rvv: import parallel vf(w)redsum hardware impl.

The number of vector FP ALUs and implementations of vf(w)redsum could be
passed as options by the following example:

"--varch=vlen:512,elen:32,slen:512,nalu:4,fredsum-impl:parallel"

By default, 4 of vector FP ALUs and ordered vector FP reduction sum
implementations are assumed.

4 files changed, 47 insertions, 14 deletions

diff --git a/riscv/insns/vfredsum_vs.h b/riscv/insns/vfredsum_vs.h
index 7b5cccc..ca8e91b 100644
--- a/riscv/insns/vfredsum_vs.h
+++ b/riscv/insns/vfredsum_vs.h
@@ -1,8 +1,8 @@
 // vfredsum: vd[0] =  sum( vs2[*] , vs1[0] )
-VI_VFP_VV_LOOP_REDUCTION
-({
-  vd_0 = f32_add(vd_0, vs2);
-},
-{
-  vd_0 = f64_add(vd_0, vs2);
-})
+
+if(p->VU.FREDSUM_IMPL == "ordered") {
+  #include "vfredosum_vs.h"
+} else if (p->VU.FREDSUM_IMPL == "parallel") {
+  #include "vfredsum_vs_parallel.h"
+} else
+  require(0);
diff --git a/riscv/insns/vfredsum_vs_parallel.h b/riscv/insns/vfredsum_vs_parallel.h
new file mode 100644
index 0000000..d611c60
--- /dev/null
+++ b/riscv/insns/vfredsum_vs_parallel.h
@@ -0,0 +1,24 @@
+// Parallel version of vfredsum
+// vfredsum: vd[0] =  sum( vs2[*] , vs1[0] )
+
+VI_CHECK_REDUCTION(false)
+VI_VFP_COMMON
+
+switch(p->VU.vsew) {
+  case e32: {
+    VI_VFP_LOOP_REDUCTIONSUM_INIT(32)
+    VI_VFP_LOOP_REDUCTIONSUM_MERGE(32)
+    VI_VFP_LOOP_REDUCTIONSUM_CLOSE(e32)
+    break;
+  }
+  case e64: {
+    VI_VFP_LOOP_REDUCTIONSUM_INIT(64)
+    VI_VFP_LOOP_REDUCTIONSUM_MERGE(64)
+    VI_VFP_LOOP_REDUCTIONSUM_CLOSE(e64)
+    break;
+  }
+  case e16:
+  default:
+    require(0);
+    break;
+};
diff --git a/riscv/insns/vfwredsum_vs.h b/riscv/insns/vfwredsum_vs.h
index 3426ef8..bed3c64 100644
--- a/riscv/insns/vfwredsum_vs.h
+++ b/riscv/insns/vfwredsum_vs.h
@@ -1,8 +1,7 @@
 // vfwredsum.vs vd, vs2, vs1
-require_vector;
-require(P.VU.vsew * 2 <= P.VU.ELEN);
-require((insn.rs2() & (P.VU.vlmul - 1)) == 0);
-VI_VFP_VV_LOOP_WIDE_REDUCTION
-({
-  vd_0 = f64_add(vd_0, vs2);
-})
+if(p->VU.FREDSUM_IMPL == "ordered") {
+  #include "vfwredosum_vs.h"
+} else if (p->VU.FREDSUM_IMPL == "parallel") {
+  #include "vfwredsum_vs_parallel.h"
+} else
+  require(0);
diff --git a/riscv/insns/vfwredsum_vs_parallel.h b/riscv/insns/vfwredsum_vs_parallel.h
new file mode 100644
index 0000000..236d4d0
--- /dev/null
+++ b/riscv/insns/vfwredsum_vs_parallel.h
@@ -0,0 +1,10 @@
+// Parallel version of vfwredsum
+require_vector;
+require(P.VU.vsew * 2 <= P.VU.ELEN);
+require((insn.rs2() & (P.VU.vlmul - 1)) == 0);
+
+VI_VFP_COMMON
+
+VI_VFP_LOOP_REDUCTIONSUM_WIDEN_INIT
+VI_VFP_LOOP_REDUCTIONSUM_MERGE(64)
+VI_VFP_LOOP_REDUCTIONSUM_CLOSE(e64)