Merge pull request #448 from howjmay/vqmovun_high

feat: Add vqmovun_high_[s16|s32|s64]
howjmay · Jul 24, 2024 · 8c6804a · 8c6804a
2 parents e4be9a7 + fe209e5
commit 8c6804a
Show file tree

Hide file tree

Showing 3 changed files with 88 additions and 9 deletions.
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -8278,11 +8278,23 @@ FORCE_INLINE uint32x2_t vqmovun_s64(int64x2_t a) {
 
 // FORCE_INLINE uint32_t vqmovund_s64(int64_t a);
 
-// FORCE_INLINE uint8x16_t vqmovun_high_s16(uint8x8_t r, int16x8_t a);
+FORCE_INLINE uint8x16_t vqmovun_high_s16(uint8x8_t r, int16x8_t a) {
+  vuint16m1_t a_non_neg = __riscv_vreinterpret_v_i16m1_u16m1(__riscv_vmax_vx_i16m1(a, 0, 8));
+  vuint8m1_t vqmovun = __riscv_vlmul_ext_v_u8mf2_u8m1(__riscv_vnclipu_wx_u8mf2(a_non_neg, 0, __RISCV_VXRM_RDN, 8));
+  return __riscv_vslideup_vx_u8m1(r, vqmovun, 8, 16);
+}
 
-// FORCE_INLINE uint16x8_t vqmovun_high_s32(uint16x4_t r, int32x4_t a);
+FORCE_INLINE uint16x8_t vqmovun_high_s32(uint16x4_t r, int32x4_t a) {
+  vuint32m1_t a_non_neg = __riscv_vreinterpret_v_i32m1_u32m1(__riscv_vmax_vx_i32m1(a, 0, 4));
+  vuint16m1_t vqmovun = __riscv_vlmul_ext_v_u16mf2_u16m1(__riscv_vnclipu_wx_u16mf2(a_non_neg, 0, __RISCV_VXRM_RDN, 4));
+  return __riscv_vslideup_vx_u16m1(r, vqmovun, 4, 8);
+}
 
-// FORCE_INLINE uint32x4_t vqmovun_high_s64(uint32x2_t r, int64x2_t a);
+FORCE_INLINE uint32x4_t vqmovun_high_s64(uint32x2_t r, int64x2_t a) {
+  vuint64m1_t a_non_neg = __riscv_vreinterpret_v_i64m1_u64m1(__riscv_vmax_vx_i64m1(a, 0, 2));
+  vuint32m1_t vqmovun = __riscv_vlmul_ext_v_u32mf2_u32m1(__riscv_vnclipu_wx_u32mf2(a_non_neg, 0, __RISCV_VXRM_RDN, 2));
+  return __riscv_vslideup_vx_u32m1(r, vqmovun, 2, 4);
+}
 
 FORCE_INLINE int16x8_t vmovl_s8(int8x8_t a) { return __riscv_vlmul_trunc_v_i16m2_i16m1(__riscv_vsext_vf2_i16m2(a, 8)); }
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -29752,11 +29752,78 @@ result_t test_vqmovuns_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { retu
 
 result_t test_vqmovund_s64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 
-result_t test_vqmovun_high_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vqmovun_high_s16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const int16_t *_a = (int16_t *)impl.test_cases_int_pointer1;
+  const uint8_t *_r = (uint8_t *)impl.test_cases_int_pointer2;
+  uint8_t _c[16];
+  for (int i = 0; i < 8; i++) {
+    if (_a[i] < 0) {
+      _c[i + 8] = 0;
+    } else if (_a[i] > UINT8_MAX) {
+      _c[i + 8] = UINT8_MAX;
+    } else {
+      _c[i + 8] = _a[i];
+    }
+    _c[i] = _r[i];
+  }
+  int16x8_t a = vld1q_s16(_a);
+  uint8x8_t r = vld1_u8(_r);
+  uint8x16_t c = vqmovun_high_s16(r, a);
+  return validate_uint8(c, _c[0], _c[1], _c[2], _c[3], _c[4], _c[5], _c[6], _c[7], _c[8], _c[9], _c[10], _c[11], _c[12],
+                        _c[13], _c[14], _c[15]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vqmovun_high_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vqmovun_high_s32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const int32_t *_a = (int32_t *)impl.test_cases_int_pointer1;
+  const uint16_t *_r = (uint16_t *)impl.test_cases_int_pointer2;
+  uint16_t _c[8];
+  for (int i = 0; i < 4; i++) {
+    if (_a[i] < 0) {
+      _c[i + 4] = 0;
+    } else if (_a[i] > UINT16_MAX) {
+      _c[i + 4] = UINT16_MAX;
+    } else {
+      _c[i + 4] = _a[i];
+    }
+    _c[i] = _r[i];
+  }
+  int32x4_t a = vld1q_s32(_a);
+  uint16x4_t r = vld1_u16(_r);
+  uint16x8_t c = vqmovun_high_s32(r, a);
+  return validate_uint16(c, _c[0], _c[1], _c[2], _c[3], _c[4], _c[5], _c[6], _c[7]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vqmovun_high_s64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vqmovun_high_s64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const int64_t *_a = (int64_t *)impl.test_cases_int_pointer1;
+  const uint32_t *_r = (uint32_t *)impl.test_cases_int_pointer2;
+  uint32_t _c[4];
+  for (int i = 0; i < 2; i++) {
+    if (_a[i] < 0) {
+      _c[i + 2] = 0;
+    } else if (_a[i] > UINT32_MAX) {
+      _c[i + 2] = UINT32_MAX;
+    } else {
+      _c[i + 2] = _a[i];
+    }
+    _c[i] = _r[i];
+  }
+  int64x2_t a = vld1q_s64(_a);
+  uint32x2_t r = vld1_u32(_r);
+  uint32x4_t c = vqmovun_high_s64(r, a);
+  return validate_uint32(c, _c[0], _c[1], _c[2], _c[3]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vmovl_s8(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #ifdef ENABLE_TEST_ALL

diff --git a/tests/impl.h b/tests/impl.h
@@ -1795,9 +1795,9 @@
   /*_(vqmovunh_s16)                                                           */ \
   /*_(vqmovuns_s32)                                                           */ \
   /*_(vqmovund_s64)                                                           */ \
-  /*_(vqmovun_high_s16)                                                       */ \
-  /*_(vqmovun_high_s32)                                                       */ \
-  /*_(vqmovun_high_s64)                                                       */ \
+  _(vqmovun_high_s16)                                                            \
+  _(vqmovun_high_s32)                                                            \
+  _(vqmovun_high_s64)                                                            \
   _(vmovl_s8)                                                                    \
   _(vmovl_s16)                                                                   \
   _(vmovl_s32)                                                                   \