Merge pull request numpy#315 from howjmay/vaddd

feat: Add vaddd_[s64|u64]
plctlab · Mar 29, 2024 · 0cbaf72 · 0cbaf72
2 parents 8d8af05 + af776f4
commit 0cbaf72
Show file tree

Hide file tree

Showing 3 changed files with 28 additions and 6 deletions.
diff --git a/neon2rvv.h b/neon2rvv.h
@@ -214,9 +214,9 @@ FORCE_INLINE float64x1_t vadd_f64(float64x1_t a, float64x1_t b) { return __riscv
 
 FORCE_INLINE float64x2_t vaddq_f64(float64x2_t a, float64x2_t b) { return __riscv_vfadd_vv_f64m1(a, b, 2); }
 
-// FORCE_INLINE int64_t vaddd_s64(int64_t a, int64_t b);
+FORCE_INLINE int64_t vaddd_s64(int64_t a, int64_t b) { return a + b; }
 
-// FORCE_INLINE uint64_t vaddd_u64(uint64_t a, uint64_t b);
+FORCE_INLINE uint64_t vaddd_u64(uint64_t a, uint64_t b) { return a + b; }
 
 FORCE_INLINE uint8x16_t vaddq_u8(uint8x16_t a, uint8x16_t b) { return __riscv_vadd_vv_u8m1(a, b, 16); }
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -545,9 +545,31 @@ result_t test_vaddq_f64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #endif  // ENABLE_TEST_ALL
 }
 
-result_t test_vaddd_s64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vaddd_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 int64_t *_b = (int64_t *)impl.test_cases_int_pointer2;
+  int64_t d0 = _a[0] + _b[0];
+
+  int64_t c = vaddd_s64(_a[0], _b[0]);
+  return c == d0 ? TEST_SUCCESS : TEST_FAIL;
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
-result_t test_vaddd_u64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) { return TEST_UNIMPL; }
+result_t test_vaddd_u64(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
+#ifdef ENABLE_TEST_ALL
+  const uint64_t *_a = (uint64_t *)impl.test_cases_int_pointer1;
+  const uint64_t *_b = (uint64_t *)impl.test_cases_int_pointer2;
+  uint64_t d0 = _a[0] + _b[0];
+
+  uint64_t c = vaddd_u64(_a[0], _b[0]);
+  return c == d0 ? TEST_SUCCESS : TEST_FAIL;
+#else
+  return TEST_UNIMPL;
+#endif  // ENABLE_TEST_ALL
+}
 
 result_t test_vaddq_u8(const NEON2RVV_TEST_IMPL &impl, uint32_t iter) {
 #ifdef ENABLE_TEST_ALL

diff --git a/tests/impl.h b/tests/impl.h
@@ -22,8 +22,8 @@
   _(vaddq_f32)                                                                   \
   _(vadd_f64)                                                                    \
   _(vaddq_f64)                                                                   \
-  /*_(vaddd_s64)                                                              */ \
-  /*_(vaddd_u64)                                                              */ \
+  _(vaddd_s64)                                                                   \
+  _(vaddd_u64)                                                                   \
   _(vaddq_u8)                                                                    \
   _(vaddq_u16)                                                                   \
   _(vaddq_u32)                                                                   \