Test wider range of float values

This requires a few fixes:
- Cleaning up tolerances of tests
- Fix infinity -> int conversion crashes
- The RVV vrnd ukernels work by converting to int and back. This loses the "special" float values (inf, NaN). This change tries to at least preserve inf, matching NaN is a little trickier and also not quite as bad to lose (but we should find a way to do so).

PiperOrigin-RevId: 734782315

BUILD.bazel

@@ -559,6 +559,7 @@ xnnpack_cxx_library(
         ":common",
         ":datatype",
         ":math",
+        ":reference_ukernels",
         ":xnnpack_h",
     ],
 )

CMakeLists.txt

@@ -870,7 +870,7 @@ IF(XNNPACK_BUILD_LIBRARY)
   TARGET_LINK_LIBRARIES(operators PRIVATE xnnpack-base allocator indirection logging microkernel-utils normalization operator-utils packing reference-ukernels datatype)
   TARGET_LINK_LIBRARIES(operator-run PRIVATE xnnpack-base logging)
   TARGET_LINK_LIBRARIES(operator-utils PRIVATE xnnpack-base logging)
-  TARGET_LINK_LIBRARIES(reference-ukernels PRIVATE xnnpack-base)
+  TARGET_LINK_LIBRARIES(reference-ukernels PRIVATE xnnpack-base datatype)
   TARGET_LINK_LIBRARIES(subgraph PRIVATE xnnpack-base allocator logging memory mutex operators operator-run datatype)
   TARGET_LINK_LIBRARIES(XNNPACK PRIVATE xnnpack-base allocator cache hardware-config indirection memory microkernel-utils microparams-init mutex normalization operators operator-run operator-utils packing microkernels-prod subgraph datatype reference-ukernels)
   TARGET_LINK_LIBRARIES(XNNPACK PUBLIC pthreadpool logging)
@@ -1754,6 +1754,7 @@ IF(XNNPACK_BUILD_TESTS)
         GTest::gmock
         GTest::gtest
         GTest::gtest_main
+        datatype
         hardware-config
         logging
         microkernels-all

src/f32-vrnd/gen/f32-vrndd-rvv-u1v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndd_ukernel__rvv_u1v(
   do {
     const size_t n = __riscv_vsetvl_e32m1(batch);
     vfloat32m1_t x_f32v = __riscv_vle32_v_f32m1(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool32_t inf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, INFINITY, n);
+    vbool32_t ninf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, -INFINITY, n);
+    vbool32_t mask_bv = __riscv_vmor_mm_b32(inf_bv, ninf_bv, n);
     vint32m1_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m1_rm(x_f32v, __RISCV_FRM_RDN, n);
     vfloat32m1_t out_f32v = __riscv_vfcvt_f_x_v_f32m1(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m1(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m1(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndd-rvv-u2v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndd_ukernel__rvv_u2v(
   do {
     const size_t n = __riscv_vsetvl_e32m2(batch);
     vfloat32m2_t x_f32v = __riscv_vle32_v_f32m2(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool16_t inf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, INFINITY, n);
+    vbool16_t ninf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, -INFINITY, n);
+    vbool16_t mask_bv = __riscv_vmor_mm_b16(inf_bv, ninf_bv, n);
     vint32m2_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m2_rm(x_f32v, __RISCV_FRM_RDN, n);
     vfloat32m2_t out_f32v = __riscv_vfcvt_f_x_v_f32m2(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m2(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m2(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndd-rvv-u4v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndd_ukernel__rvv_u4v(
   do {
     const size_t n = __riscv_vsetvl_e32m4(batch);
     vfloat32m4_t x_f32v = __riscv_vle32_v_f32m4(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool8_t inf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, INFINITY, n);
+    vbool8_t ninf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, -INFINITY, n);
+    vbool8_t mask_bv = __riscv_vmor_mm_b8(inf_bv, ninf_bv, n);
     vint32m4_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m4_rm(x_f32v, __RISCV_FRM_RDN, n);
     vfloat32m4_t out_f32v = __riscv_vfcvt_f_x_v_f32m4(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m4(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m4(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndd-rvv-u8v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndd_ukernel__rvv_u8v(
   do {
     const size_t n = __riscv_vsetvl_e32m8(batch);
     vfloat32m8_t x_f32v = __riscv_vle32_v_f32m8(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool4_t inf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, INFINITY, n);
+    vbool4_t ninf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, -INFINITY, n);
+    vbool4_t mask_bv = __riscv_vmor_mm_b4(inf_bv, ninf_bv, n);
     vint32m8_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m8_rm(x_f32v, __RISCV_FRM_RDN, n);
     vfloat32m8_t out_f32v = __riscv_vfcvt_f_x_v_f32m8(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m8(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m8(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndne-rvv-u1v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndne_ukernel__rvv_u1v(
   do {
     const size_t n = __riscv_vsetvl_e32m1(batch);
     vfloat32m1_t x_f32v = __riscv_vle32_v_f32m1(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool32_t inf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, INFINITY, n);
+    vbool32_t ninf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, -INFINITY, n);
+    vbool32_t mask_bv = __riscv_vmor_mm_b32(inf_bv, ninf_bv, n);
     vint32m1_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m1_rm(x_f32v, __RISCV_FRM_RNE, n);
     vfloat32m1_t out_f32v = __riscv_vfcvt_f_x_v_f32m1(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m1(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m1(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndne-rvv-u2v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndne_ukernel__rvv_u2v(
   do {
     const size_t n = __riscv_vsetvl_e32m2(batch);
     vfloat32m2_t x_f32v = __riscv_vle32_v_f32m2(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool16_t inf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, INFINITY, n);
+    vbool16_t ninf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, -INFINITY, n);
+    vbool16_t mask_bv = __riscv_vmor_mm_b16(inf_bv, ninf_bv, n);
     vint32m2_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m2_rm(x_f32v, __RISCV_FRM_RNE, n);
     vfloat32m2_t out_f32v = __riscv_vfcvt_f_x_v_f32m2(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m2(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m2(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndne-rvv-u4v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndne_ukernel__rvv_u4v(
   do {
     const size_t n = __riscv_vsetvl_e32m4(batch);
     vfloat32m4_t x_f32v = __riscv_vle32_v_f32m4(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool8_t inf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, INFINITY, n);
+    vbool8_t ninf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, -INFINITY, n);
+    vbool8_t mask_bv = __riscv_vmor_mm_b8(inf_bv, ninf_bv, n);
     vint32m4_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m4_rm(x_f32v, __RISCV_FRM_RNE, n);
     vfloat32m4_t out_f32v = __riscv_vfcvt_f_x_v_f32m4(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m4(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m4(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndne-rvv-u8v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndne_ukernel__rvv_u8v(
   do {
     const size_t n = __riscv_vsetvl_e32m8(batch);
     vfloat32m8_t x_f32v = __riscv_vle32_v_f32m8(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool4_t inf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, INFINITY, n);
+    vbool4_t ninf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, -INFINITY, n);
+    vbool4_t mask_bv = __riscv_vmor_mm_b4(inf_bv, ninf_bv, n);
     vint32m8_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m8_rm(x_f32v, __RISCV_FRM_RNE, n);
     vfloat32m8_t out_f32v = __riscv_vfcvt_f_x_v_f32m8(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m8(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m8(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndu-rvv-u1v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndu_ukernel__rvv_u1v(
   do {
     const size_t n = __riscv_vsetvl_e32m1(batch);
     vfloat32m1_t x_f32v = __riscv_vle32_v_f32m1(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool32_t inf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, INFINITY, n);
+    vbool32_t ninf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, -INFINITY, n);
+    vbool32_t mask_bv = __riscv_vmor_mm_b32(inf_bv, ninf_bv, n);
     vint32m1_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m1_rm(x_f32v, __RISCV_FRM_RUP, n);
     vfloat32m1_t out_f32v = __riscv_vfcvt_f_x_v_f32m1(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m1(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m1(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndu-rvv-u2v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndu_ukernel__rvv_u2v(
   do {
     const size_t n = __riscv_vsetvl_e32m2(batch);
     vfloat32m2_t x_f32v = __riscv_vle32_v_f32m2(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool16_t inf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, INFINITY, n);
+    vbool16_t ninf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, -INFINITY, n);
+    vbool16_t mask_bv = __riscv_vmor_mm_b16(inf_bv, ninf_bv, n);
     vint32m2_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m2_rm(x_f32v, __RISCV_FRM_RUP, n);
     vfloat32m2_t out_f32v = __riscv_vfcvt_f_x_v_f32m2(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m2(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m2(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndu-rvv-u4v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndu_ukernel__rvv_u4v(
   do {
     const size_t n = __riscv_vsetvl_e32m4(batch);
     vfloat32m4_t x_f32v = __riscv_vle32_v_f32m4(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool8_t inf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, INFINITY, n);
+    vbool8_t ninf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, -INFINITY, n);
+    vbool8_t mask_bv = __riscv_vmor_mm_b8(inf_bv, ninf_bv, n);
     vint32m4_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m4_rm(x_f32v, __RISCV_FRM_RUP, n);
     vfloat32m4_t out_f32v = __riscv_vfcvt_f_x_v_f32m4(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m4(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m4(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndu-rvv-u8v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndu_ukernel__rvv_u8v(
   do {
     const size_t n = __riscv_vsetvl_e32m8(batch);
     vfloat32m8_t x_f32v = __riscv_vle32_v_f32m8(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool4_t inf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, INFINITY, n);
+    vbool4_t ninf_bv = __riscv_vmfeq_vf_f32m8_b4(x_f32v, -INFINITY, n);
+    vbool4_t mask_bv = __riscv_vmor_mm_b4(inf_bv, ninf_bv, n);
     vint32m8_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m8_rm(x_f32v, __RISCV_FRM_RUP, n);
     vfloat32m8_t out_f32v = __riscv_vfcvt_f_x_v_f32m8(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m8(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m8(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndz-rvv-u1v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndz_ukernel__rvv_u1v(
   do {
     const size_t n = __riscv_vsetvl_e32m1(batch);
     vfloat32m1_t x_f32v = __riscv_vle32_v_f32m1(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool32_t inf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, INFINITY, n);
+    vbool32_t ninf_bv = __riscv_vmfeq_vf_f32m1_b32(x_f32v, -INFINITY, n);
+    vbool32_t mask_bv = __riscv_vmor_mm_b32(inf_bv, ninf_bv, n);
     vint32m1_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m1_rm(x_f32v, __RISCV_FRM_RTZ, n);
     vfloat32m1_t out_f32v = __riscv_vfcvt_f_x_v_f32m1(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m1(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m1(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndz-rvv-u2v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndz_ukernel__rvv_u2v(
   do {
     const size_t n = __riscv_vsetvl_e32m2(batch);
     vfloat32m2_t x_f32v = __riscv_vle32_v_f32m2(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool16_t inf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, INFINITY, n);
+    vbool16_t ninf_bv = __riscv_vmfeq_vf_f32m2_b16(x_f32v, -INFINITY, n);
+    vbool16_t mask_bv = __riscv_vmor_mm_b16(inf_bv, ninf_bv, n);
     vint32m2_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m2_rm(x_f32v, __RISCV_FRM_RTZ, n);
     vfloat32m2_t out_f32v = __riscv_vfcvt_f_x_v_f32m2(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m2(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m2(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);

src/f32-vrnd/gen/f32-vrndz-rvv-u4v.c

@@ -8,6 +8,7 @@
 // LICENSE file in the root directory of this source tree.
 
 #include <assert.h>
+#include <math.h>
 
 #include <riscv_vector.h>
 
@@ -31,8 +32,15 @@ void xnn_f32_vrndz_ukernel__rvv_u4v(
   do {
     const size_t n = __riscv_vsetvl_e32m4(batch);
     vfloat32m4_t x_f32v = __riscv_vle32_v_f32m4(input, n); input += n;
+    // We need to remember which values are infinity, so we can preserve them
+    // after rounding.
+    // TODO: We should also preserve NaN.
+    vbool8_t inf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, INFINITY, n);
+    vbool8_t ninf_bv = __riscv_vmfeq_vf_f32m4_b8(x_f32v, -INFINITY, n);
+    vbool8_t mask_bv = __riscv_vmor_mm_b8(inf_bv, ninf_bv, n);
     vint32m4_t x_rnd_i32v = __riscv_vfcvt_x_f_v_i32m4_rm(x_f32v, __RISCV_FRM_RTZ, n);
     vfloat32m4_t out_f32v = __riscv_vfcvt_f_x_v_f32m4(x_rnd_i32v, n);
+    out_f32v = __riscv_vmerge_vvm_f32m4(out_f32v, x_f32v, mask_bv, n);
     __riscv_vse32_v_f32m4(output, out_f32v, n); output += n;
     batch -= n;
   } while (batch != 0);