Merge pull request #61 from howjmay/malloc-free

feat: Add _mm_free and _mm_malloc
pattonkan · Jan 26, 2024 · d6dc3e5 · d6dc3e5
2 parents 52ea063 + 62609e9
commit d6dc3e5
Show file tree

Hide file tree

Showing 2 changed files with 44 additions and 21 deletions.
diff --git a/sse2rvv.h b/sse2rvv.h
@@ -1468,7 +1468,7 @@ FORCE_INLINE int _mm_extract_ps(__m128 a, const int imm8) {
 
 // FORCE_INLINE __m128 _mm_floor_ss (__m128 a, __m128 b) {}
 
-// FORCE_INLINE void _mm_free (void * mem_addr) {}
+FORCE_INLINE void _mm_free(void *mem_addr) { free(mem_addr); }
 
 // FORCE_INLINE unsigned int _MM_GET_EXCEPTION_MASK () {}
 
@@ -1826,7 +1826,20 @@ FORCE_INLINE __m128i _mm_loadu_si64(void const *mem_addr) {
 
 // FORCE_INLINE __m64 _mm_maddubs_pi16 (__m64 a, __m64 b) {}
 
-// FORCE_INLINE void* _mm_malloc (size_t size, size_t align) {}
+FORCE_INLINE void *_mm_malloc(size_t size, size_t align) {
+  void *ptr;
+  if (align == 1) {
+    return malloc(size);
+  }
+  if (align == 2 || (sizeof(void *) == 8 && align == 4)) {
+    align = sizeof(void *);
+  }
+  ptr = aligned_alloc(align, size);
+  if (ptr) {
+    return ptr;
+  }
+  return NULL;
+}
 
 // FORCE_INLINE void _mm_maskmove_si64 (__m64 a, __m64 mask, char* mem_addr) {}
 

diff --git a/tests/impl.cpp b/tests/impl.cpp
@@ -2051,12 +2051,22 @@ result_t test_mm_malloc(const SSE2RVV_TEST_IMPL &impl, uint32_t iter);
 // #ifdef ENABLE_TEST_ALL
 
 result_t test_mm_free(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   /* We verify _mm_malloc first, and there is no need to check _mm_free .
-  //   */ return test_mm_malloc(impl, iter);
-  // #else
+#ifdef ENABLE_TEST_ALL
+  /* We verify _mm_malloc first, and there is no need to check _mm_free(). */
+  const size_t *a = (const size_t *)impl.test_cases_int_pointer1;
+  const size_t *b = (const size_t *)impl.test_cases_int_pointer2;
+  size_t size = *a % (1024 * 16) + 1;
+  size_t align = 2 << (*b % 5);
+
+  void *p = _mm_malloc(size, align);
+  if (!p)
+    return TEST_FAIL;
+  result_t res = (((uintptr_t)p % align) == 0) ? TEST_SUCCESS : TEST_FAIL;
+  _mm_free(p);
+  return res;
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_get_flush_zero_mode(const SSE2RVV_TEST_IMPL &impl,
@@ -2282,21 +2292,21 @@ result_t test_mm_loadu_si64(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
 }
 
 result_t test_mm_malloc(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {
-  // #ifdef ENABLE_TEST_ALL
-  //   const size_t *a = (const size_t *)impl.test_cases_int_pointer1;
-  //   const size_t *b = (const size_t *)impl.test_cases_int_pointer2;
-  //   size_t size = *a % (1024 * 16) + 1;
-  //   size_t align = 2 << (*b % 5);
-  //
-  //   void *p = _mm_malloc(size, align);
-  //   if (!p)
-  //     return TEST_FAIL;
-  //   result_t res = (((uintptr_t)p % align) == 0) ? TEST_SUCCESS : TEST_FAIL;
-  //   _mm_free(p);
-  //   return res;
-  // #else
+#ifdef ENABLE_TEST_ALL
+  const size_t *a = (const size_t *)impl.test_cases_int_pointer1;
+  const size_t *b = (const size_t *)impl.test_cases_int_pointer2;
+  size_t size = *a % (1024 * 16) + 1;
+  size_t align = 2 << (*b % 5);
+
+  void *p = _mm_malloc(size, align);
+  if (!p)
+    return TEST_FAIL;
+  result_t res = (((uintptr_t)p % align) == 0) ? TEST_SUCCESS : TEST_FAIL;
+  _mm_free(p);
+  return res;
+#else
   return TEST_UNIMPL;
-  // #endif  // ENABLE_TEST_ALL
+#endif // ENABLE_TEST_ALL
 }
 
 result_t test_mm_maskmove_si64(const SSE2RVV_TEST_IMPL &impl, uint32_t iter) {