diff --git a/kernel/src/arch/riscv32/boot/boot_k210.asm b/kernel/src/arch/riscv32/board/k210/boot.asm
similarity index 82%
rename from kernel/src/arch/riscv32/boot/boot_k210.asm
rename to kernel/src/arch/riscv32/board/k210/boot.asm
index 0518142..e708ac7 100644
--- a/kernel/src/arch/riscv32/boot/boot_k210.asm
+++ b/kernel/src/arch/riscv32/board/k210/boot.asm
@@ -1,14 +1,14 @@
     .section .text.boot
 boot:
-    //lui x1, 0x40000
-    //jalr x0, x1, 8
-
 	csrwi mie, 0
 	csrwi mip, 0
     csrwi mscratch, 0
 	csrwi medeleg, 0
 	csrwi mideleg, 0
-    csrwi mstatus, 0
+
+    // enable float unit
+	li t0, 0x00006000 // MSTATUS_FS
+    csrw mstatus, t0
 
     // uart init
     lui x1, 0x38000
diff --git a/kernel/src/arch/riscv32/board/k210/linker.ld b/kernel/src/arch/riscv32/board/k210/linker.ld
index 4c38522..f318b99 100644
--- a/kernel/src/arch/riscv32/board/k210/linker.ld
+++ b/kernel/src/arch/riscv32/board/k210/linker.ld
@@ -41,7 +41,7 @@ SECTIONS
 
     .bss : {
         sbss = .;
-        *(.bss .bss.*)
+        *(.bss .bss.* .sbss.*)
         ebss = .;
     }
 
diff --git a/kernel/src/arch/riscv32/boot/trap.asm b/kernel/src/arch/riscv32/boot/trap.asm
index 04d206d..d32216c 100644
--- a/kernel/src/arch/riscv32/boot/trap.asm
+++ b/kernel/src/arch/riscv32/boot/trap.asm
@@ -5,6 +5,9 @@
 #   xcause
 #   xtval
 #   XRET
+#   XLENB
+#   LOAD
+#   STORE
 
 .macro SAVE_ALL
     # If coming from userspace, preserve the user stack pointer and load
@@ -19,37 +22,37 @@ _save_context:
     # provide room for trap frame
     addi sp, sp, -36 * XLENB
     # save x registers except x2 (sp)
-    sw x1, 1*XLENB(sp)
-    sw x3, 3*XLENB(sp)
+    STORE x1, 1
+    STORE x3, 3
     # tp(x4) = hartid. DON'T change.
-    # sw x4, 4*XLENB(sp)
-    sw x5, 5*XLENB(sp)
-    sw x6, 6*XLENB(sp)
-    sw x7, 7*XLENB(sp)
-    sw x8, 8*XLENB(sp)
-    sw x9, 9*XLENB(sp)
-    sw x10, 10*XLENB(sp)
-    sw x11, 11*XLENB(sp)
-    sw x12, 12*XLENB(sp)
-    sw x13, 13*XLENB(sp)
-    sw x14, 14*XLENB(sp)
-    sw x15, 15*XLENB(sp)
-    sw x16, 16*XLENB(sp)
-    sw x17, 17*XLENB(sp)
-    sw x18, 18*XLENB(sp)
-    sw x19, 19*XLENB(sp)
-    sw x20, 20*XLENB(sp)
-    sw x21, 21*XLENB(sp)
-    sw x22, 22*XLENB(sp)
-    sw x23, 23*XLENB(sp)
-    sw x24, 24*XLENB(sp)
-    sw x25, 25*XLENB(sp)
-    sw x26, 26*XLENB(sp)
-    sw x27, 27*XLENB(sp)
-    sw x28, 28*XLENB(sp)
-    sw x29, 29*XLENB(sp)
-    sw x30, 30*XLENB(sp)
-    sw x31, 31*XLENB(sp)
+    # STORE x4, 4
+    STORE x5, 5
+    STORE x6, 6
+    STORE x7, 7
+    STORE x8, 8
+    STORE x9, 9
+    STORE x10, 10
+    STORE x11, 11
+    STORE x12, 12
+    STORE x13, 13
+    STORE x14, 14
+    STORE x15, 15
+    STORE x16, 16
+    STORE x17, 17
+    STORE x18, 18
+    STORE x19, 19
+    STORE x20, 20
+    STORE x21, 21
+    STORE x22, 22
+    STORE x23, 23
+    STORE x24, 24
+    STORE x25, 25
+    STORE x26, 26
+    STORE x27, 27
+    STORE x28, 28
+    STORE x29, 29
+    STORE x30, 30
+    STORE x31, 31
 
     # get sp, sstatus, sepc, stval, scause
     # set sscratch = 0
@@ -59,16 +62,16 @@ _save_context:
     csrr s3, (xtval)
     csrr s4, (xcause)
     # store sp, sstatus, sepc, sbadvaddr, scause
-    sw s0, 2*XLENB(sp)
-    sw s1, 32*XLENB(sp)
-    sw s2, 33*XLENB(sp)
-    sw s3, 34*XLENB(sp)
-    sw s4, 35*XLENB(sp)
+    STORE s0, 2
+    STORE s1, 32
+    STORE s2, 33
+    STORE s3, 34
+    STORE s4, 35
 .endm
 
 .macro RESTORE_ALL
-    lw s1, 32*XLENB(sp)             # s1 = sstatus
-    lw s2, 33*XLENB(sp)             # s2 = sepc
+    LOAD s1, 32             # s1 = sstatus
+    LOAD s2, 33             # s2 = sepc
     andi s0, s1, 1 << 8
     bnez s0, _restore_context   # back to S-mode? (sstatus.SPP = 1)
 _save_kernel_sp:
@@ -80,38 +83,38 @@ _restore_context:
     csrw (xepc), s2
 
     # restore x registers except x2 (sp)
-    lw x1, 1*XLENB(sp)
-    lw x3, 3*XLENB(sp)
-    # lw x4, 4*XLENB(sp)
-    lw x5, 5*XLENB(sp)
-    lw x6, 6*XLENB(sp)
-    lw x7, 7*XLENB(sp)
-    lw x8, 8*XLENB(sp)
-    lw x9, 9*XLENB(sp)
-    lw x10, 10*XLENB(sp)
-    lw x11, 11*XLENB(sp)
-    lw x12, 12*XLENB(sp)
-    lw x13, 13*XLENB(sp)
-    lw x14, 14*XLENB(sp)
-    lw x15, 15*XLENB(sp)
-    lw x16, 16*XLENB(sp)
-    lw x17, 17*XLENB(sp)
-    lw x18, 18*XLENB(sp)
-    lw x19, 19*XLENB(sp)
-    lw x20, 20*XLENB(sp)
-    lw x21, 21*XLENB(sp)
-    lw x22, 22*XLENB(sp)
-    lw x23, 23*XLENB(sp)
-    lw x24, 24*XLENB(sp)
-    lw x25, 25*XLENB(sp)
-    lw x26, 26*XLENB(sp)
-    lw x27, 27*XLENB(sp)
-    lw x28, 28*XLENB(sp)
-    lw x29, 29*XLENB(sp)
-    lw x30, 30*XLENB(sp)
-    lw x31, 31*XLENB(sp)
+    LOAD x1, 1
+    LOAD x3, 3
+    # LOAD x4, 4
+    LOAD x5, 5
+    LOAD x6, 6
+    LOAD x7, 7
+    LOAD x8, 8
+    LOAD x9, 9
+    LOAD x10, 10
+    LOAD x11, 11
+    LOAD x12, 12
+    LOAD x13, 13
+    LOAD x14, 14
+    LOAD x15, 15
+    LOAD x16, 16
+    LOAD x17, 17
+    LOAD x18, 18
+    LOAD x19, 19
+    LOAD x20, 20
+    LOAD x21, 21
+    LOAD x22, 22
+    LOAD x23, 23
+    LOAD x24, 24
+    LOAD x25, 25
+    LOAD x26, 26
+    LOAD x27, 27
+    LOAD x28, 28
+    LOAD x29, 29
+    LOAD x30, 30
+    LOAD x31, 31
     # restore sp last
-    lw x2, 2*XLENB(sp)
+    LOAD x2, 2
 .endm
 
     .section .text
diff --git a/kernel/src/arch/riscv32/mod.rs b/kernel/src/arch/riscv32/mod.rs
index 32b7dd8..ac9b816 100644
--- a/kernel/src/arch/riscv32/mod.rs
+++ b/kernel/src/arch/riscv32/mod.rs
@@ -67,16 +67,28 @@ global_asm!("
 global_asm!("
     .equ XLENB,     4
     .equ XLENb,     32
+    .macro LOAD a1, a2
+        lw \\a1, \\a2*XLENB(sp)
+    .endm
+    .macro STORE a1, a2
+        sw \\a1, \\a2*XLENB(sp)
+    .endm
 ");
 #[cfg(target_pointer_width = "64")]
 global_asm!("
     .equ XLENB,     8
     .equ XLENb,     64
+    .macro LOAD a1, a2
+        ld \\a1, \\a2*XLENB(sp)
+    .endm
+    .macro STORE a1, a2
+        sd \\a1, \\a2*XLENB(sp)
+    .endm
 ");
 
 
 #[cfg(feature = "board_k210")]
-global_asm!(include_str!("boot/boot_k210.asm"));
+global_asm!(include_str!("board/k210/boot.asm"));
 global_asm!(include_str!("boot/entry.asm"));
 global_asm!(include_str!("boot/trap.asm"));