Noc Analyzer Prototype

tools/CMakeLists.txt

@@ -1,2 +1,3 @@
 add_subdirectory(ttmlir-opt)
 add_subdirectory(ttmlir-translate)
+add_subdirectory(ttmlir-noc-analyzer)

tools/ttmlir-noc-analyzer/CMakeLists.txt

@@ -0,0 +1,16 @@
+get_property(dialect_libs GLOBAL PROPERTY MLIR_DIALECT_LIBS)
+get_property(conversion_libs GLOBAL PROPERTY MLIR_CONVERSION_LIBS)
+find_package(jsoncpp REQUIRED)
+set(LIBS ${dialect_libs} ${conversion_libs} MLIROptLib MLIRTargetCpp TTMLIRStatic jsoncpp_lib)
+
+set(SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/ttmlir-noc-analyzer.cpp ${CMAKE_CURRENT_SOURCE_DIR}/noc-analytical-model.cpp)
+set(HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/noc-analytical-model.hpp)
+
+add_executable(ttmlir-noc-analyzer ${HEADERS} ${SOURCES})
+
+llvm_update_compile_flags(ttmlir-noc-analyzer)
+target_link_libraries(ttmlir-noc-analyzer PRIVATE ${LIBS})
+
+mlir_check_all_link_libraries(ttmlir-noc-analyzer)
+
+install(TARGETS ttmlir-noc-analyzer DESTINATION ${CMAKE_INSTALL_BINDIR} COMPONENT Test EXCLUDE_FROM_ALL)

tools/ttmlir-noc-analyzer/noc-analytical-model.cpp

@@ -0,0 +1,75 @@
+#include "noc-analytical-model.hpp"
+
+namespace analyzer::model {
+// Returns (total_issue_latency (no overhead), round_trip_latency, flit_latency)
+std::tuple<float, float, float>
+get_read_latency(std::map<std::string, int> noc_params, int buffer_size) {
+  int transfer_size = buffer_size <= NOC_MAX_BURST_SIZE ? buffer_size : NOC_MAX_BURST_SIZE;
+  //   float num_flits_per_transfer = transfer_size /
+  //   noc_params["noc_data_width"];
+  float num_transfer = buffer_size / transfer_size;
+  int round_trip_latency = noc_params["round_trip_latency"];
+  float issue_latency =
+      noc_params["niu_programming"] + noc_params["non_niu_programming"];
+
+  float total_issue_latency = issue_latency * num_transfer;
+  float total_flit_latency =
+      noc_params["head_flit_latency"] + noc_params["flit_latency"] * buffer_size / noc_params["noc_data_width"];
+
+//   float barrier_latency;
+//   if (num_flits_per_transfer >= issue_latency) {
+//     barrier_latency =
+//         noc_params["pre_issue_overhead"] + issue_latency + round_trip_latency +
+//         noc_params["head_flit_latency"] +
+//         noc_params["flit_latency"] * num_flits_per_transfer * num_transfer -
+//         total_issue_latency;
+//   } else {
+//     barrier_latency = round_trip_latency + noc_params["head_flit_latency"] +
+//                       noc_params["flit_latency"] * num_flits_per_transfer;
+//   }
+
+  return std::tuple<float, float, float>{total_issue_latency, round_trip_latency,
+                                         total_flit_latency};
+}
+
+std::tuple<float, float, float>
+get_write_latency(std::map<std::string, int> noc_params, int buffer_size) {
+  auto transfer_size = buffer_size <= NOC_MAX_BURST_SIZE ? buffer_size : NOC_MAX_BURST_SIZE;
+  float num_flits_per_transfer = transfer_size / noc_params["noc_data_width"];
+  float num_transfer = buffer_size / transfer_size;
+  int round_trip_latency = noc_params["round_trip_latency"];
+
+  float issue_latency =
+      noc_params["niu_programming"] + noc_params["non_niu_programming"];
+
+  float total_flit_latency =
+      noc_params["flit_latency"] * buffer_size / noc_params["noc_data_width"];
+
+  float total_issue_latency;
+  float barrier_latency;
+
+  if (num_flits_per_transfer >= issue_latency) {
+    if (num_transfer < 3) {
+      total_issue_latency = issue_latency * num_transfer;
+    } else {
+      total_issue_latency = issue_latency +
+                            noc_params["flit_latency"] *
+                                num_flits_per_transfer * (num_transfer - 2) +
+                            15;
+    }
+    barrier_latency =
+        issue_latency + round_trip_latency + noc_params["head_flit_latency"] +
+        noc_params["flit_latency"] * num_flits_per_transfer * num_transfer -
+        total_issue_latency;
+  } else {
+    total_issue_latency =
+        noc_params["pre_issue_overhead"] + issue_latency * num_transfer;
+    barrier_latency = noc_params["round_trip_latency"] +
+                      noc_params["head_flit_latency"] +
+                      noc_params["flit_latency"] * num_flits_per_transfer;
+  }
+
+  return std::tuple<float, float, float>{total_issue_latency, barrier_latency,
+                                         total_flit_latency};
+}
+} // namespace analyzer::model

tools/ttmlir-noc-analyzer/noc-analytical-model.hpp

@@ -0,0 +1,19 @@
+#include <string>
+#include <map>
+
+#define NOC_MAX_BURST_SIZE 8192
+
+namespace analyzer::model {
+
+    const std::map<std::string, int> noc_params = {
+        {"pre_issue_overhead", 17}, 
+        {"niu_programming", 6}, 
+        {"non_niu_programming", 43}, 
+        {"round_trip_latency", 96}, 
+        {"head_flit_latency", 1},
+        {"flit_latency", 1},
+        {"noc_data_width", 32}};
+
+    std::tuple<float, float, float> get_read_latency(std::map<std::string, int> noc_params, int buffer_size);
+    std::tuple<float, float, float> get_write_latency(std::map<std::string, int> noc_params, int buffer_size);
+}