Fix k = 0 edge case in power10 microkernels (#706)

Details:
- When power10 sgemm and dgemm microkernels are called with k = 0, they
  become caught in infinite loops and segfault. This is fixed now via an
  early exit in the case of k = 0.

kernels/power10/3/bli_dgemm_power10_mma.c

@@ -74,12 +74,10 @@ void bli_dgemm_power10_mma_8x8
         cntx_t*             cntx
     )
 {
-
     // Typecast local copies of integers in case dim_t and inc_t are a
     // different size than is expected by load instructions.
-    // (1 is subtracted from k0 because 1 iteration of the k loop is pulled out)
-    uint64_t k_iter = (k-1) / 4;
-    uint64_t k_left = (k-1) % 4;
+    uint64_t k_iter = k / 4;
+    uint64_t k_left = k % 4;
 
     uint64_t rs_c   = rs_c0;
 
@@ -110,6 +108,16 @@ void bli_dgemm_power10_mma_8x8
     __vector_quad acc0, acc1, acc2, acc3,
                   acc4, acc5, acc6, acc7;
 
+    // initialize the accumulators to zeros
+    __builtin_mma_xxsetaccz(&acc0);
+    __builtin_mma_xxsetaccz(&acc1);
+    __builtin_mma_xxsetaccz(&acc2);
+    __builtin_mma_xxsetaccz(&acc3);
+    __builtin_mma_xxsetaccz(&acc4);
+    __builtin_mma_xxsetaccz(&acc5);
+    __builtin_mma_xxsetaccz(&acc6);
+    __builtin_mma_xxsetaccz(&acc7);
+
     /* 2 vector pairs are necessary for a double precision outer product
        instruction. */
     __vector_pair colA_1,
@@ -141,19 +149,6 @@ void bli_dgemm_power10_mma_8x8
     */
     D_ASSEMBLE_VEC_PAIR
 
-    /* Compute accumulate outer products and override accumulators with result */
-    __builtin_mma_xvf64ger (&acc0, colA_1, rb[0]);
-    __builtin_mma_xvf64ger (&acc1, colA_1, rb[1]);
-    __builtin_mma_xvf64ger (&acc2, colA_1, rb[2]);
-    __builtin_mma_xvf64ger (&acc3, colA_1, rb[3]);
-    __builtin_mma_xvf64ger (&acc4, colA_2, rb[0]);
-    __builtin_mma_xvf64ger (&acc5, colA_2, rb[1]);
-    __builtin_mma_xvf64ger (&acc6, colA_2, rb[2]);
-    __builtin_mma_xvf64ger (&acc7, colA_2, rb[3]);
-
-    /* Move A and B pointers */
-    D_INCREMENT
-
     // k loop (unrolled by 4)
     for (int k = 0; k<k_iter; k++)
     {

kernels/power10/3/bli_sgemm_power10_mma.c

@@ -69,9 +69,8 @@ void bli_sgemm_power10_mma_8x16
 {
     // Typecast local copies of integers in case dim_t and inc_t are a
     // different size than is expected by load instructions.
-    // (1 is subtracted from k0 because 1 iteration of the k loop is pulled out)
-    uint64_t k_iter = (k-1) / 4;
-    uint64_t k_left = (k-1) % 4;
+    uint64_t k_iter = k / 4;
+    uint64_t k_left = k % 4;
 
     uint64_t rs_c   = rs_c0;
 
@@ -84,6 +83,16 @@ void bli_sgemm_power10_mma_8x16
     __vector_quad acc0, acc1, acc2, acc3,
                   acc4, acc5, acc6, acc7;
 
+    // initialize the accumulators to zeros
+    __builtin_mma_xxsetaccz(&acc0);
+    __builtin_mma_xxsetaccz(&acc1);
+    __builtin_mma_xxsetaccz(&acc2);
+    __builtin_mma_xxsetaccz(&acc3);
+    __builtin_mma_xxsetaccz(&acc4);
+    __builtin_mma_xxsetaccz(&acc5);
+    __builtin_mma_xxsetaccz(&acc6);
+    __builtin_mma_xxsetaccz(&acc7);
+
     float* restrict A0 = a;
     float* restrict B0 = b;
     float* restrict C0 = c;
@@ -95,18 +104,6 @@ void bli_sgemm_power10_mma_8x16
     vec_t *ca = (vec_t *) A0;
     vec_t *rb = (vec_t *) B0;
 
-    /* Compute accumulate outer products and override accumulators with result */
-    __builtin_mma_xvf32ger (&acc0, ca[0], rb[0]);
-    __builtin_mma_xvf32ger (&acc1, ca[0], rb[1]);
-    __builtin_mma_xvf32ger (&acc2, ca[0], rb[2]);
-    __builtin_mma_xvf32ger (&acc3, ca[0], rb[3]);
-    __builtin_mma_xvf32ger (&acc4, ca[1], rb[0]);
-    __builtin_mma_xvf32ger (&acc5, ca[1], rb[1]);
-    __builtin_mma_xvf32ger (&acc6, ca[1], rb[2]);
-    __builtin_mma_xvf32ger (&acc7, ca[1], rb[3]);
-
-    S_INCREMENT
-
     // k loop (unrolled by 4)
     for (int k = 0; k<k_iter; k++)
     {
@@ -147,4 +144,4 @@ void bli_sgemm_power10_mma_8x16
     }
 
     GEMM_UKR_FLUSH_CT( s );
-}
+}