Merge pull request #1815 from fenrus75/sgemm_beta_fix

enable the SGEMM/SKX C based kernel

kernel/x86_64/KERNEL.SKYLAKEX

@@ -1,6 +1,11 @@
 include $(KERNELDIR)/KERNEL.HASWELL
 
-SGEMMKERNEL    =  sgemm_kernel_16x4_skylakex.S
+SGEMMKERNEL    =  sgemm_kernel_16x4_skylakex.c
+
+SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
+SGEMMITCOPY    =  sgemm_tcopy_16_skylakex.c
+SGEMMONCOPY    =  sgemm_ncopy_4_skylakex.c
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
 
 DGEMMKERNEL    =  dgemm_kernel_4x8_skylakex.c
 
@@ -9,5 +14,5 @@ DGEMMITCOPY    =  dgemm_tcopy_8_skylakex.c
 DGEMMONCOPY    =  dgemm_ncopy_8_skylakex.c
 DGEMMOTCOPY    =  dgemm_tcopy_8_skylakex.c
 
-SGEMM_BETA = ../generic/gemm_beta.c
+SGEMM_BETA = sgemm_beta_skylakex.c
 DGEMM_BETA = dgemm_beta_skylakex.c

kernel/x86_64/sgemm_beta_skylakex.c

@@ -60,8 +60,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 
   if (beta == ZERO){
     __m512 z_zero;
+    __m256 y_zero;
 
     z_zero = _mm512_setzero_ps();
+    y_zero = _mm256_setzero_ps();
     j = n;
     do {
       c_offset1 = c_offset;
@@ -71,14 +73,12 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
 
       while (i > 32) {
 	  _mm512_storeu_ps(c_offset1, z_zero);
-	  _mm512_storeu_ps(c_offset1 + 8, z_zero);
 	  _mm512_storeu_ps(c_offset1 + 16, z_zero);
-	  _mm512_storeu_ps(c_offset1 + 24 , z_zero);
 	  c_offset1 += 32;
 	  i -= 32;
       }
       while (i > 8) {
-	  _mm512_storeu_ps(c_offset1, z_zero);
+	  _mm256_storeu_ps(c_offset1, y_zero);
 	  c_offset1 += 8;
 	  i -= 8;
       }

kernel/x86_64/sgemm_kernel_16x4_skylakex.c

@@ -64,419 +64,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  
 
 
-#define INIT32x8()	\
-	row0 = _mm512_setzero_ps();					\
-	row1 = _mm512_setzero_ps();					\
-	row2 = _mm512_setzero_ps();					\
-	row3 = _mm512_setzero_ps();					\
-	row4 = _mm512_setzero_ps();					\
-	row5 = _mm512_setzero_ps();					\
-	row6 = _mm512_setzero_ps();					\
-	row0b = _mm512_setzero_ps();					\
-	row1b = _mm512_setzero_ps();					\
-	row2b = _mm512_setzero_ps();					\
-	row3b = _mm512_setzero_ps();					\
-	row4b = _mm512_setzero_ps();					\
-	row5b = _mm512_setzero_ps();					\
-	row6b = _mm512_setzero_ps();					\
-	row7b = _mm512_setzero_ps();					\
-
-#define KERNEL32x8_SUB() 						\
-	zmm0   = _mm512_loadu_ps(AO);					\
-	zmm0b  = _mm512_loadu_ps(AOb);					\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 0));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 1));		\
-	row0  += zmm0 * zmm2;						\
-	row1  += zmm0 * zmm3;						\
-	row0b += zmm0b * zmm2;						\
-	row1b += zmm0b * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 2));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 3));		\
-	row2  += zmm0 * zmm2;						\
-	row3  += zmm0 * zmm3;						\
-	row2b += zmm0b * zmm2;						\
-	row3b += zmm0b * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 4));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 5));		\
-	row4  += zmm0 * zmm2;						\
-	row5  += zmm0 * zmm3;						\
-	row4b += zmm0b * zmm2;						\
-	row5b += zmm0b * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 6));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 7));		\
-	row6  += zmm0 * zmm2;						\
-	row7  += zmm0 * zmm3;						\
-	row6b += zmm0b * zmm2;						\
-	row7b += zmm0b * zmm3;						\
-	BO  += 8;							\
-	AO  += 16;							\
-	AOb += 16;
-
-
-#define SAVE32x8(ALPHA)							\
-	zmm0   = _mm512_set1_ps(ALPHA);					\
-	row0  *= zmm0;							\
-	row1  *= zmm0;							\
-	row2  *= zmm0;							\
-	row3  *= zmm0;							\
-	row4  *= zmm0;							\
-	row5  *= zmm0;							\
-	row6  *= zmm0;							\
-	row7  *= zmm0;							\
-	row0b  *= zmm0;							\
-	row1b  *= zmm0;							\
-	row2b  *= zmm0;							\
-	row3b  *= zmm0;							\
-	row4b  *= zmm0;							\
-	row5b  *= zmm0;							\
-	row6b  *= zmm0;							\
-	row7b  *= zmm0;							\
-	row0  += _mm512_loadu_ps(CO1 + 0 * ldc);			\
-	row1  += _mm512_loadu_ps(CO1 + 1 * ldc);			\
-	row2  += _mm512_loadu_ps(CO1 + 2 * ldc);			\
-	row3  += _mm512_loadu_ps(CO1 + 3 * ldc);			\
-	row4  += _mm512_loadu_ps(CO1 + 4 * ldc);			\
-	row5  += _mm512_loadu_ps(CO1 + 5 * ldc);			\
-	row6  += _mm512_loadu_ps(CO1 + 6 * ldc);			\
-	row7  += _mm512_loadu_ps(CO1 + 7 * ldc);			\
-	_mm512_storeu_ps(CO1 + 0 * ldc, row0);				\
-	_mm512_storeu_ps(CO1 + 1 * ldc, row1);				\
-	_mm512_storeu_ps(CO1 + 2 * ldc, row2);				\
-	_mm512_storeu_ps(CO1 + 3 * ldc, row3);				\
-	_mm512_storeu_ps(CO1 + 4 * ldc, row4);				\
-	_mm512_storeu_ps(CO1 + 5 * ldc, row5);				\
-	_mm512_storeu_ps(CO1 + 6 * ldc, row6);				\
-	_mm512_storeu_ps(CO1 + 7 * ldc, row7);				\
-	row0b  += _mm512_loadu_ps(CO1 + 0 * ldc + 16);			\
-	row1b  += _mm512_loadu_ps(CO1 + 1 * ldc + 16);			\
-	row2b  += _mm512_loadu_ps(CO1 + 2 * ldc + 16);			\
-	row3b  += _mm512_loadu_ps(CO1 + 3 * ldc + 16);			\
-	row4b  += _mm512_loadu_ps(CO1 + 4 * ldc + 16);			\
-	row5b  += _mm512_loadu_ps(CO1 + 5 * ldc + 16);			\
-	row6b  += _mm512_loadu_ps(CO1 + 6 * ldc + 16);			\
-	row7b  += _mm512_loadu_ps(CO1 + 7 * ldc + 16);			\
-	_mm512_storeu_ps(CO1 + 0 * ldc + 16, row0b);			\
-	_mm512_storeu_ps(CO1 + 1 * ldc + 16, row1b);			\
-	_mm512_storeu_ps(CO1 + 2 * ldc + 16, row2b);			\
-	_mm512_storeu_ps(CO1 + 3 * ldc + 16, row3b);			\
-	_mm512_storeu_ps(CO1 + 4 * ldc + 16, row4b);			\
-	_mm512_storeu_ps(CO1 + 5 * ldc + 16, row5b);			\
-	_mm512_storeu_ps(CO1 + 6 * ldc + 16, row6b);			\
-	_mm512_storeu_ps(CO1 + 7 * ldc + 16, row7b);			\
-
-
-#define INIT16x8()	\
-	row0 = _mm512_setzero_ps();					\
-	row1 = _mm512_setzero_ps();					\
-	row2 = _mm512_setzero_ps();					\
-	row3 = _mm512_setzero_ps();					\
-	row4 = _mm512_setzero_ps();					\
-	row5 = _mm512_setzero_ps();					\
-	row6 = _mm512_setzero_ps();					\
-	row7 = _mm512_setzero_ps();					\
-
-#define KERNEL16x8_SUB() 						\
-	zmm0   = _mm512_loadu_ps(AO);					\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 0));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 1));		\
-	row0  += zmm0 * zmm2;						\
-	row1  += zmm0 * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 2));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 3));		\
-	row2  += zmm0 * zmm2;						\
-	row3  += zmm0 * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 4));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 5));		\
-	row4  += zmm0 * zmm2;						\
-	row5  += zmm0 * zmm3;						\
-	zmm2   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 6));		\
-	zmm3   =  _mm512_broadcastss_ps(_mm_load_ss(BO + 7));		\
-	row6  += zmm0 * zmm2;						\
-	row7  += zmm0 * zmm3;						\
-	BO += 8;							\
-	AO += 16;
-
-
-#define SAVE16x8(ALPHA)							\
-	zmm0   = _mm512_set1_ps(ALPHA);					\
-	row0  *= zmm0;							\
-	row1  *= zmm0;							\
-	row2  *= zmm0;							\
-	row3  *= zmm0;							\
-	row4  *= zmm0;							\
-	row5  *= zmm0;							\
-	row6  *= zmm0;							\
-	row7  *= zmm0;							\
-	row0  += _mm512_loadu_ps(CO1 + 0 * ldc);			\
-	row1  += _mm512_loadu_ps(CO1 + 1 * ldc);			\
-	row2  += _mm512_loadu_ps(CO1 + 2 * ldc);			\
-	row3  += _mm512_loadu_ps(CO1 + 3 * ldc);			\
-	row4  += _mm512_loadu_ps(CO1 + 4 * ldc);			\
-	row5  += _mm512_loadu_ps(CO1 + 5 * ldc);			\
-	row6  += _mm512_loadu_ps(CO1 + 6 * ldc);			\
-	row7  += _mm512_loadu_ps(CO1 + 7 * ldc);			\
-	_mm512_storeu_ps(CO1 + 0 * ldc, row0);				\
-	_mm512_storeu_ps(CO1 + 1 * ldc, row1);				\
-	_mm512_storeu_ps(CO1 + 2 * ldc, row2);				\
-	_mm512_storeu_ps(CO1 + 3 * ldc, row3);				\
-	_mm512_storeu_ps(CO1 + 4 * ldc, row4);				\
-	_mm512_storeu_ps(CO1 + 5 * ldc, row5);				\
-	_mm512_storeu_ps(CO1 + 6 * ldc, row6);				\
-	_mm512_storeu_ps(CO1 + 7 * ldc, row7);			
-
-
-
-/*******************************************************************************************/
-
-#define INIT8x8()							\
-	row0 = _mm256_setzero_ps();					\
-	row1 = _mm256_setzero_ps();					\
-	row2 = _mm256_setzero_ps();					\
-	row3 = _mm256_setzero_ps();					\
-	row4 = _mm256_setzero_ps();					\
-	row5 = _mm256_setzero_ps();					\
-	row6 = _mm256_setzero_ps();					\
-	row7 = _mm256_setzero_ps();					\
-
-#define KERNEL8x8_SUB() 						\
-	ymm0   = _mm256_loadu_ps(AO);					\
-	ymm2   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 0));		\
-	ymm3   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 1));		\
-	row0  += ymm0 * ymm2;						\
-	row1  += ymm0 * ymm3;						\
-	ymm2   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 2));		\
-	ymm3   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 3));		\
-	row2  += ymm0 * ymm2;						\
-	row3  += ymm0 * ymm3;						\
-	ymm2   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 4));		\
-	ymm3   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 5));		\
-	row4  += ymm0 * ymm2;						\
-	row5  += ymm0 * ymm3;						\
-	ymm2   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 6));		\
-	ymm3   =  _mm256_broadcastss_ps(_mm_load_ss(BO + 7));		\
-	row6  += ymm0 * ymm2;						\
-	row7  += ymm0 * ymm3;						\
-	BO  += 8;							\
-	AO  += 8;
-
-
-#define SAVE8x8(ALPHA)							\
-	ymm0   = _mm256_set1_ps(ALPHA);					\
-	row0  *= ymm0;							\
-	row1  *= ymm0;							\
-	row2  *= ymm0;							\
-	row3  *= ymm0;							\
-	row4  *= ymm0;							\
-	row5  *= ymm0;							\
-	row6  *= ymm0;							\
-	row7  *= ymm0;							\
-	row0  += _mm256_loadu_ps(CO1 + 0 * ldc);			\
-	row1  += _mm256_loadu_ps(CO1 + 1 * ldc);			\
-	row2  += _mm256_loadu_ps(CO1 + 2 * ldc);			\
-	row3  += _mm256_loadu_ps(CO1 + 3 * ldc);			\
-	row4  += _mm256_loadu_ps(CO1 + 4 * ldc);			\
-	row5  += _mm256_loadu_ps(CO1 + 5 * ldc);			\
-	row6  += _mm256_loadu_ps(CO1 + 6 * ldc);			\
-	row7  += _mm256_loadu_ps(CO1 + 7 * ldc);			\
-	_mm256_storeu_ps(CO1 + 0 * ldc, row0);				\
-	_mm256_storeu_ps(CO1 + 1 * ldc, row1);				\
-	_mm256_storeu_ps(CO1 + 2 * ldc, row2);				\
-	_mm256_storeu_ps(CO1 + 3 * ldc, row3);				\
-	_mm256_storeu_ps(CO1 + 4 * ldc, row4);				\
-	_mm256_storeu_ps(CO1 + 5 * ldc, row5);				\
-	_mm256_storeu_ps(CO1 + 6 * ldc, row6);				\
-	_mm256_storeu_ps(CO1 + 7 * ldc, row7);				\
-
-
-
-/*******************************************************************************************/
-
-#define INIT4x8()							\
-	row0 = _mm_setzero_ps();					\
-	row1 = _mm_setzero_ps();					\
-	row2 = _mm_setzero_ps();					\
-	row3 = _mm_setzero_ps();					\
-	row4 = _mm_setzero_ps();					\
-	row5 = _mm_setzero_ps();					\
-	row6 = _mm_setzero_ps();					\
-	row7 = _mm_setzero_ps();					\
-
-
-#define KERNEL4x8_SUB() 						\
-	xmm0   = _mm_loadu_ps(AO);					\
-	xmm2   =  _mm_broadcastss_ps(_mm_load_ss(BO + 0));		\
-	xmm3   =  _mm_broadcastss_ps(_mm_load_ss(BO + 1));		\
-	row0  += xmm0 * xmm2;						\
-	row1  += xmm0 * xmm3;						\
-	xmm2   =  _mm_broadcastss_ps(_mm_load_ss(BO + 2));		\
-	xmm3   =  _mm_broadcastss_ps(_mm_load_ss(BO + 3));		\
-	row2  += xmm0 * xmm2;						\
-	row3  += xmm0 * xmm3;						\
-	xmm2   =  _mm_broadcastss_ps(_mm_load_ss(BO + 4));		\
-	xmm3   =  _mm_broadcastss_ps(_mm_load_ss(BO + 5));		\
-	row4  += xmm0 * xmm2;						\
-	row5  += xmm0 * xmm3;						\
-	xmm2   =  _mm_broadcastss_ps(_mm_load_ss(BO + 6));		\
-	xmm3   =  _mm_broadcastss_ps(_mm_load_ss(BO + 7));		\
-	row6  += xmm0 * xmm2;						\
-	row7  += xmm0 * xmm3;						\
-	BO  += 8;							\
-	AO  += 4;
-
-
-#define SAVE4x8(ALPHA)							\
-	xmm0   = _mm_set1_ps(ALPHA);					\
-	row0  *= xmm0;							\
-	row1  *= xmm0;							\
-	row2  *= xmm0;							\
-	row3  *= xmm0;							\
-	row4  *= xmm0;							\
-	row5  *= xmm0;							\
-	row6  *= xmm0;							\
-	row7  *= xmm0;							\
-	row0  += _mm_loadu_ps(CO1 + 0 * ldc);				\
-	row1  += _mm_loadu_ps(CO1 + 1 * ldc);				\
-	row2  += _mm_loadu_ps(CO1 + 2 * ldc);				\
-	row3  += _mm_loadu_ps(CO1 + 3 * ldc);				\
-	row4  += _mm_loadu_ps(CO1 + 4 * ldc);				\
-	row5  += _mm_loadu_ps(CO1 + 5 * ldc);				\
-	row6  += _mm_loadu_ps(CO1 + 6 * ldc);				\
-	row7  += _mm_loadu_ps(CO1 + 7 * ldc);				\
-	_mm_storeu_ps(CO1 + 0 * ldc, row0);				\
-	_mm_storeu_ps(CO1 + 1 * ldc, row1);				\
-	_mm_storeu_ps(CO1 + 2 * ldc, row2);				\
-	_mm_storeu_ps(CO1 + 3 * ldc, row3);				\
-	_mm_storeu_ps(CO1 + 4 * ldc, row4);				\
-	_mm_storeu_ps(CO1 + 5 * ldc, row5);				\
-	_mm_storeu_ps(CO1 + 6 * ldc, row6);				\
-	_mm_storeu_ps(CO1 + 7 * ldc, row7);				\
-
-
-/*******************************************************************************************/
-
-#define INIT2x8() 	\
-	row0a = row0b = 0; 						\
-	row1a = row1b = 0; 						\
-	row2a = row2b = 0; 						\
-	row3a = row3b = 0; 						\
-	row4a = row4b = 0; 						\
-	row5a = row5b = 0; 						\
-	row6a = row6b = 0; 						\
-	row7a = row7b = 0; 						\
-
-#define KERNEL2x8_SUB()							\
-	xmm0  = *(AO);							\
-	xmm1  = *(AO + 1);						\
-	xmm2  = *(BO + 0);						\
-	xmm3  = *(BO + 1);						\
-	row0a += xmm0 * xmm2;						\
-	row0b += xmm1 * xmm2;						\
-	row1a += xmm0 * xmm3;						\
-	row1b += xmm1 * xmm3;						\
-	xmm2 = *(BO + 2);						\
-	xmm3 = *(BO + 3);						\
-	row2a += xmm0 * xmm2;						\
-	row2b += xmm1 * xmm2;						\
-	row3a += xmm0 * xmm3;						\
-	row3b += xmm1 * xmm3;						\
-	xmm2  = *(BO + 4);						\
-	xmm3  = *(BO + 5);						\
-	row4a += xmm0 * xmm2;						\
-	row4b += xmm1 * xmm2;						\
-	row5a += xmm0 * xmm3;						\
-	row5b += xmm1 * xmm3;						\
-	xmm2 = *(BO + 6);						\
-	xmm3 = *(BO + 7);						\
-	row6a += xmm0 * xmm2;						\
-	row6b += xmm1 * xmm2;						\
-	row7a += xmm0 * xmm3;						\
-	row7b += xmm1 * xmm3;						\
-	BO += 8;							\
-	AO += 2;
-
-
-#define SAVE2x8(ALPHA)							\
-	xmm0   = ALPHA;							\
-	row0a  *= xmm0;							\
-	row0b  *= xmm0;							\
-	row1a  *= xmm0;							\
-	row1b  *= xmm0;							\
-	row2a  *= xmm0;							\
-	row2b  *= xmm0;							\
-	row3a  *= xmm0;							\
-	row3b  *= xmm0;							\
-	row4a  *= xmm0;							\
-	row4b  *= xmm0;							\
-	row5a  *= xmm0;							\
-	row5b  *= xmm0;							\
-	row6a  *= xmm0;							\
-	row6b  *= xmm0;							\
-	row7a  *= xmm0;							\
-	row7b  *= xmm0;							\
-	*(CO1 + 0 * ldc + 0) += row0a;					\
-	*(CO1 + 0 * ldc + 1) += row0b;					\
-	*(CO1 + 1 * ldc + 0) += row1a;					\
-	*(CO1 + 1 * ldc + 1) += row1b;					\
-	*(CO1 + 2 * ldc + 0) += row2a;					\
-	*(CO1 + 2 * ldc + 1) += row2b;					\
-	*(CO1 + 3 * ldc + 0) += row3a;					\
-	*(CO1 + 3 * ldc + 1) += row3b;					\
-	*(CO1 + 4 * ldc + 0) += row4a;					\
-	*(CO1 + 4 * ldc + 1) += row4b;					\
-	*(CO1 + 5 * ldc + 0) += row5a;					\
-	*(CO1 + 5 * ldc + 1) += row5b;					\
-	*(CO1 + 6 * ldc + 0) += row6a;					\
-	*(CO1 + 6 * ldc + 1) += row6b;					\
-	*(CO1 + 7 * ldc + 0) += row7a;					\
-	*(CO1 + 7 * ldc + 1) += row7b;					\
-
-
-
-/*******************************************************************************************/
-
-#define INIT1x8() \
-	row0 = row1 = row2 = row3 = row4 = row5 = row6 = row7 = 0;
-
-#define KERNEL1x8_SUB()							\
-	xmm0   = *(AO );						\
-	xmm2   = *(BO + 0);						\
-	xmm3   = *(BO + 1);						\
-	row0  += xmm0 * xmm2;						\
-	row1  += xmm0 * xmm3;						\
-	xmm2   = *(BO + 2);						\
-	xmm3   = *(BO + 3);						\
-	row2  += xmm0 * xmm2;						\
-	row3  += xmm0 * xmm3;						\
-	xmm2   = *(BO + 4);						\
-	xmm3   = *(BO + 5);						\
-	row4  += xmm0 * xmm2;						\
-	row5  += xmm0 * xmm3;						\
-	xmm2   = *(BO + 6);						\
-	xmm3   = *(BO + 7);						\
-	row6  += xmm0 * xmm2;						\
-	row7  += xmm0 * xmm3;						\
-	BO += 8;							\
-	AO += 1;
-
-
-#define SAVE1x8(ALPHA)							\
-	xmm0   = ALPHA;							\
-	row0  *= xmm0;							\
-	row1  *= xmm0;							\
-	row2  *= xmm0;							\
-	row3  *= xmm0;							\
-	row4  *= xmm0;							\
-	row5  *= xmm0;							\
-	row6  *= xmm0;							\
-	row7  *= xmm0;							\
-	*(CO1 + 0 * ldc) += row0;					\
-	*(CO1 + 1 * ldc) += row1;					\
-	*(CO1 + 2 * ldc) += row2;					\
-	*(CO1 + 3 * ldc) += row3;					\
-	*(CO1 + 4 * ldc) += row4;					\
-	*(CO1 + 5 * ldc) += row5;					\
-	*(CO1 + 6 * ldc) += row6;					\
-	*(CO1 + 7 * ldc) += row7;					\
 
 
 
@@ -1184,142 +771,6 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f
 		return 0;
 
 
-
-	// L8_0
-	while (N >= 8 && 0) {
-		float *CO1;
-		float *AO;
-		int i;
-		// L8_10
-		CO1 = C;
-		C += 8 * ldc;
-
-		AO = A;
-
-		i = m;
-
-		while (i >= 32 && 0) {
-			float *BO, *AOb;
-			// L8_11
-			__m512 zmm0, zmm0b, zmm2, zmm3, row0, row1, row2, row3, row4, row5, row6, row7, row0b, row1b, row2b, row3b, row4b, row5b, row6b, row7b;
-			BO = B;
-			int kloop = K;
-			AOb = AO + 16 * K;
-	
-			INIT32x8()
-
-			while (kloop > 0) {
-				// L12_17
-				KERNEL32x8_SUB()
-				kloop--;
-			}
-			// L8_19
-			SAVE32x8(alpha)
-			CO1 += 32;
-			AO  += 16 * K;
-	
-			i   -= 32;
-		}
-		while (i >= 16) {
-			float *BO;
-			// L8_11
-			__m512 zmm0, zmm2, zmm3, row0, row1, row2, row3, row4, row5, row6, row7;
-			BO = B;
-			int kloop = K;
-	
-			INIT16x8()
-
-			while (kloop > 0) {
-				KERNEL16x8_SUB()
-				kloop--;
-			}
-			SAVE16x8(alpha)
-			CO1 += 16;
-	
-			i -= 16;
-		}
-		while (i >= 8) {
-			float *BO;
-			// L8_11
-			__m256 ymm0, ymm2, ymm3, row0, row1, row2, row3, row4, row5, row6, row7;
-			BO = B;
-			int kloop = K;
-	
-			INIT8x8()
-
-			while (kloop > 0) {
-				// L12_17
-				KERNEL8x8_SUB()
-				kloop--;
-			}
-			// L8_19
-			SAVE8x8(alpha)
-			CO1 += 8;
-	
-			i -= 8;
-		}
-		while (i >= 4) {
-			// L8_11
-			float *BO;
-			__m128 xmm0, xmm2, xmm3, row0, row1, row2, row3, row4, row5, row6, row7;
-			BO = B;
-			int kloop = K;
-
-			INIT4x8()
-			// L8_16
-			while (kloop > 0) {
-				// L12_17
-				KERNEL4x8_SUB()
-				kloop--;
-			}
-			// L8_19
-			SAVE4x8(alpha)
-			CO1 += 4;
-
-			i -= 4;
-		}
-
-/**************************************************************************
-* Rest of M 
-***************************************************************************/
-
-		while (i >= 2) {
-			float *BO;
-			float xmm0, xmm1, xmm2, xmm3, row0a, row1a, row2a, row3a, row4a, row5a, row6a, row7a, row0b, row1b, row2b, row3b, row4b, row5b, row6b, row7b;
-			BO = B;
-
-			INIT2x8()
-			int kloop = K;
-			
-			while (kloop > 0) {
-				KERNEL2x8_SUB()
-				kloop--;
-			}
-			SAVE2x8(alpha)
-			CO1 += 2;
-			i -= 2;
-		}
-			// L13_40
-		while (i >= 1) {
-			float *BO;
-			float xmm0, xmm2, xmm3, row0, row1, row2, row3, row4, row5, row6, row7;
-			int kloop = K;
-			BO = B;
-			INIT1x8()
-				
-			while (kloop > 0) {
-				KERNEL1x8_SUB()
-				kloop--;
-			}
-			SAVE1x8(alpha)
-			CO1 += 1;
-			i -= 1;
-		}
-			
-		B += K * 8;
-		N -= 8;
-	}
-
 	while (N >= 4) {
 		float *CO1;
 		float *AO;