small fixes for separating sme and dummy parts

Author: martin-frbg

kernel/arm64/sgemm_direct_alpha_beta_arm64_sme1.c

@@ -8,11 +8,6 @@
 #include <inttypes.h>
 #include <math.h>
 #include "sme_abi.h"
-#if defined(HAVE_SME)
-
-#if defined(__ARM_FEATURE_SME) && defined(__clang__) && __clang_major__ >= 16
-#include <arm_sme.h>
-#endif
 
 #if defined(DYNAMIC_ARCH)
 #define COMBINE(a,b) a ## b
@@ -25,10 +20,17 @@
 #define SME1_PREPROCESS sgemm_direct_sme1_preprocess
 #define SME1_KERNEL2X2  sgemm_direct_alpha_beta_sme1_2VLx2VL
 #endif
+
 /* Function prototypes */
 extern void SME1_PREPROCESS(uint64_t nbr, uint64_t nbc,\
                                   const float * restrict a, float *  a_mod);
 
+#if defined(HAVE_SME)
+
+#if defined(__ARM_FEATURE_SME) && defined(__clang__) && __clang_major__ >= 16
+#include <arm_sme.h>
+#endif
+
 /* Function Definitions */
 static uint64_t sve_cntw() {
     uint64_t cnt;
@@ -210,5 +212,4 @@ void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float alpha, float * __restrict
 void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float alpha, float * __restrict A,\
             BLASLONG strideA, float * __restrict B, BLASLONG strideB ,\
             float beta, float * __restrict R, BLASLONG strideR){fprintf(stderr,"empty sgemm_direct_alpha_beta should not be called!!!\n");}
- 
 #endif

kernel/arm64/sgemm_direct_arm64_sme1.c

@@ -7,7 +7,6 @@
 #include <stdlib.h>
 #include <inttypes.h>
 #include <math.h>
-#if defined(HAVE_SME)
 #if defined(DYNAMIC_ARCH)
 #define COMBINE(a,b) a ## b
 #define COMBINE2(a,b) COMBINE(a,b)
@@ -19,6 +18,7 @@
 #define SME1_PREPROCESS sgemm_direct_sme1_preprocess
 #define SME1_DIRECT2X2 sgemm_direct_sme1_2VLx2VL
 #endif
+#if defined(HAVE_SME)
 /* Function prototypes */
 extern void SME1_PREPROCESS(uint64_t nbr, uint64_t nbc,\
                                   const float * restrict a, float *  a_mod) ;
@@ -87,6 +87,14 @@ void CNAME (BLASLONG M, BLASLONG N, BLASLONG K, float * __restrict A,\
             BLASLONG strideA, float * __restrict B, BLASLONG strideB ,\
             float * __restrict R, BLASLONG strideR){
 fprintf(stderr,"EMPTY sgemm_kernel_direct should never be called \n");
-} 
+}
+void SME1_DIRECT2X2( uint64_t M , uint64_t K, uint64_t N,\
+                               const float * restrict A_base,\
+                               const float * restrict B_base,\
+                               const float * restrict C_base){}; 
+void SME1_PREPROCESS(uint64_t nbr, uint64_t nbc,\
+                                  const float * restrict a, float *  a_mod){};
+
+
 #endif
 

kernel/arm64/sgemm_direct_performant.c

@@ -6,7 +6,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K)
 if (M<3) return 0;
 	unsigned long long mnk = M * N * K;
 	/* benchmark performance on M4 peaks around 512 and crosses the graph of the NEON SGEMM at about 3100  */
-	if (mnk >= 3100 * 3100 * 3100)
+	if (mnk >= 3100L * 3100L * 3100L)
 		return 0;
 
 	return 1;