diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index 50268c9..61ef420 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -37,7 +37,8 @@ set(KLOTSKI_CORE_SRC
 add_library(klotski_core STATIC ${KLOTSKI_CORE_SRC})
 target_compile_options(klotski_core PRIVATE -fno-rtti -fno-exceptions)  # option for `-fvisibility=hidden`
 target_include_directories(klotski_core PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
-target_link_libraries(klotski_core PRIVATE absl::flat_hash_map phmap)
+target_link_libraries(klotski_core PRIVATE absl::flat_hash_map)
+target_link_libraries(klotski_core PUBLIC phmap)
 add_library(klotski::core ALIAS klotski_core)
 
 # TODO: just for dev testing
diff --git a/src/core/benchmark/fast_cal.cc b/src/core/benchmark/fast_cal.cc
index 4d92d2a..d77be45 100644
--- a/src/core/benchmark/fast_cal.cc
+++ b/src/core/benchmark/fast_cal.cc
@@ -11,6 +11,7 @@ using klotski::codec::CommonCode;
 static void FastCalBenchmark(benchmark::State &state) {
 
     auto code = CommonCode::unsafe_create(0x1A9BF0C00).to_raw_code();
+//    auto code = CommonCode::unsafe_create(0x4FEA13400).to_raw_code();
 
     for (auto _ : state) {
 //        auto fc = FastCal(code);
diff --git a/src/core/fast_cal/fast_cal.h b/src/core/fast_cal/fast_cal.h
index 59aee58..f50d188 100644
--- a/src/core/fast_cal/fast_cal.h
+++ b/src/core/fast_cal/fast_cal.h
@@ -13,8 +13,14 @@
 #include "mover/mover.h"
 #include "raw_code/raw_code.h"
 
+#include "group/group.h"
+
 // #include <absl/container/flat_hash_map.h>
 
+#include <ranges>
+
+#include <parallel_hashmap/phmap.h>
+
 using klotski::codec::RawCode;
 using klotski::mover::MaskMover;
 
@@ -69,3 +75,61 @@ private:
 };
 
 RawCode FastCal_demo(RawCode code);
+
+namespace klotski {
+
+template <typename T>
+class LayerQueue {
+public:
+    LayerQueue(size_t reserve, std::initializer_list<T> first_layer);
+
+    void emplace(T item);
+
+    T current() const;
+
+    void next();
+
+    [[nodiscard]] bool is_ending() const;
+
+    [[nodiscard]] bool is_new_layer() const;
+
+    std::vector<T> layer_cases() const;
+
+private:
+    size_t queue_begin_ {0};
+    size_t queue_end_ {0};
+
+    size_t layer_begin_ {0};
+    size_t layer_end_ {0};
+
+    std::vector<T> data_ {};
+};
+
+} // namespace klotski
+
+namespace klotski::fast_cal {
+
+class FCDemo {
+public:
+    explicit FCDemo(RawCode raw_code);
+
+    std::optional<RawCode> DoCal();
+
+    std::vector<RawCode> DoCalMulti();
+
+    std::vector<RawCode> DoCalFurthest();
+
+private:
+    struct data_t {
+        uint64_t mask;
+        uint64_t back;
+    };
+
+    LayerQueue<uint64_t> codes_;
+    phmap::flat_hash_map<uint64_t, data_t> cases_;
+};
+
+} // namespace klotski::fast_cal
+
+#include "internal/layer_queue.inl"
+#include "internal/fast_cal.inl"
diff --git a/src/core/fast_cal/internal/demo.cc b/src/core/fast_cal/internal/demo.cc
index 973768d..fef2285 100644
--- a/src/core/fast_cal/internal/demo.cc
+++ b/src/core/fast_cal/internal/demo.cc
@@ -16,186 +16,112 @@ using klotski::cases::RangesUnion;
 using klotski::mover::MaskMover;
 using klotski::cases::GroupUnion;
 
-struct data_t {
-    uint64_t mask;
-    uint64_t back;
-};
-
-template <typename T>
-class MyQueue {
-public:
-    explicit MyQueue(size_t reserve) {
-        data_.resize(reserve);
-    }
-
-    void emplace_back(T item) {
-        data_[queue_end_] = item;
-        ++queue_end_;
-    }
-
-    T front() {
-        return data_[queue_begin_];
-    }
-
-    void pop_and_try_layer_switch() {
-        ++queue_begin_;
-        try_layer_switch();
-    }
-
-    [[nodiscard]] bool is_ending() const {
-        return queue_begin_ == queue_end_;
-    }
+using klotski::fast_cal::FCDemo;
+
+std::optional<RawCode> FCDemo::DoCal() {
+    uint64_t result = 0;
+    auto core = MaskMover([this, &result](uint64_t code, uint64_t mask) {
+        if (const auto match = cases_.find(code); match != cases_.end()) {
+            match->second.mask |= mask; // update mask
+            return;
+        }
+        cases_.emplace(code, data_t {
+            .mask = mask,
+            .back = codes_.current(),
+        });
+        codes_.emplace(code);
 
-    void try_layer_switch() {
-        if (queue_begin_ == layer_end_ && !is_ending()) {
-            layer_begin_ = layer_end_;
-            layer_end_ = queue_end_;
+        if (((code >> 39) & 0b111) == 0b100) {
+            result = code;
         }
-    }
+    });
 
-    [[nodiscard]] bool is_layer_switched() const {
-        return queue_begin_ == layer_begin_;
-    }
+    while (!codes_.is_ending()) {
+        auto curr = codes_.current();
+        core.next_cases(curr, cases_.find(curr)->second.mask);
 
-    std::vector<T> get_curr_layer() {
-        std::vector<T> layer_cases;
-        for (size_t offset = layer_begin_; offset < layer_end_; ++offset) {
-            layer_cases.emplace_back(data_[offset]);
+        codes_.next();
+        if (result != 0) {
+            return RawCode::unsafe_create(result);
         }
-        return layer_cases;
-    }
-
-private:
-    size_t queue_begin_ {0};
-    size_t queue_end_ {0};
-
-    size_t layer_begin_ {0};
-    size_t layer_end_ {1};
-
-    std::vector<T> data_ {};
-};
-
-class FCDemo {
-public:
-    explicit FCDemo(RawCode raw_code) : codes_(GroupUnion::from_raw_code(raw_code).max_group_size()) {
-        // TODO: build codes_ with reserve size
-        auto reserve = GroupUnion::from_raw_code(raw_code).max_group_size();
-//        codes_.reserve(reserve);
-//        cases_.reserve(static_cast<size_t>(reserve * 1.56));
-        cases_.reserve(static_cast<size_t>(25955 * 1.56));
-        codes_.emplace_back(raw_code.unwrap());
-        cases_.emplace(raw_code, data_t {0, 0}); // without mask
     }
+    return std::nullopt;
+}
 
-    std::optional<RawCode> DoCal() {
-        uint64_t result = 0;
-        auto core = MaskMover([this, &result](uint64_t code, uint64_t mask) {
-            if (const auto match = cases_.find(code); match != cases_.end()) {
-                match->second.mask |= mask; // update mask
-                return;
-            }
-            cases_.emplace(code, data_t {
-                .mask = mask,
-                .back = codes_.front(),
-            });
-            codes_.emplace_back(code);
-
-            if (((code >> 39) & 0b111) == 0b100) {
-                result = code;
-            }
-        });
-
-        while (!codes_.is_ending()) {
-            auto curr = codes_.front();
-            core.next_cases(curr, cases_.find(curr)->second.mask);
+std::vector<RawCode> FCDemo::DoCalMulti() {
+    bool stop_flag = false;
+    std::vector<RawCode> results {};
 
-            codes_.pop_and_try_layer_switch();
-            if (result != 0) {
-                return RawCode::unsafe_create(result);
-            }
+    auto core = MaskMover([this, &stop_flag, &results](uint64_t code, uint64_t mask) {
+        if (const auto match = cases_.find(code); match != cases_.end()) {
+            match->second.mask |= mask; // update mask
+            return;
         }
-        return std::nullopt;
-    }
-
-    std::vector<RawCode> DoCalMulti() {
-        bool stop_flag = false;
-        std::vector<RawCode> results {};
-
-        auto core = MaskMover([this, &stop_flag, &results](uint64_t code, uint64_t mask) {
-            if (const auto match = cases_.find(code); match != cases_.end()) {
-                match->second.mask |= mask; // update mask
-                return;
-            }
-            cases_.emplace(code, data_t {
-                .mask = mask,
-                .back = codes_.front(),
-            });
-            codes_.emplace_back(code);
-
-            if (((code >> 39) & 0b111) == 0b100) {
-                stop_flag = true;
-                results.emplace_back(RawCode::unsafe_create(code));
-            }
+        cases_.emplace(code, data_t {
+            .mask = mask,
+            .back = codes_.current(),
         });
+        codes_.emplace(code);
 
-        while (!codes_.is_ending()) {
-            auto curr = codes_.front();
-            core.next_cases(curr, cases_.find(curr)->second.mask);
+        if (((code >> 39) & 0b111) == 0b100) {
+            stop_flag = true;
+            results.emplace_back(RawCode::unsafe_create(code));
+        }
+    });
 
-            codes_.pop_and_try_layer_switch();
-            if (codes_.is_layer_switched() && stop_flag) {
-                return results;
-            }
+    while (!codes_.is_ending()) {
+        auto curr = codes_.current();
+        core.next_cases(curr, cases_.find(curr)->second.mask);
 
+        codes_.next();
+        if (codes_.is_new_layer() && stop_flag) {
+            // TODO: fix when solutions at last layer
+            return results;
         }
-        return {};
+
     }
+    return {};
+}
 
-    std::vector<RawCode> DoCalFurthest() {
-
-        auto core = MaskMover([this](uint64_t code, uint64_t mask) {
-            if (const auto match = cases_.find(code); match != cases_.end()) {
-                match->second.mask |= mask; // update mask
-                return;
-            }
-            cases_.emplace(code, data_t {
-                .mask = mask,
-                .back = codes_.front(),
-            });
-            codes_.emplace_back(code);
+std::vector<RawCode> FCDemo::DoCalFurthest() {
+    auto core = MaskMover([this](uint64_t code, uint64_t mask) {
+        if (const auto match = cases_.find(code); match != cases_.end()) {
+            match->second.mask |= mask; // update mask
+            return;
+        }
+        cases_.emplace(code, data_t {
+            .mask = mask,
+            .back = codes_.current(),
         });
-
-        while (true) {
-            auto curr = codes_.front();
-            core.next_cases(curr, cases_.find(curr)->second.mask);
-
-            codes_.pop_and_try_layer_switch();
-            if (codes_.is_ending()) {
-                return codes_.get_curr_layer() | std::views::transform([](uint64_t code) {
-                    return RawCode::unsafe_create(code);
-                }) | std::ranges::to<std::vector>();
-            }
+        codes_.emplace(code);
+    });
+
+    while (true) {
+        auto curr = codes_.current();
+        core.next_cases(curr, cases_.find(curr)->second.mask);
+
+        codes_.next();
+        if (codes_.is_ending()) {
+            return codes_.layer_cases() | std::views::transform([](uint64_t code) {
+                return RawCode::unsafe_create(code);
+            }) | std::ranges::to<std::vector>();
         }
     }
-
-private:
-    MyQueue<uint64_t> codes_;
-    phmap::flat_hash_map<uint64_t, data_t> cases_;
-};
+}
 
 RawCode FastCal_demo(RawCode raw_code) {
-    FCDemo fc {raw_code};
-    return fc.DoCal().value();
+    klotski::fast_cal::FCDemo fc {raw_code};
+//    return fc.DoCal().value();
 
 //    auto tmp = fc.DoCal();
-//    std::cout << tmp.to_common_code() << std::endl;
+//    std::cout << tmp.value().to_common_code() << std::endl;
 
 //    auto tmp = fc.DoCalMulti();
 //    for (const auto x : tmp) {
 //        std::cout << x.to_common_code() << std::endl;
 //    }
 
-//    auto tmp = fc.DoCalFurthest();
+    auto tmp = fc.DoCalFurthest();
 //    for (const auto x : tmp) {
 //        std::cout << x.to_common_code() << std::endl;
 //    }
diff --git a/src/core/fast_cal/internal/fast_cal.inl b/src/core/fast_cal/internal/fast_cal.inl
new file mode 100644
index 0000000..b0c40ba
--- /dev/null
+++ b/src/core/fast_cal/internal/fast_cal.inl
@@ -0,0 +1,12 @@
+#pragma once
+
+namespace klotski::fast_cal {
+
+inline FCDemo::FCDemo(RawCode raw_code) : codes_(cases::GroupUnion::from_raw_code(raw_code).max_group_size(), {raw_code.unwrap()}) {
+//    auto reserve = cases::GroupUnion::from_raw_code(raw_code).max_group_size();
+    cases_.reserve(static_cast<size_t>(25955 * 1.56));
+//    cases_.reserve(static_cast<size_t>(reserve * 1.56));
+    cases_.emplace(raw_code, data_t {0, 0}); // without mask
+}
+
+} // namespace klotski::fast_cal
diff --git a/src/core/fast_cal/internal/layer_queue.inl b/src/core/fast_cal/internal/layer_queue.inl
new file mode 100644
index 0000000..daee16b
--- /dev/null
+++ b/src/core/fast_cal/internal/layer_queue.inl
@@ -0,0 +1,53 @@
+#pragma once
+
+namespace klotski {
+
+template <typename T>
+inline LayerQueue<T>::LayerQueue(size_t reserve, std::initializer_list<T> first_layer) {
+    data_.resize(reserve);
+    for (auto x : first_layer) {
+        emplace(x);
+    }
+    layer_end_ = first_layer.size();
+}
+
+template <typename T>
+void LayerQueue<T>::emplace(T item) {
+    data_[queue_end_] = item;
+    ++queue_end_;
+}
+
+template <typename T>
+T LayerQueue<T>::current() const {
+    return data_[queue_begin_];
+}
+
+template <typename T>
+void LayerQueue<T>::next() {
+    ++queue_begin_;
+    if (queue_begin_ == layer_end_ && !is_ending()) {
+        layer_begin_ = layer_end_;
+        layer_end_ = queue_end_;
+    }
+}
+
+template <typename T>
+[[nodiscard]] bool LayerQueue<T>::is_ending() const {
+    return queue_begin_ == queue_end_;
+}
+
+template <typename T>
+[[nodiscard]] bool LayerQueue<T>::is_new_layer() const {
+    return queue_begin_ == layer_begin_;
+}
+
+template <typename T>
+std::vector<T> LayerQueue<T>::layer_cases() const {
+    std::vector<T> layer_cases;
+    for (size_t offset = layer_begin_; offset < layer_end_; ++offset) {
+        layer_cases.emplace_back(data_[offset]);
+    }
+    return layer_cases;
+}
+
+} // namespace klotski