fixed loader script stack setup

runtime/src/vx_intrinsics.S

@@ -47,6 +47,12 @@ vx_warp_gid:
 .type vx_thread_id, @function
 .global vx_thread_id
 vx_thread_id:
+	csrr a0, CSR_WTID
+	ret
+
+.type vx_thread_lid, @function
+.global vx_thread_lid
+vx_thread_lid:
 	csrr a0, CSR_LTID
 	ret
 

runtime/src/vx_spawn.c

@@ -6,7 +6,7 @@
 extern "C" {
 #endif
 
-#define NUM_CORES_MAX 8
+#define NUM_CORES_MAX 16
 
 typedef struct {
 	func_t function;

runtime/src/vx_start.S

@@ -57,18 +57,20 @@ vx_set_sp:
   # set global pointer register
   .option push
   .option norelax
-  la gp, __global_pointer$
+  la gp, __global_pointer
   .option pop
 
   # allocate stack region for a threads on the processor 
   # set stack pointer
-  csrr a1, CSR_GTID    # get global thread id
-  slli a1, a1, 10      # multiply by 1024
-  csrr a2, CSR_LTID    # get local thread id  
-  slli a2, a2, 2       # multiply by 4
-  la sp, __stack_top$  # load stack base address
+  la sp, __stack_top   # load stack base address
+  la a1, __stack_size  # stack size
+  #if SM_ENABLE
+  csrr a2, CSR_LTID    # get lobal thread id
+  #else
+  csrr a2, CSR_GTID    # get global thread id
+  #endif
+  mul  a1, a1, a2
   sub  sp, sp, a1      # sub thread block
-  add  sp, sp, a2      # reduce addr collision for perf
 
   # disable active warps except warp0
   csrr a3, CSR_LWID    # get local wid