feat: Add vrecpe[s|d]_[f32|f64]

howjmay · Jul 25, 2024 · fbb4002 · fbb4002
1 parent fd1fe49
commit fbb4002
Show file tree

Hide file tree

Showing 3 changed files with 28 additions and 6 deletions.
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -7243,9 +7243,9 @@ FORCE_INLINE float64x1_t vrecpe_f64(float64x1_t a) { return __riscv_vfrec7_v_f64
 
 FORCE_INLINE float64x2_t vrecpeq_f64(float64x2_t a) { return __riscv_vfrec7_v_f64m1(a, 2); }
 
-// FORCE_INLINE float32_t vrecpes_f32(float32_t a);
+FORCE_INLINE float32_t vrecpes_f32(float32_t a) { return 1.0f / a; }
 
-// FORCE_INLINE float64_t vrecped_f64(float64_t a);
+FORCE_INLINE float64_t vrecped_f64(float64_t a) { return 1.0f / a; }
 
 // FORCE_INLINE uint32x4_t vrecpeq_u32(uint32x4_t a);
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -25755,9 +25755,31 @@ result_t test_vrecpeq_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #endif  // ENABLE_TEST_ALL
 }
 
-result_t test_vrecpes_f32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vrecpes_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 = 1.0f / _a[0];
+
+  c = vrecpes_f32(_a[0]);
+  return validate_float_error(c, _c, 0.01f);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vrecped_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vrecped_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 = 1.0f / _a[0];
+
+  c = vrecped_f64(_a[0]);
+  return validate_double_error(c, _c, 0.01f);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vrecpeq_u32(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 

diff --git a/tests/impl.h b/tests/impl.h
@@ -1393,8 +1393,8 @@
   _(vrecpeq_f32)                                                                 \
   _(vrecpe_f64)                                                                  \
   _(vrecpeq_f64)                                                                 \
-  /*_(vrecpes_f32)                                                            */ \
-  /*_(vrecped_f64)                                                            */ \
+  _(vrecpes_f32)                                                                 \
+  _(vrecped_f64)                                                                 \
   _(vrecpeq_u32)                                                                 \
   _(vrsqrte_f32)                                                                 \
   _(vrsqrte_u32)                                                                 \