ocl: updated CMake dependencies for OpenCL kernel

* Made NULL-queue a soft-error (c_dbcsr_acc_opencl_device_synchronize).
* Support c_dbcsr_acc_stream_sync with default-stream (NULL).
* Made default definition of MAD available.
* Fixed c_dbcsr_acc_stream_destroy.
* Updated tuned parameters.

src/acc/opencl/acc_opencl.c

@@ -1077,8 +1077,10 @@ int c_dbcsr_acc_opencl_device_synchronize(int thread_id) {
   for (; i < c_dbcsr_acc_opencl_config.nstreams; ++i) {
     void* const stream = streams[i];
     if (NULL != stream) {
-      result = c_dbcsr_acc_stream_sync(stream);
-      if (EXIT_SUCCESS != result) break;
+      if (NULL != *ACC_OPENCL_STREAM(stream)) { /* soft-error? */
+        result = c_dbcsr_acc_stream_sync(stream);
+        if (EXIT_SUCCESS != result) break;
+      }
     }
 #  if defined(ACC_OPENCL_STREAM_COMPACT)
     else break;

src/acc/opencl/acc_opencl_mem.c

@@ -126,11 +126,11 @@ int c_dbcsr_acc_host_mem_allocate(void** host_mem, size_t nbytes, void* stream)
     NULL == host_ptr ? CL_MEM_ALLOC_HOST_PTR : CL_MEM_USE_HOST_PTR, nbytes, host_ptr, &result);
   assert(CL_SUCCESS == result || NULL == memory);
   if (CL_SUCCESS == result) {
-    /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-      NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+    cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+    cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                     stream);
     void* const mapped = clEnqueueMapBuffer(
       queue, memory, CL_TRUE /*blocking*/, CL_MAP_READ | CL_MAP_WRITE, 0 /*offset*/, nbytes, 0, NULL, NULL, &result);
     assert(CL_SUCCESS == result || NULL == mapped);
@@ -182,11 +182,11 @@ int c_dbcsr_acc_host_mem_deallocate(void* host_mem, void* stream) {
     c_dbcsr_acc_opencl_info_hostptr_t* const meminfo = c_dbcsr_acc_opencl_info_hostptr(host_mem);
     if (NULL != meminfo->memory) {
       const c_dbcsr_acc_opencl_info_hostptr_t info = *meminfo; /* copy meminfo prior to unmap */
-      /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-        NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+      cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+      cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                       stream);
       int result_release;
       result = clEnqueueUnmapMemObject(queue, info.memory, info.mapped, 0, NULL, NULL);
 #  if defined(CL_VERSION_2_0)
@@ -395,11 +395,11 @@ int c_dbcsr_acc_memcpy_h2d(const void* host_mem, void* dev_mem, size_t nbytes, v
 #    endif
 #  endif
     {
-      /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-        NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+      cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+      cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                       stream);
       result = clEnqueueWriteBuffer(
         queue, buffer, 0 == (1 & c_dbcsr_acc_opencl_config.async), offset, nbytes, host_mem, 0, NULL, NULL);
 #  if defined(ACC_OPENCL_STREAM_NULL)
@@ -444,11 +444,11 @@ int c_dbcsr_acc_memcpy_d2h(const void* dev_mem, void* host_mem, size_t nbytes, v
 #    endif
 #  endif
     {
-      /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-        NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+      cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+      cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                       stream);
       result = clEnqueueReadBuffer(
         queue, buffer, 0 == (2 & c_dbcsr_acc_opencl_config.async), offset, nbytes, host_mem, 0, NULL, NULL);
       if (CL_SUCCESS == result) {
@@ -506,11 +506,11 @@ int c_dbcsr_acc_memcpy_d2d(const void* devmem_src, void* devmem_dst, size_t nbyt
 #    endif
 #  endif
     {
-      /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-        NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+      cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+      cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                       stream);
       if (0 == (2 & c_dbcsr_acc_opencl_config.devcopy)) {
         result = clEnqueueCopyBuffer(queue, src, dst, src_offset, dst_offset, nbytes, 0, NULL, NULL);
       }
@@ -579,11 +579,11 @@ int c_dbcsr_acc_opencl_memset(void* dev_mem, int value, size_t offset, size_t nb
 #    endif
 #  endif
     {
-      /*const*/ cl_command_queue queue = *ACC_OPENCL_STREAM(
 #  if defined(ACC_OPENCL_STREAM_NULL)
-        NULL == stream ? c_dbcsr_acc_opencl_stream_default() :
+      cl_command_queue queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+      cl_command_queue queue = *ACC_OPENCL_STREAM(stream);
 #  endif
-                       stream);
       if (0 == (1 & c_dbcsr_acc_opencl_config.devcopy)) {
         static LIBXSMM_TLS cl_long pattern = 0;
         size_t size_of_pattern = 1;

src/acc/opencl/acc_opencl_stream.c

@@ -248,13 +248,13 @@ int c_dbcsr_acc_stream_destroy(void* stream) {
         void** const streams = c_dbcsr_acc_opencl_config.streams + tid * c_dbcsr_acc_opencl_config.nstreams;
         for (i = 0; i < c_dbcsr_acc_opencl_config.nstreams; ++i) {
           if (stream == streams[i]) {
+            int k = i;
 #  if defined(ACC_OPENCL_STREAM_COMPACT)
-            const int j = i + 1, k = c_dbcsr_acc_opencl_config.nstreams - j;
+            const int j = i + 1;
             if (j < c_dbcsr_acc_opencl_config.nstreams && NULL != streams[j]) { /* compacting streams is not thread-safe */
+              k = c_dbcsr_acc_opencl_config.nstreams - j;
               memmove(streams + i, streams + j, sizeof(void*) * k);
             }
-#  else
-            const int k = i;
 #  endif
             streams[k] = NULL;
             tid = c_dbcsr_acc_opencl_config.nthreads; /* leave outer loop */
@@ -321,6 +321,7 @@ int c_dbcsr_acc_stream_priority_range(int* least, int* greatest) {
 
 
 int c_dbcsr_acc_stream_sync(void* stream) {
+  cl_command_queue queue = NULL;
   int result = EXIT_SUCCESS;
 #  if defined(ACC_OPENCL_STREAM_PRIORITIES)
   const int* const priority = NULL;
@@ -331,8 +332,12 @@ int c_dbcsr_acc_stream_sync(void* stream) {
   static const int routine_name_len = (int)sizeof(LIBXSMM_FUNCNAME) - 1;
   c_dbcsr_timeset((const char**)&routine_name_ptr, &routine_name_len, &routine_handle);
 #  endif
-  assert(NULL != stream);
-  result = clFinish(*ACC_OPENCL_STREAM(stream));
+#  if defined(ACC_OPENCL_STREAM_NULL)
+  queue = *ACC_OPENCL_STREAM(NULL != stream ? stream : c_dbcsr_acc_opencl_stream_default());
+#  else
+  queue = *ACC_OPENCL_STREAM(stream);
+#  endif
+  result = clFinish(queue);
 #  if defined(__DBCSR_ACC) && defined(ACC_OPENCL_PROFILE)
   c_dbcsr_timestop(&routine_handle);
 #  endif

src/acc/opencl/common/opencl_common.h

@@ -15,8 +15,15 @@
 #  define UNROLL_FORCE(N)
 #endif
 
-#define MIN(A, B) ((A) < (B) ? (A) : (B))
-#define MAX(A, B) ((A) < (B) ? (B) : (A))
+#if !defined(MIN)
+#  define MIN(A, B) ((A) < (B) ? (A) : (B))
+#endif
+#if !defined(MAX)
+#  define MAX(A, B) ((A) < (B) ? (B) : (A))
+#endif
+#if !defined(MAD)
+#  define MAD fma
+#endif
 
 #if !defined(LU) || (-1 == LU)
 #  define UNROLL_OUTER(N)

src/acc/opencl/smm/CMakeLists.txt

@@ -1,6 +1,7 @@
 set(LIBSMM_ACC_HEADER_KERNELS ${CMAKE_CURRENT_SOURCE_DIR}/opencl_kernels.h)
 
 set(SMM_ACC_KERNEL_SCRIPT ${CMAKE_CURRENT_SOURCE_DIR}/../acc_opencl.sh)
+set(SMM_ACC_COMMON ../common/opencl_atomics.h ../common/opencl_common.h)
 set(SMM_ACC_KERNELS kernels/multiply.cl kernels/transpose.cl)
 list(TRANSFORM SMM_ACC_KERNELS PREPEND ${CMAKE_CURRENT_SOURCE_DIR}/)
 
@@ -18,7 +19,7 @@ add_custom_target(
   parameters ALL
   COMMAND ${SMM_ACC_KERNEL_SCRIPT} ${SMM_ACC_KERNELS} ${SMM_ACC_PARAMS}
           ${LIBSMM_ACC_HEADER_KERNELS}
-  DEPENDS ${SMM_ACC_KERNEL_SCRIPT} ${SMM_ACC_KERNELS}
+  DEPENDS ${SMM_ACC_KERNEL_SCRIPT} ${SMM_ACC_KERNELS} ${SMM_ACC_COMMON}
   BYPRODUCTS ${LIBSMM_ACC_HEADER_KERNELS}
   COMMENT "ACC/LIBSMM OpenCL: collecting tuned kernel parameters...")
 

src/acc/opencl/smm/opencl_libsmm.c

@@ -1376,7 +1376,7 @@ int libsmm_acc_process(const int* host_param_stack, const int* dev_param_stack,
                 const int slm_c = (1 != new_config.ac ? 0 : (LIBXSMM_ISPOT(m_max * typesize) + 1));
                 /* compose build parameters and flags */
                 nchar = LIBXSMM_SNPRINTF(build_params, sizeof(build_params),
-                  "-DMAD=fma -DT=%s -DINTEL=%u -DGLOBAL=%s -DSWG=%i -DSGS=%i -DFN=%s -DREPEAT=%i -DLU=%i "
+                  "-DT=%s -DINTEL=%u -DGLOBAL=%s -DSWG=%i -DSGS=%i -DFN=%s -DREPEAT=%i -DLU=%i "
                   "-DSM=%i -DSN=%i -DSK=%i -DBS=%i -DVL=%i %s -DBM=%i -DBN=%i -DBK=%i "
                   "%s %s %s %s %s %s %s %s ", /* space! */
                   tname, 0 != devinfo->intel ? devinfo->uid : 0, cmem, (int)new_config.wgsize[kernel_idx], (int)sgs, fname,