Fix bf16->f32 conversion for NEOVERSEV1 target

This fixes an issue originally introduced with the BGEMM kernel.

I've updated the tests to run with `beta=1.0` so as to test loading and
updating from C.

Alongside this, the tests now return sensible return values to reduce
the risk of them being ignored.

Author: Mousius

kernel/arm64/bgemm_kernel_2vlx4_neoversev1_impl.c

@@ -40,7 +40,8 @@
 
 #define UPDATE_C(PG, PTR, DST, SRC)                                            \
   do {                                                                         \
-    DST = svreinterpret_f32_u32(svld1uh_u32((pghalf), (uint16_t*)PTR));        \
+    svtmp16 = svld1_bf16((pghalf), (PTR));                                     \
+    DST = svreinterpret_f32(svzip1_bf16(zeros, svtmp16));                      \
     DST = svadd_z((PG), SRC, DST);                                             \
     svtmp16 = svcvt_bf16_f32_z((PG), DST);                                     \
     svtmp16 = svuzp1_bf16(svtmp16, svtmp16);                                   \
@@ -55,7 +56,8 @@
 
 #define UPDATE_C(PG, PTR, DST, SRC)                                            \
   do {                                                                         \
-    DST = svreinterpret_f32_u32(svld1uh_u32((pghalf), (uint16_t*)PTR));        \
+    svtmp16 = svld1_bf16((pghalf), (PTR));                                     \
+    DST = svreinterpret_f32(svzip1_bf16(zeros, svtmp16));                      \
     DST = svmad_z((PG), svalpha, SRC, DST);                                    \
     svtmp16 = svcvt_bf16_f32_z((PG), DST);                                     \
     svtmp16 = svuzp1_bf16(svtmp16, svtmp16);                                   \
@@ -133,6 +135,7 @@ static int bgemm_kernel_neoversev1_alpha(BLASLONG m, BLASLONG n, BLASLONG k,
     OUTPUT_FLOAT *ptr_c0, *ptr_c1, *ptr_c2, *ptr_c3;
     svfloat32_t tmp0, tmp1, tmp2, tmp3;
 #ifdef BGEMM
+    svbfloat16_t zeros = svdup_n_bf16(TO16(0.0));
     svbfloat16_t svtmp16;
 #else
     float32x2_t tmp4, tmp5, tmp6, tmp7;

kernel/arm64/sbgemm_kernel_8x4_neoversen2_impl.c

@@ -51,7 +51,8 @@
 #ifdef ALPHA_ONE
 #define UPDATE_C(PG16, PG32, PTR, SRC)                                    \
   do {                                                                    \
-    tmp32 = svreinterpret_f32_u32(svld1uh_u32((PG16), (uint16_t*)PTR));   \
+    tmp16 = svld1_bf16((PG16), (PTR));                                    \
+    tmp32 = svreinterpret_f32(svzip1_bf16(zeros, tmp16));                 \
     tmp32 = svadd_z((PG32), SRC, tmp32);                                  \
     tmp16 = svcvt_bf16_f32_z((PG32), tmp32);                              \
     tmp16 = svuzp1_bf16(tmp16, tmp16);                                    \
@@ -60,7 +61,8 @@
 #else
 #define UPDATE_C(PG16, PG32, PTR, SRC)                                     \
   do {                                                                     \
-    tmp32 = svreinterpret_f32_u32(svld1uh_u32((PG16), (uint16_t*)PTR));    \
+    tmp16 = svld1_bf16((PG16), (PTR));                                     \
+    tmp32 = svreinterpret_f32(svzip1_bf16(zeros, tmp16));                  \
     tmp32 = svmad_z((PG32), svalpha, SRC, tmp32);                          \
     tmp16 = svcvt_bf16_f32_z((PG32), tmp32);                               \
     tmp16 = svuzp1_bf16(tmp16, tmp16);                                     \
@@ -121,6 +123,7 @@ static int gemm_kernel_neoversen2_alpha(BLASLONG m, BLASLONG n, BLASLONG k, FLOA
 #ifdef BGEMM
   svbool_t pg16_first_2 = svdupq_b16(1, 1, 0, 0, 0, 0, 0, 0);
   svbool_t pg16_first_1 = svdupq_b16(1, 0, 0, 0, 0, 0, 0, 0);
+  svbfloat16_t zeros = svdup_n_bf16(vcvtah_f32_bf16(0.0));
 #endif
 
   bfloat16_t *ptr_a = (bfloat16_t *)A;

test/compare_sgemm_bgemm.c

@@ -44,7 +44,7 @@ main (int argc, char *argv[])
   int ret = 0;
   int loop = BGEMM_LARGEST;
   char transA = 'N', transB = 'N';
-  float alpha = 1.0, beta = 0.0;
+  float alpha = 1.0, beta = 1.0;
   bfloat16 alpha_bf16;
   sbstobf16_(&one, &alpha, &one, &alpha_bf16, &one);
   bfloat16 beta_bf16;
@@ -94,9 +94,15 @@ main (int argc, char *argv[])
         transB = 'T';
       }
 
-      memset(CC, 0, m * n * sizeof(bfloat16));
-      memset(DD, 0, m * n * sizeof(FLOAT));
-      memset(C, 0, m * n * sizeof(FLOAT));
+      for (j = 0; j < m; j++)
+      {
+        for (i = 0; i < n; i++)
+        {
+          C[j * n + i] = 100.0;
+          DD[j * n + i] = 100.0;
+          sbstobf16_(&one, &C[j * n + i], &one, &CC[j * n + i], &one);
+        }
+      }
 
       SGEMM (&transA, &transB, &m, &n, &k, &alpha, A,
         &m, B, &k, &beta, C, &m);
@@ -152,7 +158,8 @@ main (int argc, char *argv[])
   }
 
   if (ret != 0) {
-    fprintf (stderr, "FATAL ERROR BGEMM - Return code: %d\n", ret);
+    fprintf(stderr, "BGEMM FAILURES: %d\n", ret);
+    return 1;
   }
 
   return ret;

test/compare_sgemm_sbgemm.c

@@ -140,7 +140,8 @@ main (int argc, char *argv[])
   }
 
   if (ret != 0) {
-    fprintf (stderr, "FATAL ERROR SBGEMM - Return code: %d\n", ret);
+    fprintf(stderr, "SBGEMM FAILURES: %d\n", ret);
+    return 1;
   }
 
   return ret;

test/compare_sgemv_bgemv.c

@@ -147,7 +147,10 @@ int main(int argc, char *argv[])
     } // alpha
   } // beta
 
-  if (ret != 0)
-    fprintf(stderr, "FATAL ERROR BGEMV - Return code: %d\n", ret);
+  if (ret != 0) {
+    fprintf(stderr, "BGEMV FAILURES: %d\n", ret);
+    return 1;
+  }
+
   return ret;
 }

test/compare_sgemv_sbgemv.c

@@ -122,7 +122,10 @@ main (int argc, char *argv[])
   } // alpha
   } // beta
 
-  if (ret != 0)
-    fprintf (stderr, "FATAL ERROR SBGEMV - Return code: %d\n", ret);
+  if (ret != 0) {
+    fprintf(stderr, "SBGEMV FAILURES: %d\n", ret);
+    return 1;
+  }
+
   return ret;
 }