many fixes

runtime/Makefile

@@ -10,7 +10,7 @@ CFLAGS += -I./include -I../hw
 
 PROJECT = libvortexrt
 
-SRCS = ./src/vx_start.S ./src/vx_print.S ./src/vx_print.c ./src/vx_spawn.c ./src/vx_tex.c 
+SRCS = ./src/vx_start.S ./src/vx_print.S ./src/vx_print.c ./src/vx_spawn.c
 
 OBJS := $(addsuffix .o, $(notdir $(SRCS)))
 

runtime/include/vx_intrinsics.h

@@ -74,16 +74,9 @@ inline void vx_join() {
 
 // Warp Barrier
 inline void vx_barrier(unsigned barried_id, unsigned num_warps) {
-    asm volatile (".insn s 0x6b, 4, %1, 0cd (%0)" :: "r"(barried_id), "r"(num_warps));
+    asm volatile (".insn s 0x6b, 4, %1, 0(%0)" :: "r"(barried_id), "r"(num_warps));
 }
 
-// Texture load
-#define vx_tex_ld(unit, u, v, lod)	({              \
-	register unsigned result;				    \
-    asm volatile (".insn r4 0x6b, 5, " __ASM_STR(unit) ", %0, %1, %2, %3" :: "r"(result), "r"(u), "r"(v), "r"(lod)); \
-	result;							            \
-})
-
 // Return active warp's thread id 
 inline int vx_thread_id() {
     int result;

runtime/include/vx_tex.h

@@ -1,14 +1,16 @@
-#ifndef VX_API_H
-#define VX_API_H
-
-#include <stdint.h>
-#include <stdio.h>
+#ifndef VX_TEX_H
+#define VX_TEX_H
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-int vx_tex(unsigned t, unsigned u, unsigned v, unsigned lod);
+unsigned vx_tex(unsigned unit, unsigned u, unsigned v, unsigned lod) {
+    unsigned result;
+    unsigned lod_unit = (unit << 24) | lod;
+    asm volatile (".insn r4 0x6b, 5, 0, %0, %1, %2, %3" : "=r"(result) : "r"(u), "r"(v), "r"(lod_unit));
+    return result;
+}
 
 #ifdef __cplusplus
 }

runtime/libvortexrt.dump

@@ -1028,46 +1028,3 @@ Disassembly of section .riscv.attributes:
   1c:	326d                	jal	fffff9c6 <.L50+0xfffff79e>
   1e:	3070                	fld	fa2,224(s0)
   20:	665f 7032 0030      	0x307032665f
-
-vx_tex.c.o:     file format elf32-littleriscv
-
-
-Disassembly of section .text.vx_tex:
-
-00000000 <vx_tex>:
-   0:	00869693          	slli	a3,a3,0x8
-   4:	00a6e6b3          	or	a3,a3,a0
-   8:	00000513          	li	a0,0
-   c:	6ac5d56b          	0x6ac5d56b
-  10:	00008067          	ret
-
-Disassembly of section .comment:
-
-00000000 <.comment>:
-   0:	4700                	lw	s0,8(a4)
-   2:	203a4343          	fmadd.s	ft6,fs4,ft3,ft4,rmm
-   6:	4728                	lw	a0,72(a4)
-   8:	554e                	lw	a0,240(sp)
-   a:	2029                	jal	14 <vx_tex+0x14>
-   c:	2e39                	jal	32a <vx_tex+0x32a>
-   e:	2e32                	fld	ft8,264(sp)
-  10:	0030                	addi	a2,sp,8
-
-Disassembly of section .riscv.attributes:
-
-00000000 <.riscv.attributes>:
-   0:	2541                	jal	680 <vx_tex+0x680>
-   2:	0000                	unimp
-   4:	7200                	flw	fs0,32(a2)
-   6:	7369                	lui	t1,0xffffa
-   8:	01007663          	bgeu	zero,a6,14 <vx_tex+0x14>
-   c:	0000001b          	0x1b
-  10:	1004                	addi	s1,sp,32
-  12:	7205                	lui	tp,0xfffe1
-  14:	3376                	fld	ft6,376(sp)
-  16:	6932                	flw	fs2,12(sp)
-  18:	7032                	flw	ft0,44(sp)
-  1a:	5f30                	lw	a2,120(a4)
-  1c:	326d                	jal	fffff9c6 <vx_tex+0xfffff9c6>
-  1e:	3070                	fld	fa2,224(s0)
-  20:	665f 7032 0030      	0x307032665f

runtime/src/vx_tex.S

@@ -1,9 +0,0 @@
-#include <VX_config.h>
-
-@ .type vx_tex_ld, @function
-@ .global vx_tex_ld
-@ vx_tex_ld:
-@     slli	a1,a1,0x8
-@     or	a1,a1,a0
-@     .word 0x5ae7952b
-@     ret

runtime/src/vx_tex.c

@@ -1,13 +0,0 @@
-#include <vx_tex.h>
-#include <vx_intrinsics.h>
-#include <inttypes.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define NUM_CORES_MAX 32
-
-int vx_tex(unsigned t, unsigned u, unsigned v, unsigned lod){
-    return vx_tex_ld(t,u,v,lod);
-}