#6748: Clean up function footprints

tenstorrent · Oct 3, 2024 · f52272a · f52272a
1 parent 03e6cff
commit f52272a
Show file tree

Hide file tree

Showing 6 changed files with 13 additions and 36 deletions.
diff --git a/tt_metal/hostdevcommon/profiler_common.h b/tt_metal/hostdevcommon/profiler_common.h
@@ -41,6 +41,7 @@ namespace kernel_profiler{
         NOC_X,
         NOC_Y,
         FLAT_ID,
+        CORE_COUNT_PER_DRAM,
         DROPPED_ZONES,
         PROFILER_DONE,
     };

diff --git a/tt_metal/impl/device/device.cpp b/tt_metal/impl/device/device.cpp
@@ -2911,9 +2911,7 @@ void Device::generate_device_headers(const std::string &path) const
         dram_noc_coord_per_bank,
         dram_offsets_per_bank,
         l1_noc_coord_per_bank,
-        l1_offset_per_bank,
-        soc_d.profiler_ceiled_core_count_perf_dram_bank,
-        soc_d.physical_routing_to_profiler_flat_id
+        l1_offset_per_bank
     );
 }
 

diff --git a/tt_metal/jit_build/genfiles.cpp b/tt_metal/jit_build/genfiles.cpp
@@ -524,9 +524,7 @@ std::string generate_bank_to_noc_coord_descriptor_string(
     std::vector<CoreCoord>& dram_bank_map,
     std::vector<int32_t>& dram_bank_offset_map,
     std::vector<CoreCoord>& l1_bank_map,
-    std::vector<int32_t>& l1_bank_offset_map,
-    int core_count_per_dram,
-    const std::map<CoreCoord, int32_t>& profiler_flat_id_map) {
+    std::vector<int32_t>& l1_bank_offset_map) {
     stringstream ss;
     bool is_dram_pow2 = ceil(log2(dram_bank_map.size())) == log2(dram_bank_map.size());
     bool is_l1_pow2 = ceil(log2(l1_bank_map.size())) == log2(l1_bank_map.size());
@@ -577,13 +575,6 @@ std::string generate_bank_to_noc_coord_descriptor_string(
     ss << "extern int32_t bank_to_dram_offset[NUM_DRAM_BANKS];" << endl;
     ss << "extern uint16_t l1_bank_to_noc_xy[NUM_NOCS][NUM_L1_BANKS];" << endl;
     ss << "extern int32_t bank_to_l1_offset[NUM_L1_BANKS];" << endl;
-#if defined(TRACY_ENABLE)
-    ss << "#if defined(PROFILE_KERNEL) && (defined(COMPILE_FOR_BRISC) || defined(COMPILE_FOR_NCRISC) || "
-          "defined(COMPILE_FOR_ERISC))"
-       << endl;
-    ss << "extern uint16_t profiler_core_count_per_dram;" << endl;
-    ss << "#endif" << endl;
-#endif
 
     ss << endl;
     ss << "#else // !KERNEL_BUILD (FW_BUILD)" << endl;
@@ -611,17 +602,6 @@ std::string generate_bank_to_noc_coord_descriptor_string(
     ss << "};" << endl;
     ss << endl;
 
-#if defined(TRACY_ENABLE)
-    ss << "#if defined(PROFILE_KERNEL) && (defined(COMPILE_FOR_BRISC) || defined(COMPILE_FOR_NCRISC) || "
-          "defined(COMPILE_FOR_ERISC))"
-       << endl;
-    ss << "uint16_t profiler_core_count_per_dram __attribute__((used)) = ";
-    ss << core_count_per_dram << ";" << endl;
-    ss << endl;
-    ss << "#endif" << endl;
-
-#endif
-
     ss << "uint16_t l1_bank_to_noc_xy[NUM_NOCS][NUM_L1_BANKS] __attribute__((used)) = {" << endl;
     for (unsigned int noc = 0; noc < 2; noc++) {
         ss << "    {"
@@ -654,17 +634,13 @@ void jit_build_genfiles_bank_to_noc_coord_descriptor(
     std::vector<CoreCoord>& dram_bank_map,
     std::vector<int32_t>& dram_bank_offset_map,
     std::vector<CoreCoord>& l1_bank_map,
-    std::vector<int32_t>& l1_bank_offset_map,
-    int core_count_per_dram,
-    const std::map<CoreCoord, int32_t>& profiler_flat_id_map) {
+    std::vector<int32_t>& l1_bank_offset_map) {
     string output_string = generate_bank_to_noc_coord_descriptor_string(
         grid_size,
         dram_bank_map,
         dram_bank_offset_map,
         l1_bank_map,
-        l1_bank_offset_map,
-        core_count_per_dram,
-        profiler_flat_id_map);
+        l1_bank_offset_map);
 
     fs::create_directories(path + "/brisc");
     ofstream file_stream_br(path + "/brisc/generated_bank_to_noc_coord_mapping.h");

diff --git a/tt_metal/jit_build/genfiles.hpp b/tt_metal/jit_build/genfiles.hpp
@@ -27,9 +27,7 @@ void jit_build_genfiles_bank_to_noc_coord_descriptor(
     std::vector<CoreCoord>& dram_bank_map,
     std::vector<int32_t>& dram_bank_offset_map,
     std::vector<CoreCoord>& l1_bank_map,
-    std::vector<int32_t>& l1_bank_offset_map,
-    int core_count_per_dram,
-    const std::map<CoreCoord, int32_t>& profiler_flat_id_map);
+    std::vector<int32_t>& l1_bank_offset_map);
 
 void jit_build_genfiles_descriptors(const JitBuildEnv& env, JitBuildOptions& options);
 

diff --git a/tt_metal/tools/profiler/kernel_profiler.hpp b/tt_metal/tools/profiler/kernel_profiler.hpp
@@ -187,11 +187,12 @@ namespace kernel_profiler{
         if (profiler_control_buffer[PROFILER_DONE] == 1){
             return;
         }
-        uint32_t pageSize =
-            PROFILER_FULL_HOST_BUFFER_SIZE_PER_RISC * MAX_RISCV_PER_CORE * profiler_core_count_per_dram;
-
         while (!profiler_control_buffer[DRAM_PROFILER_ADDRESS]);
         uint32_t core_flat_id = profiler_control_buffer[FLAT_ID];
+        uint32_t profiler_core_count_per_dram = profiler_control_buffer[CORE_COUNT_PER_DRAM];
+
+        uint32_t pageSize =
+            PROFILER_FULL_HOST_BUFFER_SIZE_PER_RISC * MAX_RISCV_PER_CORE * profiler_core_count_per_dram;
 
         for (uint32_t riscID = 0; riscID < PROFILER_RISC_COUNT; riscID ++)
 	{
@@ -267,6 +268,7 @@ namespace kernel_profiler{
 
         while (!profiler_control_buffer[DRAM_PROFILER_ADDRESS]);
         uint32_t core_flat_id = profiler_control_buffer[FLAT_ID];
+        uint32_t profiler_core_count_per_dram = profiler_control_buffer[CORE_COUNT_PER_DRAM];
 
         profiler_data_buffer[myRiscID][ID_LH] = ((core_flat_id & 0xFF) << 3) | myRiscID;
 

diff --git a/tt_metal/tools/profiler/tt_metal_profiler.cpp b/tt_metal/tools/profiler/tt_metal_profiler.cpp
@@ -62,6 +62,8 @@ void setControlBuffer(uint32_t device_id, std::vector<uint32_t>& control_buffer)
 #if defined(TRACY_ENABLE)
     const metal_SocDescriptor& soc_d = tt::Cluster::instance().get_soc_desc(device_id);
 
+    control_buffer[kernel_profiler::CORE_COUNT_PER_DRAM] = soc_d.profiler_ceiled_core_count_perf_dram_bank;
+
     auto ethCores = soc_d.get_physical_ethernet_cores() ;
     for (auto &core : soc_d.physical_routing_to_profiler_flat_id)
     {