ocl: make backend less verbose at termination (#789)

ocl: make backend less verbose at termination * Made OpenCL common/shared code/definitions available. * Updated tuned parameters.
cp2k · May 13, 2024 · 954815d · 954815d
1 parent 317a3c4
commit 954815d
Show file tree

Hide file tree

Showing 4 changed files with 475 additions and 468 deletions.
diff --git a/src/acc/opencl/acc_opencl.c b/src/acc/opencl/acc_opencl.c
@@ -679,7 +679,7 @@ int c_dbcsr_acc_finalize(void) {
 #  endif
   assert(c_dbcsr_acc_opencl_config.ndevices < ACC_OPENCL_MAXNDEVS);
   if (0 != c_dbcsr_acc_opencl_config.ndevices && NULL != cleanup) {
-    if (0 != c_dbcsr_acc_opencl_config.verbosity) {
+    if (2 <= c_dbcsr_acc_opencl_config.verbosity || 0 > c_dbcsr_acc_opencl_config.verbosity) {
       int d;
       fprintf(stderr, "INFO ACC/OpenCL: pid=%u nthreads=%i", libxsmm_get_pid(), c_dbcsr_acc_opencl_config.nthreads);
       if (NULL != c_dbcsr_acc_opencl_config.device.context &&

diff --git a/src/acc/opencl/common/opencl_common.h b/src/acc/opencl/common/opencl_common.h
@@ -24,16 +24,6 @@
 #  define UNROLL_AUTO
 #endif
 
-#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)
 #  define UNROLL(N)
@@ -48,4 +38,21 @@
 #  define UNROLL(N) UNROLL_FORCE(N)
 #endif
 
+#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
+
+#define DIVUP(A, B) (((A) + (B) - 1) / (B))
+#define NUP(N, UP) (DIVUP(N, UP) * (UP))
+#define BLR(N, BN) (NUP(N, BN) - (N))
+
+#define IDX(I, J, M, N) ((int)(I) * (N) + (J))
+#define IDT(I, J, M, N) IDX(J, I, N, M)
+
 #endif /*OPENCL_COMMON_H*/
diff --git a/src/acc/opencl/smm/kernels/multiply.cl b/src/acc/opencl/smm/kernels/multiply.cl
@@ -9,13 +9,13 @@
 #include "../../common/opencl_atomics.h"
 
 #if !defined(AL) || (SM != SN) || (SM != BM) || (SN != SK) || (1 == BS)
-#  define ADX(M, K) adata[SM * K + M + a0] /* transposed */
-#  define BDX(K, N) bdata[SN * K + N + b0] /* linear */
-#  define CDX(M, N) cdata[SM * N + M + c0] /* transposed */
+#  define ADX(M, K) adata[IDT(M, K, SM, SK) + a0] /* transposed */
+#  define BDX(K, N) bdata[IDX(K, N, SK, SN) + b0] /* linear */
+#  define CDX(M, N) cdata[IDT(M, N, SM, SN) + c0] /* transposed */
 #else
-#  define ADX(M, K) adata[SK * M + K + a0] /* linear */
-#  define BDX(K, N) bdata[SK * N + K + b0] /* transposed */
-#  define CDX(M, N) cdata[SN * M + N + c0] /* linear */
+#  define ADX(M, K) adata[IDX(M, K, SM, SK) + a0] /* linear */
+#  define BDX(K, N) bdata[IDT(K, N, SK, SN) + b0] /* transposed */
+#  define CDX(M, N) cdata[IDX(M, N, SM, SN) + c0] /* linear */
 #endif
 
 #if defined(SLM_A)
@@ -63,8 +63,8 @@
 #  define REPEAT 1
 #endif
 
-#define NBM ((SM + BM - 1) / BM)
-#define NBN ((SN + BN - 1) / BN)
+#define NBM DIVUP(SM, BM)
+#define NBN DIVUP(SN, BN)
 #define WRK (NBM * NBN)
 
 #define UM (SM / BK)