[ZARCH] Fix iamax/imax single precision

mtk-i · web-flow · commit c7143c1019d7 · 2019-01-28T17:52:23.000+02:00
diff --git a/kernel/zarch/icamax.c b/kernel/zarch/icamax.c
@@ -248,6 +248,8 @@ static BLASLONG icamax_kernel_32(BLASLONG n, FLOAT *x, FLOAT *amax)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v2,%%v0           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
diff --git a/kernel/zarch/icamin.c b/kernel/zarch/icamin.c
@@ -248,6 +248,8 @@ static BLASLONG icamin_kernel_32(BLASLONG n, FLOAT *x, FLOAT *amin)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v0,%%v2           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
diff --git a/kernel/zarch/isamax.c b/kernel/zarch/isamax.c
@@ -216,6 +216,8 @@ static BLASLONG isamax_kernel_64(BLASLONG n, FLOAT *x, FLOAT *amax)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v2,%%v0           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
diff --git a/kernel/zarch/isamin.c b/kernel/zarch/isamin.c
@@ -216,6 +216,8 @@ static BLASLONG isamin_kernel_64(BLASLONG n, FLOAT *x, FLOAT *amin)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v0,%%v2           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
diff --git a/kernel/zarch/ismax.c b/kernel/zarch/ismax.c
@@ -192,6 +192,8 @@ static BLASLONG ismax_kernel_64(BLASLONG n, FLOAT *x, FLOAT *max)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v2,%%v0           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
diff --git a/kernel/zarch/ismin.c b/kernel/zarch/ismin.c
@@ -192,6 +192,8 @@ static BLASLONG ismin_kernel_64(BLASLONG n, FLOAT *x, FLOAT *min)
         "j 2f                            \n\t"
         "1:                              \n\t"
         "wfchsb %%v4,%%v0,%%v2           \n\t"
+        "vesrlg %%v4,%%v4,32             \n\t"
+        "vsegf  %%v4,%%v4                \n\t"
         "vsel   %%v1,%%v3,%%v1,%%v4      \n\t"
         "vsel   %%v0,%%v2,%%v0,%%v4      \n\t"
         "ste    %%f0,%1                  \n\t"
