diff --git a/kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c b/kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c
index 6e7b06884d..f83abb2339 100644
--- a/kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c
+++ b/kernel/riscv64/sbgemm_kernel_16x8_zvl256b.c
@@ -1,22 +1,90 @@
 #include "common.h"
 #include <riscv_vector.h>
 
+#define BF16_WIDEN_ONE  // Convert pre-hand and do operations in FP32
+#define USE_BF16_CVT    // Comment out for pre-RVA23 systems like BananaPi
+
+#ifdef BF16_WIDEN_ONE
+#define FORCEINLINE      inline __attribute__((always_inline))
+#define B_UNROLL         64
+
+// Convert from BF16 to FP32
+static void FORCEINLINE B_CONV(__bf16 *BB, FLOAT *CONV, BLASLONG count)
+{
+    BLASLONG count2 = (count & (B_UNROLL - 1));
+    count &= -B_UNROLL;
+    while (count) {
+        vbfloat16m4_t B00 = __riscv_vle16_v_bf16m4(BB, B_UNROLL);
+#ifdef USE_BF16_CVT
+        vfloat32m8_t B0 = __riscv_vfwcvtbf16_f_f_v_f32m8(B00, B_UNROLL);
+#else
+	vfloat32m8_t B0 = __riscv_vreinterpret_v_u32m8_f32m8(__riscv_vsll_vx_u32m8(
+            __riscv_vwcvtu_x_x_v_u32m8(__riscv_vreinterpret_v_bf16m4_u16m4(B00), B_UNROLL), 16, B_UNROLL));
+#endif
+        __riscv_vse32_v_f32m8(CONV, B0, B_UNROLL);
+        BB += B_UNROLL;
+        CONV += B_UNROLL;
+        count -= B_UNROLL;
+    }
+    if (count2) {
+        BLASLONG gvl2 = __riscv_vsetvl_e16m4(count2);
+        vbfloat16m4_t B00 = __riscv_vle16_v_bf16m4(BB, gvl2);
+#ifdef USE_BF16_CVT
+        vfloat32m8_t B0 = __riscv_vfwcvtbf16_f_f_v_f32m8(B00, gvl2);
+#else
+        vfloat32m8_t B0 = __riscv_vreinterpret_v_u32m8_f32m8(__riscv_vsll_vx_u32m8(
+            __riscv_vwcvtu_x_x_v_u32m8(__riscv_vreinterpret_v_bf16m4_u16m4(B00), gvl2), 16, gvl2));
+#endif
+        __riscv_vse32_v_f32m8(CONV, B0, gvl2);
+    }
+}
+#endif
+
+#ifndef VECTORIZE_MEMSET
+#define memset_zero(ptr, size, dir)  memset(ptr, 0, size)
+#else
+void memset_zero(void *input, BLASLONG size, bool dir);
+#endif
+
 int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
 {
     BLASLONG gvl = 0;
     BLASLONG m_top = 0;
     BLASLONG n_top = 0;
+#if !defined(BF16_WIDEN_ONE) || !defined(BF16_DONT_CONV)
     __bf16 *BB = (__bf16 *)(B);
     __bf16 *AA = (__bf16 *)(A);
+#endif
+
+#ifdef BF16_WIDEN_ONE
+    FLOAT *CONV = (FLOAT *)(malloc((K * (8 + M)) * sizeof(FLOAT)));
+    if (!CONV) return 1;
+#ifndef BF16_DONT_CONV
+    B_CONV(AA, CONV + (K * 8), K * M);
+#else
+    memset_zero(CONV, (K * (8 + M)) * sizeof(FLOAT), false);
+#endif
+#endif
 
     // -- MAIN PASS
     for (BLASLONG j=0; j<N/8; j+=1) {
         m_top = 0;
         BLASLONG gvl = __riscv_vsetvl_e16m1(16);
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 8;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
@@ -28,6 +96,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result7 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                float B4 = CONV[bi2+4];
+                float B5 = CONV[bi2+5];
+                float B6 = CONV[bi2+6];
+                float B7 = CONV[bi2+7];
+                bi2 += 8;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+		ai2 += 16;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f32m2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f32m2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f32m2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f32m2(result7, B7, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -49,6 +140,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmaccbf16_vf_f32m2(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmaccbf16_vf_f32m2(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmaccbf16_vf_f32m2(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -89,8 +181,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 8 ) {
             gvl = __riscv_vsetvl_e16mf2(8);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -102,6 +198,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                float B4 = CONV[bi2+4];
+                float B5 = CONV[bi2+5];
+                float B6 = CONV[bi2+6];
+                float B7 = CONV[bi2+7];
+                bi2 += 8;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -123,6 +242,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -161,8 +281,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16mf2(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -174,6 +298,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                float B4 = CONV[bi2+4];
+                float B5 = CONV[bi2+5];
+                float B6 = CONV[bi2+6];
+                float B7 = CONV[bi2+7];
+                bi2 += 8;
+
+                vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vle32_v_f32mf2(&CONV[ai2], gvl));
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -195,6 +342,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -254,10 +402,34 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result13 = 0;
             float result14 = 0;
             float result15 = 0;
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                result4+=CONV[ai2+0]*CONV[bi2+2];
+                result5+=CONV[ai2+1]*CONV[bi2+2];
+                result6+=CONV[ai2+0]*CONV[bi2+3];
+                result7+=CONV[ai2+1]*CONV[bi2+3];
+                result8+=CONV[ai2+0]*CONV[bi2+4];
+                result9+=CONV[ai2+1]*CONV[bi2+4];
+                result10+=CONV[ai2+0]*CONV[bi2+5];
+                result11+=CONV[ai2+1]*CONV[bi2+5];
+                result12+=CONV[ai2+0]*CONV[bi2+6];
+                result13+=CONV[ai2+1]*CONV[bi2+6];
+                result14+=CONV[ai2+0]*CONV[bi2+7];
+                result15+=CONV[ai2+1]*CONV[bi2+7];
+                ai2+=2;
+                bi2+=8;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
@@ -276,6 +448,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result15+=(float)(AA[ai+1])*(float)(BB[bi+7]);
                 ai+=2;
                 bi+=8;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -312,10 +485,26 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result6 = 0;
             float result7 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                result2+=CONV[ai2+0]*CONV[bi2+2];
+                result3+=CONV[ai2+0]*CONV[bi2+3];
+                result4+=CONV[ai2+0]*CONV[bi2+4];
+                result5+=CONV[ai2+0]*CONV[bi2+5];
+                result6+=CONV[ai2+0]*CONV[bi2+6];
+                result7+=CONV[ai2+0]*CONV[bi2+7];
+                ai2+=1;
+                bi2+=8;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
@@ -326,6 +515,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result7+=(float)(AA[ai+0])*(float)(BB[bi+7]);
                 ai+=1;
                 bi+=8;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -346,9 +536,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
 
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 4;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
+
         for (BLASLONG i=0; i<M/16; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
@@ -356,6 +558,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result3 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                bi2 += 4;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 16;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -369,6 +586,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmaccbf16_vf_f32m2(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmaccbf16_vf_f32m2(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -394,8 +612,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 8 ) {
             gvl = __riscv_vsetvl_e16mf2(8);
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -403,6 +625,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                bi2 += 4;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -416,6 +653,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -442,8 +680,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16mf2(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -451,6 +693,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                bi2 += 4;
+
+                vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vle32_v_f32mf2(&CONV[ai2], gvl));
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -464,6 +721,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -499,10 +757,26 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result5 = 0;
             float result6 = 0;
             float result7 = 0;
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                result4+=CONV[ai2+0]*CONV[bi2+2];
+                result5+=CONV[ai2+1]*CONV[bi2+2];
+                result6+=CONV[ai2+0]*CONV[bi2+3];
+                result7+=CONV[ai2+1]*CONV[bi2+3];
+                ai2+=2;
+                bi2+=4;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
@@ -513,6 +787,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result7+=(float)(AA[ai+1])*(float)(BB[bi+3]);
                 ai+=2;
                 bi+=4;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -535,16 +810,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result2 = 0;
             float result3 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                result2+=CONV[ai2+0]*CONV[bi2+2];
+                result3+=CONV[ai2+0]*CONV[bi2+3];
+                ai2+=1;
+                bi2+=4;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
                 result3+=(float)(AA[ai+0])*(float)(BB[bi+3]);
                 ai+=1;
                 bi+=4;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -562,15 +850,37 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
     if ( N & 2 ) {
         gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 2;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                bi2 += 2;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 16;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 bi += 2;
@@ -580,6 +890,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -598,13 +909,28 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 8 ) {
             gvl = __riscv_vsetvl_e16mf2(8);
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                bi2 += 2;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 bi += 2;
@@ -614,6 +940,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -634,13 +961,28 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16mf2(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                bi2 += 2;
+
+                vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vle32_v_f32mf2(&CONV[ai2], gvl));
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 bi += 2;
@@ -650,6 +992,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -673,16 +1016,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result1 = 0;
             float result2 = 0;
             float result3 = 0;
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                ai2+=2;
+                bi2+=2;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result3+=(float)(AA[ai+1])*(float)(BB[bi+1]);
                 ai+=2;
                 bi+=2;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -699,14 +1055,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result0 = 0;
             float result1 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                ai2+=1;
+                bi2+=2;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 ai+=1;
                 bi+=2;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -722,14 +1089,35 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
     if ( N & 1 ) {
         gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 1;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
+
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                bi2 += 1;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 16;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 bi += 1;
 
@@ -737,6 +1125,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 ai += 16;
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -753,12 +1142,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 8 ) {
             gvl = __riscv_vsetvl_e16mf2(8);
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                bi2 += 1;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 bi += 1;
 
@@ -766,6 +1168,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 ai += 8;
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -783,12 +1186,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16mf2(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                bi2 += 1;
+
+                vfloat32m1_t A0 = __riscv_vlmul_ext_v_f32mf2_f32m1(__riscv_vle32_v_f32mf2(&CONV[ai2], gvl));
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 bi += 1;
 
@@ -796,6 +1212,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 ai += 4;
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -812,14 +1229,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 2 ) {
             float result0 = 0;
             float result1 = 0;
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                ai2+=2;
+                bi2+=1;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 ai+=2;
                 bi+=1;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -833,13 +1261,23 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             float result0 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                ai2+=1;
+                bi2+=1;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 ai+=1;
                 bi+=1;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -849,5 +1287,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         n_top += 1;
     }
+#ifdef BF16_WIDEN_ONE
+    free(CONV);
+#endif
     return 0;
 }
diff --git a/kernel/riscv64/sbgemm_kernel_8x8_zvl128b.c b/kernel/riscv64/sbgemm_kernel_8x8_zvl128b.c
index 74679f8cbf..8cf6d64890 100644
--- a/kernel/riscv64/sbgemm_kernel_8x8_zvl128b.c
+++ b/kernel/riscv64/sbgemm_kernel_8x8_zvl128b.c
@@ -1,22 +1,91 @@
 #include "common.h"
 #include <riscv_vector.h>
 
+#define BF16_WIDEN_ONE  // Convert pre-hand and do operations in FP32
+#define USE_BF16_CVT    // Comment out for pre-RVA23 systems
+
+#ifdef BF16_WIDEN_ONE
+#define FORCEINLINE      inline __attribute__((always_inline))
+#define B_UNROLL         32
+
+// Convert from BF16 to FP32
+static void FORCEINLINE B_CONV(__bf16 *BB, FLOAT *CONV, BLASLONG count)
+{
+    BLASLONG count2 = (count & (B_UNROLL - 1));
+    count &= -B_UNROLL;
+    while (count) {
+        vbfloat16m4_t B00 = __riscv_vle16_v_bf16m4(BB, B_UNROLL);
+#ifdef USE_BF16_CVT
+        vfloat32m8_t B0 = __riscv_vfwcvtbf16_f_f_v_f32m8(B00, B_UNROLL);
+#else
+        vfloat32m8_t B0 = __riscv_vreinterpret_v_u32m8_f32m8(__riscv_vsll_vx_u32m8(
+            __riscv_vwcvtu_x_x_v_u32m8(__riscv_vreinterpret_v_bf16m4_u16m4(B00), B_UNROLL), 16, B_UNROLL));
+#endif
+        __riscv_vse32_v_f32m8(CONV, B0, B_UNROLL);
+        BB += B_UNROLL;
+        CONV += B_UNROLL;
+        count -= B_UNROLL;
+    }
+    if (count2) {
+        BLASLONG gvl2 = __riscv_vsetvl_e16m4(count2);
+        vbfloat16m4_t B00 = __riscv_vle16_v_bf16m4(BB, gvl2);
+#ifdef USE_BF16_CVT
+        vfloat32m8_t B0 = __riscv_vfwcvtbf16_f_f_v_f32m8(B00, gvl2);
+#else
+        vfloat32m8_t B0 = __riscv_vreinterpret_v_u32m8_f32m8(__riscv_vsll_vx_u32m8(
+            __riscv_vwcvtu_x_x_v_u32m8(__riscv_vreinterpret_v_bf16m4_u16m4(B00), gvl2), 16, gvl2));
+#endif
+        __riscv_vse32_v_f32m8(CONV, B0, gvl2);
+    }
+}
+#endif
+
+#ifndef VECTORIZE_MEMSET
+#define memset_zero(ptr, size, dir)  memset(ptr, 0, size)
+#else
+void memset_zero(void *input, BLASLONG size, bool dir);
+#endif
+
 int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
 {
     BLASLONG gvl = 0;
     BLASLONG m_top = 0;
     BLASLONG n_top = 0;
+#if !defined(BF16_WIDEN_ONE) || !defined(BF16_DONT_CONV)
     __bf16 *BB = (__bf16 *)(B);
     __bf16 *AA = (__bf16 *)(A);
+#endif
+
+#ifdef BF16_WIDEN_ONE
+    FLOAT *CONV = (FLOAT *)(malloc((K * (8 + M)) * sizeof(FLOAT)));
+    if (!CONV) return 1;
+#ifndef BF16_DONT_CONV
+    B_CONV(AA, CONV + (K * 8), K * M);
+#else
+    memset_zero(CONV, (K * (8 + M)) * sizeof(FLOAT), false);
+#endif
+#endif
 
     // -- MAIN PASS
     for (BLASLONG j=0; j<N/8; j+=1) {
         m_top = 0;
         BLASLONG gvl = __riscv_vsetvl_e16m1(8);
 
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 8;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
+
         for (BLASLONG i=0; i<M/8; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
@@ -28,6 +97,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result7 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                float B4 = CONV[bi2+4];
+                float B5 = CONV[bi2+5];
+                float B6 = CONV[bi2+6];
+                float B7 = CONV[bi2+7];
+                bi2 += 8;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f32m2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f32m2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f32m2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f32m2(result7, B7, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -49,6 +141,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmaccbf16_vf_f32m2(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmaccbf16_vf_f32m2(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmaccbf16_vf_f32m2(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -89,8 +182,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16m1(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -102,6 +199,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result7 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                float B4 = CONV[bi2+4];
+                float B5 = CONV[bi2+5];
+                float B6 = CONV[bi2+6];
+                float B7 = CONV[bi2+7];
+                bi2 += 8;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f32m1(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f32m1(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f32m1(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f32m1(result7, B7, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -123,6 +243,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmaccbf16_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmaccbf16_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmaccbf16_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -167,8 +288,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 2 ) {
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             float result0 = 0;
             float result1 = 0;
@@ -188,6 +313,26 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result15 = 0;
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                result4+=CONV[ai2+0]*CONV[bi2+2];
+                result5+=CONV[ai2+1]*CONV[bi2+2];
+                result6+=CONV[ai2+0]*CONV[bi2+3];
+                result7+=CONV[ai2+1]*CONV[bi2+3];
+                result8+=CONV[ai2+0]*CONV[bi2+4];
+                result9+=CONV[ai2+1]*CONV[bi2+4];
+                result10+=CONV[ai2+0]*CONV[bi2+5];
+                result11+=CONV[ai2+1]*CONV[bi2+5];
+                result12+=CONV[ai2+0]*CONV[bi2+6];
+                result13+=CONV[ai2+1]*CONV[bi2+6];
+                result14+=CONV[ai2+0]*CONV[bi2+7];
+                result15+=CONV[ai2+1]*CONV[bi2+7];
+                ai2+=2;
+                bi2+=8;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
@@ -206,6 +351,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result15+=(float)(AA[ai+1])*(float)(BB[bi+7]);
                 ai+=2;
                 bi+=8;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -240,10 +386,26 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result6 = 0;
             float result7 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                result2+=CONV[ai2+0]*CONV[bi2+2];
+                result3+=CONV[ai2+0]*CONV[bi2+3];
+                result4+=CONV[ai2+0]*CONV[bi2+4];
+                result5+=CONV[ai2+0]*CONV[bi2+5];
+                result6+=CONV[ai2+0]*CONV[bi2+6];
+                result7+=CONV[ai2+0]*CONV[bi2+7];
+                ai2+=1;
+                bi2+=8;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
@@ -254,6 +416,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result7+=(float)(AA[ai+0])*(float)(BB[bi+7]);
                 ai+=1;
                 bi+=8;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -276,9 +439,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         gvl = __riscv_vsetvl_e16m1(8);
         m_top = 0;
 
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 4;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
+
         for (BLASLONG i=0; i<M/8; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
@@ -286,6 +461,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result3 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                bi2 += 4;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m2(result3, B3, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -299,6 +489,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmaccbf16_vf_f32m2(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmaccbf16_vf_f32m2(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -325,8 +516,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16m1(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
@@ -334,6 +529,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result3 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                float B2 = CONV[bi2+2];
+                float B3 = CONV[bi2+3];
+                bi2 += 4;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f32m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f32m1(result3, B3, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 __bf16 B2 = BB[bi+2];
@@ -347,6 +557,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmaccbf16_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmaccbf16_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -375,8 +586,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 2 ) {
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             float result0 = 0;
             float result1 = 0;
@@ -388,6 +603,18 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result7 = 0;
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                result4+=CONV[ai2+0]*CONV[bi2+2];
+                result5+=CONV[ai2+1]*CONV[bi2+2];
+                result6+=CONV[ai2+0]*CONV[bi2+3];
+                result7+=CONV[ai2+1]*CONV[bi2+3];
+                ai2+=2;
+                bi2+=4;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
@@ -398,6 +625,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result7+=(float)(AA[ai+1])*(float)(BB[bi+3]);
                 ai+=2;
                 bi+=4;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -420,16 +648,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result2 = 0;
             float result3 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                result2+=CONV[ai2+0]*CONV[bi2+2];
+                result3+=CONV[ai2+0]*CONV[bi2+3];
+                ai2+=1;
+                bi2+=4;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+2]);
                 result3+=(float)(AA[ai+0])*(float)(BB[bi+3]);
                 ai+=1;
                 bi+=4;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -447,15 +688,37 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
     if ( N & 2 ) {
         gvl = __riscv_vsetvl_e16m1(8);
         m_top = 0;
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 2;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
 
         for (BLASLONG i=0; i<M/8; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
             vfloat32m2_t result1 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                bi2 += 2;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m2(result1, B1, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 bi += 2;
@@ -465,6 +728,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmaccbf16_vf_f32m2(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -485,13 +749,28 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16m1(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
             vfloat32m1_t result1 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                float B1 = CONV[bi2+1];
+                bi2 += 2;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f32m1(result1, B1, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 __bf16 B1 = BB[bi+1];
                 bi += 2;
@@ -501,6 +780,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmaccbf16_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -521,8 +801,12 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 2 ) {
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             float result0 = 0;
             float result1 = 0;
@@ -530,12 +814,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result3 = 0;
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                result2+=CONV[ai2+0]*CONV[bi2+1];
+                result3+=CONV[ai2+1]*CONV[bi2+1];
+                ai2+=2;
+                bi2+=2;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 result2+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 result3+=(float)(AA[ai+1])*(float)(BB[bi+1]);
                 ai+=2;
                 bi+=2;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -553,14 +846,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result0 = 0;
             float result1 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+0]*CONV[bi2+1];
+                ai2+=1;
+                bi2+=2;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+0])*(float)(BB[bi+1]);
                 ai+=1;
                 bi+=2;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -576,21 +880,42 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
     if ( N & 1 ) {
         gvl = __riscv_vsetvl_e16m1(8);
         m_top = 0;
+#ifdef BF16_WIDEN_ONE
+        BLASLONG bi2 = K * 1;
+#ifndef BF16_DONT_CONV
+        B_CONV(BB + (n_top*K), CONV, bi2);
+#endif
+        BLASLONG ai2 = K * 8;
+#endif
 
         for (BLASLONG i=0; i<M/8; i+=1) {
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m2_t result0 = __riscv_vfmv_v_f_f32m2(0.0f, gvl);
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                bi2 += 1;
+
+                vfloat32m2_t A0 = __riscv_vle32_v_f32m2(&CONV[ai2], gvl);
+                ai2 += 8;
+
+                result0 = __riscv_vfmacc_vf_f32m2(result0, B0, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 bi += 1;
 
-                vbfloat16m1_t A0 = __riscv_vle16_v_bf16m1( &AA[ai+0*gvl], gvl );
+                vbfloat16mf2_t A0 = __riscv_vle16_v_bf16mf2( &AA[ai+0*gvl], gvl );
                 ai += 8;
 
-                result0 = __riscv_vfwmaccbf16_vf_f32m2(result0, B0, A0, gvl);
+                result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -608,12 +933,25 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         if ( M & 4 ) {
             gvl = __riscv_vsetvl_e16m1(4);
 
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
             BLASLONG ai=m_top*K;
             BLASLONG bi=n_top*K;
+#endif
 
             vfloat32m1_t result0 = __riscv_vfmv_v_f_f32m1(0.0f, gvl);
 
             for (BLASLONG k=0; k < K; ++k) {
+#ifdef BF16_WIDEN_ONE
+                float B0 = CONV[bi2+0];
+                bi2 += 1;
+
+                vfloat32m1_t A0 = __riscv_vle32_v_f32m1(&CONV[ai2], gvl);
+                ai2 += 4;
+
+                result0 = __riscv_vfmacc_vf_f32m1(result0, B0, A0, gvl);
+#else
                 __bf16 B0 = BB[bi+0];
                 bi += 1;
 
@@ -621,6 +959,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 ai += 4;
 
                 result0 = __riscv_vfwmaccbf16_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -636,17 +975,28 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if ( M & 2 ) {
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             float result0 = 0;
             float result1 = 0;
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                result1+=CONV[ai2+1]*CONV[bi2+0];
+                ai2+=2;
+                bi2+=1;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 result1+=(float)(AA[ai+1])*(float)(BB[bi+0]);
                 ai+=2;
                 bi+=1;
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -660,13 +1010,23 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             float result0 = 0;
 
-            BLASLONG ai = m_top * K;
-            BLASLONG bi = n_top * K;
+#ifdef BF16_WIDEN_ONE
+            bi2 = 0;
+#else
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+#endif
 
             for (BLASLONG k=0; k<K; k++) {
+#ifdef BF16_WIDEN_ONE
+                result0+=CONV[ai2+0]*CONV[bi2+0];
+                ai2+=1;
+                bi2+=1;
+#else
                 result0+=(float)(AA[ai+0])*(float)(BB[bi+0]);
                 ai+=1;
                 bi+=1;
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -676,6 +1036,8 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         n_top += 1;
     }
-
+#ifdef BF16_WIDEN_ONE
+    free(CONV);
+#endif
     return 0;
 }
diff --git a/kernel/riscv64/shgemm_kernel_16x8_zvl256b.c b/kernel/riscv64/shgemm_kernel_16x8_zvl256b.c
index 746e07d55f..8bb466e338 100644
--- a/kernel/riscv64/shgemm_kernel_16x8_zvl256b.c
+++ b/kernel/riscv64/shgemm_kernel_16x8_zvl256b.c
@@ -1,18 +1,121 @@
-
 #include "common.h"
 #include <riscv_vector.h>
+
+#define FP16_NARROW     // Accumulate in FP16
+
 int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
 {
     BLASLONG gvl = 0;
     BLASLONG m_top = 0;
     BLASLONG n_top = 0;
+#ifdef FP16_NARROW
+    IFLOAT alpha16 = (IFLOAT)(alpha);
+#endif
 
     // -- MAIN PASS
     for (BLASLONG j=0; j<N/8; j+=1) {
         m_top = 0;
+#ifdef FP16_NARROW
+        BLASLONG gvl = __riscv_vsetvl_e16m2(32);
+
+        for (BLASLONG i=0; i<M/32; i+=1) {
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+
+            _Float16 B0 = B[bi+0];
+            _Float16 B1 = B[bi+1];
+            _Float16 B2 = B[bi+2];
+            _Float16 B3 = B[bi+3];
+            _Float16 B4 = B[bi+4];
+            _Float16 B5 = B[bi+5];
+            _Float16 B6 = B[bi+6];
+            _Float16 B7 = B[bi+7];
+            bi += 8;
+
+            vfloat16m1_t A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+            vfloat16m1_t A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+            vfloat16m2_t A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+            ai += 16;
+
+            vfloat16m2_t result0 = __riscv_vfmul_vf_f16m2( A0, B0, gvl);
+            vfloat16m2_t result1 = __riscv_vfmul_vf_f16m2( A0, B1, gvl);
+            vfloat16m2_t result2 = __riscv_vfmul_vf_f16m2( A0, B2, gvl);
+            vfloat16m2_t result3 = __riscv_vfmul_vf_f16m2( A0, B3, gvl);
+            vfloat16m2_t result4 = __riscv_vfmul_vf_f16m2( A0, B4, gvl);
+            vfloat16m2_t result5 = __riscv_vfmul_vf_f16m2( A0, B5, gvl);
+            vfloat16m2_t result6 = __riscv_vfmul_vf_f16m2( A0, B6, gvl);
+            vfloat16m2_t result7 = __riscv_vfmul_vf_f16m2( A0, B7, gvl);
+
+            for(BLASLONG k=1; k<K; k++) {
+                B0 = B[bi+0];
+                B1 = B[bi+1];
+                B2 = B[bi+2];
+                B3 = B[bi+3];
+                B4 = B[bi+4];
+                B5 = B[bi+5];
+                B6 = B[bi+6];
+                B7 = B[bi+7];
+                bi += 8;
+                A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+                A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+                A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+                ai += 16;
+
+                result0 = __riscv_vfmacc_vf_f16m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16m2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16m2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16m2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16m2(result7, B7, A0, gvl);
+            }
+
+            BLASLONG ci=n_top*ldc+m_top;
+
+            vfloat32m4_t c0 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c1 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c2 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c3 = __riscv_vle32_v_f32m4( &C[ci], gvl);
+
+            ci-=ldc*3;
+
+            c0 = __riscv_vfwmacc_vf_f32m4(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m4(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m4(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m4(c3, alpha16, result3, gvl);
+
+            __riscv_vse32_v_f32m4( &C[ci], c0, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c1, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c2, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c3, gvl); ci += ldc-gvl*0;
+
+            vfloat32m4_t c4 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c5 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c6 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c7 = __riscv_vle32_v_f32m4( &C[ci], gvl);
+
+            ci-=ldc*3;
+
+            c4 = __riscv_vfwmacc_vf_f32m4(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m4(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m4(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m4(c7, alpha16, result7, gvl);
+
+            __riscv_vse32_v_f32m4( &C[ci], c4, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c5, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c6, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c7, gvl);
+            m_top += 32;
+        }
+
+        if (M & 16) {
+          gvl = __riscv_vsetvl_e16m1(16);
+#else
         BLASLONG gvl = __riscv_vsetvl_e16m1(16);
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#endif
             BLASLONG ai=m_top*K;	
             BLASLONG bi=n_top*K;	
 
@@ -28,7 +131,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 16;
-            
+
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+            vfloat16m1_t result2 = __riscv_vfmul_vf_f16m1( A0, B2, gvl);
+            vfloat16m1_t result3 = __riscv_vfmul_vf_f16m1( A0, B3, gvl);
+            vfloat16m1_t result4 = __riscv_vfmul_vf_f16m1( A0, B4, gvl);
+            vfloat16m1_t result5 = __riscv_vfmul_vf_f16m1( A0, B5, gvl);
+            vfloat16m1_t result6 = __riscv_vfmul_vf_f16m1( A0, B6, gvl);
+            vfloat16m1_t result7 = __riscv_vfmul_vf_f16m1( A0, B7, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
             vfloat32m2_t result2 = __riscv_vfwmul_vf_f32m2( A0, B2, gvl);
@@ -37,6 +150,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result5 = __riscv_vfwmul_vf_f32m2( A0, B5, gvl);
             vfloat32m2_t result6 = __riscv_vfwmul_vf_f32m2( A0, B6, gvl);
             vfloat32m2_t result7 = __riscv_vfwmul_vf_f32m2( A0, B7, gvl);
+#endif
             
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -50,7 +164,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 bi += 8;
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 16;
-                
+
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m1(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16m1(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16m1(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16m1(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16m1(result7, B7, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m2(result2, B2, A0, gvl);
@@ -59,6 +183,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmacc_vf_f32m2(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmacc_vf_f32m2(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmacc_vf_f32m2(result7, B7, A0, gvl);
+#endif
             }
             
             BLASLONG ci=n_top*ldc+m_top;
@@ -71,7 +196,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t c5 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c6 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c7 = __riscv_vle32_v_f32m2( &C[ci], gvl);
-            
+
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m2(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m2(c3, alpha16, result3, gvl);
+            c4 = __riscv_vfwmacc_vf_f32m2(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m2(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m2(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m2(c7, alpha16, result7, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m2(c2, alpha, result2, gvl);
@@ -80,6 +215,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             c5 = __riscv_vfmacc_vf_f32m2(c5, alpha, result5, gvl);
             c6 = __riscv_vfmacc_vf_f32m2(c6, alpha, result6, gvl);
             c7 = __riscv_vfmacc_vf_f32m2(c7, alpha, result7, gvl);
+#endif
 
             ci=n_top*ldc+m_top;
 
@@ -116,6 +252,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+            vfloat16mf2_t result4 = __riscv_vfmul_vf_f16mf2( A0, B4, gvl);
+            vfloat16mf2_t result5 = __riscv_vfmul_vf_f16mf2( A0, B5, gvl);
+            vfloat16mf2_t result6 = __riscv_vfmul_vf_f16mf2( A0, B6, gvl);
+            vfloat16mf2_t result7 = __riscv_vfmul_vf_f16mf2( A0, B7, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
@@ -124,6 +270,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result5 = __riscv_vfwmul_vf_f32m1( A0, B5, gvl);
             vfloat32m1_t result6 = __riscv_vfwmul_vf_f32m1( A0, B6, gvl);
             vfloat32m1_t result7 = __riscv_vfwmul_vf_f32m1( A0, B7, gvl);
+#endif
 
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -138,8 +285,18 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 8;
-                
-		        result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
+
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16mf2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16mf2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16mf2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16mf2(result7, B7, A0, gvl);
+#else
+                result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m1(result3, B3, A0, gvl);
@@ -147,6 +304,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmacc_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmacc_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmacc_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
@@ -159,7 +317,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c5 = __riscv_vle32_v_f32m1( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m1_t c6 = __riscv_vle32_v_f32m1( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m1_t c7 = __riscv_vle32_v_f32m1( &C[ci], gvl);
-            
+
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+            c4 = __riscv_vfwmacc_vf_f32m1(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m1(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m1(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m1(c7, alpha16, result7, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
@@ -168,6 +336,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             c5 = __riscv_vfmacc_vf_f32m1(c5, alpha, result5, gvl);
             c6 = __riscv_vfmacc_vf_f32m1(c6, alpha, result6, gvl);
             c7 = __riscv_vfmacc_vf_f32m1(c7, alpha, result7, gvl);
+#endif
 
             ci=n_top*ldc+m_top;
 
@@ -201,6 +370,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+            vfloat16mf2_t result4 = __riscv_vfmul_vf_f16mf2( A0, B4, gvl);
+            vfloat16mf2_t result5 = __riscv_vfmul_vf_f16mf2( A0, B5, gvl);
+            vfloat16mf2_t result6 = __riscv_vfmul_vf_f16mf2( A0, B6, gvl);
+            vfloat16mf2_t result7 = __riscv_vfmul_vf_f16mf2( A0, B7, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
@@ -209,6 +388,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result5 = __riscv_vfwmul_vf_f32m1( A0, B5, gvl);
             vfloat32m1_t result6 = __riscv_vfwmul_vf_f32m1( A0, B6, gvl);
             vfloat32m1_t result7 = __riscv_vfwmul_vf_f32m1( A0, B7, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -224,6 +404,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16mf2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16mf2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16mf2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16mf2(result7, B7, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
@@ -232,6 +422,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmacc_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmacc_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmacc_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -251,6 +442,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c6 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c7 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+            c4 = __riscv_vfwmacc_vf_f32m1(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m1(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m1(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m1(c7, alpha16, result7, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
@@ -259,6 +461,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             c5 = __riscv_vfmacc_vf_f32m1(c5, alpha, result5, gvl);
             c6 = __riscv_vfmacc_vf_f32m1(c6, alpha, result6, gvl);
             c7 = __riscv_vfmacc_vf_f32m1(c7, alpha, result7, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -275,6 +478,24 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
         }
 
         if( M & 2 ) {
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+            _Float16 result8 = 0;
+            _Float16 result9 = 0;
+            _Float16 result10 = 0;
+            _Float16 result11 = 0;
+            _Float16 result12 = 0;
+            _Float16 result13 = 0;
+            _Float16 result14 = 0;
+            _Float16 result15 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -291,10 +512,29 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result13 = 0;
             float result14 = 0;
             float result15 = 0;
+#endif
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
             
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+                result4+=A[ai+0]*B[bi+2];
+                result5+=A[ai+1]*B[bi+2];
+                result6+=A[ai+0]*B[bi+3];
+                result7+=A[ai+1]*B[bi+3];
+                result8+=A[ai+0]*B[bi+4];
+                result9+=A[ai+1]*B[bi+4];
+                result10+=A[ai+0]*B[bi+5];
+                result11+=A[ai+1]*B[bi+5];
+                result12+=A[ai+0]*B[bi+6];
+                result13+=A[ai+1]*B[bi+6];
+                result14+=A[ai+0]*B[bi+7];
+                result15+=A[ai+1]*B[bi+7];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
@@ -311,12 +551,31 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result13+=(float)(A[ai+1])*(float)(B[bi+6]);
                 result14+=(float)(A[ai+0])*(float)(B[bi+7]);
                 result15+=(float)(A[ai+1])*(float)(B[bi+7]);
+#endif
                 ai+=2;
                 bi+=8;
             }
 
             BLASLONG ci=n_top*ldc+m_top;
 
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 2 * ldc + 1] += alpha16 * (float)(result5);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 3 * ldc + 1] += alpha16 * (float)(result7);
+            C[ci + 4 * ldc + 0] += alpha16 * (float)(result8);
+            C[ci + 4 * ldc + 1] += alpha16 * (float)(result9);
+            C[ci + 5 * ldc + 0] += alpha16 * (float)(result10);
+            C[ci + 5 * ldc + 1] += alpha16 * (float)(result11);
+            C[ci + 6 * ldc + 0] += alpha16 * (float)(result12);
+            C[ci + 6 * ldc + 1] += alpha16 * (float)(result13);
+            C[ci + 7 * ldc + 0] += alpha16 * (float)(result14);
+            C[ci + 7 * ldc + 1] += alpha16 * (float)(result15);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
@@ -333,6 +592,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 6 * ldc + 1] += alpha * result13;
             C[ci + 7 * ldc + 0] += alpha * result14;
             C[ci + 7 * ldc + 1] += alpha * result15;
+#endif
 
             m_top+=2;
         }
@@ -340,6 +600,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -348,11 +618,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result5 = 0;
             float result6 = 0;
             float result7 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+                result2+=A[ai+0]*B[bi+2];
+                result3+=A[ai+0]*B[bi+3];
+                result4+=A[ai+0]*B[bi+4];
+                result5+=A[ai+0]*B[bi+5];
+                result6+=A[ai+0]*B[bi+6];
+                result7+=A[ai+0]*B[bi+7];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+2]);
@@ -361,11 +642,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5+=(float)(A[ai+0])*(float)(B[bi+5]);
                 result6+=(float)(A[ai+0])*(float)(B[bi+6]);
                 result7+=(float)(A[ai+0])*(float)(B[bi+7]);
+#endif
                 ai+=1;
                 bi+=8;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result3);
+            C[ci + 4 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 5 * ldc + 0] += alpha16 * (float)(result5);
+            C[ci + 6 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 7 * ldc + 0] += alpha16 * (float)(result7);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
             C[ci + 2 * ldc + 0] += alpha * result2;
@@ -374,16 +666,83 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 5 * ldc + 0] += alpha * result5;
             C[ci + 6 * ldc + 0] += alpha * result6;
             C[ci + 7 * ldc + 0] += alpha * result7;
+#endif
             m_top+=1;
         }
         n_top += 8;
     }
 
     if( N & 4 ) {
-        gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
+#ifdef FP16_NARROW
+        gvl = __riscv_vsetvl_e16m2(32);
+
+        for (BLASLONG i=0; i<M/32; i+=1) {
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+
+            _Float16 B0 = B[bi+0];
+            _Float16 B1 = B[bi+1];
+            _Float16 B2 = B[bi+2];
+            _Float16 B3 = B[bi+3];
+            bi += 4;
+
+            vfloat16m1_t A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+            vfloat16m1_t A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+            vfloat16m2_t A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+            ai += 16;
+
+            vfloat16m2_t result0 = __riscv_vfmul_vf_f16m2( A0, B0, gvl);
+            vfloat16m2_t result1 = __riscv_vfmul_vf_f16m2( A0, B1, gvl);
+            vfloat16m2_t result2 = __riscv_vfmul_vf_f16m2( A0, B2, gvl);
+            vfloat16m2_t result3 = __riscv_vfmul_vf_f16m2( A0, B3, gvl);
+
+            for(BLASLONG k=1; k<K; k++) {
+                B0 = B[bi+0];
+                B1 = B[bi+1];
+                B2 = B[bi+2];
+                B3 = B[bi+3];
+                bi += 4;
+
+                A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+                A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+                A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+                ai += 16;
+
+                result0 = __riscv_vfmacc_vf_f16m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m2(result3, B3, A0, gvl);
+            }
+
+            BLASLONG ci=n_top*ldc+m_top;
+
+            vfloat32m4_t c0 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c1 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c2 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c3 = __riscv_vle32_v_f32m4( &C[ci], gvl);
+
+            c0 = __riscv_vfwmacc_vf_f32m4(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m4(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m4(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m4(c3, alpha16, result3, gvl);
+
+            ci=n_top*ldc+m_top;
+
+            __riscv_vse32_v_f32m4( &C[ci], c0, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c1, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c2, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c3, gvl);
+            m_top += 32;
+        }
+
+        if (M & 16) {
+          gvl = __riscv_vsetvl_e16m1(16);
+#else
+        gvl = __riscv_vsetvl_e16m1(16);
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#endif
             BLASLONG ai=m_top*K;	
             BLASLONG bi=n_top*K;	
             
@@ -396,11 +755,18 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 16;
-            
+
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+            vfloat16m1_t result2 = __riscv_vfmul_vf_f16m1( A0, B2, gvl);
+            vfloat16m1_t result3 = __riscv_vfmul_vf_f16m1( A0, B3, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
             vfloat32m2_t result2 = __riscv_vfwmul_vf_f32m2( A0, B2, gvl);
             vfloat32m2_t result3 = __riscv_vfwmul_vf_f32m2( A0, B3, gvl);
+#endif
             
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -412,10 +778,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 16;
                 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m1(result3, B3, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m2(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m2(result3, B3, A0, gvl);
+#endif
             }
             
             BLASLONG ci=n_top*ldc+m_top;
@@ -424,11 +797,18 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t c1 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c2 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c3 = __riscv_vle32_v_f32m2( &C[ci], gvl);
-            
+
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m2(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m2(c3, alpha16, result3, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m2(c2, alpha, result2, gvl);
             c3 = __riscv_vfmacc_vf_f32m2(c3, alpha, result3, gvl);
+#endif
 
             ci=n_top*ldc+m_top;
 
@@ -453,10 +833,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
             vfloat32m1_t result3 = __riscv_vfwmul_vf_f32m1( A0, B3, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -468,10 +855,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
             
 
@@ -482,10 +876,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c2 = __riscv_vle32_v_f32m1( &C[ci], gvl); ci += ldc - gvl * 0;
             vfloat32m1_t c3 = __riscv_vle32_v_f32m1( &C[ci], gvl);
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
             c3 = __riscv_vfmacc_vf_f32m1(c3, alpha, result3, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -510,10 +911,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
             vfloat32m1_t result3 = __riscv_vfwmul_vf_f32m1( A0, B3, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -525,10 +933,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -540,10 +955,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c2 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c3 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
             c3 = __riscv_vfmacc_vf_f32m1(c3, alpha, result3, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -557,6 +979,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
 
         if( M & 2 ) {
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -565,10 +997,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result5 = 0;
             float result6 = 0;
             float result7 = 0;
+#endif
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+                result4+=A[ai+0]*B[bi+2];
+                result5+=A[ai+1]*B[bi+2];
+                result6+=A[ai+0]*B[bi+3];
+                result7+=A[ai+1]*B[bi+3];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
@@ -577,12 +1020,23 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5+=(float)(A[ai+1])*(float)(B[bi+2]);
                 result6+=(float)(A[ai+0])*(float)(B[bi+3]);
                 result7+=(float)(A[ai+1])*(float)(B[bi+3]);
+#endif
                 ai+=2;
                 bi+=4;
             }
             
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 2 * ldc + 1] += alpha16 * (float)(result5);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 3 * ldc + 1] += alpha16 * (float)(result7);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
@@ -591,6 +1045,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 2 * ldc + 1] += alpha * result5;
             C[ci + 3 * ldc + 0] += alpha * result6;
             C[ci + 3 * ldc + 1] += alpha * result7;
+#endif
 
             m_top += 2;
         }
@@ -598,28 +1053,49 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
             float result3 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+                result2+=A[ai+0]*B[bi+2];
+                result3+=A[ai+0]*B[bi+3];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+2]);
                 result3+=(float)(A[ai+0])*(float)(B[bi+3]);
+#endif
                 ai+=1;
                 bi+=4;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result3);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
             C[ci + 2 * ldc + 0] += alpha * result2;
             C[ci + 3 * ldc + 0] += alpha * result3;
+#endif
             m_top += 1;
         }
 
@@ -630,10 +1106,61 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
     // -- tails for N=2
     if( N & 2 ) {
-        gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
+#ifdef FP16_NARROW
+        gvl = __riscv_vsetvl_e16m2(32);
+
+        for (BLASLONG i=0; i<M/32; i+=1) {
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+
+            _Float16 B0 = B[bi+0];
+            _Float16 B1 = B[bi+1];
+            bi += 2;
+
+            vfloat16m1_t A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+            vfloat16m1_t A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+            vfloat16m2_t A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+            ai += 16;
+
+            vfloat16m2_t result0 = __riscv_vfmul_vf_f16m2( A0, B0, gvl);
+            vfloat16m2_t result1 = __riscv_vfmul_vf_f16m2( A0, B1, gvl);
+
+            for(BLASLONG k=1; k<K; k++) {
+                B0 = B[bi+0];
+                B1 = B[bi+1];
+                bi += 2;
+
+                A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+                A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+                A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+                ai += 16;
+
+                result0 = __riscv_vfmacc_vf_f16m2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m2(result1, B1, A0, gvl);
+            }
+
+            BLASLONG ci=n_top*ldc+m_top;
+
+            vfloat32m4_t c0 = __riscv_vle32_v_f32m4( &C[ci], gvl); ci += ldc-gvl*0;
+            vfloat32m4_t c1 = __riscv_vle32_v_f32m4( &C[ci], gvl);
+            c0 = __riscv_vfwmacc_vf_f32m4(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m4(c1, alpha16, result1, gvl);
+
+            ci=n_top*ldc+m_top;
+
+            __riscv_vse32_v_f32m4( &C[ci], c0, gvl); ci += ldc-gvl*0;
+            __riscv_vse32_v_f32m4( &C[ci], c1, gvl);
+            m_top += 32;
+        }
+
+        if (M & 16) {
+          gvl = __riscv_vsetvl_e16m1(16);
+#else
+        gvl = __riscv_vsetvl_e16m1(16);
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#endif
             BLASLONG ai=m_top*K;	
             BLASLONG bi=n_top*K;	
             
@@ -645,8 +1172,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 16;
             
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
+#endif
             
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -656,16 +1188,26 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 16;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci=n_top*ldc+m_top;
 
             vfloat32m2_t c0 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c1 = __riscv_vle32_v_f32m2( &C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
+#endif
 
             ci=n_top*ldc+m_top;
 
@@ -686,8 +1228,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -697,8 +1244,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
 
@@ -706,9 +1258,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1( &C[ci], gvl); ci += ldc - gvl * 0;
             vfloat32m1_t c1 = __riscv_vle32_v_f32m1( &C[ci], gvl); 
-            
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -729,8 +1285,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -740,8 +1301,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -749,8 +1315,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c1 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -762,27 +1333,48 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
 
         if( M & 2 ) {
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
             float result3 = 0;
+#endif
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result3+=(float)(A[ai+1])*(float)(B[bi+1]);
+#endif
                 ai+=2;
                 bi+=2;
             }
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
             C[ci + 1 * ldc + 1] += alpha * result3;
+#endif
 
             m_top += 2;
         }
@@ -790,22 +1382,37 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
+#endif
                 ai+=1;
                 bi+=2;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
+#endif
             m_top += 1;
         }
 
@@ -816,10 +1423,54 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
     // -- tails for N=1
     if( N & 1 ) {
-        gvl = __riscv_vsetvl_e16m1(16);
         m_top = 0;
+#ifdef FP16_NARROW
+        gvl = __riscv_vsetvl_e16m2(32);
+
+        for (BLASLONG i=0; i<M/32; i+=1) {
+            BLASLONG ai=m_top*K;
+            BLASLONG bi=n_top*K;
+            _Float16 B0 = B[bi+0];
+            bi += 1;
+
+            vfloat16m1_t A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+            vfloat16m1_t A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+            vfloat16m2_t A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+            ai += 16;
+
+            vfloat16m2_t result0 = __riscv_vfmul_vf_f16m2( A0, B0, gvl);
+
+            for(BLASLONG k=1; k<K; k++) {
+                B0 = B[bi+0];
+                bi += 1;
+
+                A00 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], 16 );
+                A01 = __riscv_vle16_v_f16m1( &A[ai+0*gvl+16*K], 16 );
+                A0 = __riscv_vcreate_v_f16m1_f16m2(A00, A01);
+                ai += 16;
+
+                result0 = __riscv_vfmacc_vf_f16m2(result0, B0, A0, gvl);
+            }
+
+            BLASLONG ci=n_top*ldc+m_top;
+
+            vfloat32m4_t c0 = __riscv_vle32_v_f32m4( &C[ci], gvl);
+
+            c0 = __riscv_vfwmacc_vf_f32m4(c0, alpha16, result0, gvl);
+
+            ci=n_top*ldc+m_top;
+
+            __riscv_vse32_v_f32m4( &C[ci], c0, gvl);
+            m_top += 32;
+        }
+
+        if (M & 16) {
+          gvl = __riscv_vsetvl_e16m1(16);
+#else
+        gvl = __riscv_vsetvl_e16m1(16);
 
         for (BLASLONG i=0; i<M/16; i+=1) {
+#endif
             BLASLONG ai=m_top*K;	
             BLASLONG bi=n_top*K;	
             _Float16 B0 = B[bi+0];
@@ -828,7 +1479,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 16;
 
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
+#endif
 
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -837,14 +1492,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 16;
                 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
+#endif
             }
             
             BLASLONG ci=n_top*ldc+m_top;
 
             vfloat32m2_t c0 = __riscv_vle32_v_f32m2( &C[ci], gvl);
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
+#endif
 
             ci=n_top*ldc+m_top;
 
@@ -863,7 +1526,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -872,7 +1539,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
 
@@ -880,7 +1551,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1( &C[ci], gvl);
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -899,7 +1574,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -908,13 +1587,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2( &A[ai+0*gvl], gvl );
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
 
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -924,22 +1611,37 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
 
         if( M & 2 ) {
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
+#endif
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
+#endif
                 ai+=2;
                 bi+=1;
             }
             
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
+#endif
 
             m_top += 2;
         }
@@ -947,19 +1649,31 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+#else
             float result0 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
+#endif
                 ai+=1;
                 bi+=1;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
+#endif
             m_top += 1;
         }
 
diff --git a/kernel/riscv64/shgemm_kernel_8x8_zvl128b.c b/kernel/riscv64/shgemm_kernel_8x8_zvl128b.c
index 3197f7c005..03bdcc240e 100644
--- a/kernel/riscv64/shgemm_kernel_8x8_zvl128b.c
+++ b/kernel/riscv64/shgemm_kernel_8x8_zvl128b.c
@@ -1,12 +1,16 @@
-
 #include "common.h"
 #include <riscv_vector.h>
 
+#define FP16_NARROW     // Accumulate in FP16
+
 int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B, FLOAT *C, BLASLONG ldc)
 {
     BLASLONG gvl = 0;
     BLASLONG m_top = 0;
     BLASLONG n_top = 0;
+#ifdef FP16_NARROW
+    IFLOAT alpha16 = (IFLOAT)(alpha);
+#endif
 
     // -- MAIN PASS
     for (BLASLONG j=0; j<N/8; j+=1) {
@@ -30,6 +34,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+            vfloat16m1_t result2 = __riscv_vfmul_vf_f16m1( A0, B2, gvl);
+            vfloat16m1_t result3 = __riscv_vfmul_vf_f16m1( A0, B3, gvl);
+            vfloat16m1_t result4 = __riscv_vfmul_vf_f16m1( A0, B4, gvl);
+            vfloat16m1_t result5 = __riscv_vfmul_vf_f16m1( A0, B5, gvl);
+            vfloat16m1_t result6 = __riscv_vfmul_vf_f16m1( A0, B6, gvl);
+            vfloat16m1_t result7 = __riscv_vfmul_vf_f16m1( A0, B7, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
             vfloat32m2_t result2 = __riscv_vfwmul_vf_f32m2( A0, B2, gvl);
@@ -38,6 +52,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t result5 = __riscv_vfwmul_vf_f32m2( A0, B5, gvl);
             vfloat32m2_t result6 = __riscv_vfwmul_vf_f32m2( A0, B6, gvl);
             vfloat32m2_t result7 = __riscv_vfwmul_vf_f32m2( A0, B7, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -54,6 +69,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 ai += 8;
 
                 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m1(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16m1(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16m1(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16m1(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16m1(result7, B7, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m2(result2, B2, A0, gvl);
@@ -62,6 +87,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmacc_vf_f32m2(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmacc_vf_f32m2(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmacc_vf_f32m2(result7, B7, A0, gvl);
+#endif
             }
             
 
@@ -75,7 +101,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t c5 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c6 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
             vfloat32m2_t c7 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc-gvl*0;
-            
+
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m2(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m2(c3, alpha16, result3, gvl);
+            c4 = __riscv_vfwmacc_vf_f32m2(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m2(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m2(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m2(c7, alpha16, result7, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m2(c2, alpha, result2, gvl);
@@ -84,6 +120,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             c5 = __riscv_vfmacc_vf_f32m2(c5, alpha, result5, gvl);
             c6 = __riscv_vfmacc_vf_f32m2(c6, alpha, result6, gvl);
             c7 = __riscv_vfmacc_vf_f32m2(c7, alpha, result7, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -118,6 +155,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+            vfloat16mf2_t result4 = __riscv_vfmul_vf_f16mf2( A0, B4, gvl);
+            vfloat16mf2_t result5 = __riscv_vfmul_vf_f16mf2( A0, B5, gvl);
+            vfloat16mf2_t result6 = __riscv_vfmul_vf_f16mf2( A0, B6, gvl);
+            vfloat16mf2_t result7 = __riscv_vfmul_vf_f16mf2( A0, B7, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
@@ -126,6 +173,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t result5 = __riscv_vfwmul_vf_f32m1( A0, B5, gvl);
             vfloat32m1_t result6 = __riscv_vfwmul_vf_f32m1( A0, B6, gvl);
             vfloat32m1_t result7 = __riscv_vfwmul_vf_f32m1( A0, B7, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -141,6 +189,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+                result4 = __riscv_vfmacc_vf_f16mf2(result4, B4, A0, gvl);
+                result5 = __riscv_vfmacc_vf_f16mf2(result5, B5, A0, gvl);
+                result6 = __riscv_vfmacc_vf_f16mf2(result6, B6, A0, gvl);
+                result7 = __riscv_vfmacc_vf_f16mf2(result7, B7, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
@@ -149,6 +207,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5 = __riscv_vfwmacc_vf_f32m1(result5, B5, A0, gvl);
                 result6 = __riscv_vfwmacc_vf_f32m1(result6, B6, A0, gvl);
                 result7 = __riscv_vfwmacc_vf_f32m1(result7, B7, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -168,6 +227,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c6 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c7 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+            c4 = __riscv_vfwmacc_vf_f32m1(c4, alpha16, result4, gvl);
+            c5 = __riscv_vfwmacc_vf_f32m1(c5, alpha16, result5, gvl);
+            c6 = __riscv_vfwmacc_vf_f32m1(c6, alpha16, result6, gvl);
+            c7 = __riscv_vfwmacc_vf_f32m1(c7, alpha16, result7, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
@@ -176,6 +245,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             c5 = __riscv_vfmacc_vf_f32m1(c5, alpha, result5, gvl);
             c6 = __riscv_vfmacc_vf_f32m1(c6, alpha, result6, gvl);
             c7 = __riscv_vfmacc_vf_f32m1(c7, alpha, result7, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -197,6 +267,24 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+            _Float16 result8 = 0;
+            _Float16 result9 = 0;
+            _Float16 result10 = 0;
+            _Float16 result11 = 0;
+            _Float16 result12 = 0;
+            _Float16 result13 = 0;
+            _Float16 result14 = 0;
+            _Float16 result15 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -213,8 +301,27 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result13 = 0;
             float result14 = 0;
             float result15 = 0;
+#endif
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+                result4+=A[ai+0]*B[bi+2];
+                result5+=A[ai+1]*B[bi+2];
+                result6+=A[ai+0]*B[bi+3];
+                result7+=A[ai+1]*B[bi+3];
+                result8+=A[ai+0]*B[bi+4];
+                result9+=A[ai+1]*B[bi+4];
+                result10+=A[ai+0]*B[bi+5];
+                result11+=A[ai+1]*B[bi+5];
+                result12+=A[ai+0]*B[bi+6];
+                result13+=A[ai+1]*B[bi+6];
+                result14+=A[ai+0]*B[bi+7];
+                result15+=A[ai+1]*B[bi+7];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
@@ -231,12 +338,31 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result13+=(float)(A[ai+1])*(float)(B[bi+6]);
                 result14+=(float)(A[ai+0])*(float)(B[bi+7]);
                 result15+=(float)(A[ai+1])*(float)(B[bi+7]);
+#endif
                 ai+=2;
                 bi+=8;
             }
             
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 2 * ldc + 1] += alpha16 * (float)(result5);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 3 * ldc + 1] += alpha16 * (float)(result7);
+            C[ci + 4 * ldc + 0] += alpha16 * (float)(result8);
+            C[ci + 4 * ldc + 1] += alpha16 * (float)(result9);
+            C[ci + 5 * ldc + 0] += alpha16 * (float)(result10);
+            C[ci + 5 * ldc + 1] += alpha16 * (float)(result11);
+            C[ci + 6 * ldc + 0] += alpha16 * (float)(result12);
+            C[ci + 6 * ldc + 1] += alpha16 * (float)(result13);
+            C[ci + 7 * ldc + 0] += alpha16 * (float)(result14);
+            C[ci + 7 * ldc + 1] += alpha16 * (float)(result15);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
@@ -253,6 +379,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 6 * ldc + 1] += alpha * result13;
             C[ci + 7 * ldc + 0] += alpha * result14;
             C[ci + 7 * ldc + 1] += alpha * result15;
+#endif
 
             m_top+=2;
         }
@@ -260,6 +387,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -268,11 +405,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result5 = 0;
             float result6 = 0;
             float result7 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+                result2+=A[ai+0]*B[bi+2];
+                result3+=A[ai+0]*B[bi+3];
+                result4+=A[ai+0]*B[bi+4];
+                result5+=A[ai+0]*B[bi+5];
+                result6+=A[ai+0]*B[bi+6];
+                result7+=A[ai+0]*B[bi+7];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+2]);
@@ -281,11 +429,22 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5+=(float)(A[ai+0])*(float)(B[bi+5]);
                 result6+=(float)(A[ai+0])*(float)(B[bi+6]);
                 result7+=(float)(A[ai+0])*(float)(B[bi+7]);
+#endif
                 ai+=1;
                 bi+=8;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result3);
+            C[ci + 4 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 5 * ldc + 0] += alpha16 * (float)(result5);
+            C[ci + 6 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 7 * ldc + 0] += alpha16 * (float)(result7);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
             C[ci + 2 * ldc + 0] += alpha * result2;
@@ -294,6 +453,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 5 * ldc + 0] += alpha * result5;
             C[ci + 6 * ldc + 0] += alpha * result6;
             C[ci + 7 * ldc + 0] += alpha * result7;
+#endif
             m_top+=1;
         }
 
@@ -318,10 +478,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+            vfloat16m1_t result2 = __riscv_vfmul_vf_f16m1( A0, B2, gvl);
+            vfloat16m1_t result3 = __riscv_vfmul_vf_f16m1( A0, B3, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
             vfloat32m2_t result2 = __riscv_vfwmul_vf_f32m2( A0, B2, gvl);
             vfloat32m2_t result3 = __riscv_vfwmul_vf_f32m2( A0, B3, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -333,10 +500,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16m1(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16m1(result3, B3, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m2(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m2(result3, B3, A0, gvl);
+#endif
             }
             
 
@@ -347,10 +521,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t c2 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc - gvl * 0;
             vfloat32m2_t c3 = __riscv_vle32_v_f32m2( &C[ci], gvl);
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m2(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m2(c3, alpha16, result3, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m2(c2, alpha, result2, gvl);
             c3 = __riscv_vfmacc_vf_f32m2(c3, alpha, result3, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -375,10 +556,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+            vfloat16mf2_t result2 = __riscv_vfmul_vf_f16mf2( A0, B2, gvl);
+            vfloat16mf2_t result3 = __riscv_vfmul_vf_f16mf2( A0, B3, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
             vfloat32m1_t result2 = __riscv_vfwmul_vf_f32m1( A0, B2, gvl);
             vfloat32m1_t result3 = __riscv_vfwmul_vf_f32m1( A0, B3, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -390,10 +578,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+                result2 = __riscv_vfmacc_vf_f16mf2(result2, B2, A0, gvl);
+                result3 = __riscv_vfmacc_vf_f16mf2(result3, B3, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
                 result2 = __riscv_vfwmacc_vf_f32m1(result2, B2, A0, gvl);
                 result3 = __riscv_vfwmacc_vf_f32m1(result3, B3, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -405,10 +600,17 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c2 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c3 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+            c2 = __riscv_vfwmacc_vf_f32m1(c2, alpha16, result2, gvl);
+            c3 = __riscv_vfwmacc_vf_f32m1(c3, alpha16, result3, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
             c2 = __riscv_vfmacc_vf_f32m1(c2, alpha, result2, gvl);
             c3 = __riscv_vfmacc_vf_f32m1(c3, alpha, result3, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -426,6 +628,16 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+            _Float16 result4 = 0;
+            _Float16 result5 = 0;
+            _Float16 result6 = 0;
+            _Float16 result7 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
@@ -434,8 +646,19 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             float result5 = 0;
             float result6 = 0;
             float result7 = 0;
+#endif
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+                result4+=A[ai+0]*B[bi+2];
+                result5+=A[ai+1]*B[bi+2];
+                result6+=A[ai+0]*B[bi+3];
+                result7+=A[ai+1]*B[bi+3];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
@@ -444,12 +667,23 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 result5+=(float)(A[ai+1])*(float)(B[bi+2]);
                 result6+=(float)(A[ai+0])*(float)(B[bi+3]);
                 result7+=(float)(A[ai+1])*(float)(B[bi+3]);
+#endif
                 ai+=2;
                 bi+=4;
             }
             
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result4);
+            C[ci + 2 * ldc + 1] += alpha16 * (float)(result5);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result6);
+            C[ci + 3 * ldc + 1] += alpha16 * (float)(result7);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
@@ -458,6 +692,7 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             C[ci + 2 * ldc + 1] += alpha * result5;
             C[ci + 3 * ldc + 0] += alpha * result6;
             C[ci + 3 * ldc + 1] += alpha * result7;
+#endif
 
             m_top += 2;
         }
@@ -465,28 +700,49 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
             float result3 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+                result2+=A[ai+0]*B[bi+2];
+                result3+=A[ai+0]*B[bi+3];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+2]);
                 result3+=(float)(A[ai+0])*(float)(B[bi+3]);
+#endif
                 ai+=1;
                 bi+=4;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+            C[ci + 2 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 3 * ldc + 0] += alpha16 * (float)(result3);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
             C[ci + 2 * ldc + 0] += alpha * result2;
             C[ci + 3 * ldc + 0] += alpha * result3;
+#endif
             m_top += 1;
         }
 
@@ -511,8 +767,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+            vfloat16m1_t result1 = __riscv_vfmul_vf_f16m1( A0, B1, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
             vfloat32m2_t result1 = __riscv_vfwmul_vf_f32m2( A0, B1, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -522,8 +783,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16m1(result1, B1, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m2(result1, B1, A0, gvl);
+#endif
             }
 
 
@@ -532,8 +798,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m2_t c0 = __riscv_vle32_v_f32m2( &C[ci], gvl); ci += ldc - gvl * 0;
             vfloat32m2_t c1 = __riscv_vle32_v_f32m2( &C[ci], gvl); 
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m2(c1, alpha16, result1, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m2(c1, alpha, result1, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -554,8 +825,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+            vfloat16mf2_t result1 = __riscv_vfmul_vf_f16mf2( A0, B1, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
             vfloat32m1_t result1 = __riscv_vfwmul_vf_f32m1( A0, B1, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -565,8 +841,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+                result1 = __riscv_vfmacc_vf_f16mf2(result1, B1, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
                 result1 = __riscv_vfwmacc_vf_f32m1(result1, B1, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
@@ -574,8 +855,13 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1(&C[ci], gvl);
             ci += ldc - gvl * 0;
             vfloat32m1_t c1 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+            c1 = __riscv_vfwmacc_vf_f32m1(c1, alpha16, result1, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
             c1 = __riscv_vfmacc_vf_f32m1(c1, alpha, result1, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -591,25 +877,46 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+            _Float16 result2 = 0;
+            _Float16 result3 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
             float result2 = 0;
             float result3 = 0;
+#endif
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+                result2+=A[ai+0]*B[bi+1];
+                result3+=A[ai+1]*B[bi+1];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
                 result2+=(float)(A[ai+0])*(float)(B[bi+1]);
                 result3+=(float)(A[ai+1])*(float)(B[bi+1]);
+#endif
                 ai+=2;
                 bi+=2;
             }
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result2);
+            C[ci + 1 * ldc + 1] += alpha16 * (float)(result3);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
             C[ci + 1 * ldc + 0] += alpha * result2;
             C[ci + 1 * ldc + 1] += alpha * result3;
+#endif
 
             m_top += 2;
         }
@@ -617,22 +924,37 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+0]*B[bi+1];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+0])*(float)(B[bi+1]);
+#endif
                 ai+=1;
                 bi+=2;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 1 * ldc + 0] += alpha16 * (float)(result1);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 1 * ldc + 0] += alpha * result1;
+#endif
             m_top += 1;
         }
 
@@ -656,7 +978,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16m1_t A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
             ai += 8;
 
+#ifdef FP16_NARROW
+            vfloat16m1_t result0 = __riscv_vfmul_vf_f16m1( A0, B0, gvl);
+#else
             vfloat32m2_t result0 = __riscv_vfwmul_vf_f32m2( A0, B0, gvl);
+#endif
 		
             for(BLASLONG k=1; k<K; k++) {
                 B0 = B[bi+0];
@@ -665,7 +991,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16m1( &A[ai+0*gvl], gvl );
                 ai += 8;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16m1(result0, B0, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m2(result0, B0, A0, gvl);
+#endif
             }
 
 
@@ -673,7 +1003,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
             vfloat32m2_t c0 = __riscv_vle32_v_f32m2( &C[ci], gvl);
             
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m2(c0, alpha16, result0, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m2(c0, alpha, result0, gvl);
+#endif
 
             ci = n_top * ldc + m_top;
 
@@ -692,7 +1026,11 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             vfloat16mf2_t A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
             ai += 4;
 
+#ifdef FP16_NARROW
+            vfloat16mf2_t result0 = __riscv_vfmul_vf_f16mf2( A0, B0, gvl);
+#else
             vfloat32m1_t result0 = __riscv_vfwmul_vf_f32m1( A0, B0, gvl);
+#endif
 
             for(BLASLONG k=1; k < K; ++k) {
                 B0 = B[bi+0];
@@ -701,13 +1039,21 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
                 A0 = __riscv_vle16_v_f16mf2(&A[ai + 0 * gvl], gvl);
                 ai += 4;
 
+#ifdef FP16_NARROW
+                result0 = __riscv_vfmacc_vf_f16mf2(result0, B0, A0, gvl);
+#else
                 result0 = __riscv_vfwmacc_vf_f32m1(result0, B0, A0, gvl);
+#endif
             }
 
             BLASLONG ci = n_top * ldc + m_top;
 
             vfloat32m1_t c0 = __riscv_vle32_v_f32m1(&C[ci], gvl);
+#ifdef FP16_NARROW
+            c0 = __riscv_vfwmacc_vf_f32m1(c0, alpha16, result0, gvl);
+#else
             c0 = __riscv_vfmacc_vf_f32m1(c0, alpha, result0, gvl);
+#endif
 
             ci= n_top * ldc + m_top;
 
@@ -721,20 +1067,35 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+            _Float16 result1 = 0;
+#else
             float result0 = 0;
             float result1 = 0;
+#endif
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+                result1+=A[ai+1]*B[bi+0];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
                 result1+=(float)(A[ai+1])*(float)(B[bi+0]);
+#endif
                 ai+=2;
                 bi+=1;
             }
             
             
             BLASLONG ci=n_top*ldc+m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+            C[ci + 0 * ldc + 1] += alpha16 * (float)(result1);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
             C[ci + 0 * ldc + 1] += alpha * result1;
+#endif
 
             m_top += 2;
         }
@@ -742,19 +1103,31 @@ int CNAME(BLASLONG M, BLASLONG N, BLASLONG K, FLOAT alpha, IFLOAT *A, IFLOAT *B,
 
         if( M & 1 ) {
             
+#ifdef FP16_NARROW
+            _Float16 result0 = 0;
+#else
             float result0 = 0;
+#endif
             
             BLASLONG ai = m_top * K;
             BLASLONG bi = n_top * K;
 
             for(BLASLONG k=0; k<K; k++) {
+#ifdef FP16_NARROW
+                result0+=A[ai+0]*B[bi+0];
+#else
                 result0+=(float)(A[ai+0])*(float)(B[bi+0]);
+#endif
                 ai+=1;
                 bi+=1;
             }
 
             BLASLONG ci = n_top * ldc + m_top;
+#ifdef FP16_NARROW
+            C[ci + 0 * ldc + 0] += alpha16 * (float)(result0);
+#else
             C[ci + 0 * ldc + 0] += alpha * result0;
+#endif
             m_top += 1;
         }