rtl refactoring

runtime/intrinsics/vx_intrinsics.S

@@ -0,0 +1,69 @@
+#include "../config.h"
+
+.section .text
+
+.type vx_wspawn, @function
+.global vx_wspawn
+vx_wspawn:
+	.word 0x00b5106b # wspawn a0(numWarps), a1(PC SPAWN)
+	ret
+
+.type vx_tmc, @function
+.global vx_tmc
+vx_tmc:
+	.word 0x0005006b    # tmc a0
+	ret
+
+.type vx_barrier, @function
+.global vx_barrier
+vx_barrier:
+	.word 0x00b5406b # barrier a0(barrier id), a1(numWarps)
+	ret
+
+.type vx_split, @function
+.global vx_split
+vx_split:
+	.word 0x0005206b    # split a0
+	ret
+
+.type vx_join, @function
+.global vx_join
+vx_join:
+	.word 0x0000306b    #join
+	ret
+
+.type vx_warp_id, @function
+.global vx_warp_id
+vx_warp_id:
+	csrr a0, CSR_LWID # read warp index
+	ret
+
+.type vx_warp_gid, @function
+.global vx_warp_gid
+vx_warp_gid:
+	csrr a0, CSR_GWID # read warp index
+	ret
+
+.type vx_thread_id, @function
+.global vx_thread_id
+vx_thread_id:
+	csrr a0, CSR_LTID # read thread index
+	ret
+
+.type vx_thread_gid, @function
+.global vx_thread_gid
+vx_thread_gid:
+	csrr a0, CSR_GTID # read thread index
+	ret
+	
+.type vx_num_cycles, @function
+.global vx_num_cycles
+vx_num_cycles:
+    csrr a0, CSR_CYCLL
+    ret
+    
+.type vx_num_instrs, @function
+.global vx_num_instrs
+vx_num_instrs:
+    csrr a0, CSR_INSTL
+    ret

runtime/intrinsics/vx_intrinsics.h

@@ -7,35 +7,38 @@
 extern "C" {
 #endif
 
-// Spawns Warps
+// Spawn warps
 void vx_wspawn(unsigned numWarps, unsigned PC_spawn);
 
-// Changes thread mask (activated/deactivates threads)
+// Set thread mask
 void vx_tmc(unsigned numThreads);
 
 // Warp Barrier
 void vx_barrier(unsigned barriedID, unsigned numWarps);
 
-// split on a predicate
+// Split on a predicate
 void vx_split(unsigned predicate);
 
 // Join
 void vx_join(void);
 
-// Get Hardware thread ID
-unsigned vx_threadID(void);
+// Return the warp thread index
+unsigned vx_thread_id(void);
 
-// Get hardware warp ID
-unsigned vx_warpID(void);
+// Return the core warp index
+unsigned vx_warp_id(void);
 
-// Get global warp number
-unsigned vx_warpNum(void);
+// Return processsor unique thread id
+unsigned vx_thread_gid(void);
 
-// Get Number cycles/Inst
-unsigned vx_getCycles(void);
-unsigned vx_getInst(void);
+// Return processsor unique warp id
+unsigned vx_warp_gid(void);
 
-void vx_resetStack(void);
+// Return number cycles
+unsigned vx_num_cycles(void);
+
+// Return number instructions
+unsigned vx_num_instrs(void);
 
 #define __if(b) vx_split(b); \
                 if (b) 

runtime/intrinsics/vx_intrinsics.s

@@ -1,85 +0,0 @@
-.section .text
-
-.type vx_wspawn, @function
-.global vx_wspawn
-vx_wspawn:
-	.word 0x00b5106b # wspawn a0(numWarps), a1(PC SPAWN)
-	ret
-
-.type vx_tmc, @function
-.global vx_tmc
-vx_tmc:
-	.word 0x0005006b    # tmc a0
-	ret
-
-.type vx_barrier, @function
-.global vx_barrier
-vx_barrier:
-	.word 0x00b5406b # barrier a0(barrier id), a1(numWarps)
-	ret
-
-.type vx_split, @function
-.global vx_split
-vx_split:
-	.word 0x0005206b    # split a0
-	ret
-
-.type vx_join, @function
-.global vx_join
-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:
-    csrr a0, 0x22 # read warp IDs
-    ret
-
-.type vx_threadID, @function
-.global vx_threadID
-vx_threadID:
-	csrr a0, 0x20 # read thread IDs
-	ret
-	
-.type vx_getCycles, @function
-.global vx_getCycles
-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:
-    li a0, 4
-    .word 0x0005006b    # tmc 4
-
-    csrr a3, 0x21        # get wid
-    slli a3, a3, 15      # shift by wid
-    csrr a2, 0x20        # get tid
-    slli a1, a2, 10      # multiply tid by 1024
-    slli a2, a2, 2       # multiply tid by 4
-    lui  sp, 0x6ffff     # load base sp
-    sub  sp, sp, a1      # sub sp - (1024*tid)
-    sub  sp, sp, a3      # shoft per warp
-    add  sp, sp, a2      # shift sp for better performance
-
-    csrr a3, 0x21        # get wid
-    beqz a3, RETURN
-    li a0, 0
-    .word 0x0005006b    # tmc 0
-RETURN:
-    ret
-