pipeline refactoring: centralized issue buffer

runtime/src/vx_intrinsics.S

@@ -83,11 +83,11 @@ vx_num_cores:
 .type vx_num_cycles, @function
 .global vx_num_cycles
 vx_num_cycles:
-    csrr a0, CSR_CYCLL
+    csrr a0, CSR_CYCLE_L
     ret
     
 .type vx_num_instrs, @function
 .global vx_num_instrs
 vx_num_instrs:
-    csrr a0, CSR_INSTL
+    csrr a0, CSR_INSTR_L
     ret

runtime/src/vx_start.S

@@ -49,7 +49,7 @@ vx_set_sp:
   slli a1, a1, 10      # multiply by 1024
   csrr a2, CSR_LTID    # get local thread id  
   slli a2, a2, 2       # multiply by 4
-  lui  sp, STACK_BASE_ADDR # load base sp
+  lui  sp, (SHARED_MEM_BASE_ADDR>>12) # load base sp
   sub  sp, sp, a1      # sub thread block
   add  sp, sp, a2      # reduce addr collision for perf
 

runtime/tests/simple/Makefile

@@ -8,7 +8,7 @@ CP = $(RISCV_TOOLCHAIN_PATH)/bin/riscv32-unknown-elf-objcopy
 
 CFLAGS += -march=rv32im -mabi=ilp32 -O3 -Wl,-Bstatic,-T,$(VORTEX_RT_PATH)/linker/vx_link.ld
 CFLAGS += -nostartfiles -ffreestanding -fno-exceptions -Wl,--gc-sections 
-CFLAGS += -I$(VORTEX_RT_PATH)/include
+CFLAGS += -I$(VORTEX_RT_PATH)/include -I../../../hw
 
 LDFLAGS += $(VORTEX_RT_PATH)/libvortexrt.a
 

runtime/tests/simple/main.c

@@ -3,6 +3,7 @@
 #include <vx_intrinsics.h>
 #include <vx_print.h>
 #include <vx_spawn.h>
+#include <VX_config.h>
 
 
 typedef struct
@@ -100,7 +101,7 @@ int main()
 	test_wsapwn();
 
 	vx_print_str("Shared Memory test\n");
-	unsigned * ptr = (unsigned *) 0xFFFF0000;
+	unsigned * ptr = (unsigned *) SHARED_MEM_BASE_ADDR;
 	unsigned value = 0;
 	for (int i = 0; i < 5; i++)
 	{
@@ -112,7 +113,6 @@ int main()
 		vx_print_str("-------------------\n");
 		value++;
 		ptr++;
-
 	}
 
 	vx_print_str("vx_spawn_warps mat_add_kernel\n");

runtime/tests/simple/tests.c

@@ -125,16 +125,4 @@ void intrinsics_tests()
 	// Test wspawn
 	vx_print_str("test_spawn\n");
 	test_wsapwn();
-}
-
-
-
-
-
-
-
-
-
-
-
-
+}

runtime/tests/simple/vx_simple.dump

@@ -81,14 +81,14 @@ Disassembly of section .text:
 80000114:	3a8000ef          	jal	ra,800004bc <test_wsapwn>
 80000118:	80001537          	lui	a0,0x80001
 8000011c:	ddc50513          	addi	a0,a0,-548 # 80000ddc <__global_pointer$+0xfffff5d4>
-80000120:	ffff0437          	lui	s0,0xffff0
+80000120:	6ffff437          	lui	s0,0x6ffff
 80000124:	628000ef          	jal	ra,8000074c <vx_print_str>
 80000128:	00000493          	li	s1,0
 8000012c:	80001b37          	lui	s6,0x80001
 80000130:	80001ab7          	lui	s5,0x80001
 80000134:	80001a37          	lui	s4,0x80001
 80000138:	800019b7          	lui	s3,0x80001
-8000013c:	01440913          	addi	s2,s0,20 # ffff0014 <__global_pointer$+0x7ffee80c>
+8000013c:	01440913          	addi	s2,s0,20 # 6ffff014 <_start-0x10000fec>
 80000140:	00942023          	sw	s1,0(s0)
 80000144:	00040593          	mv	a1,s0
 80000148:	df0b0513          	addi	a0,s6,-528 # 80000df0 <__global_pointer$+0xfffff5e8>

runtime/tests/simple/vx_simple.hex

@@ -17,7 +17,7 @@
 :1000E80037150080130585DBEF00C0651305400058
 :1000F800EF00405EEF00402E13051000EF00805D1A
 :10010800371500801305C5DCEF00C063EF00803AA7
-:10011800371500801305C5DD3704FFFFEF00806247
+:10011800371500801305C5DD37F4FF6FEF008062E7
 :1001280093040000371B0080B71A0080371A00803C
 :10013800B719008013094401232094009305040093
 :1001480013050BDFEF00006E9385040013858ADF2B