Add CTS for scratch register reading.

conformance_tests/tools/debug/src/test_debug.cpp

@@ -8,6 +8,7 @@
 
 #include "test_debug.hpp"
 #include "test_debug_utils.hpp"
+#include "test_harness/zet_intel_gpu_debug.h"
 
 namespace lzt = level_zero_tests;
 
@@ -1376,6 +1377,143 @@ void zetDebugReadWriteRegistersTest::run_read_write_registers_test(
   }
 }
 
+void zetDebugReadWriteRegistersTest::run_read_registers_test(
+    std::vector<ze_device_handle_t> &devices, bool use_sub_devices) {
+  for (auto &device : devices) {
+    print_device(device);
+    if (!is_debug_supported(device))
+      continue;
+
+    synchro->clear_debugger_signal();
+    debugHelper = launch_process(LONG_RUNNING_KERNEL_INTERRUPTED_SCRATCH,
+                                 device, use_sub_devices);
+
+    zet_debug_event_t module_event;
+    attach_and_get_module_event(debugHelper.id(), synchro, device, debugSession,
+                                module_event);
+
+    if (module_event.flags & ZET_DEBUG_EVENT_FLAG_NEED_ACK) {
+      LOG_DEBUG << "[Debugger] Acking event: "
+                << lzt::debuggerEventTypeString[module_event.type];
+      lzt::debug_ack_event(debugSession, &module_event);
+    }
+
+    uint64_t gpu_buffer_va = 0;
+    synchro->wait_for_application_signal();
+    if (!synchro->get_app_gpu_buffer_address(gpu_buffer_va)) {
+      FAIL() << "[Debugger] Could not get a valid GPU buffer VA";
+    }
+    synchro->clear_application_signal();
+
+    zet_debug_memory_space_desc_t memorySpaceDesc;
+    memorySpaceDesc.type = ZET_DEBUG_MEMORY_SPACE_TYPE_DEFAULT;
+    int sizeToRead = 512;
+    uint8_t *kernel_buffer = new uint8_t[sizeToRead];
+    // set buffer[0] to 0 to break the loop. See debug_loop_slm.cl
+    kernel_buffer[0] = 0;
+    memorySpaceDesc.address = gpu_buffer_va;
+
+    ze_device_thread_t device_threads = {};
+    device_threads.slice = UINT32_MAX;
+    device_threads.subslice = UINT32_MAX;
+    device_threads.eu = UINT32_MAX;
+    device_threads.thread = UINT32_MAX;
+
+    LOG_INFO << "[Debugger] Stopping all device threads";
+    // give time to app to launch the kernel
+    std::this_thread::sleep_for(std::chrono::seconds(6));
+    lzt::debug_interrupt(debugSession, device_threads);
+
+    std::vector<ze_device_thread_t> stopped_threads;
+    if (!find_stopped_threads(debugSession, device, device_threads, true,
+                              stopped_threads)) {
+      delete[] kernel_buffer;
+      FAIL() << "[Debugger] Did not find stopped threads";
+    }
+
+    bool is_heapless_mode = false;
+    LOG_INFO << "[Debugger] Reading/Writing Thread Scratch Register on "
+                "interrupted threads";
+
+    for (auto &stopped_thread : stopped_threads) {
+      uint8_t *mode_values = nullptr;
+      auto register_set_properties = lzt::get_register_set_properties(device);
+      for (auto &register_set : register_set_properties){
+
+        if (register_set.type == ZET_DEBUG_REGSET_TYPE_MODE_FLAGS_INTEL_GPU){
+          auto reg_size_in_bytes = register_set.count * register_set.byteSize;
+          mode_values = new uint8_t[reg_size_in_bytes];
+          ASSERT_EQ(
+              zetDebugReadRegisters(debugSession, stopped_thread,
+                                    ZET_DEBUG_REGSET_TYPE_MODE_FLAGS_INTEL_GPU,
+                                    0, register_set.count, mode_values),
+              ZE_RESULT_SUCCESS);
+
+          uint32_t *uint32_t_values = (uint32_t *)mode_values;
+          LOG_DEBUG << "[Debugger] mode value: %u " << uint32_t_values[0];
+          is_heapless_mode =
+              (uint32_t_values[0] & ZET_DEBUG_MODE_FLAG_HEAPLESS);
+        }
+      }
+
+      if (is_heapless_mode) {
+        for (auto &register_set : register_set_properties){
+          if ((register_set.type ==
+               ZET_DEBUG_REGSET_TYPE_THREAD_SCRATCH_INTEL_GPU) &&
+              (register_set.generalFlags & ZET_DEBUG_REGSET_FLAG_READABLE)) {
+            LOG_DEBUG << "[Debugger] Register set type " << register_set.type
+                      << " is readable";
+            size_t reg_size_in_bytes =
+                register_set.count * register_set.byteSize;
+
+            uint64_t *thread_scratch_reg_values =
+                new uint64_t[reg_size_in_bytes];
+            ASSERT_EQ(zetDebugReadRegisters(
+                          debugSession, stopped_thread,
+                          ZET_DEBUG_REGSET_TYPE_DEBUG_SCRATCH_INTEL_GPU, 0,
+                          register_set.count, thread_scratch_reg_values),
+                      ZE_RESULT_SUCCESS);
+          } else {
+            LOG_INFO << "[Debugger] Register set type " << register_set.type
+                     << " is NOT readable";
+          }
+          if (register_set.generalFlags & ZET_DEBUG_REGSET_FLAG_WRITEABLE) {
+          } else {
+            LOG_INFO << "[Debugger] Register set " << register_set.type
+                     << " type is NOT writeable";
+          }
+        }
+      }
+    }
+
+    lzt::debug_write_memory(debugSession, device_threads, memorySpaceDesc, 1,
+                            kernel_buffer);
+    delete[] kernel_buffer;
+
+    LOG_INFO << "[Debugger] resuming interrupted threads";
+    lzt::debug_resume(debugSession, device_threads);
+    debugHelper.wait();
+
+    std::vector<zet_debug_event_type_t> expectedEvents = {
+        ZET_DEBUG_EVENT_TYPE_MODULE_UNLOAD, ZET_DEBUG_EVENT_TYPE_PROCESS_EXIT};
+
+    if (!check_events(debugSession, expectedEvents)) {
+      FAIL() << "[Debugger] Did not receive expected events";
+    }
+
+    lzt::debug_detach(debugSession);
+    ASSERT_EQ(debugHelper.exit_code(), 0);
+  }
+}
+
+TEST_F(
+    zetDebugReadWriteRegistersTest,
+    GivenActiveDebugSessionWhenReadingScratchRegistersThenDataReadIsDoneSuccessfully) {
+  auto driver = lzt::get_default_driver();
+  auto devices = lzt::get_devices(driver);
+  run_read_registers_test(devices, false);
+}
+
 TEST_F(
     zetDebugReadWriteRegistersTest,
     GivenActiveDebugSessionWhenReadingAndWritingRegistersThenValidDataReadAndDataWrittenSuccessfully) {

conformance_tests/tools/debug/src/test_debug.hpp

@@ -255,6 +255,8 @@ class zetDebugReadWriteRegistersTest : public zetDebugMemAccessTest {
   void TearDown() override { zetDebugMemAccessTest::TearDown(); }
   void run_read_write_registers_test(std::vector<ze_device_handle_t> &devices,
                                      bool use_sub_devices);
+  void run_read_registers_test(std::vector<ze_device_handle_t> &devices,
+                          bool use_sub_devices);
 };
 
 class zetDebugThreadControlTest : public zetDebugBaseSetup {

conformance_tests/tools/debug/src/test_debug_common.hpp

@@ -48,6 +48,7 @@ typedef enum {
   ATTACH_AFTER_MODULE_DESTROYED,
   LONG_RUNNING_KERNEL_INTERRUPTED,
   LONG_RUNNING_KERNEL_INTERRUPTED_SLM,
+  LONG_RUNNING_KERNEL_INTERRUPTED_SCRATCH,
   PAGE_FAULT,
   MULTIPLE_THREADS,
   MULTIPLE_CQ,

conformance_tests/tools/debug/src/test_debug_helper.cpp

@@ -583,6 +583,133 @@ void run_long_kernel(ze_context_handle_t context, ze_device_handle_t device,
   }
 }
 
+void run_long_kernel_scratch(ze_context_handle_t context,
+                             ze_device_handle_t device,
+                             process_synchro &synchro, debug_options &options) {
+
+  auto command_list = lzt::create_command_list(device);
+  auto command_queue = lzt::create_command_queue(device);
+  std::string module_name = options.module_name_in;
+
+  std::string kernel_name = "long_kernel_slm";
+  size_t slm_buffer_size = 512; // NOTE: Not all SKUs have same SLM so can go too big.
+
+  synchro.wait_for_debugger_signal();
+  const char *build_flags ="-g -igc_opts 'VISAOptions=-forcespills'";
+  auto module =
+      lzt::create_module(device, module_name, ZE_MODULE_FORMAT_IL_SPIRV,
+                         build_flags /* include debug symbols*/, nullptr);
+
+  auto kernel = lzt::create_function(module, kernel_name);
+  auto size = slm_buffer_size;
+
+  ze_kernel_properties_t kernel_properties = {
+      ZE_STRUCTURE_TYPE_KERNEL_PROPERTIES, nullptr};
+  EXPECT_EQ(ZE_RESULT_SUCCESS,
+            zeKernelGetProperties(kernel, &kernel_properties));
+  int threadCount = std::ceil(size / kernel_properties.maxSubgroupSize);
+
+  LOG_INFO << "[Application] Problem size: " << size
+           << ". Kernel maxSubGroupSize: " << kernel_properties.maxSubgroupSize
+           << ". GPU thread count: ceil (P size/maxSubGroupSize) = "
+           << threadCount;
+
+  auto dest_buffer_d =
+      lzt::allocate_device_memory(size, size, 0, 0, device, context);
+  auto dest_buffer_s =
+      lzt::allocate_shared_memory(size, size, 0, 0, device, context);
+  auto src_buffer_d =
+      lzt::allocate_device_memory(size, size, 0, 0, device, context);
+  auto src_buffer_s =
+      lzt::allocate_shared_memory(size, size, 0, 0, device, context);
+
+  void *slm_output_s = nullptr;
+  slm_output_s = lzt::allocate_shared_memory(slm_buffer_size, slm_buffer_size,
+                                             0, 0, device, context);
+
+  unsigned long loop_max = 1000000000;
+
+  auto loop_counter_d = lzt::allocate_device_memory(
+      loop_counter_alloc_size, loop_counter_alloc_size, 0, 0, device, context);
+  auto loop_counter_s = lzt::allocate_shared_memory(
+      loop_counter_alloc_size, loop_counter_alloc_size, 0, 0, device, context);
+
+  LOG_DEBUG << "[Application] Allocated source device memory at: " << std::hex
+            << src_buffer_d;
+  LOG_DEBUG << "[Application] Allocated destination device memory at: "
+            << std::hex << dest_buffer_d;
+
+  std::memset(dest_buffer_s, 1, size);
+  std::memset(src_buffer_s, 0, size);
+  std::memset(loop_counter_s, 0, loop_counter_alloc_size);
+  for (size_t i = 0; i < size; i++) {
+    static_cast<uint8_t *>(src_buffer_s)[i] = (i + 1 & 0xFF);
+  }
+
+  lzt::set_argument_value(kernel, 0, sizeof(dest_buffer_d), &dest_buffer_d);
+  lzt::set_argument_value(kernel, 1, sizeof(src_buffer_d), &src_buffer_d);
+  lzt::set_argument_value(kernel, 2, sizeof(loop_counter_d), &loop_counter_d);
+  lzt::set_argument_value(kernel, 3, sizeof(loop_max), &loop_max);
+  lzt::set_argument_value(kernel, 4, sizeof(slm_output_s), &slm_output_s);
+
+  uint32_t group_size_x = 1;
+  uint32_t group_size_y = 1;
+  uint32_t group_size_z = 1;
+  lzt::suggest_group_size(kernel, size, 1, 1, group_size_x, group_size_y,
+                          group_size_z);
+  lzt::set_group_size(kernel, group_size_x, 1, 1);
+  ze_group_count_t group_count = {};
+  group_count.groupCountX = size / group_size_x;
+  group_count.groupCountY = 1;
+  group_count.groupCountZ = 1;
+
+  lzt::append_memory_copy(command_list, src_buffer_d, src_buffer_s, size);
+  lzt::append_barrier(command_list);
+  lzt::append_launch_function(command_list, kernel, &group_count, nullptr, 0,
+                              nullptr);
+  lzt::append_barrier(command_list);
+  lzt::append_memory_copy(command_list, dest_buffer_s, dest_buffer_d, size);
+  lzt::append_memory_copy(command_list, loop_counter_s, loop_counter_d,
+                          loop_counter_alloc_size);
+  lzt::close_command_list(command_list);
+
+  LOG_DEBUG << "[Application] launching execution of " << kernel_name;
+
+  synchro.update_gpu_buffer_address(reinterpret_cast<uint64_t>(src_buffer_d));
+  synchro.notify_debugger();
+
+  lzt::execute_command_lists(command_queue, 1, &command_list, nullptr);
+  lzt::synchronize(command_queue, UINT64_MAX);
+
+  for (size_t i = 1; i < size; i++) {
+    EXPECT_EQ(static_cast<uint8_t *>(dest_buffer_s)[i],
+              static_cast<uint8_t *>(src_buffer_s)[i]);
+    if (static_cast<uint8_t *>(dest_buffer_s)[i] !=
+        static_cast<uint8_t *>(src_buffer_s)[i]) {
+      LOG_ERROR << "[Application] Buffer Sanity check did not pass";
+      break;
+    }
+  }
+
+  // cleanup
+  lzt::free_memory(context, dest_buffer_s);
+  lzt::free_memory(context, dest_buffer_d);
+  lzt::free_memory(context, src_buffer_s);
+  lzt::free_memory(context, src_buffer_d);
+  lzt::free_memory(context, loop_counter_s);
+  lzt::free_memory(context, loop_counter_d);
+  lzt::free_memory(context, slm_output_s);
+
+  lzt::destroy_function(kernel);
+  lzt::destroy_module(module);
+  lzt::destroy_command_list(command_list);
+  lzt::destroy_command_queue(command_queue);
+
+  if (::testing::Test::HasFailure()) {
+    exit(1);
+  }
+}
+
 void run_multiple_threads(ze_context_handle_t context,
                           ze_device_handle_t device, process_synchro &synchro,
                           debug_options &options) {
@@ -1227,6 +1354,11 @@ int main(int argc, char **argv) {
     options.kernel_name_in = "long_kernel_slm";
     run_long_kernel(context, device, synchro, options);
     break;
+  case LONG_RUNNING_KERNEL_INTERRUPTED_SCRATCH:
+    options.use_custom_module = true;
+    options.module_name_in = "debug_loop_slm.spv";
+    run_long_kernel_scratch(context, device, synchro, options);
+    break;
   case MULTIPLE_THREADS:
     run_multiple_threads(context, device, synchro, options);
     break;