From 039e56d9f895ef9c938b2748a26f881537d2f1c4 Mon Sep 17 00:00:00 2001
From: Yang Hau <yuanyanghau@gmail.com>
Date: Thu, 25 Jul 2024 16:47:40 +0800
Subject: [PATCH] feat: Add vrsqrte[s|d]_[f32|f64]

---
 neon2rvv.h     |  4 ++--
 tests/impl.cpp | 26 ++++++++++++++++++++++++--
 tests/impl.h   |  4 ++--
 3 files changed, 28 insertions(+), 6 deletions(-)

diff --git a/neon2rvv.h b/neon2rvv.h
index 245fbde3..313e58c0 100644
--- a/neon2rvv.h
+++ b/neon2rvv.h
@@ -7259,9 +7259,9 @@ FORCE_INLINE float64x1_t vrsqrte_f64(float64x1_t a) { return __riscv_vfrsqrt7_v_
 
 FORCE_INLINE float64x2_t vrsqrteq_f64(float64x2_t a) { return __riscv_vfrsqrt7_v_f64m1(a, 2); }
 
-// FORCE_INLINE float32_t vrsqrtes_f32(float32_t a);
+FORCE_INLINE float32_t vrsqrtes_f32(float32_t a) { return 1 / sqrtf(a); }
 
-// FORCE_INLINE float64_t vrsqrted_f64(float64_t a);
+FORCE_INLINE float64_t vrsqrted_f64(float64_t a) { return 1 / sqrt(a); }
 
 // FORCE_INLINE uint32x4_t vrsqrteq_u32(uint32x4_t a);
 
diff --git a/tests/impl.cpp b/tests/impl.cpp
index b3a58a8b..f01084f8 100644
--- a/tests/impl.cpp
+++ b/tests/impl.cpp
@@ -25845,9 +25845,31 @@ result_t test_vrsqrteq_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #endif  // ENABLE_TEST_ALL
 }
 
-result_t test_vrsqrtes_f32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vrsqrtes_f32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const float *_a = impl.test_cases_float_pointer1;
+  float _c, c;
+  _c = 1 / sqrtf(_a[0]);
+
+  c = vrsqrtes_f32(_a[0]);
+  return validate_float_error(c, _c, 0.01f);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vrsqrted_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vrsqrted_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const double *_a = (const double *)impl.test_cases_float_pointer1;
+  double _c, c;
+  _c = 1 / sqrt(_a[0]);
+
+  c = vrsqrted_f64(_a[0]);
+  return validate_double_error(c, _c, 0.01f);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vrsqrteq_u32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 
diff --git a/tests/impl.h b/tests/impl.h
index 47d95dc5..f7ae14bf 100644
--- a/tests/impl.h
+++ b/tests/impl.h
@@ -1401,8 +1401,8 @@
   _(vrsqrteq_f32)                                                                \
   _(vrsqrte_f64)                                                                 \
   _(vrsqrteq_f64)                                                                \
-  /*_(vrsqrtes_f32)                                                           */ \
-  /*_(vrsqrted_f64)                                                           */ \
+  _(vrsqrtes_f32)                                                                \
+  _(vrsqrted_f64)                                                                \
   _(vrsqrteq_u32)                                                                \
   _(vget_lane_s8)                                                                \
   _(vget_lane_s16)                                                               \