minor updates

driver/tests/tex_demo/kernel.c

@@ -1,6 +1,5 @@
 #include <stdint.h>
 #include <vx_intrinsics.h>
-#include <vx_tex.h>
 #include "common.h"
 
 uint32_t ilog2 (uint32_t value) {

runtime/include/vx_intrinsics.h

@@ -52,6 +52,13 @@ extern "C" {
 	__asm__ __volatile__ ("csrc " __ASM_STR(csr) ", %0"	:: "rK" (__v) : "memory"); \
 })
 
+// Texture load
+#define vx_tex(unit, u, v, l)    ({             \
+	register unsigned __r;		                \
+    __asm__ __volatile__ (".insn r4 0x6b, 5, " __ASM_STR(unit) ", %0, %1, %2, %3" : "=r"(__r) : "r"(u), "r"(v), "r"(l)); \
+	__r;							            \
+})
+
 // Set thread mask
 inline void vx_tmc(unsigned num_threads) {
     asm volatile (".insn s 0x6b, 0, x0, 0(%0)" :: "r"(num_threads));

runtime/include/vx_tex.h

@@ -1,19 +0,0 @@
-#ifndef VX_TEX_H
-#define VX_TEX_H
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-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
-}
-#endif
-
-#endif