STRSM optimization for MIPS P5600 and I6400 using MSA

Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>

CONTRIBUTORS.md

@@ -160,3 +160,4 @@ In chronological order:
 
 * Kaustubh Raste <https://github.com/ksraste/>
   * [2016-05-09] DTRSM optimization for MIPS P5600 and I6400 using MSA
+  * [2016-05-20] STRSM optimization for MIPS P5600 and I6400 using MSA

kernel/mips/KERNEL.P5600

@@ -113,10 +113,10 @@ ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
 ZGEMMONCOPYOBJ = zgemm_oncopy.o
 ZGEMMOTCOPYOBJ = zgemm_otcopy.o
 
-STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
+STRSMKERNEL_LN = ../mips/strsm_kernel_LN_8x8_msa.c
+STRSMKERNEL_LT = ../mips/strsm_kernel_LT_8x8_msa.c
+STRSMKERNEL_RN = ../mips/strsm_kernel_RN_8x8_msa.c
+STRSMKERNEL_RT = ../mips/strsm_kernel_RT_8x8_msa.c
 
 DTRSMKERNEL_LN = ../mips/dtrsm_kernel_LN_8x4_msa.c
 DTRSMKERNEL_LT = ../mips/dtrsm_kernel_LT_8x4_msa.c

kernel/mips/dtrsm_kernel_LN_8x4_msa.c

@@ -1170,7 +1170,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
 
     for (j = (n >> 2); j--;)
     {
-        kk = m;
+        kk = m + offset;
 
         if (m & 7)
         {
@@ -1233,7 +1233,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
     {
         if (n & 2)
         {
-            kk = m;
+            kk = m + offset;
 
             if (m & 7)
             {
@@ -1291,7 +1291,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
 
         if (n & 1)
         {
-            kk = m;
+            kk = m + offset;
 
             if (m & 7)
             {

kernel/mips/dtrsm_kernel_LT_8x4_msa.c

@@ -1182,7 +1182,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
 
     for (j = (n >> 2); j--;)
     {
-        kk = 0;
+        kk = offset;
         aa = a;
         cc = c;
 
@@ -1233,7 +1233,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
     {
         if (n & 2)
         {
-            kk = 0;
+            kk = offset;
             aa = a;
             cc = c;
 
@@ -1282,7 +1282,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
 
         if (n & 1)
         {
-            kk = 0;
+            kk = offset;
             aa = a;
             cc = c;
 

kernel/mips/dtrsm_kernel_RN_8x4_msa.c

@@ -809,7 +809,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
     BLASLONG i, j, kk;
     FLOAT *aa, *cc;
 
-    kk = 0;
+    kk = -offset;
 
     for (j = (n >> 2); j--;)
     {

kernel/mips/dtrsm_kernel_RT_8x4_msa.c

@@ -865,7 +865,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1, FLOAT *a, FLOAT *b,
     BLASLONG i, j, kk;
     FLOAT *aa, *cc, *bb;
 
-    kk = n;
+    kk = n - offset;
     c += n * ldc;
     b += n * k;
 

kernel/mips/macros_msa.h

@@ -137,6 +137,30 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 }
 #define ILVRL_D2_DP(...) ILVRL_D2(v2f64, __VA_ARGS__)
 
+/* Description : Indexed word element values are replicated to all
+                 elements in output vector
+   Arguments   : Inputs  - in, stidx
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : 'stidx' element value from 'in' vector is replicated to all
+                 elements in 'out0' vector
+                 'stidx + 1' element value from 'in' vector is replicated to all
+                 elements in 'out1' vector
+                 Valid index range for word operation is 0-3
+*/
+#define SPLATI_W2(RTYPE, in, stidx, out0, out1)            \
+{                                                          \
+    out0 = (RTYPE) __msa_splati_w((v4i32) in, stidx);      \
+    out1 = (RTYPE) __msa_splati_w((v4i32) in, (stidx+1));  \
+}
+
+#define SPLATI_W4(RTYPE, in, out0, out1, out2, out3)  \
+{                                                     \
+    SPLATI_W2(RTYPE, in, 0, out0, out1);              \
+    SPLATI_W2(RTYPE, in, 2, out2, out3);              \
+}
+#define SPLATI_W4_SP(...) SPLATI_W4(v4f32, __VA_ARGS__)
+
 /* Description : Transpose 4x4 block with word elements in vectors
    Arguments   : Inputs  - in0, in1, in2, in3
                  Outputs - out0, out1, out2, out3