              m, n, k, MAX_KERNEL_DIM, 1 /*homogeneous*/, stream, stream),
+      CHECK(libsmm_acc_process(stack_hst, stack_dev, stack_size, DBCSR_TYPE(ELEM_TYPE), amat_dev, bmat_dev, cmat_dev, m, n, k,
+              MAX_KERNEL_DIM, 1 /*homogeneous*/, stream, stream),
         &result);
     }
     CHECK(c_dbcsr_acc_memset_zero(cmat_dev, 0 /*offset*/, sizeof(ELEM_TYPE) * mn * nc, stream), &result);
@@ -381,19 +391,19 @@ int main(int argc, char* argv[]) {
 #endif
     for (r = 0; r < nrepeat; ++r) {
       /* GPU-kernel is limited to C += Ai * Bi^T, i.e., NT (for NN, all Bi must be transposed upfront) */
-      CHECK(libsmm_acc_process(stack_hst, stack_dev, stack_size, 3 /*nparams*/, DBCSR_TYPE(ELEM_TYPE), amat_dev, bmat_dev, cmat_dev,
-              m, n, k, MAX_KERNEL_DIM, 1 /*homogeneous*/, stream, stream),
+      CHECK(libsmm_acc_process(stack_hst, stack_dev, stack_size, DBCSR_TYPE(ELEM_TYPE), amat_dev, bmat_dev, cmat_dev, m, n, k,
+              MAX_KERNEL_DIM, 1 /*homogeneous*/, stream, stream),
         &result);
     }
 #if defined(USE_LIBXSMM)
     CHECK(c_dbcsr_acc_stream_sync(stream), &result);
     duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
     if (EXIT_SUCCESS == result) {
 #  if defined(TRANSPOSE)
-      printf("transpose: %.2g ms %.1f GFLOPS/s\n", 1000.0 * (duration + transpose) / (nrepeat * smm_nrepeat),
+      PRINTF("transpose: %.2g ms %.1f GFLOPS/s\n", 1000.0 * (duration + transpose) / (nrepeat * smm_nrepeat),
         1E-9 * ((size_t)2 * m * n * k * stack_size * nrepeat * smm_nrepeat) / (duration + transpose));
 #  endif
-      printf("device: %.2g ms %.1f GFLOPS/s\n", 1000.0 * duration / (nrepeat * smm_nrepeat),
+      PRINTF("device: %.2g ms %.1f GFLOPS/s\n", 1000.0 * duration / (nrepeat * smm_nrepeat),
         1E-9 * ((size_t)2 * m * n * k * stack_size * nrepeat * smm_nrepeat) / duration);
     }
 #  if defined(VALIDATE)
@@ -425,7 +435,7 @@ int main(int argc, char* argv[]) {
             stack_hst + 2, stack_size);
         }
         duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
-        printf("host: %.2g ms %.1f GFLOPS/s\n", 1000.0 * duration / (nrepeat * smm_nrepeat),
+        PRINTF("host: %.2g ms %.1f GFLOPS/s\n", 1000.0 * duration / (nrepeat * smm_nrepeat),
           1E-9 * ((size_t)2 * m * n * k * stack_size * nrepeat * smm_nrepeat) / duration);
         /* validate correctness in case of successful result code/status */
         if (EXIT_SUCCESS == result) {
@@ -439,18 +449,18 @@ int main(int argc, char* argv[]) {
             result = libxsmm_matdiff(&diff, LIBXSMM_DATATYPE(ELEM_TYPE), mn, nc, gold_hst, cmat_hst, &mn, &mn);
             if (EXIT_SUCCESS == result) {
               const double relerror = 1.0 - diff.rsq;
-              printf("rel.error: %g", relerror);
+              PRINTF("rel.error: %g", relerror);
               if (maxerror < relerror && NULL != file) maxerror = relerror;
               if (0 < relerror) {
                 if (LIBXSMM_NOTNAN(diff.v_tst)) {
-                  printf(" (%g != %g)\n", diff.v_ref, diff.v_tst);
+                  PRINTF(" (%g != %g)\n", diff.v_ref, diff.v_tst);
                 }
                 else {
-                  printf(" (%g)\n", diff.v_tst);
+                  PRINTF(" (%g)\n", diff.v_tst);
                 }
               }
               else {
-                printf("\n");
+                PRINTF("\n");
               }
               if (0 < check && check < relerror) result = EXIT_FAILURE;
             }
@@ -474,13 +484,16 @@ int main(int argc, char* argv[]) {
     CHECK(c_dbcsr_acc_dev_mem_deallocate(cmat_dev), NULL);
     CHECK(c_dbcsr_acc_stream_destroy(stream), NULL);
     if (EXIT_SUCCESS == result) {
-      ++nok;
       parse_params(argc, argv, &file, &snr, &sss, &ssm, &ssn, &ssk, &snc, &sna, &snb);
-      if (NULL != file) {
-        printf("\n");
-      }
-      else break;
+      if (NULL != file) PRINTF("\n");
+      ++nok;
     }
+#if defined(USE_LIBXSMM)
+    LIBXSMM_STDIO_ACQUIRE();
+    fputs(print_buffer, stdout);
+    LIBXSMM_STDIO_RELEASE();
+#endif
+    if (EXIT_SUCCESS == result && NULL == file) break;
   }
   free(rnd); /* release array of random numbers */
 #if !defined(__CUDA)

src/acc/acc_bench_smm.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+####################################################################################################
+# Copyright (C) by the DBCSR developers group - All rights reserved                                #
+# This file is part of the DBCSR library.                                                          #
+#                                                                                                  #
+# For information on the license, see the LICENSE file.                                            #
+# For further information please visit https://dbcsr.cp2k.org                                      #
+# SPDX-License-Identifier: GPL-2.0+                                                                #
+####################################################################################################
+
+HERE=$(cd "$(dirname "$0")" && pwd -P)
+SED=$(command -v gsed)
+
+# GNU sed is desired (macOS)
+if [ ! "${SED}" ]; then
+  SED=$(command -v sed)
+fi
+
+if [ "${SED}" ] && [ -x "${HERE}/acc_bench_smm" ]; then
+  LIBXSMM_PEXEC=${LIBXSMM_ROOT:-${HOME}/libxsmm}/scripts/tool_pexec.sh
+  if [ -x "${LIBXSMM_PEXEC}" ] && [ -e "$1" ]; then
+    NDEVICES=$(ACC_OPENCL_VERBOSE=1 CHECK=0 "${HERE}/acc_bench_smm" 1 1 1 2>&1 >/dev/null \
+             | ${SED} -n "s/INFO ACC\/OpenCL: ndevices=\([0-9][0-9]*\) ..*$/\1/p")
+  fi
+  if [ "${NDEVICES}" ] && [ "0" != "$((1<NDEVICES))" ]; then
+    NLINES=0
+    while read -r LINE; do
+      echo "ACC_OPENCL_DEVICE=$((NLINES%NDEVICES)) ${HERE}/acc_bench_smm ${LINE}"
+      NLINES=$((NLINES+1))
+    done <"$1" | ${LIBXSMM_PEXEC} "${NDEVICES}"
+  else
+    "${HERE}/acc_bench_smm" "$*"
+  fi
+else
+  >&2 echo "ERROR: missing prerequisites!"
+  exit 1
+fi

src/acc/acc_libsmm.h

@@ -42,9 +42,9 @@ c_dbcsr_acc_bool_t libsmm_acc_is_thread_safe(void);
 int libsmm_acc_transpose(const int* dev_trs_stack, int offset, int stack_size, void* dev_data, libsmm_acc_data_t datatype, int m,
   int n, int max_kernel_dim, void* stream);
 
-int libsmm_acc_process(const int* host_param_stack, const int* dev_param_stack, int stack_size, int nparams,
-  libsmm_acc_data_t datatype, const void* dev_a_data, const void* dev_b_data, void* dev_c_data, int m_max, int n_max, int k_max,
-  int max_kernel_dim, c_dbcsr_acc_bool_t def_mnk, void* stack_stream, void* c_stream);
+int libsmm_acc_process(const int* host_param_stack, const int* dev_param_stack, int stack_size, libsmm_acc_data_t datatype,
+  const void* dev_a_data, const void* dev_b_data, void* dev_c_data, int m_max, int n_max, int k_max, int max_kernel_dim,
+  c_dbcsr_acc_bool_t def_mnk, void* stack_stream, void* c_stream);
 
 static const char libsmm_acc_transpose_routine_name_str[] = "jit_kernel_transpose";
 static const char* const libsmm_acc_transpose_routine_name_ptr = libsmm_acc_transpose_routine_name_str;

src/acc/acc_triplets.sh

@@ -14,7 +14,7 @@ SED=$(command -v gsed)
 CUT=$(command -v cut)
 
 # GNU sed is desired (macOS)
-if [ "" = "${SED}" ]; then
+if [ ! "${SED}" ]; then
   SED=$(command -v sed)
 fi
 

src/acc/cuda/Makefile

@@ -232,9 +232,13 @@ $(DIRSMM)/parameters.h: $(MAKDIR)/Makefile $(DIRSMM)/generate_parameters.py $(PA
 $(DIRSMM)/smm_acc_kernels.h: $(GPUSMM) $(MAKDIR)/Makefile $(DIRSMM)/generate_kernels.py $(PARAMS)
 	@cd $(DIRSMM) && $(PYTHON) ../libsmm_acc/generate_kernels.py ../libsmm_acc/kernels
 
+.PHONY: backend
+backend: $(ACCDIR)/dbcsr_acc.a
 $(ACCDIR)/dbcsr_acc.a: $(OBJACC) $(DIRSMM)/libsmm_acc_init.o
 	$(AR) -rs $@ $^
 
+.PHONY: libsmm
+libsmm: $(ACCDIR)/dbcsr_acc_smm.a
 $(ACCDIR)/dbcsr_acc_smm.a: $(OBJSMM)
 	$(AR) -rs $@ $^
 

src/acc/cuda_hip/acc_mem.cpp

@@ -48,6 +48,8 @@ extern "C" int c_dbcsr_acc_dev_mem_deallocate(void* dev_mem) {
 extern "C" int c_dbcsr_acc_host_mem_allocate(void** host_mem, size_t n, void* stream) {
   unsigned int flag = ACC(HostAllocDefault);
 
+  DBCSR_MARK_USED(stream);
+
   ACC_API_CALL(HostAlloc, ((void**)host_mem, (size_t)n, flag));
   if (host_mem == NULL) return -2;
   if (verbose_print) printf("Allocating %zd bytes of host pinned memory at %p\n", n, *host_mem);
@@ -58,6 +60,8 @@ extern "C" int c_dbcsr_acc_host_mem_allocate(void** host_mem, size_t n, void* st
 
 /****************************************************************************/
 extern "C" int c_dbcsr_acc_host_mem_deallocate(void* host_mem, void* stream) {
+  DBCSR_MARK_USED(stream);
+
   if (verbose_print) printf("Host pinned deallocation address %p\n", host_mem);
   ACC_API_CALL(FreeHost, ((void*)host_mem));
 

src/acc/cuda_hip/acc_stream.cpp

@@ -57,6 +57,8 @@ extern "C" int c_dbcsr_acc_stream_create(void** stream_p, const char* name, int
 
 #if defined(__CUDA_PROFILING)
   nvtxNameCudaStreamA(*acc_stream, name);
+#else
+  DBCSR_MARK_USED(name);
 #endif
 
   return 0;

src/acc/libsmm_acc/libsmm_acc.cpp

@@ -224,7 +224,7 @@ kernel_map_iterator add_kernel_handle_to_jitted_kernels(
 
 //===========================================================================
 int libsmm_acc_process_blas(const int* param_stack, int stack_size, ACC_DRV(stream) stream, int m, int n, int k, int max_kernel_dim,
-  const double* a_data, const double* b_data, double* c_data, std::vector<ACC_BLAS(Handle_t) *> handles = acc_blashandles) {
+  const double* a_data, const double* b_data, double* c_data) {
 #if defined _OPENMP
   int ithread = omp_get_thread_num();
 #else
@@ -287,9 +287,9 @@ int libsmm_acc_process_d(const int* param_stack, int stack_size, ACC_DRV(stream)
 }
 
 //===========================================================================
-int libsmm_acc_process(const int* param_stack_host, const int* param_stack_dev, int stack_size, int nparams,
-  libsmm_acc_data_t datatype, const void* a_data, const void* b_data, void* c_data, int m, int n, int k, int max_kernel_dim,
-  int def_mnk, void* stack_stream, void* c_stream) {
+int libsmm_acc_process(const int* param_stack_host, const int* param_stack_dev, int stack_size, libsmm_acc_data_t datatype,
+  const void* a_data, const void* b_data, void* c_data, int m, int n, int k, int max_kernel_dim, int def_mnk, void* stack_stream,
+  void* c_stream) {
   if (def_mnk != 1) return -1; // inhomogeneous stacks not supported
   if (datatype == dbcsr_type_real_8) {
     if (m > max_kernel_dim || n > max_kernel_dim || k > max_kernel_dim)

src/acc/libsmm_acc/libsmm_acc.h

@@ -38,7 +38,7 @@ typedef std::unordered_map<Triplet, kernel_launcher>::iterator kernel_map_iterat
 static std::unordered_map<Triplet, kernel_launcher> kernel_handles;
 
 int libsmm_acc_process_blas(const int* param_stack_host, int stack_size, ACC_DRV(stream) stream, int m, int n, int k,
-  const double* a_data, const double* b_data, double* c_data, ACC_BLAS(Handle_t) * handle);
+  const double* a_data, const double* b_data, double* c_data);
 
 int libsmm_acc_process_d(const int* param_stack_dev, int stack_size, ACC_DRV(stream) stream, int m, int n, int k,
   const double* a_data, const double* b_data, double* c_data);