diff --git a/src/layer/riscv/convolution_sgemm_packn_fp16s.h b/src/layer/riscv/convolution_sgemm_packn_fp16s.h
index 3a8ef2c39..82086acea 100644
--- a/src/layer/riscv/convolution_sgemm_packn_fp16s.h
+++ b/src/layer/riscv/convolution_sgemm_packn_fp16s.h
@@ -55,20 +55,35 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
                 for (int k = 0; k < maxk; k++)
                 {
 #if RVV_SPEC_0_7
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr[4] = img0[l + packn * 4];
-                        tmpptr[5] = img0[l + packn * 5];
-                        tmpptr[6] = img0[l + packn * 6];
-                        tmpptr[7] = img0[l + packn * 7];
-                        tmpptr += 8;
-                    }
+                    asm volatile(
+                        "mv        t3,     %[LEN]  \n\t"
+                        "mv        t1,     %[SRC]  \n\t"
+                        "mv        t2,     %[TMP]  \n\t"
+                        "slli      t3,     t3,     1       \n\t"
+                        "vle.v     v0,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v1,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v2,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v3,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v4,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v5,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v6,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v7,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vsseg8e.v v0,     (t2)    \n\t"
+                        :
+                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
+                        : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "t1", "t2", "t3");
 
                     img0 += size * packn;
+                    tmpptr += packn * 8;
+
 #else
                     vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
                     vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
@@ -104,16 +119,25 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
                 for (int k = 0; k < maxk; k++)
                 {
 #if RVV_SPEC_0_7
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr[2] = img0[l + packn * 2];
-                        tmpptr[3] = img0[l + packn * 3];
-                        tmpptr += 4;
-                    }
+                    asm volatile(
+                        "mv        t3,     %[LEN]  \n\t"
+                        "mv        t1,     %[SRC]  \n\t"
+                        "mv        t2,     %[TMP]  \n\t"
+                        "slli      t3,     t3,     1       \n\t"
+                        "vle.v     v0,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v1,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v2,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v3,     (t1)    \n\t"
+                        "vsseg4e.v v0,     (t2)    \n\t"
+                        :
+                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
+                        : "cc", "memory", "v0", "v1", "v2", "v3", "t1", "t2", "t3");
 
                     img0 += size * packn;
+                    tmpptr += packn * 4;
 #else
                     vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
                     vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);
@@ -146,14 +170,21 @@ static void im2col_sgemm_packn_fp16sa_rvv(const Mat& bottom_im2col, Mat& top_blo
                 for (int k = 0; k < maxk; k++)
                 {
 #if RVV_SPEC_0_7
-                    for (int l = 0; l < packn; l++)
-                    {
-                        tmpptr[0] = img0[l];
-                        tmpptr[1] = img0[l + packn];
-                        tmpptr += 2;
-                    }
+                    asm volatile(
+                        "mv        t3,     %[LEN]  \n\t"
+                        "mv        t1,     %[SRC]  \n\t"
+                        "mv        t2,     %[TMP]  \n\t"
+                        "slli      t3,     t3,     1       \n\t"
+                        "vle.v     v0,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        "vle.v     v1,     (t1)    \n\t"
+                        "add       t1,     t1,     t3      \n\t"
+                        :
+                        : [LEN] "r"(packn), [SRC] "r"(img0), [TMP] "r"(tmpptr)
+                        : "cc", "memory", "v0", "v1", "t1", "t2", "t3");
 
                     img0 += size * packn;
+                    tmpptr += packn * 2;
 #else
                     vfloat16m1_t _val0 = vle16_v_f16m1(img0, vl);
                     vfloat16m1_t _val1 = vle16_v_f16m1(img0 + packn, vl);