diff --git a/tests/FPGAIntegrationTest.cpp b/tests/FPGAIntegrationTest.cpp
index 956ad33d2..b5a385ce3 100644
--- a/tests/FPGAIntegrationTest.cpp
+++ b/tests/FPGAIntegrationTest.cpp
@@ -1225,4 +1225,85 @@ TEST_CASE("HLS_C_Simulation_internal_packet_generator_spot_finder_snr_threshold"
     REQUIRE (spot_finder_result.count_threshold == 0);
     REQUIRE (spot_finder_result.strong_pixel[0] == (1<<0));
     REQUIRE (spot_finder_result.strong_pixel[(89*1024 + 300) / 8] == (1<<4));  // 300 % 8 == 4
-}
\ No newline at end of file
+}
+
+TEST_CASE("HLS_C_Simulation_upside_down", "[FPGA][Full]") {
+    const uint16_t nmodules = 4;
+
+    DiffractionExperiment x((DetectorGeometry(nmodules)));
+
+    std::vector<uint16_t> input_frame(FRAME_GENERATOR_MODULES*RAW_MODULE_SIZE);
+    std::vector<uint16_t> input_frame_transformed(FRAME_GENERATOR_MODULES*RAW_MODULE_SIZE);
+
+    std::mt19937 g1(1387);
+    std::uniform_int_distribution<uint16_t> dist(0, 65535);
+    for (int n = 0; n < nmodules; n++) {
+        for (int line = 0; line < RAW_MODULE_LINES; line++) {
+            size_t line_transformed = RAW_MODULE_LINES - 1 - line;
+            for (int col = 0; col < RAW_MODULE_COLS; col++) {
+                uint16_t tmp = dist(g1);
+                input_frame[n * RAW_MODULE_SIZE + line * RAW_MODULE_COLS + col] = tmp;
+                input_frame_transformed[n * RAW_MODULE_SIZE + line_transformed * RAW_MODULE_COLS + col] = tmp;
+            }
+        }
+    }
+
+    x.Mode(DetectorMode::Raw);
+    x.UseInternalPacketGenerator(true).ImagesPerTrigger(4).PedestalG0Frames(0);
+
+    HLSSimulatedDevice test(0, 64);
+    test.SetInternalGeneratorFrame(input_frame);
+
+    REQUIRE_NOTHROW(test.StartAction(x, MODE_MODULE_UPSIDE_DOWN));
+    REQUIRE_NOTHROW(test.WaitForActionComplete());
+
+    REQUIRE(test.OutputStream().size() == 1);
+
+    REQUIRE(test.GetBytesReceived() == 128 * nmodules * 4 * JUNGFRAU_PACKET_SIZE_BYTES);
+
+    for (int image = 0; image < 4; image++) {
+        for (int m = 0; m < nmodules; m++) {
+            auto imageBuf = (uint16_t *) test.GetDeviceOutput(image, m)->pixels;
+            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
+                REQUIRE(imageBuf[i] == input_frame_transformed[m * RAW_MODULE_SIZE + i]);
+            }
+        }
+    }
+}
+
+TEST_CASE("HLS_C_Simulation_bitsuffle", "[FPGA][Full]") {
+    const uint16_t nmodules = 4;
+
+    DiffractionExperiment x((DetectorGeometry(nmodules)));
+
+    std::vector<uint16_t> input_frame(FRAME_GENERATOR_MODULES*RAW_MODULE_SIZE);
+
+    std::mt19937 g1(1387);
+    std::uniform_int_distribution<uint16_t> dist(0, 65535);
+    for (auto &i: input_frame)
+        i = dist(g1);
+
+    x.Mode(DetectorMode::Raw);
+    x.UseInternalPacketGenerator(true).ImagesPerTrigger(4).PedestalG0Frames(0);
+
+    HLSSimulatedDevice test(0, 64);
+    test.SetInternalGeneratorFrame(input_frame);
+
+    REQUIRE_NOTHROW(test.StartAction(x, MODE_BITSHUFFLE_FPGA));
+    REQUIRE_NOTHROW(test.WaitForActionComplete());
+
+    REQUIRE(test.OutputStream().size() == 1);
+
+    REQUIRE(test.GetBytesReceived() == 128 * nmodules * 4 * JUNGFRAU_PACKET_SIZE_BYTES);
+
+    for (int image = 0; image < 4; image++) {
+        for (int m = 0; m < nmodules; m++) {
+            auto imageBuf = (uint16_t *) test.GetDeviceOutput(image, m)->pixels;
+            std::vector<uint16_t> image_out_unshuf(RAW_MODULE_SIZE);
+            bshuf_bitunshuffle(imageBuf, image_out_unshuf.data(), RAW_MODULE_SIZE, sizeof(uint16_t), 2048);
+            for (int i = 0; i < RAW_MODULE_SIZE; i++) {
+                REQUIRE(image_out_unshuf[i] == input_frame[m * RAW_MODULE_SIZE + i]);
+            }
+        }
+    }
+}