[soc_dump] add doc

Change-Id: Icaf313113415a9bf0ad9c75abdcb609d661c815b

docs/developer_manual/source_en/Appx.02_tpulang_support_op.rst

@@ -3930,9 +3930,11 @@ Definition
             save_path: str = "",
             out_fixed: bool = False,
             dump_cmd_info: bool = True,
-            cmodel_skip_check: bool = True,  # disable CMODEL data_check to increase processing speed
+            skip_check: bool = True,  # disable data_check to increase processing speed
+            run_by_op: bool = False, # enable to run_by_op, may cause timeout error when some OPs contain too many atomic cmds
             is_soc: bool = False,  # soc mode ONLY support {reference_data_fn=xxx.npz, dump_file=True}
-            enable_soc_log: bool = False,
+            using_memory_opt: bool = False, # required when is_soc=True
+            enable_soc_log: bool = False, # required when is_soc=True
             tmp_path: str = "/tmp",  # required when is_soc=True
             tools_path: str = "/soc_infer",  # required when is_soc=True
             hostname: str = None,  # required when is_soc=True
@@ -3955,9 +3957,12 @@ Parameters
 * dump_file: Bool type, representing whether save results as file.
 * save_path: String type, representing the abs path of saving results on host.
 * out_fixed: Bool type, representing whether to get results in fixed number.
-* dump_cmd_info: Bool type, enable to save atomic cmd info at `save_path`. 
+* dump_cmd_info: Bool type, enable to save atomic cmd info at `save_path`.
+* skip_check: Bool tyoe, set to True to disable data check to decrease time cost for CMODEL/PCIE mode.
+* run_by_op: Bool type, enable to run_by_op, decrease time cost but may cause timeout error when some OPs contain too many atomic cmds.
 * cmodel_skip_check: Bool type, enable this to skip data check to speed up inference.
 * is_soc: Bool type, representing whether to use in soc mode.
+* using_memory_opt: Bool type, enable to use memory opt, decrease memory usage at the expense of increasing time cost. Suggest to enable when running large model.
 * enable_soc_log: Bool type, enable to print and save log at `save_path`.
 * tmp_path: String type, representing the abs path of tmp files on device in soc mode.
 * tools_path: String type, representing the dir of soc_infer tools on device in soc mode.

docs/developer_manual/source_zh/Appx.02_tpulang_support_op.rst

@@ -4802,9 +4802,11 @@ bmodel_inference_combine
           save_path: str = "",
           out_fixed: bool = False,
           dump_cmd_info: bool = True,
-          cmodel_skip_check: bool = True,  # disable CMODEL data_check to increase processing speed
+          skip_check: bool = True,  # disable data_check to increase processing speed
+          run_by_op: bool = False, # enable to run_by_op, may cause timeout error when some OPs contain too many atomic cmds
           is_soc: bool = False,  # soc mode ONLY support {reference_data_fn=xxx.npz, dump_file=True}
-          enable_soc_log: bool = False,  
+          using_memory_opt: bool = False, # required when is_soc=True
+          enable_soc_log: bool = False, # required when is_soc=True
           tmp_path: str = "/tmp",  # required when is_soc=True
           tools_path: str = "/soc_infer",  # required when is_soc=True
           hostname: str = None,  # required when is_soc=True
@@ -4828,8 +4830,10 @@ bmodel_inference_combine
 * save_path: String类型，表示 `dump_file=True` 时的主机(host)端保存逐层推理的.npz文件的绝对路径。
 * out_fixed: Bool类型，表示逐层Tensor数据输出是否保持为定点格式。
 * dump_cmd_info: Bool类型，表示将当前bmodel中包含的所有原子指令对应的final.mlir的信息保存成txt文件，保存路径在save_path下。
-* cmodel_skip_check: Bool类型， 在cmodel模式下启用此项可禁用数据对比，提高推理速度。soc模式下默认不进行数据对比。
+* skip_check: Bool类型，启用此项可禁用数据对比，提高推理速度。soc模式下默认不进行数据对比。
+* run_by_op: Bool类型，启用后按OP粒度运行，禁用时为按原子指令粒度运行。按OP粒度运行速度较快，但当一个OP中包含过多原子指令时可能会引发timeout错误。
 * is_soc: Bool类型，表示是否启用soc模式进行推理。
+* using_memory_opt: Bool类型，启用后会减小在device端的内存消耗，但会增加耗时。推荐在大模型时启用。
 * enable_soc_log: Bool类型，启用此项打印并在save_path下保存log日志。
 * tmp_path: String类型，表示soc模式下，板卡(device)端存放临时文件的绝对路径。
 * tools_path: String类型，表示soc模式下，device端存放工具的文件夹名称。

python/transform/TpuLang.py

@@ -297,10 +297,10 @@ def bmodel_inference_combine(
     out_fixed: bool = False,
     dump_cmd_info: bool = True,
     skip_check: bool = True,  # disable data_check to increase processing speed
-    run_by_op: bool = False,
+    run_by_op: bool = False, # enable to run_by_op, may cause timeout error when some OPs contain too many atomic cmds
     is_soc: bool = False,  # soc mode ONLY support {reference_data_fn=xxx.npz, dump_file=True}
-    using_memory_opt: bool = False,
-    enable_soc_log: bool = False,
+    using_memory_opt: bool = False, # required when is_soc=True
+    enable_soc_log: bool = False, # required when is_soc=True
     tmp_path: str = "/tmp",  # required when is_soc=True
     tools_path: str = "/soc_infer",  # required when is_soc=True
     hostname: str = None,  # required when is_soc=True
@@ -405,7 +405,7 @@ def progress_get(file_name, local_path="", remote_path="", progress=None):
         remote_bmodel = os.path.basename(bmodel_file)
         remote_input = os.path.basename(input_data_fn)
         remote_ref = os.path.basename(reference_data_fn)
-        exec_command = f"cd {tools_path} && source envsetup.sh && python3 soc_bmodel_infer.py --path {tmp_path} --bmodel {remote_bmodel} --input {remote_input} --ref {remote_ref} --tool_path {tools_path}"
+        exec_command = f"cd {tools_path} && source envsetup.sh && nohup python3 soc_bmodel_infer.py --path {tmp_path} --bmodel {remote_bmodel} --input {remote_input} --ref {remote_ref} --tool_path {tools_path}"
         if out_fixed:
             exec_command += " --out_fixed"
         if enable_soc_log:
@@ -414,6 +414,7 @@ def progress_get(file_name, local_path="", remote_path="", progress=None):
             exec_command += " --using_memory_opt"
         if not run_by_op:
             exec_command += " --run_by_atomic"
+        exec_command += " &"
         print(f"soc execute command: {exec_command}")
 
         client.get_transport().set_keepalive(30)