X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=benchmarks%2Fmm%2Frb.h;h=421c9648fb47118a886dadd14b29955f943b21b3;hb=795c81a287bde70c6d8029a812f0a76e83823d68;hp=c5d58907c1652fe0eb20b66e46f169628837f250;hpb=83ed3f519de9929b6551b98677047228a8ab4d0c;p=riscv-tests.git

diff --git a/benchmarks/mm/rb.h b/benchmarks/mm/rb.h
index c5d5890..421c964 100644
--- a/benchmarks/mm/rb.h
+++ b/benchmarks/mm/rb.h
@@ -1,5 +1,5 @@
 static const int RBM = 4, RBN = 5, RBK = 6;
-static const int CBM = 36, CBN = 35, CBK = 36;
+static const int CBM = 24, CBN = 25, CBK = 24;
 static inline void kloop(size_t p, t* a0, size_t lda, t* b0, size_t ldb, t* c, size_t ldc)
 {
   t* c_0 = &c[ldc*0];
@@ -39,124 +39,124 @@ static inline void kloop(size_t p, t* a0, size_t lda, t* b0, size_t ldb, t* c, s
     t* b_4 = &b[ldb*4];
     t* b_5 = &b[ldb*5];
     c_0_0 = fma(a_0[0], b_0[0], c_0_0);
-    c_0_0 = fma(a_0[1], b_1[0], c_0_0);
-    c_0_0 = fma(a_0[2], b_2[0], c_0_0);
-    c_0_0 = fma(a_0[3], b_3[0], c_0_0);
-    c_0_0 = fma(a_0[4], b_4[0], c_0_0);
-    c_0_0 = fma(a_0[5], b_5[0], c_0_0);
     c_0_1 = fma(a_0[0], b_0[1], c_0_1);
-    c_0_1 = fma(a_0[1], b_1[1], c_0_1);
-    c_0_1 = fma(a_0[2], b_2[1], c_0_1);
-    c_0_1 = fma(a_0[3], b_3[1], c_0_1);
-    c_0_1 = fma(a_0[4], b_4[1], c_0_1);
-    c_0_1 = fma(a_0[5], b_5[1], c_0_1);
     c_0_2 = fma(a_0[0], b_0[2], c_0_2);
-    c_0_2 = fma(a_0[1], b_1[2], c_0_2);
-    c_0_2 = fma(a_0[2], b_2[2], c_0_2);
-    c_0_2 = fma(a_0[3], b_3[2], c_0_2);
-    c_0_2 = fma(a_0[4], b_4[2], c_0_2);
-    c_0_2 = fma(a_0[5], b_5[2], c_0_2);
     c_0_3 = fma(a_0[0], b_0[3], c_0_3);
-    c_0_3 = fma(a_0[1], b_1[3], c_0_3);
-    c_0_3 = fma(a_0[2], b_2[3], c_0_3);
-    c_0_3 = fma(a_0[3], b_3[3], c_0_3);
-    c_0_3 = fma(a_0[4], b_4[3], c_0_3);
-    c_0_3 = fma(a_0[5], b_5[3], c_0_3);
     c_0_4 = fma(a_0[0], b_0[4], c_0_4);
-    c_0_4 = fma(a_0[1], b_1[4], c_0_4);
-    c_0_4 = fma(a_0[2], b_2[4], c_0_4);
-    c_0_4 = fma(a_0[3], b_3[4], c_0_4);
-    c_0_4 = fma(a_0[4], b_4[4], c_0_4);
-    c_0_4 = fma(a_0[5], b_5[4], c_0_4);
     c_1_0 = fma(a_1[0], b_0[0], c_1_0);
-    c_1_0 = fma(a_1[1], b_1[0], c_1_0);
-    c_1_0 = fma(a_1[2], b_2[0], c_1_0);
-    c_1_0 = fma(a_1[3], b_3[0], c_1_0);
-    c_1_0 = fma(a_1[4], b_4[0], c_1_0);
-    c_1_0 = fma(a_1[5], b_5[0], c_1_0);
     c_1_1 = fma(a_1[0], b_0[1], c_1_1);
-    c_1_1 = fma(a_1[1], b_1[1], c_1_1);
-    c_1_1 = fma(a_1[2], b_2[1], c_1_1);
-    c_1_1 = fma(a_1[3], b_3[1], c_1_1);
-    c_1_1 = fma(a_1[4], b_4[1], c_1_1);
-    c_1_1 = fma(a_1[5], b_5[1], c_1_1);
     c_1_2 = fma(a_1[0], b_0[2], c_1_2);
-    c_1_2 = fma(a_1[1], b_1[2], c_1_2);
-    c_1_2 = fma(a_1[2], b_2[2], c_1_2);
-    c_1_2 = fma(a_1[3], b_3[2], c_1_2);
-    c_1_2 = fma(a_1[4], b_4[2], c_1_2);
-    c_1_2 = fma(a_1[5], b_5[2], c_1_2);
     c_1_3 = fma(a_1[0], b_0[3], c_1_3);
-    c_1_3 = fma(a_1[1], b_1[3], c_1_3);
-    c_1_3 = fma(a_1[2], b_2[3], c_1_3);
-    c_1_3 = fma(a_1[3], b_3[3], c_1_3);
-    c_1_3 = fma(a_1[4], b_4[3], c_1_3);
-    c_1_3 = fma(a_1[5], b_5[3], c_1_3);
     c_1_4 = fma(a_1[0], b_0[4], c_1_4);
-    c_1_4 = fma(a_1[1], b_1[4], c_1_4);
-    c_1_4 = fma(a_1[2], b_2[4], c_1_4);
-    c_1_4 = fma(a_1[3], b_3[4], c_1_4);
-    c_1_4 = fma(a_1[4], b_4[4], c_1_4);
-    c_1_4 = fma(a_1[5], b_5[4], c_1_4);
     c_2_0 = fma(a_2[0], b_0[0], c_2_0);
-    c_2_0 = fma(a_2[1], b_1[0], c_2_0);
-    c_2_0 = fma(a_2[2], b_2[0], c_2_0);
-    c_2_0 = fma(a_2[3], b_3[0], c_2_0);
-    c_2_0 = fma(a_2[4], b_4[0], c_2_0);
-    c_2_0 = fma(a_2[5], b_5[0], c_2_0);
     c_2_1 = fma(a_2[0], b_0[1], c_2_1);
-    c_2_1 = fma(a_2[1], b_1[1], c_2_1);
-    c_2_1 = fma(a_2[2], b_2[1], c_2_1);
-    c_2_1 = fma(a_2[3], b_3[1], c_2_1);
-    c_2_1 = fma(a_2[4], b_4[1], c_2_1);
-    c_2_1 = fma(a_2[5], b_5[1], c_2_1);
     c_2_2 = fma(a_2[0], b_0[2], c_2_2);
-    c_2_2 = fma(a_2[1], b_1[2], c_2_2);
-    c_2_2 = fma(a_2[2], b_2[2], c_2_2);
-    c_2_2 = fma(a_2[3], b_3[2], c_2_2);
-    c_2_2 = fma(a_2[4], b_4[2], c_2_2);
-    c_2_2 = fma(a_2[5], b_5[2], c_2_2);
     c_2_3 = fma(a_2[0], b_0[3], c_2_3);
-    c_2_3 = fma(a_2[1], b_1[3], c_2_3);
-    c_2_3 = fma(a_2[2], b_2[3], c_2_3);
-    c_2_3 = fma(a_2[3], b_3[3], c_2_3);
-    c_2_3 = fma(a_2[4], b_4[3], c_2_3);
-    c_2_3 = fma(a_2[5], b_5[3], c_2_3);
     c_2_4 = fma(a_2[0], b_0[4], c_2_4);
-    c_2_4 = fma(a_2[1], b_1[4], c_2_4);
-    c_2_4 = fma(a_2[2], b_2[4], c_2_4);
-    c_2_4 = fma(a_2[3], b_3[4], c_2_4);
-    c_2_4 = fma(a_2[4], b_4[4], c_2_4);
-    c_2_4 = fma(a_2[5], b_5[4], c_2_4);
     c_3_0 = fma(a_3[0], b_0[0], c_3_0);
-    c_3_0 = fma(a_3[1], b_1[0], c_3_0);
-    c_3_0 = fma(a_3[2], b_2[0], c_3_0);
-    c_3_0 = fma(a_3[3], b_3[0], c_3_0);
-    c_3_0 = fma(a_3[4], b_4[0], c_3_0);
-    c_3_0 = fma(a_3[5], b_5[0], c_3_0);
     c_3_1 = fma(a_3[0], b_0[1], c_3_1);
-    c_3_1 = fma(a_3[1], b_1[1], c_3_1);
-    c_3_1 = fma(a_3[2], b_2[1], c_3_1);
-    c_3_1 = fma(a_3[3], b_3[1], c_3_1);
-    c_3_1 = fma(a_3[4], b_4[1], c_3_1);
-    c_3_1 = fma(a_3[5], b_5[1], c_3_1);
     c_3_2 = fma(a_3[0], b_0[2], c_3_2);
-    c_3_2 = fma(a_3[1], b_1[2], c_3_2);
-    c_3_2 = fma(a_3[2], b_2[2], c_3_2);
-    c_3_2 = fma(a_3[3], b_3[2], c_3_2);
-    c_3_2 = fma(a_3[4], b_4[2], c_3_2);
-    c_3_2 = fma(a_3[5], b_5[2], c_3_2);
     c_3_3 = fma(a_3[0], b_0[3], c_3_3);
-    c_3_3 = fma(a_3[1], b_1[3], c_3_3);
-    c_3_3 = fma(a_3[2], b_2[3], c_3_3);
-    c_3_3 = fma(a_3[3], b_3[3], c_3_3);
-    c_3_3 = fma(a_3[4], b_4[3], c_3_3);
-    c_3_3 = fma(a_3[5], b_5[3], c_3_3);
     c_3_4 = fma(a_3[0], b_0[4], c_3_4);
+    c_0_0 = fma(a_0[1], b_1[0], c_0_0);
+    c_0_1 = fma(a_0[1], b_1[1], c_0_1);
+    c_0_2 = fma(a_0[1], b_1[2], c_0_2);
+    c_0_3 = fma(a_0[1], b_1[3], c_0_3);
+    c_0_4 = fma(a_0[1], b_1[4], c_0_4);
+    c_1_0 = fma(a_1[1], b_1[0], c_1_0);
+    c_1_1 = fma(a_1[1], b_1[1], c_1_1);
+    c_1_2 = fma(a_1[1], b_1[2], c_1_2);
+    c_1_3 = fma(a_1[1], b_1[3], c_1_3);
+    c_1_4 = fma(a_1[1], b_1[4], c_1_4);
+    c_2_0 = fma(a_2[1], b_1[0], c_2_0);
+    c_2_1 = fma(a_2[1], b_1[1], c_2_1);
+    c_2_2 = fma(a_2[1], b_1[2], c_2_2);
+    c_2_3 = fma(a_2[1], b_1[3], c_2_3);
+    c_2_4 = fma(a_2[1], b_1[4], c_2_4);
+    c_3_0 = fma(a_3[1], b_1[0], c_3_0);
+    c_3_1 = fma(a_3[1], b_1[1], c_3_1);
+    c_3_2 = fma(a_3[1], b_1[2], c_3_2);
+    c_3_3 = fma(a_3[1], b_1[3], c_3_3);
     c_3_4 = fma(a_3[1], b_1[4], c_3_4);
+    c_0_0 = fma(a_0[2], b_2[0], c_0_0);
+    c_0_1 = fma(a_0[2], b_2[1], c_0_1);
+    c_0_2 = fma(a_0[2], b_2[2], c_0_2);
+    c_0_3 = fma(a_0[2], b_2[3], c_0_3);
+    c_0_4 = fma(a_0[2], b_2[4], c_0_4);
+    c_1_0 = fma(a_1[2], b_2[0], c_1_0);
+    c_1_1 = fma(a_1[2], b_2[1], c_1_1);
+    c_1_2 = fma(a_1[2], b_2[2], c_1_2);
+    c_1_3 = fma(a_1[2], b_2[3], c_1_3);
+    c_1_4 = fma(a_1[2], b_2[4], c_1_4);
+    c_2_0 = fma(a_2[2], b_2[0], c_2_0);
+    c_2_1 = fma(a_2[2], b_2[1], c_2_1);
+    c_2_2 = fma(a_2[2], b_2[2], c_2_2);
+    c_2_3 = fma(a_2[2], b_2[3], c_2_3);
+    c_2_4 = fma(a_2[2], b_2[4], c_2_4);
+    c_3_0 = fma(a_3[2], b_2[0], c_3_0);
+    c_3_1 = fma(a_3[2], b_2[1], c_3_1);
+    c_3_2 = fma(a_3[2], b_2[2], c_3_2);
+    c_3_3 = fma(a_3[2], b_2[3], c_3_3);
     c_3_4 = fma(a_3[2], b_2[4], c_3_4);
+    c_0_0 = fma(a_0[3], b_3[0], c_0_0);
+    c_0_1 = fma(a_0[3], b_3[1], c_0_1);
+    c_0_2 = fma(a_0[3], b_3[2], c_0_2);
+    c_0_3 = fma(a_0[3], b_3[3], c_0_3);
+    c_0_4 = fma(a_0[3], b_3[4], c_0_4);
+    c_1_0 = fma(a_1[3], b_3[0], c_1_0);
+    c_1_1 = fma(a_1[3], b_3[1], c_1_1);
+    c_1_2 = fma(a_1[3], b_3[2], c_1_2);
+    c_1_3 = fma(a_1[3], b_3[3], c_1_3);
+    c_1_4 = fma(a_1[3], b_3[4], c_1_4);
+    c_2_0 = fma(a_2[3], b_3[0], c_2_0);
+    c_2_1 = fma(a_2[3], b_3[1], c_2_1);
+    c_2_2 = fma(a_2[3], b_3[2], c_2_2);
+    c_2_3 = fma(a_2[3], b_3[3], c_2_3);
+    c_2_4 = fma(a_2[3], b_3[4], c_2_4);
+    c_3_0 = fma(a_3[3], b_3[0], c_3_0);
+    c_3_1 = fma(a_3[3], b_3[1], c_3_1);
+    c_3_2 = fma(a_3[3], b_3[2], c_3_2);
+    c_3_3 = fma(a_3[3], b_3[3], c_3_3);
     c_3_4 = fma(a_3[3], b_3[4], c_3_4);
+    c_0_0 = fma(a_0[4], b_4[0], c_0_0);
+    c_0_1 = fma(a_0[4], b_4[1], c_0_1);
+    c_0_2 = fma(a_0[4], b_4[2], c_0_2);
+    c_0_3 = fma(a_0[4], b_4[3], c_0_3);
+    c_0_4 = fma(a_0[4], b_4[4], c_0_4);
+    c_1_0 = fma(a_1[4], b_4[0], c_1_0);
+    c_1_1 = fma(a_1[4], b_4[1], c_1_1);
+    c_1_2 = fma(a_1[4], b_4[2], c_1_2);
+    c_1_3 = fma(a_1[4], b_4[3], c_1_3);
+    c_1_4 = fma(a_1[4], b_4[4], c_1_4);
+    c_2_0 = fma(a_2[4], b_4[0], c_2_0);
+    c_2_1 = fma(a_2[4], b_4[1], c_2_1);
+    c_2_2 = fma(a_2[4], b_4[2], c_2_2);
+    c_2_3 = fma(a_2[4], b_4[3], c_2_3);
+    c_2_4 = fma(a_2[4], b_4[4], c_2_4);
+    c_3_0 = fma(a_3[4], b_4[0], c_3_0);
+    c_3_1 = fma(a_3[4], b_4[1], c_3_1);
+    c_3_2 = fma(a_3[4], b_4[2], c_3_2);
+    c_3_3 = fma(a_3[4], b_4[3], c_3_3);
     c_3_4 = fma(a_3[4], b_4[4], c_3_4);
+    c_0_0 = fma(a_0[5], b_5[0], c_0_0);
+    c_0_1 = fma(a_0[5], b_5[1], c_0_1);
+    c_0_2 = fma(a_0[5], b_5[2], c_0_2);
+    c_0_3 = fma(a_0[5], b_5[3], c_0_3);
+    c_0_4 = fma(a_0[5], b_5[4], c_0_4);
+    c_1_0 = fma(a_1[5], b_5[0], c_1_0);
+    c_1_1 = fma(a_1[5], b_5[1], c_1_1);
+    c_1_2 = fma(a_1[5], b_5[2], c_1_2);
+    c_1_3 = fma(a_1[5], b_5[3], c_1_3);
+    c_1_4 = fma(a_1[5], b_5[4], c_1_4);
+    c_2_0 = fma(a_2[5], b_5[0], c_2_0);
+    c_2_1 = fma(a_2[5], b_5[1], c_2_1);
+    c_2_2 = fma(a_2[5], b_5[2], c_2_2);
+    c_2_3 = fma(a_2[5], b_5[3], c_2_3);
+    c_2_4 = fma(a_2[5], b_5[4], c_2_4);
+    c_3_0 = fma(a_3[5], b_5[0], c_3_0);
+    c_3_1 = fma(a_3[5], b_5[1], c_3_1);
+    c_3_2 = fma(a_3[5], b_5[2], c_3_2);
+    c_3_3 = fma(a_3[5], b_5[3], c_3_3);
     c_3_4 = fma(a_3[5], b_5[4], c_3_4);
   }
   for (t *a = a0 + p/RBK*RBK, *b = b0 + p/RBK*RBK*ldb; a < a0 + p; a++, b += ldb)