Merge branch 'main' into tyf/addSeed

Author: Bonbon-Tang

.github/workflows/main.yml

@@ -353,6 +353,21 @@ jobs:
           source scripts/ci/ascend/ci_ascend_env.sh
           bash scripts/ci/ascend/ci_ascend_script.sh build_dipu \
           || ( cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/ && rm -rf ${GITHUB_JOB} && exit 1 )
+  
+  Build-Ascend-910b-with-autocompare:
+    name: Build-dipu-ascend-910b-with-autocompare
+    needs: [Build-PyTorch-For-Ascend-910b]
+    runs-on: tps-ascend-ci-910b
+    steps:
+      - name: Build dipu
+        run: |
+          set -ex
+          export USE_COVERAGE=ON
+          export USE_AUTOCOMPARE=ON
+          cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/ && rm -rf ${GITHUB_JOB} && cp -R source ${GITHUB_JOB}  && cd ${GITHUB_JOB}/dipu
+          source scripts/ci/ascend/ci_ascend_env.sh
+          bash scripts/ci/ascend/ci_ascend_script.sh build_dipu \
+          || ( cd ${DEEPLINK_PATH}/${GITHUB_RUN_NUMBER}/ && rm -rf ${GITHUB_JOB} && exit 1 )
 
   Test-Ascend-910b:
     name: Test-dipu-ascend-910b

dipu/CMakeLists.txt

@@ -5,6 +5,8 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
 
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
 set(CMAKE_BUILD_TYPE "Release" CACHE STRING "default build type is Release")
 option(TESTS "Whether to build unit tests" OFF)
 option(LIBS "Whether to build dipu lib, default on" ON)

dipu/QuickStart.md

@@ -183,21 +183,12 @@ python -c "import torch_dipu"
 
 #### 算子精度自动对比功能介绍
 
-由于该功能默认不开启，使用该功能时需要打开该功能并重新编译 DIPU。
-
-如在寒武纪设备上，可将 `dipu/torch_dipu/csrc_dipu/CMakeLists.txt` 中的 `autocompare` 修改为 `True`。
-
-```cmake
-add_custom_command(
-  OUTPUT "${DIPU_AUTOGENED_KERNELS_CPP}"
-  COMMAND
-    python "${DIPU_AUTOGEN_DIOPI_WRAPPER_SCRIPT}" --config
-    "${DIPU_AUTOGEN_DIOPI_WRAPPER_CONFIG}" --out "${DIPU_AUTOGENED_KERNELS_CPP}"
-    --use_diopi_adapter False --autocompare True --print_func_call_info True
-    --print_op_arg True --fun_config_dict
-    '{\"current_device\": \"${UsedVendor}\"}'
-  DEPENDS ${DIPU_AUTOGEN_DIOPI_WRAPPER_SCRIPT}
-          ${DIPU_AUTOGEN_DIOPI_WRAPPER_CONFIG})
+由于该功能默认不开启，使用该功能时需要打开该功能并重新编译DIPU。
+
+可以通过设置环境变量USE_AUTOCOMPARE=ON，来开启该功能，然后需要重新编译DIPU。
+
+```shell
+export USE_AUTOCOMPARE=ON
 ```
 
 以上方法是对所有算子开启自动精度对比。如果只需要对特定算子做精度对比，也可只给需要的算子做精度对比，只需要在相关的配置文件（如 `dipu/scripts/autogen_diopi_wrapper/diopi_functions.yaml`）给相应的算子添加 `autocompare: True` 即可。

dipu/SupportedDiopiFunctions.txt

@@ -218,6 +218,7 @@ diopiSigmoidBackward
 diopiSign
 diopiSilu
 diopiSiluBackward
+diopiSiluInp
 diopiSin
 diopiSinInp
 diopiSmoothL1Loss

dipu/scripts/autogen_diopi_wrapper/autogen_diopi_wrapper.py

@@ -114,36 +114,41 @@ def create_return_code_frome_schema(schema, allow_return_ref=True):
 def create_transform_input_to_cpu_code(fun_config):
     input_process_code = ""
     schema = fun_config["schema"]
+    opname = get_op_name_from_schema(schema)
     inputs = re.findall("Tensor +([\w\d_]+)", schema[: schema.find("->")])
     for input in inputs:
         input_process_code += (
-            f"at::Tensor {input}_cpu = to_cpu_without_diopi({input});\n"
+            f"at::Tensor {input}_cpu = toCpuTensorWithoutDiopiCopy({input});\n"
         )
 
     optional_inputs = re.findall("Tensor *\? +([\w\d_]+)", schema[: schema.find("->")])
     for input in optional_inputs:
-        input_process_code += f"\nc10::optional<at::Tensor> {input}_cpu = {input}.has_value() && {input}.value().defined() ? c10::make_optional<at::Tensor>(to_cpu_without_diopi({input}.value())) : {input};\n"
+        input_process_code += f"\nc10::optional<at::Tensor> {input}_cpu = {input}.has_value() && {input}.value().defined() ? c10::make_optional<at::Tensor>(toCpuTensorWithoutDiopiCopy({input}.value())) : {input};\n"
 
     optional_tensor_list_inputs = re.findall(
         "Tensor *\? *\[ *\] +([\w\d_]+)", schema[: schema.find("->")]
     )
     for input in optional_tensor_list_inputs:
         input_process_code += f"\nc10::List<c10::optional<at::Tensor>> {input}_cpu;\n"
         input_process_code += f"for (int i = 0; i < {input}.size();++i)" + " {\n"
-        input_process_code += f"  {input}_cpu.push_back({input}[i].has_value() && {input}[i].value().defined() ? c10::make_optional<at::Tensor>(({input}[i].value())) : {input}[i]);\n"
+        input_process_code += f"  {input}_cpu.push_back({input}[i].has_value() && {input}[i].value().defined() ? c10::make_optional<at::Tensor>(toCpuTensorWithoutDiopiCopy({input}[i].value())) : {input}[i]);\n"
         input_process_code += "}\n"
 
     outputs = re.findall(
         "Tensor\([a-z]!\)[ ]+([\w\d_]+){1}", schema[: schema.find("->")]
     )
     for output in outputs:
-        if output.strip().endswith("?"):
-            output = output.replace("?", "")
-            input_process_code += f"\nc10::optional<at::Tensor> {output}_cpu = {output}.has_value() && {output}.value().defined() ? c10::make_optional<at::Tensor>(to_cpu_without_diopi({output}.value()) : {output};\n"
-        else:
-            input_process_code += (
-                f"at::Tensor {output}_cpu = to_cpu_without_diopi({output});\n"
-            )
+        input_process_code += (
+            f"at::Tensor {output}_cpu = toCpuTensorWithoutDiopiCopy({output});\n"
+        )
+        if ".out" in opname or "_out" in opname:
+            for i in range(len(inputs)):
+                input_process_code += (
+                    f"if (({inputs[i]}.data_ptr()) == {output}.data_ptr())"
+                )
+                input_process_code += "{\n\t"
+                input_process_code += f"{inputs[i]}_cpu = {output}_cpu;\n\t"
+                input_process_code += "}\n"
 
     tensors_arrays = re.findall(
         "Tensor *\[ *\] * +([\w\d_]+)", schema[: schema.find("->")]
@@ -161,9 +166,8 @@ def create_transform_input_to_cpu_code(fun_config):
             )
             input_process_code += (
                 f"std::transform({tensors_arg}.begin(), {tensors_arg}.end(), {tensors_arg}_cpu.begin(), [](const at::Tensor& tensor)"
-                + "{return to_cpu_without_diopi(tensor);});\n"
+                + "{return toCpuTensorWithoutDiopiCopy(tensor);});\n"
             )
-
     return input_process_code
 
 
@@ -487,6 +491,9 @@ def create_call_aten_cpu_cpp_function_code_from_config(fun_config):
             code,
         )
 
+    if "device" in code:
+        code = code.replace("device", "at::kCPU")
+
     inputs = re.findall("Tensor +([\w\d_]+)", schema[: schema.find("->")])
     optional_inputs = re.findall("Tensor *\? +([\w\d_]+)", schema[: schema.find("->")])
     outputs = re.findall(
@@ -550,7 +557,6 @@ def create_result_compare_code(fun_config):
     for i in range(len(inputs)):
         code += separator_code
         code += f'std::cout << "autocompare:\t{op_name}\t{inputs[i]}: " << std::endl << allclose_autocompare({inputs[i]}_cpu, {inputs[i]}) << std::endl;\n'
-
     return code
 
 
@@ -972,9 +978,12 @@ def functions_code_gen(fun_config):
 
 
 def boolean_string(s):
-    if s not in {"False", "True"}:
-        raise ValueError("Not a valid boolean string")
-    return s == "True"
+    if s.lower() in ["true", "on"]:
+        return True
+    elif s.lower() in ["false", "off"]:
+        return False
+    else:
+        raise ValueError("Not a valid boolean string.")
 
 
 def parse_args():

dipu/scripts/autogen_diopi_wrapper/autogen_wrapped_code.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+DIPU_DIR=$(dirname $(dirname $(dirname "$0")))
+USE_AUTOCOMPARE="$1"
+UsedVendor="$2"
+Torch_VERSION="$3"
+
+GENERATED_KERNELS=${DIPU_DIR}/torch_dipu/csrc_dipu/aten/ops/AutoGenedKernels.cpp
+GENERATED_KERNELS_SCRIPT=${DIPU_DIR}/scripts/autogen_diopi_wrapper/autogen_diopi_wrapper.py
+GENERATED_KERNELS_CONFIG=${DIPU_DIR}/scripts/autogen_diopi_wrapper/diopi_functions.yaml
+GENERATED_KERNELS_VENDOR=${DIPU_DIR}/third_party/DIOPI/impl/${UsedVendor}/convert_config.yaml
+
+PYTHON_CMD="python3 ${GENERATED_KERNELS_SCRIPT} --out=${GENERATED_KERNELS} --config=${GENERATED_KERNELS_CONFIG} \
+    --autocompare=${USE_AUTOCOMPARE} --print_op_arg=True --use_diopi_adapter=False --print_func_call_info=True \
+    --fun_config_dict='{\"current_device\":\"${UsedVendor}\",\"current_torch_ver\":\"${Torch_VERSION}\"}'"
+
+if [ -f "$GENERATED_KERNELS_VENDOR" ]; then
+    PYTHON_CMD="$PYTHON_CMD --convert_config=${GENERATED_KERNELS_VENDOR}"
+fi
+
+eval "$PYTHON_CMD"

dipu/scripts/autogen_diopi_wrapper/diopi_functions.yaml

@@ -1648,6 +1648,9 @@
       auto out = nodispatch::empty_like(self);
   interface: diopiSiluBackward(ctx, out, grad_output, self)
 
+- schema: "silu_(Tensor(a!) self) -> Tensor(a!)"
+  interface: diopiSiluInp(ctx, self)
+
 - schema: "reciprocal.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"
   interface: diopiReciprocal(ctx, out, self)
 

dipu/scripts/ci/ascend/ci_ascend_script.sh

@@ -10,47 +10,53 @@ function build_diopi_lib() {
 }
 
 function config_dipu_ascend_cmake() {
-    mkdir -p build && cd ./build && rm -rf ./*
-    cmake ../  -DCMAKE_BUILD_TYPE=Release \
-        -DDEVICE=ascend \
-        -DWITH_DIOPI_LIBRARY=DISABLE
+    mkdir -p build && cd ./build
+    cmake_args="-DCMAKE_BUILD_TYPE=Release -DDEVICE=ascend -DWITH_DIOPI_LIBRARY=DISABLE"
+    if [ -n "$USE_AUTOCOMPARE" ]; then
+        cmake_args+=" -DUSE_AUTOCOMPARE=${USE_AUTOCOMPARE}"
+    fi
+    cmake ../ $cmake_args
     cd ../
 }
 
 function config_all_ascend_cmake() {
-    mkdir -p build && cd ./build && rm -rf ./*
-    cmake ../  -DCMAKE_BUILD_TYPE=Release \
-        -DDEVICE=ascend \
-        -DENABLE_COVERAGE=${USE_COVERAGE} \
-        -DWITH_DIOPI=INTERNAL
+    mkdir -p build && cd ./build
+    cmake_args="-DCMAKE_BUILD_TYPE=Release -DDEVICE=ascend -DENABLE_COVERAGE=${USE_COVERAGE} -DWITH_DIOPI=INTERNAL"
+    if [ -n "$USE_AUTOCOMPARE" ]; then
+        cmake_args+=" -DUSE_AUTOCOMPARE=${USE_AUTOCOMPARE}"
+    fi
+    cmake ../ $cmake_args
     cd ../
 }
 
 function build_dipu_without_diopi() {
     echo "building dipu_lib without diopi:$(pwd)"
     config_dipu_ascend_cmake 2>&1 | tee ./build1.log
-    cd build && make -j8  2>&1 | tee ./build1.log &&  cd ..
+    cd build && make -j8 2>&1 | tee ./build1.log && cd ..
 }
 
 function build_all() {
     echo "building dipu_lib:$(pwd)"
-    echo  "DIOPI_ROOT:${DIOPI_ROOT}"
+    echo "DIOPI_ROOT:${DIOPI_ROOT}"
     config_all_ascend_cmake 2>&1 | tee ./build1.log
-    cd build && make -j8  2>&1 | tee ./build1.log &&  cd ..
+    cd build && make -j8 2>&1 | tee ./build1.log && cd ..
 }
 
 case $1 in
-    build_dipu)
-        (
-            build_all
-        ) \
-        || exit -1;;
-    build_dipu_without_diopi)
-        (
-            build_dipu_without_diopi
-        ) \
-        || exit -1;;
-    *)
-        echo -e "[ERROR] Incorrect option:" $1;
+build_dipu)
+    (
+        build_all
+    ) ||
+        exit -1
+    ;;
+build_dipu_without_diopi)
+    (
+        build_dipu_without_diopi
+    ) ||
+        exit -1
+    ;;
+*)
+    echo -e "[ERROR] Incorrect option:" $1
+    ;;
 esac
 exit 0

dipu/scripts/op_capture/op_capture.py

@@ -48,6 +48,7 @@ def extract_op_arg(op):
         name = arg[0:index].strip()
         attrs = arg[index + 1 :]
         attrs = re.sub(", *data_ptr: 0x[\da-f]+", "", attrs)
+        attrs = re.sub(", *storage_data_ptr: 0x[\da-f]+", "", attrs)
         args_info.append(name + ":[" + attrs + "] ")
 
     return args_info

dipu/third_party/DIOPI

@@ -1,4 +1,5 @@
 #[[ Dependencies ]]
+option(USE_AUTOCOMPARE "whether to use USE_AUTOCOMPARE" OFF)
 
 # Import Python3::Python, Python3_EXECUTABLE
 # Also see https://cmake.org/cmake/help/latest/module/FindPython3.html
@@ -43,9 +44,8 @@ endif()
 
 # Auto generated code.
 # Consider moving it to other files.
+set(AUTOGEN_CODE_SH "${PROJECT_SOURCE_DIR}/scripts/autogen_diopi_wrapper/autogen_wrapped_code.sh")
 set(GENERATED_KERNELS "${CMAKE_CURRENT_SOURCE_DIR}/aten/ops/AutoGenedKernels.cpp")
-set(GENERATED_KERNELS_SCRIPT "${PROJECT_SOURCE_DIR}/scripts/autogen_diopi_wrapper/autogen_diopi_wrapper.py")
-set(GENERATED_KERNELS_CONFIG "${PROJECT_SOURCE_DIR}/scripts/autogen_diopi_wrapper/diopi_functions.yaml")
 set(GENERATED_KERNELS_VENDOR "${PROJECT_SOURCE_DIR}/third_party/DIOPI/impl/${UsedVendor}/convert_config.yaml")
 
 if(NOT EXISTS "${GENERATED_KERNELS_VENDOR}")
@@ -54,21 +54,13 @@ endif()
 
 add_custom_command(
   OUTPUT "${GENERATED_KERNELS}"
-  COMMAND "${Python3_EXECUTABLE}"
-    "${GENERATED_KERNELS_SCRIPT}"
-    "--out=${GENERATED_KERNELS}"
-    "--config=${GENERATED_KERNELS_CONFIG}"
-    "--autocompare=False"
-    "--print_op_arg=True"
-    "--use_diopi_adapter=False"
-    "--print_func_call_info=True"
-    "--fun_config_dict='{\"current_device\":\"${UsedVendor}\",\"current_torch_ver\":\"${Torch_VERSION}\"}'"
-    "$<$<BOOL:${GENERATED_KERNELS_VENDOR}>:--convert_config=${GENERATED_KERNELS_VENDOR}>"
+  COMMAND bash -c "${AUTOGEN_CODE_SH} ${USE_AUTOCOMPARE} ${UsedVendor} ${Torch_VERSION}"
+  COMMENT "Generating ${GENERATED_KERNELS}$<$<BOOL:${GENERATED_KERNELS_VENDOR}>: with ${GENERATED_KERNELS_VENDOR}>"
   DEPENDS
     "${GENERATED_KERNELS_SCRIPT}"
     "${GENERATED_KERNELS_CONFIG}"
-  WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
-  COMMENT "Generating ${GENERATED_KERNELS}$<$<BOOL:${GENERATED_KERNELS_VENDOR}>: with ${GENERATED_KERNELS_VENDOR}>")
+    "${AUTOGEN_CODE_SH}"
+  )
 
 # Collect source files.
 # You may use: find . -name "*.cpp" to update list.

dipu/torch_dipu/csrc_dipu/CMakeLists.txt

@@ -42,6 +42,7 @@ inline at::Tensor toCpuTensorWithoutDiopiCopy(const at::Tensor& in) {
   at::Tensor out = at::empty_strided(in.sizes(), in.strides(),
                                      in.options().device(c10::Device("cpu")));
   if (in.nbytes() > 0) {
+    dipu::getCurrentDIPUStream().synchronize();
     dipu::devapis::memCopyD2H(out.storage().nbytes(), out.data_ptr(),
                               in.data_ptr());
   }

dipu/torch_dipu/csrc_dipu/aten/ops/OpUtils.hpp

@@ -136,6 +136,8 @@ EventStatus getEventStatus(deviceEvent_t event) {
     ::tangGetLastError(); /* reset internal error state*/
     return devapis::EventStatus::PENDING;
   } else {
+    printf("call a tangrt function (tangEventQuery) failed. return code=%d",
+           ret);
     throw std::runtime_error("dipu device error");
   }
 }

dipu/torch_dipu/csrc_dipu/vendor/droplet/deviceimpl.cpp

@@ -29,8 +29,6 @@
 # as "gpu" or "cuda" (mock cuda is another problem)
 # only partially aligned now,
 __all__ = [
-    # resume initialize flag after random generator ready
-    # "is_initialized",
     # device
     "can_device_access_peer",
     "current_device",
@@ -42,6 +40,7 @@
     "get_device_properties",
     "get_device_capability",
     "is_available",
+    "is_initialized",
     "set_device",
     "GetDeviceProxy",
     "GetDeviceStaticProxy",