commit high-performance mm (scalar and vector versions)

3 files changed, 14 insertions, 18 deletions

diff --git a/benchmarks/common/crt.S b/benchmarks/common/crt.S
index b273900..82cad93 100644
--- a/benchmarks/common/crt.S
+++ b/benchmarks/common/crt.S
@@ -8,14 +8,6 @@
 # define SREG sw
 #endif
 
-  .data
-  .globl _heapend
-  .globl environ
-_heapend:
-  .word 0
-environ:
-  .word 0
-
   .text
   .globl _start
 
@@ -62,10 +54,13 @@ _start:
   li a0, SR_EF | SR_EA
   csrs status, a0
 
-  ## if that didn't stick, we don't have an FPU, so don't initialize it
   csrr t0, status
-  and t0, t0, SR_EF
-  beqz t0, 1f
+  and t1, t0, SR_EA
+  sw t1, have_vec, t2
+
+  ## if that didn't stick, we don't have an FPU, so don't initialize it
+  and t1, t0, SR_EF
+  beqz t1, 1f
 
   fssr    x0
   fmv.s.x f0, x0
diff --git a/benchmarks/common/syscalls.c b/benchmarks/common/syscalls.c
index a168ebf..e9f04e2 100644
--- a/benchmarks/common/syscalls.c
+++ b/benchmarks/common/syscalls.c
@@ -8,6 +8,9 @@
 
 #define SYS_stats 1234
 
+// initialized in crt.S
+int have_vec;
+
 static long handle_frontend_syscall(long which, long arg0, long arg1, long arg2)
 {
   volatile uint64_t magic_mem[8] __attribute__((aligned(64)));
@@ -150,8 +153,8 @@ void _init(int cid, int nc)
 #undef putchar
 int putchar(int ch)
 {
-  static char buf[64] __attribute__((aligned(64)));
-  static int buflen = 0;
+  static __thread char buf[64] __attribute__((aligned(64)));
+  static __thread int buflen = 0;
 
   buf[buflen++] = ch;
 
diff --git a/benchmarks/common/util.h b/benchmarks/common/util.h
index 7c2a048..6c4f963 100644
--- a/benchmarks/common/util.h
+++ b/benchmarks/common/util.h
@@ -31,6 +31,8 @@ static void setStats(int enable) {}
 extern void setStats(int enable);
 #endif
 
+extern int have_vec;
+
 #define static_assert(cond) switch(0) { case 0: case !!(long)(cond): ; }
 
 static void printArray(const char name[], int n, const int arr[])
@@ -87,11 +89,7 @@ static int verifyDouble(int n, const volatile double* test, const double* verify
   return 0;
 }
 
-#ifndef ncores
-#define ncores 1
-#endif
-
-static void __attribute__((noinline)) barrier()
+static void __attribute__((noinline)) barrier(int ncores)
 {
   static volatile int sense;
   static volatile int count;