Add convergent attribute to vx_barrier

Note this attribute is only supported by Clang, so this will only be
applied to the kernel binary but not runtime.

kernel/include/gemmini_mmio.h

@@ -9,6 +9,8 @@
 // #define SMEM_SIZE 0x4000
 // 64KB
 // #define SMEM_SIZE 0x10000
+// 128KB
+// #define SMEM_SIZE 0x20000
 // 256KB
 #define SMEM_SIZE 0x40000
 

kernel/include/vx_intrinsics.h

@@ -149,6 +149,7 @@ inline void vx_join(unsigned stack_ptr) {
 }
 
 // Warp Barrier
+__attribute__((convergent))
 inline void vx_barrier(unsigned barried_id, unsigned num_warps) {
     asm volatile (".insn r %0, 4, 0, x0, %1, %2" :: "i"(RISCV_CUSTOM0), "r"(barried_id), "r"(num_warps));
 }