Using HSA API for hipMemsetAsync (#1346)

src/hip_memory.cpp

@@ -1557,6 +1557,13 @@ hipError_t ihipMemPtrGetInfo(void* ptr, size_t* size) {
 
 template <typename T>
 void ihipMemsetKernel(hipStream_t stream, T* ptr, T val, size_t count) {
+    // Just Use count, instead of dividing by 4, the calling API already does it
+    if (sizeof(T) == sizeof(uint32_t) && (count % sizeof(uint32_t) == 0) &&
+        !hsa_amd_memory_fill(ptr, reinterpret_cast<const std::uint32_t&>(val), count)) {
+        // Only return if the execution completes without error
+        // if error occured, try the normal version
+        return;
+    }
     static constexpr uint32_t block_dim = 256;
 
     const uint32_t grid_dim = clamp_integer<size_t>(count / block_dim, 1, UINT32_MAX);