[SPM] Clean up obsolete SPM logics and refine some SPM-related definition (#157)

* Clean up obsolete SPM logics
* Add PERFCOUNTER_SELECT1 to CounterRegInfo
* Add RLC_SPM_PERFMON_SAMPLE_DELAY_MAX

[ROCm/aqlprofile commit: 6f236ffb5f]

projects/aqlprofile/src/pm4/gfx10_cmd_builder.h

@@ -330,7 +330,7 @@ class Gfx10CmdBuilder : public CmdBuilder {
   }
 
   void BuildNopPacket(CmdBuffer* cmdbuf, uint32_t num_dwords) {
-    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords);
+    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords * sizeof(uint32_t));
     APPEND_COMMAND_WRAPPER(cmdbuf, header);
     if (num_dwords > 1) {
       std::vector<uint32_t> data_block((num_dwords - 1), 0);

projects/aqlprofile/src/pm4/gfx11_cmd_builder.h

@@ -304,7 +304,7 @@ class Gfx11CmdBuilder : public CmdBuilder {
   }
 
   void BuildNopPacket(CmdBuffer* cmdbuf, uint32_t num_dwords) {
-    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords);
+    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords * sizeof(uint32_t));
     APPEND_COMMAND_WRAPPER(cmdbuf, header);
     if (num_dwords > 1) {
       std::vector<uint32_t> data_block((num_dwords - 1), 0);

projects/aqlprofile/src/pm4/gfx9_cmd_builder.h

@@ -325,7 +325,7 @@ class Gfx9CmdBuilder : public CmdBuilder {
   }
 
   void BuildNopPacket(CmdBuffer* cmdbuf, uint32_t num_dwords) {
-    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords);
+    uint32_t header = MakePacket3Header(PACKET3_NOP, num_dwords * sizeof(uint32_t));
     APPEND_COMMAND_WRAPPER(cmdbuf, header);
     if (num_dwords > 1) {
       std::vector<uint32_t> data_block((num_dwords - 1), 0);

projects/aqlprofile/src/pm4/spm_builder.h

@@ -104,19 +104,6 @@ class GpuSpmBuilder : public SpmBuilder, protected Primitives {
     // 3. Program the RLC_PERFMON_SEGMENT_SIZE register.
     builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::RLC_SPM_PERFMON_CNTL__ADDR,
                                        Primitives::rlc_spm_perfmon_cntl_value(sampling_rate));
-    if (!config->spm_kfd_mode) {
-      builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::RLC_SPM_PERFMON_RING_BASE_LO__ADDR,
-                                         buffer_ptr);
-      builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::RLC_SPM_PERFMON_RING_BASE_HI__ADDR,
-                                         buffer_ptr >> 32);
-      builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::RLC_SPM_PERFMON_RING_SIZE__ADDR,
-                                         buffer_size);
-    }
-
-    // Setting VMID
-    if (!config->spm_kfd_mode)
-      builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::RLC_SPM_MC_CNTL__ADDR,
-                                         Primitives::rlc_spm_mc_cntl_value());
 
     // Iterate through the list of blocks to create PM4 packets to read counter values
     // Below pair.first is the block id of a counter event and pair.second is the index into
@@ -272,17 +259,16 @@ class GpuSpmBuilder : public SpmBuilder, protected Primitives {
           for (size_t j = 0; j < block_info->instance_count; ++j) {
             builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::GRBM_GFX_INDEX_ADDR,
                                                Primitives::grbm_inst_se_sh_index_value(j, 0, 0));
-            builder.BuildWriteUConfigRegPacket(cmd_buffer, block_info->delay_info[j].reg,
+            builder.BuildWriteUConfigRegPacket(cmd_buffer, block_info->delay_info.reg,
                                                Primitives::get_spm_global_delay(counter_des, j));
           }
         } else {
-          for (size_t i = 0; i < config->spm_se_number_total; ++i) {
+          for (size_t i = 0; i < config->se_number; ++i) {
             for (size_t j = 0; j < block_info->instance_count; ++j) {
-              int delay_index = i * block_info->instance_count + j;
               builder.BuildWriteUConfigRegPacket(cmd_buffer, Primitives::GRBM_GFX_INDEX_ADDR,
                                                  Primitives::grbm_inst_se_index_value(j, i));
               builder.BuildWriteUConfigRegPacket(cmd_buffer,
-                                                 block_info->delay_info[delay_index].reg,
+                                                 block_info->delay_info.reg,
                                                  Primitives::get_spm_se_delay(counter_des, i, j));
             }
           }
@@ -341,7 +327,7 @@ class GpuSpmBuilder : public SpmBuilder, protected Primitives {
     builder.BuildWriteUConfigRegPacket(
         cmd_buffer, Primitives::RLC_SPM_PERFMON_SEGMENT_SIZE__ADDR,
         Primitives::rlc_spm_perfmon_segment_size_value(global_count, se_count));
-    if (config->mi100) {
+    if (config->spm_has_core1) {
       builder.BuildWriteUConfigRegPacket(
           cmd_buffer, Primitives::RLC_SPM_PERFMON_SEGMENT_SIZE_CORE1__ADDR,
           Primitives::rlc_spm_perfmon_segment_size_core1_value(se_count));

projects/aqlprofile/src/pm4/trace_config.h

@@ -42,18 +42,20 @@ struct TraceConfig {
   uint32_t perfMASK = ~0u;
   uint32_t perfCTRL = 0;
   std::vector<std::pair<size_t, size_t>> perfcounters{};
+  // GC configurations used by both TT and SPM
+  uint32_t se_number = 0;
+  uint32_t sa_number = 0;
+  uint32_t xcc_number = 0;
   // SPM mode
   bool spm_sq_32bit_mode = true;
-  bool spm_kfd_mode = true;
-  bool mi100 = false;
+  bool spm_has_core1 = false;
+  uint32_t spm_sample_delay_max = 0;
 
   void* control_buffer_ptr = nullptr;
   uint32_t control_buffer_size = 0;
   void* data_buffer_ptr = nullptr;
   uint32_t data_buffer_size = 0;
 
-  // SE number for tracing
-  uint32_t spm_se_number_total = 0;
   // concurrent kernels mode
   uint32_t concurrent = 0;
   // SE mask for tracing; note -> replicated for all XCCs