qemu system mode (#1201)

* close non-equal split

* fix typeinfer of conv2d

* fix fold binary

* do not swap binary args

* update setup.py

* do not pre-preprocess onnx with external data

* imgnore .mono in git

* restore SwapBinaryArgs

* fix onnx test runner

---------

Co-authored-by: sunnycase <sunnycase@live.cn>

.gitignore

@@ -307,3 +307,6 @@ cmake-build-*
 *.ipynb_checkpoints*
 # Auto generated files
 # generated/
+
+.mono/
+.history/

conanfile.py

@@ -54,7 +54,7 @@ class nncaseConan(ConanFile):
             self.requires('rapidjson/1.1.x')
 
         if self.options.python:
-            self.requires('pybind11/2.6.1')
+            self.requires('pybind11/2.12.0')
 
         if not self.options.runtime:
             self.requires('abseil/20220623.1')

setup.py

@@ -83,8 +83,7 @@ class InstallCMakeLibs(install_lib):
                 os.walk(os.path.join(bin_dir, 'sharplibs')) for _lib in files if
                 os.path.isfile(os.path.join(root, _lib)) and
                 (os.path.splitext(_lib)[-1] in [".dll", ".so", ".dylib", ".json"] or
-                _lib.startswith("lib"))
-                and not _lib.endswith(".deps.json")]
+                _lib.startswith("lib"))]
 
         for lib in sharp_libs:
             shutil.move(lib, os.path.join(self.build_dir,
@@ -204,7 +203,7 @@ class BuildCMakeExt(build_ext):
             extdir += os.path.sep
 
         bin_dir = os.path.abspath(os.path.join(self.build_temp, 'install'))
-        cmake_args = ['-G', 'Ninja', '-DDOTNET_INIT_FOR_CONFIG=ON']
+        cmake_args = ['-G', 'Ninja', '-DDOTNET_INIT_FOR_CONFIG=OFF']
         if platform.system() == 'Windows':
             cmake_args += ['-DCMAKE_C_COMPILER=clang-cl']
             cmake_args += ['-DCMAKE_CXX_COMPILER=clang-cl']

src/Native/include/nncase/runtime/char_array_buffer.h

@@ -24,21 +24,21 @@ class char_array_buffer : public std::streambuf {
         : begin_(data.begin()), end_(data.end()), current_(data.data()) {}
 
   private:
-    int_type underflow() {
+    int_type underflow() override {
         if (current_ == end_)
             return traits_type::eof();
 
         return traits_type::to_int_type(*current_);
     }
 
-    int_type uflow() {
+    int_type uflow() override {
         if (current_ == end_)
             return traits_type::eof();
 
         return traits_type::to_int_type(*current_++);
     }
 
-    int_type pbackfail(int_type ch) {
+    int_type pbackfail(int_type ch) override {
         if (current_ == begin_ ||
             (ch != traits_type::eof() && ch != current_[-1]))
             return traits_type::eof();
@@ -46,13 +46,14 @@ class char_array_buffer : public std::streambuf {
         return traits_type::to_int_type(*--current_);
     }
 
-    std::streamsize showmanyc() {
+    std::streamsize showmanyc() override {
         assert(std::less_equal<const char *>()(current_, end_));
         return end_ - current_;
     }
 
-    std::streampos seekoff(std::streamoff off, std::ios_base::seekdir way,
-                           [[maybe_unused]] std::ios_base::openmode which) {
+    std::streampos
+    seekoff(std::streamoff off, std::ios_base::seekdir way,
+            [[maybe_unused]] std::ios_base::openmode which) override {
         if (way == std::ios_base::beg) {
             current_ = begin_ + off;
         } else if (way == std::ios_base::cur) {
@@ -67,8 +68,9 @@ class char_array_buffer : public std::streambuf {
         return current_ - begin_;
     }
 
-    std::streampos seekpos(std::streampos sp,
-                           [[maybe_unused]] std::ios_base::openmode which) {
+    std::streampos
+    seekpos(std::streampos sp,
+            [[maybe_unused]] std::ios_base::openmode which) override {
         current_ = begin_ + sp;
 
         if (current_ < begin_ || current_ > end_)
@@ -77,6 +79,17 @@ class char_array_buffer : public std::streambuf {
         return current_ - begin_;
     }
 
+    std::streamsize xsgetn(char_type *s, std::streamsize count) override {
+        std::streamsize available =
+            static_cast<std::streamsize>(end_ - current_);
+        std::streamsize n = (count > available) ? available : count;
+        if (n > 0) {
+            traits_type::copy(s, current_, static_cast<size_t>(n));
+            current_ += n;
+        }
+        return n;
+    }
+
     const char *const begin_;
     const char *const end_;
     const char *current_;

src/Nncase.Core/Utilities/DistributedUtility.cs

@@ -17,7 +17,7 @@ public static class DistributedUtility
             var ndsbp = new List<SBP>();
             for (int axis = 0; axis < tensorType.Shape.Rank; axis++)
             {
-                if (tensorType.Shape[axis] is { IsFixed: true, Value: int s } && IsDivideBy(s, placement.Hierarchy[i]))
+                if (tensorType.Shape[axis] is { IsFixed: true, Value: int s } && IsDivideExactly(s, placement.Hierarchy[i]))
                 {
                     ndsbp.Add(SBP.S(axis));
                 }
@@ -50,7 +50,7 @@ public static class DistributedUtility
                 candidateNdsbps[i].Add(SBP.B);
                 for (int axis = 0; axis < tensorType.Shape.Rank; axis++)
                 {
-                    if (tensorType.Shape[axis] is { IsFixed: true, Value: int s } && IsDivideBy(s, placement.Hierarchy[i]) && !innerSplitedAxes.Contains(axis))
+                    if (tensorType.Shape[axis] is { IsFixed: true, Value: int s } && IsDivideExactly(s, placement.Hierarchy[i]) && !innerSplitedAxes.Contains(axis))
                     {
                         candidateNdsbps[i].Add(SBP.S(axis));
                     }
@@ -73,7 +73,7 @@ public static class DistributedUtility
         }
 
         var divisors = GetDivisors(new DistributedType(tensorType, new IRArray<SBP>(ndsbp.ToArray()), placement));
-        return divisors.Select((d, axis) => (d, axis)).All(p => p.d == 0 ? true : IsDivideBy(tensorType.Shape[p.axis].FixedValue, p.d));
+        return divisors.Select((d, axis) => (d, axis)).All(p => p.d == 0 ? true : IsDivideExactly(tensorType.Shape[p.axis].FixedValue, p.d));
     }
 
     public static IReadOnlyList<int> GetDivisors(DistributedType distributedType)

src/Nncase.Evaluator/NN/Conv2D.cs

@@ -129,7 +129,7 @@ public class Conv2DEvaluator : IEvaluator<Conv2D>, ITypeInferencer<Conv2D>, ICos
             return new InvalidType(string.Empty);
         }
 
-        if (input.Placement != weights.Placement)
+        if (input.Placement != weights.Placement || input.Placement != bias.Placement)
         {
             return new InvalidType("placement not equal");
         }

tests/onnx_test_runner.py

@@ -58,11 +58,13 @@ class OnnxTestRunner(TestRunner):
         elif model_file.startswith('onnx-models'):
             model_file = os.path.join(os.getenv('ONNX_MODELS_DIR'),
                                       model_file[len('onnx-models/'):])
+        has_external_data = False
         if self.case_dir != os.path.dirname(model_file):
             new_file = os.path.join(self.case_dir, 'test.onnx')
             shutil.copy(model_file, new_file)
             for tensor in external_data_helper._get_all_tensors(onnx.load(model_file, load_external_data=False)):
                 if external_data_helper.uses_external_data(tensor):
+                    has_external_data = True
                     info = external_data_helper.ExternalDataInfo(tensor)
                     file_location = external_data_helper._sanitize_path(info.location)
                     external_data_src_path = os.path.join(
@@ -76,7 +78,8 @@ class OnnxTestRunner(TestRunner):
         if not self.inputs:
             self.parse_model(model_file)
 
-        model_file = self.do_preprocess(model_file)
+        if not has_external_data:
+            model_file = self.do_preprocess(model_file)
 
         super().run(model_file)