Merge pull request #5091 from goplanid/develop

Small gemm kernel improvements for AArch64

Author: martin-frbg

kernel/arm64/dgemm_small_kernel_tn_sve.c

@@ -213,7 +213,7 @@ CNAME(BLASLONG M,
   const BLASLONG n2 = N & -2;
   const BLASLONG n8 = N & -8;
 
-  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  const int pack_a = M >= v_size2 && N >= 8 ? 1 : 0;
   FLOAT* packed_a =
     (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
 

kernel/arm64/dgemm_small_kernel_tt_sve.c

@@ -219,7 +219,7 @@ CNAME(BLASLONG M,
   const BLASLONG n4 = N & -4;
   const BLASLONG n2 = N & -2;
 
-  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  const int pack_a = M >= v_size2 && N >= 8 ? 1 : 0;
   FLOAT* packed_a =
     (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
 

kernel/arm64/sgemm_small_kernel_tn_sve.c

@@ -222,7 +222,7 @@ CNAME(BLASLONG M,
   const BLASLONG n8 = N & -8;
   const BLASLONG n4 = N & -4;
 
-  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  const int pack_a = M >= v_size2 && N >= 8 ? 1 : 0;
   FLOAT* packed_a =
     (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;
 

kernel/arm64/sgemm_small_kernel_tt_sve.c

@@ -223,7 +223,7 @@ CNAME(BLASLONG M,
   const BLASLONG n8 = N & -8;
   const BLASLONG n4 = N & -4;
 
-  const int pack_a = M >= v_size2 && N >= 8 && K >= 8 ? 1 : 0;
+  const int pack_a = M >= v_size2 && N >= 8 ? 1 : 0;
   FLOAT* packed_a =
     (pack_a) ? packed_a = (FLOAT*)malloc(K * v_size2 * sizeof(FLOAT)) : NULL;