redesigned driver demo, fixed startup code, removed --cpu from simx,

runtime/intrinsics/vx_intrinsics.h

@@ -8,29 +8,27 @@ extern "C" {
 #endif
 
 // Spawns Warps
-void vx_wspawn (unsigned numWarps, unsigned PC_spawn);
+void vx_wspawn(unsigned numWarps, unsigned PC_spawn);
 
 // Changes thread mask (activated/deactivates threads)
-void vx_tmc    (unsigned numThreads);
+void vx_tmc(unsigned numThreads);
 
 // Warp Barrier
 void vx_barrier(unsigned barriedID, unsigned numWarps);
 
 // split on a predicate
-void vx_split  (unsigned predicate);
-
+void vx_split(unsigned predicate);
 
 // Join
-void vx_join   (void);
-
+void vx_join(void);
 
 // Get Hardware thread ID
 unsigned vx_threadID(void);
 
-
 // Get hardware warp ID
 unsigned vx_warpID(void);
 
+// Get global warp number
 unsigned vx_warpNum(void);
 
 // Get Number cycles/Inst
@@ -39,16 +37,13 @@ unsigned vx_getInst(void);
 
 void vx_resetStack(void);
 
-
 #define __if(b) vx_split(b); \
                 if (b) 
 
 #define __else else
 
-
 #define __endif vx_join();
 
-
 #ifdef __cplusplus
 }
 #endif

runtime/intrinsics/vx_intrinsics.s

@@ -1,9 +1,5 @@
-
-
-
 .section .text
 
-
 .type vx_wspawn, @function
 .global vx_wspawn
 vx_wspawn:
@@ -16,7 +12,6 @@ vx_tmc:
 	.word 0x0005006b    # tmc a0
 	ret
 
-
 .type vx_barrier, @function
 .global vx_barrier
 vx_barrier:
@@ -35,12 +30,12 @@ vx_join:
 	.word 0x0000306b    #join
 	ret
 
-
 .type vx_warpID, @function
 .global vx_warpID
 vx_warpID:
 	csrr a0, 0x21 # read warp IDs
 	ret
+
 .type vx_warpNum, @function
 .global vx_warpNum
 vx_warpNum:
@@ -59,14 +54,12 @@ vx_getCycles:
     csrr a0, 0x26 # read thread IDs
     ret
     
-
 .type vx_getInst, @function
 .global vx_getInst
 vx_getInst:
     csrr a0, 0x25 # read thread IDs
     ret
     
-
 .type vx_resetStack, @function
 .global vx_resetStack
 vx_resetStack:
@@ -89,4 +82,4 @@ vx_resetStack:
     .word 0x0005006b    # tmc 0
 RETURN:
     ret
-
+