Merge pull request #50 from howjmay/movepi64

feat: Add _mm_movepi64_[e]pi64
pattonkan · Jan 20, 2024 · 5f8e62c · 5f8e62c
2 parents 8552de4 + 44b48a9
commit 5f8e62c
Show file tree

Hide file tree

Showing 2 changed files with 30 additions and 24 deletions.
diff --git a/sse2rvv.h b/sse2rvv.h
@@ -2032,9 +2032,15 @@ FORCE_INLINE int _mm_movemask_ps(__m128 a) {
   return (int)(__riscv_vmv_x_s_u8m1_u8(nonzeros) & 0xf);
 }
 
-// FORCE_INLINE __m64 _mm_movepi64_pi64 (__m128i a) {}
+FORCE_INLINE __m64 _mm_movepi64_pi64(__m128i a) {
+  return vreinterpretq_i32_m64(vreinterpretq_m128i_i32(a));
+}
 
-// FORCE_INLINE __m128i _mm_movpi64_epi64 (__m64 a) {}
+FORCE_INLINE __m128i _mm_movpi64_epi64(__m64 a) {
+  vint32m1_t _a = vreinterpretq_m128i_i32(a);
+  vint32m1_t zeros = __riscv_vmv_v_x_i32m1(0, 4);
+  return vreinterpretq_i32_m64(__riscv_vslideup_vx_i32m1(zeros, _a, 0, 2));
+}
 
 // FORCE_INLINE __m128i _mm_mpsadbw_epu8 (__m128i a, __m128i b, const int imm8)
 // {}

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -6084,33 +6084,33 @@ result_t test_mm_movemask_pd(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
 }
 
 result_t test_mm_movepi64_pi64(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   const int64_t *_a = (const int64_t *)impl.test_cases_int_pointer1;
-  //
-  //   int64_t d0 = _a[0];
-  //
-  //   __m128i a = load_m128i(_a);
-  //   __m64 c = _mm_movepi64_pi64(a);
-  //
-  //   return validate_int64(c, d0);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  const int64_t *_a = (const int64_t *)impl.test_cases_int_pointer1;
+
+  int64_t d0 = _a[0];
+
+  __m128i a = load_m128i(_a);
+  __m64 c = _mm_movepi64_pi64(a);
+
+  return validate_int64(c, d0);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_movpi64_epi64(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   const int64_t *_a = (const int64_t *)impl.test_cases_int_pointer1;
-  //
-  //   int64_t d0 = _a[0];
-  //
-  //   __m64 a = load_m64(_a);
-  //   __m128i c = _mm_movpi64_epi64(a);
-  //
-  //   return validate_int64(c, d0, 0);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  const int64_t *_a = (const int64_t *)impl.test_cases_int_pointer1;
+
+  int64_t d0 = _a[0];
+
+  __m64 a = load_m64(_a);
+  __m128i c = _mm_movpi64_epi64(a);
+
+  return validate_int64(c, d0, 0);
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_mul_epu32(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {