Move the device timer conversion in cvk_device instead of commands

kpet · Mar 11, 2024 · 0d4475e · 0d4475e
1 parent 0c6a3e3
commit 0d4475e
Show file tree

Hide file tree

Showing 3 changed files with 24 additions and 23 deletions.
diff --git a/src/device.cpp b/src/device.cpp
@@ -1196,11 +1196,15 @@ cl_int cvk_device::get_device_host_timer(cl_ulong* device_timestamp,
     return CL_SUCCESS;
 }
 
-cl_ulong cvk_device::device_timer_to_host(cl_ulong dev, cl_ulong sync_dev,
-                                          cl_ulong sync_host) const {
-    if (sync_host > sync_dev) {
-        return (sync_host - sync_dev) + dev;
+cl_ulong cvk_device::device_timer_to_host(cl_ulong dev) {
+    if (dev > m_sync_dev) {
+        if (get_device_host_timer(&m_sync_dev, &m_sync_host) != CL_SUCCESS) {
+            return dev;
+        }
+    }
+    if (m_sync_host > m_sync_dev) {
+        return (m_sync_host - m_sync_dev) + dev;
     } else {
-        return dev - (sync_dev - sync_host);
+        return dev - (m_sync_dev - m_sync_host);
     }
 }
diff --git a/src/device.hpp b/src/device.hpp
@@ -514,8 +514,7 @@ struct cvk_device : public _cl_device_id,
 
     CHECK_RETURN cl_int get_device_host_timer(cl_ulong* dev_ts,
                                               cl_ulong* host_ts) const;
-    cl_ulong device_timer_to_host(cl_ulong dev, cl_ulong sync_dev,
-                                  cl_ulong sync_host) const;
+    cl_ulong device_timer_to_host(cl_ulong dev);
 
     uint64_t timestamp_to_ns(uint64_t ts) const {
         double ns_per_tick = vulkan_limits().timestampPeriod;
@@ -708,6 +707,9 @@ struct cvk_device : public _cl_device_id,
 
     cl_uint m_preferred_subgroup_size{};
 
+    cl_ulong m_sync_host{};
+    cl_ulong m_sync_dev{};
+
     spv_target_env m_vulkan_spirv_env;
 
     std::unique_ptr<cvk_device_properties> m_clvk_properties;

diff --git a/src/queue.hpp b/src/queue.hpp
@@ -673,16 +673,18 @@ struct cvk_command_batchable : public cvk_command {
     CHECK_RETURN cl_int do_action() override;
     CHECK_RETURN virtual cl_int do_post_action() { return CL_SUCCESS; }
 
-    CHECK_RETURN cl_int set_profiling_info_end(cl_ulong sync_dev,
-                                               cl_ulong sync_host) {
+    CHECK_RETURN cl_int set_profiling_info_end() {
+        // If it has already been set, don't override it
+        if (m_event->get_profiling_info(CL_PROFILING_COMMAND_END) != 0) {
+            return CL_SUCCESS;
+        }
         cl_ulong start, end;
         auto perr = get_timestamp_query_results(&start, &end);
         if (perr != CL_COMPLETE) {
             return perr;
         }
-        start =
-            m_queue->device()->device_timer_to_host(start, sync_dev, sync_host);
-        end = m_queue->device()->device_timer_to_host(end, sync_dev, sync_host);
+        start = m_queue->device()->device_timer_to_host(start);
+        end = m_queue->device()->device_timer_to_host(end);
         m_event->set_profiling_info(CL_PROFILING_COMMAND_START, start);
         m_event->set_profiling_info(CL_PROFILING_COMMAND_END, end);
         return CL_SUCCESS;
@@ -698,12 +700,10 @@ struct cvk_command_batchable : public cvk_command {
             pinfo == CL_PROFILING_COMMAND_SUBMIT) {
             return cvk_command::set_profiling_info(pinfo);
         } else if (pinfo == CL_PROFILING_COMMAND_START) {
-            return m_queue->device()->get_device_host_timer(&m_sync_dev,
-                                                            &m_sync_host);
+            return CL_SUCCESS;
         } else {
             CVK_ASSERT(pinfo == CL_PROFILING_COMMAND_END);
-            CVK_ASSERT(m_sync_dev != 0 && m_sync_host != 0);
-            return set_profiling_info_end(m_sync_dev, m_sync_host);
+            return set_profiling_info_end();
         }
     }
 
@@ -714,8 +714,6 @@ struct cvk_command_batchable : public cvk_command {
     static const int NUM_POOL_QUERIES_PER_COMMAND = 2;
     static const int POOL_QUERY_CMD_START = 0;
     static const int POOL_QUERY_CMD_END = 1;
-
-    cl_ulong m_sync_dev{}, m_sync_host{};
 };
 
 struct cvk_ndrange {
@@ -845,14 +843,12 @@ struct cvk_command_batch : public cvk_command {
         cl_int status = cvk_command::set_profiling_info(pinfo);
         if (m_queue->profiling_on_device()) {
             if (pinfo == CL_PROFILING_COMMAND_START) {
-                return m_queue->device()->get_device_host_timer(&m_sync_dev,
-                                                                &m_sync_host);
+                return status;
             } else {
                 for (auto& cmd : m_commands) {
                     cl_int err;
                     if (pinfo == CL_PROFILING_COMMAND_END) {
-                        err = cmd->set_profiling_info_end(m_sync_dev,
-                                                          m_sync_host);
+                        err = cmd->set_profiling_info_end();
                     } else {
                         err = cmd->set_profiling_info(pinfo);
                     }
@@ -880,7 +876,6 @@ struct cvk_command_batch : public cvk_command {
 private:
     std::vector<std::unique_ptr<cvk_command_batchable>> m_commands;
     std::unique_ptr<cvk_command_buffer> m_command_buffer;
-    cl_ulong m_sync_dev, m_sync_host;
 };
 
 struct cvk_command_map_buffer final : public cvk_command_buffer_base_region {