Clean up canonical mt benchmarks and reorganize extra versions in /mt. All versions...

[riscv-tests.git] / mt / cy_matmul.c

diff --git a/mt/cy_matmul.c b/mt/cy_matmul.c
new file mode 100644 (file)
index 0000000..b0713cd
--- /dev/null
+++ b/mt/cy_matmul.c
@@ -0,0 +1,108 @@
+#include "stdlib.h"
+
+#include "util.h"
+
+#include "dataset.h"\r
+void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda,  const data_t A[], const data_t B[], data_t C[] )\r
+{\r
+   if(coreid > 1) return; 
+  static __thread int i, j, k;\r
+  static __thread data_t tempA0, tempA1, tempA2, tempA3, tempA4, tempA5, tempA6, tempA7;\r
+  static __thread data_t tempC0, tempC1, tempC2, tempC3, tempC4, tempC5, tempC6, tempC7, tempC8, tempC9, tempC10, tempC11, tempC12, tempC13, tempC14, tempC15;\r
+\r
+  static __thread int start, end, jStride, jToRow, jToCol;\r
+  static data_t A1[1024], B1[1024];;\r
+  \r
+  start = coreid << 9;\r
+  end = ((ncores == 1) ? 2 :(coreid+1)) << 9;\r
+  jStride = 8;\r
+\r
+  if (coreid == 0) { \r
+    for (j=start; j < end; j+=jStride) {\r
+      jToRow = (j>>5)<<5;\r
+      jToCol = j%32;\r
+      tempC0  = 0;\r
+      tempC1  = 0;\r
+      tempC2  = 0;\r
+      tempC3  = 0;\r
+      tempC4  = 0;\r
+      tempC5  = 0;\r
+      tempC6  = 0;\r
+      tempC7  = 0;\r
+      for ( i=0; i < lda; i+=2 ) {\r
+        tempA0 = A[i   + jToRow];\r
+        tempA1 = A[i+1 + jToRow];\r
+        tempC0  += tempA0 * B[(jToCol   ) + (i<<5)];\r
+        tempC1  += tempA0 * B[(jToCol+1 ) + (i<<5)];\r
+        tempC2  += tempA0 * B[(jToCol+2 ) + (i<<5)];\r
+        tempC3  += tempA0 * B[(jToCol+3 ) + (i<<5)];\r
+        tempC4  += tempA0 * B[(jToCol+4 ) + (i<<5)];\r
+        tempC5  += tempA0 * B[(jToCol+5 ) + (i<<5)];\r
+        tempC6  += tempA0 * B[(jToCol+6 ) + (i<<5)];\r
+        tempC7  += tempA0 * B[(jToCol+7 ) + (i<<5)];\r
+        tempC0  += tempA1 * B[(jToCol   ) + ((i+1)<<5)];\r
+        tempC1  += tempA1 * B[(jToCol+1 ) + ((i+1)<<5)];\r
+        tempC2  += tempA1 * B[(jToCol+2 ) + ((i+1)<<5)];\r
+        tempC3  += tempA1 * B[(jToCol+3 ) + ((i+1)<<5)];\r
+        tempC4  += tempA1 * B[(jToCol+4 ) + ((i+1)<<5)];\r
+        tempC5  += tempA1 * B[(jToCol+5 ) + ((i+1)<<5)];\r
+        tempC6  += tempA1 * B[(jToCol+6 ) + ((i+1)<<5)];\r
+        tempC7  += tempA1 * B[(jToCol+7 ) + ((i+1)<<5)];\r
+      }\r
+      C[j] =tempC0;\r
+      C[j + 1 ]=tempC1;\r
+      C[j + 2 ]=tempC2;\r
+      C[j + 3 ]=tempC3;\r
+      C[j + 4 ]=tempC4;\r
+      C[j + 5 ]=tempC5;\r
+      C[j + 6 ]=tempC6;\r
+      C[j + 7 ]=tempC7;\r
+    }\r
+  }\r
+  else { \r
+    for (i = 0; i < 1024; i++) {\r
+      A1[i] = A[i];\r
+      B1[i] = B[i];\r
+    }\r
+    for (j=start; j < end; j+=jStride) {\r
+      jToRow = (j>>5)<<5;\r
+      jToCol = j%32;\r
+      tempC0  = 0;\r
+      tempC1  = 0;\r
+      tempC2  = 0;\r
+      tempC3  = 0;\r
+      tempC4  = 0;\r
+      tempC5  = 0;\r
+      tempC6  = 0;\r
+      tempC7  = 0;\r
+      for ( i=0; i < lda; i+=2 ) {\r
+        tempA0 = A1[i   + jToRow];\r
+        tempA1 = A1[i+1 + jToRow];\r
+        tempC0  += tempA0 * B1[(jToCol   ) + (i<<5)];\r
+        tempC1  += tempA0 * B1[(jToCol+1 ) + (i<<5)];\r
+        tempC2  += tempA0 * B1[(jToCol+2 ) + (i<<5)];\r
+        tempC3  += tempA0 * B1[(jToCol+3 ) + (i<<5)];\r
+        tempC4  += tempA0 * B1[(jToCol+4 ) + (i<<5)];\r
+        tempC5  += tempA0 * B1[(jToCol+5 ) + (i<<5)];\r
+        tempC6  += tempA0 * B1[(jToCol+6 ) + (i<<5)];\r
+        tempC7  += tempA0 * B1[(jToCol+7 ) + (i<<5)];\r
+        tempC0  += tempA1 * B1[(jToCol   ) + ((i+1)<<5)];\r
+        tempC1  += tempA1 * B1[(jToCol+1 ) + ((i+1)<<5)];\r
+        tempC2  += tempA1 * B1[(jToCol+2 ) + ((i+1)<<5)];\r
+        tempC3  += tempA1 * B1[(jToCol+3 ) + ((i+1)<<5)];\r
+        tempC4  += tempA1 * B1[(jToCol+4 ) + ((i+1)<<5)];\r
+        tempC5  += tempA1 * B1[(jToCol+5 ) + ((i+1)<<5)];\r
+        tempC6  += tempA1 * B1[(jToCol+6 ) + ((i+1)<<5)];\r
+        tempC7  += tempA1 * B1[(jToCol+7 ) + ((i+1)<<5)];\r
+      }\r
+      C[j] =tempC0;\r
+      C[j + 1 ]=tempC1;\r
+      C[j + 2 ]=tempC2;\r
+      C[j + 3 ]=tempC3;\r
+      C[j + 4 ]=tempC4;\r
+      C[j + 5 ]=tempC5;\r
+      C[j + 6 ]=tempC6;\r
+      C[j + 7 ]=tempC7;\r
+    }\r
+  }\r
+}\r