SWDEV-232903 - Move hipmemset dword optimization into ROCclr.

Change-Id: I1af437de2771f6a48d06ce643ade6b79f551a55c


[ROCm/clr commit: 4d1195f0a7]

projects/clr/hipamd/src/hip_memory.cpp

@@ -2238,55 +2238,10 @@ hipError_t ihipMemsetCommand(std::vector<amd::Command*>& commands, void* dst, in
   int64_t value64 = 0;
   amd::Command* command;
 
-  if (sizeBytes / sizeof(int64_t) > 0) {
-    n_head_bytes = static_cast<uint8_t*>(aligned_dst) - static_cast<uint8_t*>(dst);
-    n_tail_bytes = ((sizeBytes - n_head_bytes) % sizeof(int64_t));
-    offset = offset + n_head_bytes;
-    size_t n_bytes = sizeBytes - n_tail_bytes - n_head_bytes;
-    if (n_bytes > 0) {
-      if (valueSize == sizeof(int8_t)) {
-        value = value & 0xff;
-        value64 = ((value << 56) | (value << 48) | (value << 40) | (value << 32) | (value << 24) |
-                   (value << 16) | (value << 8) | (value));
-      } else if (valueSize == sizeof(int16_t)) {
-        value = value & 0xffff;
-        value64 = ((value << 48) | (value << 32) | (value << 16) | (value));
-      } else if (valueSize == sizeof(int32_t)) {
-        value = value & 0xffffffff;
-        value64 = ((value << 32) | (value));
-      } else if (valueSize == sizeof(int64_t)) {
-        value64 = value;
-      } else {
-        LogPrintfError("Unsupported Pattern size: %u \n", valueSize);
-        return hipErrorInvalidValue;
-      }
-      // If n_tail_bytes is != 0 then we will do a second fillBuffer Command
-      // on the same stream below, dont wait, do the first call async.
-      hip_error =
-          packFillMemoryCommand(command, memory, offset, value64, sizeof(int64_t), n_bytes, queue);
-      commands.push_back(command);
-    }
-    if (hip_error != hipSuccess) {
-      return hip_error;
-    }
-  } else {
-    n_head_bytes = sizeBytes;
-  }
+  hip_error = packFillMemoryCommand(command, memory, offset, value, valueSize, sizeBytes,
+                                    queue);
+  commands.push_back(command);
 
-  if (n_head_bytes != 0) {
-    memory = getMemoryObject(dst, offset);
-    hip_error =
-        packFillMemoryCommand(command, memory, offset, value, valueSize, n_head_bytes, queue);
-    commands.push_back(command);
-  }
-
-  if (n_tail_bytes != 0) {
-    void* new_dst = (reinterpret_cast<address>(dst) + sizeBytes) - n_tail_bytes;
-    memory = getMemoryObject(new_dst, offset);
-    hip_error =
-        packFillMemoryCommand(command, memory, offset, value, valueSize, n_tail_bytes, queue);
-    commands.push_back(command);
-  }
   return hip_error;
 }