#8223: Clean up magic numbers and fix first time stamp calculation

tenstorrent · May 28, 2024 · 49ec429 · 49ec429
1 parent 9c5fe88
commit 49ec429
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 24 deletions.
diff --git a/tt_metal/impl/dispatch/kernels/cq_prefetch.cpp b/tt_metal/impl/dispatch/kernels/cq_prefetch.cpp
@@ -658,6 +658,7 @@ uint32_t process_relay_linear_cmd(uint32_t cmd_ptr,
 
 uint32_t process_stall(uint32_t cmd_ptr) {
 
+    //DeviceZoneScopedN("PROCESS-STALL");
     static uint32_t count = 0;
 
     count++;
@@ -776,7 +777,6 @@ uint32_t process_exec_buf_cmd(uint32_t cmd_ptr_outer,
 
     bool done = false;
     while (!done) {
-        //DeviceZoneScopedMainN("PROC-MAIN");
         uint32_t cmd_ptr = cmddat_q_base;
 
         paged_read_into_cmddat_q(cmd_ptr);
@@ -1012,7 +1012,6 @@ void kernel_main_h() {
 
     bool done = false;
     while (!done) {
-        //DeviceZoneScopedMainN("KERNEL-MAIN-H");
         fetch_q_get_cmds<sizeof(CQPrefetchHToPrefetchDHeader)>(fence, cmd_ptr, pcie_read_ptr);
 
         volatile CQPrefetchCmd tt_l1_ptr *cmd = (volatile CQPrefetchCmd tt_l1_ptr *)(cmd_ptr + sizeof(CQPrefetchHToPrefetchDHeader));
@@ -1051,7 +1050,6 @@ void kernel_main_d() {
 
     bool done = false;
     while (!done) {
-        //DeviceZoneScopedMainN("KERNEL-MAIN-D");
         // cmds come in packed batches based on HostQ reads in prefetch_h
         // once a packed batch ends, we need to jump to the next page
         uint32_t length = relay_cb_get_cmds(fence, cmd_ptr);

diff --git a/tt_metal/tools/profiler/kernel_profiler.hpp b/tt_metal/tools/profiler/kernel_profiler.hpp
@@ -44,6 +44,8 @@ namespace kernel_profiler{
     extern uint32_t sums[SUM_COUNT];
     extern uint32_t sumIDs[SUM_COUNT];
 
+    constexpr uint32_t QUICK_PUSH_MARKER_COUNT = 2;
+
 #if defined(COMPILE_FOR_BRISC)
     constexpr uint32_t profilerBuffer = PROFILER_L1_BUFFER_BR;
     constexpr uint32_t deviceBufferEndIndex = DEVICE_BUFFER_END_INDEX_BR;
@@ -398,7 +400,7 @@ namespace kernel_profiler{
         uint64_t dram_bank_dst_noc_addr = s.get_noc_addr(core_flat_id / profiler_core_count_per_dram, dram_offset);
 
         mark_end_at_index_inlined(wIndex, hash, get_end_timer_id(hash));
-        wIndex += 2 * PROFILER_L1_MARKER_UINT32_SIZE;
+        wIndex += QUICK_PUSH_MARKER_COUNT * PROFILER_L1_MARKER_UINT32_SIZE;
 
         uint32_t currEndIndex = profiler_control_buffer[HOST_BUFFER_END_INDEX_NC] + wIndex;
 
@@ -425,7 +427,7 @@ namespace kernel_profiler{
         bool start_marked = false;
         inline __attribute__((always_inline)) profileScope ()
         {
-            if (wIndex < (PROFILER_L1_VECTOR_SIZE - stackSize - (2 * PROFILER_L1_MARKER_UINT32_SIZE)))
+            if (wIndex < (PROFILER_L1_VECTOR_SIZE - stackSize - (QUICK_PUSH_MARKER_COUNT * PROFILER_L1_MARKER_UINT32_SIZE)))
             {
                 stackSize += PROFILER_L1_MARKER_UINT32_SIZE;
                 start_marked = true;
@@ -443,7 +445,7 @@ namespace kernel_profiler{
                 start_marked = false;
                 stackSize -= PROFILER_L1_MARKER_UINT32_SIZE;
             }
-            if (wIndex >= (PROFILER_L1_VECTOR_SIZE - (2 * PROFILER_L1_MARKER_UINT32_SIZE)))
+            if (wIndex >= (PROFILER_L1_VECTOR_SIZE - (QUICK_PUSH_MARKER_COUNT * PROFILER_L1_MARKER_UINT32_SIZE)))
             {
                 quick_push();
             }

diff --git a/tt_metal/tools/profiler/profiler.cpp b/tt_metal/tools/profiler/profiler.cpp
@@ -380,24 +380,6 @@ void DeviceProfiler::dumpResults (
 
         }
 
-        for (const auto &worker_core : worker_cores) {
-            std::pair<uint32_t, CoreCoord> device_core = {device_id, worker_core};
-            if (device_tracy_contexts.find(device_core) == device_tracy_contexts.end())
-            {
-                auto tracyCtx = TracyTTContext();
-                std::string tracyTTCtxName = fmt::format("Device: {}, Core ({},{})", device_id, worker_core.x, worker_core.y);
-                TracyTTContextPopulate(tracyCtx, smallest_timestamp, 1000.f / (float)device_core_frequency);
-                TracyTTContextName(tracyCtx, tracyTTCtxName.c_str(), tracyTTCtxName.size());
-
-                device_tracy_contexts.emplace(
-                        device_core,
-                        tracyCtx
-                    );
-            }
-        }
-
-        //std::sort (device_events.begin(), device_events.end());
-
     }
     else
     {
@@ -412,6 +394,38 @@ void DeviceProfiler::pushTracyDeviceResults()
     ZoneScoped;
     std::set<std::pair<uint32_t, CoreCoord>> device_cores_set;
     std::vector<std::pair<uint32_t, CoreCoord>> device_cores;
+    for (auto& event: device_events)
+    {
+        std::pair<uint32_t, CoreCoord> device_core = {event.chip_id, (CoreCoord){event.core_x,event.core_y}};
+        auto ret = device_cores_set.insert(device_core);
+        if (ret.second )
+        {
+            device_cores.push_back(device_core);
+        }
+    }
+
+    for (auto& device_core: device_cores)
+    {
+        int device_id = device_core.first;
+        CoreCoord worker_core = device_core.second;
+
+
+        if (device_tracy_contexts.find(device_core) == device_tracy_contexts.end())
+        {
+            auto tracyCtx = TracyTTContext();
+            std::string tracyTTCtxName = fmt::format("Device: {}, Core ({},{})", device_id, worker_core.x, worker_core.y);
+
+	    TracyTTContextPopulate(tracyCtx, smallest_timestamp, 1000.f / (float)device_core_frequency);
+
+            TracyTTContextName(tracyCtx, tracyTTCtxName.c_str(), tracyTTCtxName.size());
+
+            device_tracy_contexts.emplace(
+                    device_core,
+                    tracyCtx
+                );
+        }
+    }
+
     for (auto& event: device_events)
     {
         std::pair<uint32_t, CoreCoord> device_core = {event.chip_id, (CoreCoord){event.core_x,event.core_y}};