diff --git a/Makefrag b/Makefrag
index fea556b..01ad89d 100644
--- a/Makefrag
+++ b/Makefrag
@@ -462,16 +462,24 @@ tests = \
   vloxseg2ei32_v-0 \
   vloxseg2ei32_v-1 \
   vloxseg2ei64_v-0 \
+  vloxseg2ei64_v-1 \
   vloxseg2ei8_v-0 \
   vloxseg2ei8_v-1 \
   vloxseg3ei16_v-0 \
+  vloxseg3ei16_v-1 \
   vloxseg3ei32_v-0 \
+  vloxseg3ei32_v-1 \
   vloxseg3ei64_v-0 \
   vloxseg3ei8_v-0 \
+  vloxseg3ei8_v-1 \
   vloxseg4ei16_v-0 \
+  vloxseg4ei16_v-1 \
   vloxseg4ei32_v-0 \
+  vloxseg4ei32_v-1 \
   vloxseg4ei64_v-0 \
+  vloxseg4ei64_v-1 \
   vloxseg4ei8_v-0 \
+  vloxseg4ei8_v-1 \
   vloxseg5ei16_v-0 \
   vloxseg5ei32_v-0 \
   vloxseg5ei64_v-0 \
@@ -553,17 +561,25 @@ tests = \
   vlseg8e8_v-0 \
   vlseg8e8ff_v-0 \
   vlsseg2e16_v-0 \
+  vlsseg2e16_v-1 \
   vlsseg2e32_v-0 \
+  vlsseg2e32_v-1 \
   vlsseg2e64_v-0 \
+  vlsseg2e64_v-1 \
   vlsseg2e8_v-0 \
+  vlsseg2e8_v-1 \
   vlsseg3e16_v-0 \
   vlsseg3e32_v-0 \
   vlsseg3e64_v-0 \
   vlsseg3e8_v-0 \
   vlsseg4e16_v-0 \
+  vlsseg4e16_v-1 \
   vlsseg4e32_v-0 \
+  vlsseg4e32_v-1 \
   vlsseg4e64_v-0 \
+  vlsseg4e64_v-1 \
   vlsseg4e8_v-0 \
+  vlsseg4e8_v-1 \
   vlsseg5e16_v-0 \
   vlsseg5e32_v-0 \
   vlsseg5e64_v-0 \
@@ -592,16 +608,24 @@ tests = \
   vluxseg2ei32_v-0 \
   vluxseg2ei32_v-1 \
   vluxseg2ei64_v-0 \
+  vluxseg2ei64_v-1 \
   vluxseg2ei8_v-0 \
   vluxseg2ei8_v-1 \
   vluxseg3ei16_v-0 \
+  vluxseg3ei16_v-1 \
   vluxseg3ei32_v-0 \
+  vluxseg3ei32_v-1 \
   vluxseg3ei64_v-0 \
   vluxseg3ei8_v-0 \
+  vluxseg3ei8_v-1 \
   vluxseg4ei16_v-0 \
+  vluxseg4ei16_v-1 \
   vluxseg4ei32_v-0 \
+  vluxseg4ei32_v-1 \
   vluxseg4ei64_v-0 \
+  vluxseg4ei64_v-1 \
   vluxseg4ei8_v-0 \
+  vluxseg4ei8_v-1 \
   vluxseg5ei16_v-0 \
   vluxseg5ei32_v-0 \
   vluxseg5ei64_v-0 \
@@ -1173,6 +1197,7 @@ tests = \
   vsoxseg2ei16_v-0 \
   vsoxseg2ei16_v-1 \
   vsoxseg2ei32_v-0 \
+  vsoxseg2ei32_v-1 \
   vsoxseg2ei64_v-0 \
   vsoxseg2ei8_v-0 \
   vsoxseg2ei8_v-1 \
@@ -1181,9 +1206,12 @@ tests = \
   vsoxseg3ei64_v-0 \
   vsoxseg3ei8_v-0 \
   vsoxseg4ei16_v-0 \
+  vsoxseg4ei16_v-1 \
   vsoxseg4ei32_v-0 \
+  vsoxseg4ei32_v-1 \
   vsoxseg4ei64_v-0 \
   vsoxseg4ei8_v-0 \
+  vsoxseg4ei8_v-1 \
   vsoxseg5ei16_v-0 \
   vsoxseg5ei32_v-0 \
   vsoxseg5ei64_v-0 \
@@ -1224,12 +1252,16 @@ tests = \
   vsrl_vx-5 \
   vsse16_v-0 \
   vsse16_v-1 \
+  vsse16_v-2 \
   vsse32_v-0 \
   vsse32_v-1 \
+  vsse32_v-2 \
   vsse64_v-0 \
   vsse64_v-1 \
+  vsse64_v-2 \
   vsse8_v-0 \
   vsse8_v-1 \
+  vsse8_v-2 \
   vsseg2e16_v-0 \
   vsseg2e32_v-0 \
   vsseg2e64_v-0 \
@@ -1365,13 +1397,21 @@ tests = \
   vssseg2e8_v-0 \
   vssseg2e8_v-1 \
   vssseg3e16_v-0 \
+  vssseg3e16_v-1 \
   vssseg3e32_v-0 \
+  vssseg3e32_v-1 \
   vssseg3e64_v-0 \
+  vssseg3e64_v-1 \
   vssseg3e8_v-0 \
+  vssseg3e8_v-1 \
   vssseg4e16_v-0 \
+  vssseg4e16_v-1 \
   vssseg4e32_v-0 \
+  vssseg4e32_v-1 \
   vssseg4e64_v-0 \
+  vssseg4e64_v-1 \
   vssseg4e8_v-0 \
+  vssseg4e8_v-1 \
   vssseg5e16_v-0 \
   vssseg5e32_v-0 \
   vssseg5e64_v-0 \
@@ -1428,6 +1468,7 @@ tests = \
   vsuxseg2ei16_v-0 \
   vsuxseg2ei16_v-1 \
   vsuxseg2ei32_v-0 \
+  vsuxseg2ei32_v-1 \
   vsuxseg2ei64_v-0 \
   vsuxseg2ei8_v-0 \
   vsuxseg2ei8_v-1 \
@@ -1436,9 +1477,12 @@ tests = \
   vsuxseg3ei64_v-0 \
   vsuxseg3ei8_v-0 \
   vsuxseg4ei16_v-0 \
+  vsuxseg4ei16_v-1 \
   vsuxseg4ei32_v-0 \
+  vsuxseg4ei32_v-1 \
   vsuxseg4ei64_v-0 \
   vsuxseg4ei8_v-0 \
+  vsuxseg4ei8_v-1 \
   vsuxseg5ei16_v-0 \
   vsuxseg5ei32_v-0 \
   vsuxseg5ei64_v-0 \
diff --git a/generator/insn_util.go b/generator/insn_util.go
index e662c26..74d0899 100644
--- a/generator/insn_util.go
+++ b/generator/insn_util.go
@@ -117,7 +117,7 @@ func getEEW(name string) SEW {
 	return SEW(eew)
 }
 
-func getNfieldsRoundedUp(name string) int {
+func getNfields(name string) int {
 	s := regexp.MustCompile(`v.+?seg(\d)e.+?\.v`)
 	subs := s.FindStringSubmatch(name)
 	if len(subs) < 2 {
@@ -127,19 +127,7 @@ func getNfieldsRoundedUp(name string) int {
 	if err != nil {
 		return 1
 	}
-	switch nfields {
-	case 1:
-		return 1
-	case 2:
-		return 2
-	case 3, 4:
-		return 4
-	case 5, 6, 7, 8:
-		return 8
-	default:
-		log.Fatalln("unreachable")
-		return 1
-	}
+	return nfields
 }
 
 func iff[T any](condition bool, t T, f T) T {
diff --git a/generator/insn_vdrs1mrs2vm.go b/generator/insn_vdrs1mrs2vm.go
index 35de7dc..5864fb7 100644
--- a/generator/insn_vdrs1mrs2vm.go
+++ b/generator/insn_vdrs1mrs2vm.go
@@ -2,12 +2,11 @@ package generator
 
 import (
 	"fmt"
-	"math"
 	"strings"
 )
 
 func (i *Insn) genCodeVdRs1mRs2Vm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -23,11 +22,12 @@ func (i *Insn) genCodeVdRs1mRs2Vm(pos int) []string {
 		builder.WriteString(i.gWriteRandomData(LMUL(1)))
 		builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-		lmul1 := LMUL(math.Max(float64(c.LMUL1)*float64(nfields), 1))
-		vd, _, _ := getVRegs(lmul1, false, i.Name)
-		builder.WriteString(i.gWriteRandomData(lmul1))
-		builder.WriteString(i.gLoadDataIntoRegisterGroup(vd, lmul1, c.SEW))
-		builder.WriteString(i.gWriteIntegerTestData(lmul1, c.SEW, 0))
+		vd := int(c.LMUL1)
+		builder.WriteString(i.gWriteRandomData(c.LMUL1))
+		for nf := 0; nf < nfields; nf++ {
+			builder.WriteString(i.gLoadDataIntoRegisterGroup(vd+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+		}
+		builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), c.SEW, 0))
 
 		builder.WriteString("# -------------- TEST BEGIN --------------\n")
 		builder.WriteString(i.gVsetvli(c.Vl, c.SEW, c.LMUL))
@@ -35,15 +35,16 @@ func (i *Insn) genCodeVdRs1mRs2Vm(pos int) []string {
 		builder.WriteString("# -------------- TEST END   --------------\n")
 
 		builder.WriteString(i.gResultDataAddr())
-		builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd, lmul1, c.SEW))
-		builder.WriteString(i.gMagicInsn(vd, lmul1))
+		builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd, c.LMUL1, c.SEW))
+		builder.WriteString(i.gMagicInsn(vd, c.LMUL1))
 
 		for _, stride := range []int{minStride, 0, 1, maxStride} {
 			stride = stride * int(c.SEW) / 8
-			builder.WriteString(i.gWriteRandomData(lmul1))
-			builder.WriteString(i.gLoadDataIntoRegisterGroup(vd, lmul1, c.SEW))
-			builder.WriteString(i.gWriteIntegerTestData(lmul1*strides, c.SEW, 0))
-
+			builder.WriteString(i.gWriteRandomData(c.LMUL1))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gLoadDataIntoRegisterGroup(vd+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+			}
+			builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields*stride), c.SEW, 0))
 			builder.WriteString(fmt.Sprintf("li a5, %d\n", -minStride*i.vlenb()*int(c.LMUL1)))
 			builder.WriteString("add a0, a0, a5\n")
 
@@ -54,8 +55,10 @@ func (i *Insn) genCodeVdRs1mRs2Vm(pos int) []string {
 			builder.WriteString("# -------------- TEST END   --------------\n")
 
 			builder.WriteString(i.gResultDataAddr())
-			builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd, lmul1, c.SEW))
-			builder.WriteString(i.gMagicInsn(vd, lmul1))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+				builder.WriteString(i.gMagicInsn(vd+(nf*int(c.LMUL1)), c.LMUL1))
+			}
 		}
 		res = append(res, builder.String())
 	}
diff --git a/generator/insn_vdrs1mvm.go b/generator/insn_vdrs1mvm.go
index 446023b..09ba259 100644
--- a/generator/insn_vdrs1mvm.go
+++ b/generator/insn_vdrs1mvm.go
@@ -2,12 +2,11 @@ package generator
 
 import (
 	"fmt"
-	"math"
 	"strings"
 )
 
 func (i *Insn) genCodeVdRs1mVm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -23,11 +22,12 @@ func (i *Insn) genCodeVdRs1mVm(pos int) []string {
 		builder.WriteString(i.gWriteRandomData(LMUL(1)))
 		builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-		lmul1 := LMUL(math.Max(float64(c.LMUL1)*float64(nfields), 1))
-		vd, _, _ := getVRegs(lmul1, false, i.Name)
-		builder.WriteString(i.gWriteRandomData(lmul1))
-		builder.WriteString(i.gLoadDataIntoRegisterGroup(vd, lmul1, c.SEW))
-		builder.WriteString(i.gWriteIntegerTestData(lmul1, c.SEW, 0))
+		vd := int(c.LMUL1)
+		builder.WriteString(i.gWriteRandomData(c.LMUL1))
+		for nf := 0; nf < nfields; nf++ {
+			builder.WriteString(i.gLoadDataIntoRegisterGroup(vd+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+		}
+		builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), c.SEW, 0))
 
 		builder.WriteString("# -------------- TEST BEGIN --------------\n")
 		builder.WriteString(i.gVsetvli(c.Vl, c.SEW, c.LMUL))
@@ -35,8 +35,10 @@ func (i *Insn) genCodeVdRs1mVm(pos int) []string {
 		builder.WriteString("# -------------- TEST END   --------------\n")
 
 		builder.WriteString(i.gResultDataAddr())
-		builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd, lmul1, c.SEW))
-		builder.WriteString(i.gMagicInsn(vd, lmul1))
+		for nf := 0; nf < nfields; nf++ {
+			builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+			builder.WriteString(i.gMagicInsn(vd+(nf*int(c.LMUL1)), c.LMUL1))
+		}
 
 		res = append(res, builder.String())
 	}
diff --git a/generator/insn_vdrs1mvs2vm.go b/generator/insn_vdrs1mvs2vm.go
index 9fad7f5..881b0a1 100644
--- a/generator/insn_vdrs1mvs2vm.go
+++ b/generator/insn_vdrs1mvs2vm.go
@@ -7,7 +7,7 @@ import (
 )
 
 func (i *Insn) genCodeVdRs1mVs2Vm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -27,21 +27,29 @@ func (i *Insn) genCodeVdRs1mVs2Vm(pos int) []string {
 			if emul > 8 || emul < 1./8 {
 				continue
 			}
-			emul = math.Max(emul, 1)
+			emul = math.Max(emul, 1) // offset lmul
 			builder.WriteString(c.initialize())
 
 			builder.WriteString(i.gWriteRandomData(LMUL(1)))
 			builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-			lmul1 := LMUL(math.Max(float64(c.LMUL)*float64(nfields), 1))
-			vd := int(lmul1)
-			vs2 := 2 * int(math.Max(emul, float64(int(c.LMUL1)*nfields)))
+			vd := int(c.LMUL1)
 
-			builder.WriteString(i.gWriteIndexData(lmul1, LMUL(emul), c.Vl, sew, c.SEW))
+			vs2 := 0
+			for k := vd+int(c.LMUL1)*nfields; true; k += 1 {
+				if k % int(emul) == 0 {
+					vs2 = k
+					break
+				}
+			}
+
+			builder.WriteString(i.gWriteIndexData(c.LMUL1, LMUL(emul), c.Vl, sew, c.SEW))
 			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs2, LMUL(emul), c.SEW))
-			builder.WriteString(i.gWriteRandomData(lmul1))
-			builder.WriteString(i.gLoadDataIntoRegisterGroup(vd, lmul1, sew))
-			builder.WriteString(i.gWriteIntegerTestData(lmul1, sew, 0))
+			builder.WriteString(i.gWriteRandomData(c.LMUL1))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gLoadDataIntoRegisterGroup(vd+(nf*int(c.LMUL1)), c.LMUL1, sew))
+			}
+			builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), sew, 0))
 
 			builder.WriteString("# -------------- TEST BEGIN --------------\n")
 			builder.WriteString(i.gVsetvli(c.Vl, sew, c.LMUL))
@@ -49,8 +57,10 @@ func (i *Insn) genCodeVdRs1mVs2Vm(pos int) []string {
 			builder.WriteString("# -------------- TEST END   --------------\n")
 
 			builder.WriteString(i.gResultDataAddr())
-			builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd, lmul1, sew))
-			builder.WriteString(i.gMagicInsn(vd, lmul1))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gStoreRegisterGroupIntoResultData(vd+(nf*int(c.LMUL1)), c.LMUL1, sew))
+				builder.WriteString(i.gMagicInsn(vd+(nf*int(c.LMUL1)), c.LMUL1))
+			}
 		}
 		res = append(res, builder.String())
 	}
diff --git a/generator/insn_vs3rs1mrs2vm.go b/generator/insn_vs3rs1mrs2vm.go
index d075f43..4b46bf4 100644
--- a/generator/insn_vs3rs1mrs2vm.go
+++ b/generator/insn_vs3rs1mrs2vm.go
@@ -2,12 +2,11 @@ package generator
 
 import (
 	"fmt"
-	"math"
 	"strings"
 )
 
 func (i *Insn) genCodeVs3Rs1mRs2Vm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -23,12 +22,16 @@ func (i *Insn) genCodeVs3Rs1mRs2Vm(pos int) []string {
 		builder.WriteString(i.gWriteRandomData(LMUL(1)))
 		builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-		lmul1 := LMUL(math.Max(float64(c.LMUL1)*float64(nfields), 1))
-		vs3, _, _ := getVRegs(lmul1, false, i.Name)
+		vs3 := int(c.LMUL1)
 		for _, s := range []int{minStride, 0, 1, maxStride} {
 			stride := s * int(c.SEW) / 8
-			builder.WriteString(i.gWriteIntegerTestData(lmul1, c.SEW, 0))
-			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, c.SEW))
+
+			builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), c.SEW, 0))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+				builder.WriteString(fmt.Sprintf("li a5, %d\n", stride*i.vlenb()*int(c.LMUL1)))
+				builder.WriteString("add a0, a0, a5\n")
+			}
 
 			builder.WriteString(i.gResultDataAddr())
 			builder.WriteString(fmt.Sprintf("li a5, %d\n", -minStride*i.vlenb()*int(c.LMUL1)))
@@ -45,8 +48,8 @@ func (i *Insn) genCodeVs3Rs1mRs2Vm(pos int) []string {
 
 			builder.WriteString("mv a4, a0\n")
 			for a := 0; a < strides; a++ {
-				builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, c.SEW))
-				builder.WriteString(i.gMagicInsn(vs3, lmul1))
+				builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, c.LMUL1, c.SEW))
+				builder.WriteString(i.gMagicInsn(vs3, c.LMUL1))
 				builder.WriteString(fmt.Sprintf("li a5, %d\n", i.vlenb()*int(c.LMUL1)))
 				builder.WriteString(fmt.Sprintf("add a4, a4, a5\n"))
 				builder.WriteString("mv a0, a4\n")
diff --git a/generator/insn_vs3rs1mvm.go b/generator/insn_vs3rs1mvm.go
index 4b9afa6..bd2cd4c 100644
--- a/generator/insn_vs3rs1mvm.go
+++ b/generator/insn_vs3rs1mvm.go
@@ -2,12 +2,11 @@ package generator
 
 import (
 	"fmt"
-	"math"
 	"strings"
 )
 
 func (i *Insn) genCodeVs3Rs1mVm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -23,10 +22,14 @@ func (i *Insn) genCodeVs3Rs1mVm(pos int) []string {
 		builder.WriteString(i.gWriteRandomData(LMUL(1)))
 		builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-		lmul1 := LMUL(math.Max(float64(c.LMUL1)*float64(nfields), 1))
-		vs3, _, _ := getVRegs(lmul1, false, i.Name)
-		builder.WriteString(i.gWriteIntegerTestData(lmul1, c.SEW, 0))
-		builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, c.SEW))
+		vs3 := int(c.LMUL1)
+
+		builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), c.SEW, 0))
+		for nf := 0; nf < nfields; nf++ {
+			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3+(nf*int(c.LMUL1)), c.LMUL1, c.SEW))
+			builder.WriteString(fmt.Sprintf("li a5, %d\n", i.vlenb()*int(c.LMUL1)))
+			builder.WriteString("add a0, a0, a5\n")
+		}
 
 		builder.WriteString(i.gResultDataAddr())
 
@@ -35,8 +38,8 @@ func (i *Insn) genCodeVs3Rs1mVm(pos int) []string {
 		builder.WriteString(fmt.Sprintf("%s v%d, (a0)%s\n", i.Name, vs3, v0t(c.Mask)))
 		builder.WriteString("# -------------- TEST END   --------------\n")
 
-		builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, c.SEW))
-		builder.WriteString(i.gMagicInsn(vs3, lmul1))
+		builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, c.LMUL1, c.SEW))
+		builder.WriteString(i.gMagicInsn(vs3, c.LMUL1))
 
 		res = append(res, builder.String())
 	}
diff --git a/generator/insn_vs3rs1mvs2vm.go b/generator/insn_vs3rs1mvs2vm.go
index c23c8f6..9a9ac41 100644
--- a/generator/insn_vs3rs1mvs2vm.go
+++ b/generator/insn_vs3rs1mvs2vm.go
@@ -7,7 +7,7 @@ import (
 )
 
 func (i *Insn) genCodeVs3Rs1mVs2Vm(pos int) []string {
-	nfields := getNfieldsRoundedUp(i.Name)
+	nfields := getNfields(i.Name)
 	combinations := i.combinations(
 		nfieldsLMULs(nfields),
 		[]SEW{getEEW(i.Name)},
@@ -28,18 +28,30 @@ func (i *Insn) genCodeVs3Rs1mVs2Vm(pos int) []string {
 			if emul > 8 || emul < 1./8 {
 				continue
 			}
-			emul = math.Max(emul, 1)
+			emul = math.Max(emul, 1) // offset lmul
 			builder.WriteString(c.initialize())
 
 			builder.WriteString(i.gWriteRandomData(LMUL(1)))
 			builder.WriteString(i.gLoadDataIntoRegisterGroup(0, LMUL(1), SEW(32)))
 
-			lmul1 := LMUL(math.Max(float64(c.LMUL1)*float64(nfields), 1))
-			vs3 := int(lmul1)
-			vs1 := 2 * int(math.Max(emul, float64(int(c.LMUL1)*nfields)))
-			builder.WriteString(i.gWriteIntegerTestData(lmul1, sew, 0))
-			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, sew))
-			builder.WriteString(i.gWriteIndexData(lmul1, LMUL(emul), c.Vl, sew, c.SEW))
+			vs3 := int(c.LMUL1)
+
+			vs1 := 0
+			for k := vs3+int(c.LMUL1)*nfields; true; k += 1 {
+				if k % int(emul) == 0 {
+					vs1 = k
+					break
+				}
+			}
+
+			builder.WriteString(i.gWriteIntegerTestData(c.LMUL1*LMUL(nfields), sew, 0))
+			for nf := 0; nf < nfields; nf++ {
+				builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3+(nf*int(c.LMUL1)), c.LMUL1, sew))
+				builder.WriteString(fmt.Sprintf("li a5, %d\n", i.vlenb()*int(c.LMUL1)))
+				builder.WriteString("add a0, a0, a5\n")
+			}
+
+			builder.WriteString(i.gWriteIndexData(c.LMUL1, LMUL(emul), c.Vl, sew, c.SEW))
 			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs1, LMUL(emul), c.SEW))
 
 			builder.WriteString(i.gResultDataAddr())
@@ -48,8 +60,8 @@ func (i *Insn) genCodeVs3Rs1mVs2Vm(pos int) []string {
 			builder.WriteString(fmt.Sprintf("%s v%d, (a0), v%d%s\n", i.Name, vs3, vs1, v0t(c.Mask)))
 			builder.WriteString("# -------------- TEST END   --------------\n")
 
-			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, lmul1, sew))
-			builder.WriteString(i.gMagicInsn(vs3, lmul1))
+			builder.WriteString(i.gLoadDataIntoRegisterGroup(vs3, c.LMUL1, sew))
+			builder.WriteString(i.gMagicInsn(vs3, c.LMUL1))
 		}
 		res = append(res, builder.String())
 	}