Fix remaining min/max

control/context.c

@@ -301,7 +301,7 @@ int MORSE_Set(MORSE_enum param, int value)
                 morse_warning("MORSE_Set", "autotuning has been automatically disable\n");
             }
             /* Limit ib to nb */
-            morse->ib = min( morse->nb, morse->ib );
+            morse->ib = chameleon_min( morse->nb, morse->ib );
             break;
         case MORSE_INNER_BLOCK_SIZE:
             if (value <= 0) {

cudablas/include/cudablas.h

@@ -88,19 +88,6 @@
  **/
 #define cudablas_error(k, str) fprintf(stderr, "%s: Parameter %d / %s\n", __func__, k, str)
 
-/*******************************************************************************
- *  Global utilities
- **/
-#ifndef max
-#define max(a, b) ((a) > (b) ? (a) : (b))
-#endif
-#ifndef min
-#define min(a, b) ((a) < (b) ? (a) : (b))
-#endif
-#ifndef roundup
-#define roundup(a, b) (b <= 0) ? (a) : (((a) + (b)-1) & ~((b)-1))
-#endif
-
 /** ****************************************************************************
  *  LAPACK Constants
  **/

runtime/starpu/codelets/codelet_zgelqt.c

@@ -149,7 +149,7 @@ static void cl_zgelqt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda, &ldt, &h_work);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
     CORE_zgelqt(m, n, ib, A, lda, T, ldt, TAU, WORK);
 }
 #endif /* !defined(CHAMELEON_SIMULATION) */
@@ -182,7 +182,7 @@ static void cl_zgelqt_cuda_func(void *descr[], void *cl_arg)
     /* Gather pointer to scratch data on host */
     h_T   = h_A   + ib*n;
     h_TAU = h_T   + ib*ib;
-    h_W   = h_TAU + max(m,n);
+    h_W   = h_TAU + chameleon_max(m,n);
     h_D   = h_W   + ib*ib;
 
     stream = starpu_cuda_get_local_stream();

runtime/starpu/codelets/codelet_zgeqrt.c

@@ -150,7 +150,7 @@ static void cl_zgeqrt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda, &ldt, &h_work);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
     CORE_zgeqrt(m, n, ib, A, lda, T, ldt, TAU, WORK);
 }
 
@@ -182,7 +182,7 @@ static void cl_zgeqrt_cuda_func(void *descr[], void *cl_arg)
     /* Gather pointer to scratch data on host */
     h_T   = h_A   + m*ib;
     h_TAU = h_T   + ib*ib;
-    h_W   = h_TAU + max(m,n);
+    h_W   = h_TAU + chameleon_max(m,n);
     h_D   = h_W   + ib*ib;
 
     stream = starpu_cuda_get_local_stream();

runtime/starpu/codelets/codelet_zgetrf_incpiv.c

@@ -162,7 +162,7 @@ static void cl_zgetrf_incpiv_cpu_func(void *descr[], void *cl_arg)
 
             L += ib;
             for (i=0; i<n; i+=ib) {
-                sb = min( ib, n-i );
+                sb = chameleon_min( ib, n-i );
                 CORE_zlacpy(MorseUpperLower, sb, sb, A+(i*lda+i), lda, L+(i*ldl), ldl );
 
                 CORE_ztrtri( MorseLower, MorseUnit, sb, L+(i*ldl), ldl, &info );

runtime/starpu/codelets/codelet_ztslqt.c

@@ -170,7 +170,7 @@ static void cl_ztslqt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda1, &lda2, &ldt, &h_work);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
     CORE_ztslqt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
 }
 
@@ -200,7 +200,7 @@ static void cl_ztslqt_cuda_func(void *descr[], void *cl_arg)
     h_A2  = (cuDoubleComplex*)RUNTIME_starpu_ws_getlocal(h_work);
     h_T   = h_A2  + ib*n;
     h_TAU = h_T   + ib*n;
-    h_W   = h_TAU + max(m,n);
+    h_W   = h_TAU + chameleon_max(m,n);
     h_D   = h_W   + ib*m;
 
     stream = starpu_cuda_get_local_stream();

runtime/starpu/codelets/codelet_ztsqrt.c

@@ -160,7 +160,7 @@ static void cl_ztsqrt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda1, &lda2, &ldt, &h_work);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
     CORE_ztsqrt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
 }
 
@@ -190,7 +190,7 @@ static void cl_ztsqrt_cuda_func(void *descr[], void *cl_arg)
     h_A2  = (cuDoubleComplex*)RUNTIME_starpu_ws_getlocal(h_work);
     h_T   = h_A2  + m*ib;
     h_TAU = h_T   + ib*ib;
-    h_W   = h_TAU + max(m,n);
+    h_W   = h_TAU + chameleon_max(m,n);
     h_D   = h_W   + ib*n;
 
     stream = starpu_cuda_get_local_stream();

runtime/starpu/codelets/codelet_ztstrf.c

@@ -187,7 +187,7 @@ static void cl_ztstrf_cpu_func(void *descr[], void *cl_arg)
     {
         int i, sb;
         for (i=0; i<n; i+=ib) {
-            sb = min( ib, n-i );
+            sb = chameleon_min( ib, n-i );
             CORE_zlacpy(MorseUpperLower, sb, sb, L+(i*ldl), ldl, L+(i*ldl)+ib, ldl );
 
             CORE_ztrtri( MorseLower, MorseUnit, sb, L+(i*ldl)+ib, ldl, &info );

runtime/starpu/codelets/codelet_zttlqt.c

@@ -166,7 +166,7 @@ static void cl_zttlqt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda1, &lda2, &ldt);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
 
     CORE_zttlqt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
 }

runtime/starpu/codelets/codelet_zttqrt.c

@@ -166,7 +166,7 @@ static void cl_zttqrt_cpu_func(void *descr[], void *cl_arg)
 
     starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda1, &lda2, &ldt);
 
-    WORK = TAU + max( m, n );
+    WORK = TAU + chameleon_max( m, n );
 
     CORE_zttqrt(m, n, ib, A1, lda1, A2, lda2, T, ldt, TAU, WORK);
 }