Merge pull request #5942 from HecaiYuan/develop

loongarch64: fix segfaults in copy kernels and adjust dsyrk block size

Author: martin-frbg

driver/level3/level3_syrk_threaded.c

@@ -37,6 +37,11 @@
 /* or implied, of The University of Texas at Austin.                 */
 /*********************************************************************/
 
+#if defined(__loongarch__) && defined(LA464) && defined(DOUBLE)
+#undef GEMM_P
+#define GEMM_P 96
+#endif
+
 #ifndef CACHE_LINE_SIZE
 #define CACHE_LINE_SIZE 8
 #endif

kernel/loongarch64/dgemm_ncopy_4_lsx.S

@@ -172,6 +172,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     blt        ZERO, I,    .L_2II1
 .L_N1:
     move      S1,    TS
+    andi      J,     N,   0x01
+    beq       ZERO,  J,   .L_N0
     beq       ZERO,  M,   .L_N0
 .L_M1:
     fld.d     F0,    S1,  0x00

kernel/loongarch64/sgemm_ncopy_16_lasx.S

@@ -453,6 +453,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .align 5
 .L_N1:
     move      S1,    TS
+    andi      J,     N,   0x01
+    beq       ZERO,  J,   .L_N0
     beq       ZERO,  M,   .L_N0
 .L_N1_M1:
     fld.s       F0,    S1,  0x00

kernel/loongarch64/sgemm_ncopy_8_lasx.S

@@ -288,7 +288,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .align 5
 .L_N1:
     move      S1,    TS
+    andi      J,     N,   0x01
+    beq       ZERO,  J,   .L_N0
     beq       ZERO,  M,   .L_N0
+
 .L_N1_M1:
     fld.s       F0,    S1,  0x00
     PTR_ADDI    S1,    S1,  0x04