feat: Add vld1[q]_dup_f64

howjmay · Jul 30, 2024 · a8de6b0 · a8de6b0
1 parent 2a46988
commit a8de6b0
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 6 deletions.
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -13254,9 +13254,9 @@ FORCE_INLINE float32x4_t vld1q_dup_f32(const float32_t *a) { return vdupq_n_f32(
 
 // FORCE_INLINE poly16x8_t vld1q_dup_p16(poly16_t const * ptr);
 
-// FORCE_INLINE float64x1_t vld1_dup_f64(float64_t const * ptr);
+FORCE_INLINE float64x1_t vld1_dup_f64(float64_t const *a) { return vdup_n_f64(a[0]); }
 
-// FORCE_INLINE float64x2_t vld1q_dup_f64(float64_t const * ptr);
+FORCE_INLINE float64x2_t vld1q_dup_f64(float64_t const *a) { return vdupq_n_f64(a[0]); }
 
 FORCE_INLINE uint8x16_t vld1q_dup_u8(const uint8_t *a) { return vdupq_n_u8(a[0]); }
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -43751,9 +43751,25 @@ result_t test_vld1_dup_p16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { retu
 
 result_t test_vld1q_dup_p16(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
 
-result_t test_vld1_dup_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vld1_dup_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const double *_a = (const double *)impl.test_cases_float_pointer1;
+  float64x1_t c = vld1_dup_f64(_a);
+  return validate_double(c, _a[0]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vld1q_dup_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vld1q_dup_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const double *_a = (const double *)impl.test_cases_float_pointer1;
+  float64x2_t c = vld1q_dup_f64(_a);
+  return validate_double(c, _a[0], _a[0]);
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vld1q_dup_u8(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #ifdef ENABLE_TEST_ALL

diff --git a/tests/impl.h b/tests/impl.h
@@ -2662,8 +2662,8 @@
   _(vld1q_dup_u32)                                                               \
   /*_(vld1q_dup_p8)                                                           */ \
   /*_(vld1q_dup_p16)                                                          */ \
-  /*_(vld1_dup_f64)                                                           */ \
-  /*_(vld1q_dup_f64)                                                          */ \
+  _(vld1_dup_f64)                                                                \
+  _(vld1q_dup_f64)                                                               \
   /*_(vld1q_dup_p64)                                                          */ \
   _(vld1q_dup_s64)                                                               \
   _(vld1q_dup_u64)                                                               \