ocl: updated tuned parameters

- Disable code controlling Intel OpenCL kernel-cache (ACC_OPENCL_CACHEDIR).
- Disable kernel-cache when auto-tuning kernels (NEO_CACHE_PERSISTENT=0).
- Link OpenCL ICD-loader in a version-specific fashion (opencl/Makefile).
- Allow to explicitly disabled checks (in order to tune larger kernels).
- Improved MAXLU, and unrolling (1<LU).
- Updated LIBXSMM (Daint-CI).

.ci/daint.cscs.ch/ocl.build.sh

@@ -27,7 +27,7 @@ if [ ! -d "${HOME}/libxsmm" ]; then
 fi
 cd "${HOME}/libxsmm"
 git fetch
-git checkout ab992663b97571c9306ac01f6ca9a4a76bb5b559
+git checkout b5228c71ccdec143eb8461f40213d41676ea245a
 make -j
 cd ..
 

src/acc/opencl/Makefile

@@ -218,7 +218,7 @@ else
   ifneq (,$(wildcard $(OPENCL_LIB)))
     LDFLAGS += $(OPENCL_LIB)
   else
-    LDFLAGS += -lOpenCL
+    LDFLAGS += -l:libOpenCL.so.1
   endif
 endif
 

src/acc/opencl/acc_opencl.c

@@ -31,7 +31,7 @@
 #  if !defined(ACC_OPENCL_TEMPDIR) && 1
 #    define ACC_OPENCL_TEMPDIR "/tmp"
 #  endif
-#  if !defined(ACC_OPENCL_CACHEDIR) && 1
+#  if !defined(ACC_OPENCL_CACHEDIR) && 0
 #    define ACC_OPENCL_CACHEDIR ".cl_cache"
 #  endif
 #  if !defined(ACC_OPENCL_CPPBIN) && 1

src/acc/opencl/smm/kernels/multiply.cl

@@ -15,11 +15,11 @@
 #define MIN(A, B) ((A) < (B) ? (A) : (B))
 #define MAX(A, B) ((A) < (B) ? (B) : (A))
 
-#if !defined(LU) || (-1 == LU) || (1 < LU)
+#if !defined(LU) || (-1 == LU)
 #  define UNROLL_OUTER(N)
 #  define UNROLL(N)
 #else
-#  if (1 == LU)
+#  if (1 <= LU)
 #    define UNROLL_OUTER(N) UNROLL_FORCE(1)
 #  else
 #    define UNROLL_OUTER(N) UNROLL_FORCE(N)