added debug print states or rtl

runtime/intrinsics/vx_intrinsics.S

@@ -5,7 +5,7 @@
 .type vx_wspawn, @function
 .global vx_wspawn
 vx_wspawn:
-	.word 0x00b5106b # wspawn a0(numWarps), a1(PC SPAWN)
+	.word 0x00b5106b # wspawn a0(num_warps), a1(func_ptr)
 	ret
 
 .type vx_tmc, @function
@@ -17,7 +17,7 @@ vx_tmc:
 .type vx_barrier, @function
 .global vx_barrier
 vx_barrier:
-	.word 0x00b5406b # barrier a0(barrier id), a1(numWarps)
+	.word 0x00b5406b # barrier a0(barrier_id), a1(num_warps)
 	ret
 
 .type vx_split, @function

runtime/intrinsics/vx_intrinsics.h

@@ -1,20 +1,18 @@
-
-#ifndef VX_INTRINSICS
-
-#define VX_INTRINSICS
+#ifndef VX_INTRINSICS_H
+#define VX_INTRINSICS_H
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 // Spawn warps
-void vx_wspawn(int numWarps, int PC_spawn);
+void vx_wspawn(int num_warps, unsigned func_ptr);
 
 // Set thread mask
-void vx_tmc(int numThreads);
+void vx_tmc(int num_threads);
 
 // Warp Barrier
-void vx_barrier(int barriedID, int numWarps);
+void vx_barrier(int barried_id, int num_warps);
 
 // Split on a predicate
 void vx_split(int predicate);

runtime/io/vx_io.h

@@ -1,5 +1,5 @@
-
-#pragma once
+#ifndef VX_IO_H
+#define VX_IO_H
 
 #include <stdbool.h>
 
@@ -15,7 +15,8 @@ void vx_printf(const char *, unsigned);
 void vx_print_str(const char *);
 void vx_printc(unsigned, char c);
 
-
 #ifdef __cplusplus
 }
+#endif
+
 #endif

runtime/tests/simple/vx_simple_main.dump

@@ -625,7 +625,7 @@ Disassembly of section .text:
 8000083c:	28d1a023          	sw	a3,640(gp) # 80016a88 <global_argument_struct>
 80000840:	26b1ae23          	sw	a1,636(gp) # 80016a84 <global_num_threads>
 80000844:	00100793          	li	a5,1
-80000848:	00a7fa63          	bgeu	a5,a0,8000085c <vx_spawn_warps+0x38>
+80000848:	00a7da63          	bge	a5,a0,8000085c <vx_spawn_warps+0x38>
 8000084c:	800005b7          	lui	a1,0x80000
 80000850:	7a058593          	addi	a1,a1,1952 # 800007a0 <__BSS_END__+0xfffe9c78>
 80000854:	d55ff0ef          	jal	ra,800005a8 <vx_wspawn>

runtime/tests/simple/vx_simple_main.hex

@@ -132,7 +132,7 @@
 :1008180013351500130101016FF01FD9130101FFF2
 :100828002324810023229100232021012326110063
 :1008380023A2C12823A0D12823AEB12693071000F4
-:1008480063FAA700B70500809385057AEFF05FD5B6
+:1008480063DAA700B70500809385057AEFF05FD5D6
 :1008580083A5C12713850500EFF01FD503A501283F
 :1008680083A74128E7800700EFF01FD60324810003
 :100878008320C1008324410003290100133515009A

runtime/vx_api/vx_api.c

@@ -9,7 +9,7 @@ extern "C" {
 
 func_t global_function_pointer;
 void * global_argument_struct;
-unsigned global_num_threads;
+int global_num_threads;
 
 void spawn_warp_runonce() {
 	// active all threads
@@ -19,12 +19,12 @@ void spawn_warp_runonce() {
 	global_function_pointer(global_argument_struct);
 
 	// resume single-thread execution on exit
-	unsigned wid = vx_warp_id();
+	int wid = vx_warp_id();
 	unsigned tmask = (0 == wid) ? 0x1 : 0x0; 
 	vx_tmc(tmask);
 }
 
-void vx_spawn_warps(unsigned numWarps, unsigned numThreads, func_t func_ptr, void * args) {
+void vx_spawn_warps(int numWarps, int numThreads, func_t func_ptr, void * args) {
 	global_function_pointer = func_ptr;
 	global_argument_struct  = args;
 	global_num_threads      = numThreads;
@@ -34,7 +34,7 @@ void vx_spawn_warps(unsigned numWarps, unsigned numThreads, func_t func_ptr, voi
 	spawn_warp_runonce();
 }
 
-unsigned               pocl_threads;
+int                    pocl_threads;
 struct context_t *     pocl_ctx;
 vx_pocl_workgroup_func pocl_pfn;
 const void *           pocl_args;

runtime/vx_api/vx_api.h

@@ -1,6 +1,5 @@
-
-#ifndef VX_API_
-#define VX_API_
+#ifndef VX_API_H
+#define VX_API_H
 
 #include <stdint.h>
 #include <stdio.h>
@@ -11,7 +10,7 @@ extern "C" {
 
 typedef void (*func_t)(void *);
 
-void vx_spawn_warps(unsigned numWarps, unsigned numThreads, func_t func_ptr , void * args);
+void vx_spawn_warps(int num_warps, int num_threads, func_t func_ptr , void * args);
 
 struct context_t {
   uint32_t num_groups[3];