diff --git a/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c b/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c
index 74f30d44a..e9edd4571 100644
--- a/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c
+++ b/kernel/x86_64/sbgemm_tcopy_4_cooperlake.c
@@ -29,6 +29,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <immintrin.h>
 #include "common.h"
 
+#define STORE_VEC(Bx, By, vec) \
+	if (By == 0) asm("vmovdqu16 %0, (%1)": : "v"(vec), "r"(boffset##Bx)); \
+	else asm("vmovdqu16 %0, (%1, %2, %c3)": : "v"(vec), "r"(boffset##Bx), "r"(blk_size), "n"(By * 2));
+
 int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
 	BLASLONG i, j;
 
@@ -36,13 +40,106 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
 
 	boffset0   = b;
 
+	BLASLONG n24 = n - (n % 24);
 	BLASLONG n8 = n & ~7;
+	BLASLONG m8 = m & ~7;
 	BLASLONG m4 = m & ~3;
 	BLASLONG m2 = m & ~1;
 
-	for (j = 0; j < n8; j += 8) {
+	int permute_table[] = {
+		0x0, 0x1, 0x2, 0x3, 0x10, 0x11, 0x12, 0x13, 0x8, 0x9, 0xa, 0xb, 0x18, 0x19, 0x1a, 0x1b,
+		0x4, 0x5, 0x6, 0x7, 0x14, 0x15, 0x16, 0x17, 0xc, 0xd, 0xe, 0xf, 0x1c, 0x1d, 0x1e, 0x1f,
+		0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
+		0x8, 0x9, 0xa, 0xb, 0xc, 0xd, 0xe, 0xf, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
+	};
+
+	j = 0;
+	if (n > 23) {
+		/* n = 24 is the max width in current blocking setting */
+		__m512i idx_lo_128 = _mm512_loadu_si512(permute_table);
+		__m512i idx_hi_128 = _mm512_loadu_si512(permute_table + 16);
+		__m512i idx_lo_256 = _mm512_loadu_si512(permute_table + 32);
+		__m512i idx_hi_256 = _mm512_loadu_si512(permute_table + 48);
+		__mmask32 mask24 = (1UL << 24) - 1;
+		BLASLONG blk_size = m * 4;
+		BLASLONG stride = blk_size * 3;
+
+		for (; j < n24; j += 24) {
+			boffset1 = boffset0 + stride;
+			for (i = 0; i < m8; i += 8) {
+				__m512i r0, r1, r2, r3, r4, r5, r6, r7;
+				__m512i t0, t1, t2, t3, t4, t5, t6, t7;
+				r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]);
+				r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]);
+				r2 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 2)*lda + j]);
+				r3 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 3)*lda + j]);
+				r4 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 4)*lda + j]);
+				r5 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 5)*lda + j]);
+				r6 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 6)*lda + j]);
+				r7 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 7)*lda + j]);
+
+				t0 = _mm512_unpacklo_epi16(r0, r1);
+				t1 = _mm512_unpackhi_epi16(r0, r1);
+				t2 = _mm512_unpacklo_epi16(r2, r3);
+				t3 = _mm512_unpackhi_epi16(r2, r3);
+				t4 = _mm512_unpacklo_epi16(r4, r5);
+				t5 = _mm512_unpackhi_epi16(r4, r5);
+				t6 = _mm512_unpacklo_epi16(r6, r7);
+				t7 = _mm512_unpackhi_epi16(r6, r7);
+
+				r0 = _mm512_permutex2var_epi32(t0, idx_lo_128, t2);
+				r1 = _mm512_permutex2var_epi32(t1, idx_lo_128, t3);
+				r2 = _mm512_permutex2var_epi32(t4, idx_lo_128, t6);
+				r3 = _mm512_permutex2var_epi32(t5, idx_lo_128, t7);
+				r4 = _mm512_permutex2var_epi32(t0, idx_hi_128, t2);
+				r5 = _mm512_permutex2var_epi32(t1, idx_hi_128, t3);
+				r6 = _mm512_permutex2var_epi32(t4, idx_hi_128, t6);
+				r7 = _mm512_permutex2var_epi32(t5, idx_hi_128, t7);
+
+				t0 = _mm512_permutex2var_epi32(r0, idx_lo_256, r2);
+				t1 = _mm512_permutex2var_epi32(r1, idx_lo_256, r3);
+				t2 = _mm512_permutex2var_epi32(r4, idx_lo_256, r6);
+				t3 = _mm512_permutex2var_epi32(r5, idx_lo_256, r7);
+				t4 = _mm512_permutex2var_epi32(r0, idx_hi_256, r2);
+				t5 = _mm512_permutex2var_epi32(r1, idx_hi_256, r3);
+
+				STORE_VEC(0, 0, t0); STORE_VEC(0, 1, t1); STORE_VEC(0, 2, t2);
+				STORE_VEC(1, 0, t3); STORE_VEC(1, 1, t4); STORE_VEC(1, 2, t5);
+				boffset0 += 32;
+				boffset1 += 32;
+			}
+			for (; i < m2; i += 2) {
+				__m512i r0, r1, t0, t1;
+				r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]);
+				r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]);
+				t0 = _mm512_unpacklo_epi16(r0, r1);
+				t1 = _mm512_unpackhi_epi16(r0, r1);
+				STORE_VEC(0, 0, _mm512_extracti32x4_epi32(t0, 0));
+				STORE_VEC(0, 1, _mm512_extracti32x4_epi32(t1, 0));
+				STORE_VEC(0, 2, _mm512_extracti32x4_epi32(t0, 1));
+				STORE_VEC(1, 0, _mm512_extracti32x4_epi32(t1, 1));
+				STORE_VEC(1, 1, _mm512_extracti32x4_epi32(t0, 2));
+				STORE_VEC(1, 2, _mm512_extracti32x4_epi32(t1, 2));
+				boffset0 += 8;
+				boffset1 += 8;
+			}
+			for (; i < m; i++) {
+				*(uint64_t *)(boffset0 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 0];
+				*(uint64_t *)(boffset0 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 4];
+				*(uint64_t *)(boffset0 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 8];
+				*(uint64_t *)(boffset1 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 12];
+				*(uint64_t *)(boffset1 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 16];
+				*(uint64_t *)(boffset1 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 20];
+				boffset0 += 4;
+				boffset1 += 4;
+			}
+			boffset0 += stride * 2;
+		}
+	}
+
+	for (; j < n8; j += 8) {
 		boffset1 = boffset0 + m * 4;
-		for (i = 0; i < m4; i +=4) {
+		for (i = 0; i < m4; i += 4) {
 			__m128i a0 = _mm_loadu_si128((void *)&a[(i + 0)*lda + j]);
 			__m128i a1 = _mm_loadu_si128((void *)&a[(i + 1)*lda + j]);
 			__m128i a2 = _mm_loadu_si128((void *)&a[(i + 2)*lda + j]);