Use the skylake sgemm beta code also for haswell

with a few small changes it's possible to use the skylake sgemm code
also for haswell, this gives a modest gain (10% range) for smallish
matrixes but does wonders for very skinny matrixes

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 848de38..2aec600 100644 (file)
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -33,6 +33,7 @@ ZAXPYKERNEL = zaxpy.c
 
 STRMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
+SGEMM_BETA     =  sgemm_beta_skylakex.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_16.c
 SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c

diff --git a/kernel/x86_64/sgemm_beta_skylakex.c b/kernel/x86_64/sgemm_beta_skylakex.c
index 498c46f..e865311 100644 (file)
--- a/kernel/x86_64/sgemm_beta_skylakex.c
+++ b/kernel/x86_64/sgemm_beta_skylakex.c
@@ -61,11 +61,11 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
   c_offset = c;
 
   if (beta == ZERO){
-    __m512 z_zero;
-    __m256 y_zero;
+#ifdef __AVX512CD__
+    __m512 z_zero = _mm512_setzero_ps();
+#endif
+    __m256 y_zero = _mm256_setzero_ps();
 
-    z_zero = _mm512_setzero_ps();
-    y_zero = _mm256_setzero_ps();
     j = n;
     do {
       c_offset1 = c_offset;
@@ -74,8 +74,15 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
       i = m;
 
       while (i >= 32) {
+#ifdef __AVX512CD__
          _mm512_storeu_ps(c_offset1, z_zero);
          _mm512_storeu_ps(c_offset1 + 16, z_zero);
+#else
+         _mm256_storeu_ps(c_offset1, y_zero);
+         _mm256_storeu_ps(c_offset1 + 8, y_zero);
+         _mm256_storeu_ps(c_offset1 + 16, y_zero);
+         _mm256_storeu_ps(c_offset1 + 24, y_zero);
+#endif
          c_offset1 += 32;
          i -= 32;
       }