Merge branch 'master' into shape-bucket-opt

README.md

@@ -2,8 +2,9 @@
 <img src="docs/logo.png" width="400" alt="nncase" />
 </div>
 
-[![License](https://img.shields.io/badge/license-Apache%202-blue)](https://raw.githubusercontent.com/kendryte/nncase/master/LICENSE)
-[![compiler-build](https://github.com/kendryte/nncase/actions/workflows/compiler-build.yml/badge.svg)](https://github.com/kendryte/nncase/actions/workflows/compiler-build.yml)
+[![GitHub repository](https://img.shields.io/badge/github-repository-blue?logo=github&style=plastic)](https://github.com/kendryte/nncase)
+[![Gitee repository](https://img.shields.io/badge/gitee-repository-blue?logo=gitee&style=plastic)](https://gitee.com/kendryte/nncase)
+[![GitHub release](https://img.shields.io/github/v/release/kendryte/nncase?color=brightgreen&display_name=tag&logo=github&style=plastic)](https://github.com/kendryte/nncase/releases)
 
 `nncase` is a neural network compiler for AI accelerators.
 

conanfile.py

@@ -51,6 +51,7 @@ def requirements(self):
         if self.options.tests:
             self.requires('gtest/1.10.0')
             self.requires('ortki/0.0.2')
+            self.requires('rapidjson/1.1.x')
 
         if self.options.python:
             self.requires('pybind11/2.6.1')

pyproject.toml

@@ -24,7 +24,8 @@ dependencies = ["numpy"]
 homepage = "https://github.com/kendryte/nncase"
 
 [build-system]
-requires = ["setuptools>=42", "wheel", "conan<=1.59", "ninja"]
+requires = ["setuptools>=42", "wheel", "conan<=1.59", "ninja", "gitpython"]
+build-backend = "setuptools.build_meta"
 
 [tool.cibuildwheel]
 build = ["cp37*", "cp38*", "cp39*", "cp310*"]

requirements.test.txt

@@ -17,4 +17,4 @@ pytest-xdist
 pyyaml
 pythonnet==3.0.1
 clr_loader==0.2.4
-toml==0.10.2
+toml==0.10.2

setup.py

@@ -12,7 +12,8 @@
 import io
 import re
 import time
-
+import subprocess
+from git.repo import Repo
 # See ref: https://stackoverflow.com/a/51575996
 
 
@@ -277,8 +278,19 @@ def find_version():
     version_prefix = re.findall(r"NNCASE_VERSION \"(.+)\"", version_file)
 
     if version_prefix:
+        repo_path = os.getcwd()
+        repo = Repo(repo_path)
+        if repo.tags:
+            latest_commit = subprocess.check_output(
+                ['git', 'rev-parse', 'HEAD']).decode('utf-8').strip()
+            tagged_commit = subprocess.check_output(
+                ['git', 'rev-list', '-n', '1', repo.tags[-1].name]).decode('utf-8').strip()
+            if latest_commit == tagged_commit:
+                return version_prefix[0]
+
         version_suffix = time.strftime("%Y%m%d", time.localtime())
         return version_prefix[0] + "." + version_suffix
+
     raise RuntimeError("Unable to find version string.")
 
 

src/Native/src/kernels/stackvm/reference/batch_to_space.cpp

@@ -104,7 +104,7 @@ dims_t infer_shape(gsl::span<const size_t> origin_in_shape,
                    gsl::span<const size_t> block_shape,
                    const paddings_t &crops) {
     auto d4 = fixed_dims(0, 2, 3, 1);
-    auto d3 = fixed_dims(1, 2, 0);
+    auto d3 = fixed_dims(0, 2, 1);
     auto inPerm = origin_in_shape.size() == 4
                       ? gsl::span<const size_t>{d4.data(), d4.size()}
                       : gsl::span<const size_t>{d3.data(), d3.size()};
@@ -123,7 +123,7 @@ dims_t infer_shape(gsl::span<const size_t> origin_in_shape,
                          in_shape.end());
     }
     auto outd4 = fixed_dims(0, 3, 1, 2);
-    auto outd3 = fixed_dims(2, 0, 1);
+    auto outd3 = fixed_dims(0, 2, 1);
     auto outPerm = origin_in_shape.size() == 4
                        ? gsl::span<const size_t>{outd4.data(), outd4.size()}
                        : gsl::span<const size_t>{outd3.data(), outd3.size()};

src/Native/src/kernels/stackvm/reference/matmul.cpp

@@ -52,8 +52,17 @@ result<void> matmul_unit_impl(const T *input_a, const T *input_b, T *output,
 
 template <typename T>
 result<void> matmul_impl(const T *input_a, const T *input_b, T *output,
-                         gsl::span<const size_t> in_a_shape,
-                         gsl::span<const size_t> in_b_shape) noexcept {
+                         gsl::span<const size_t> in_a_shape_,
+                         gsl::span<const size_t> in_b_shape_) noexcept {
+    dims_t in_a_shape = in_a_shape_;
+    dims_t in_b_shape = in_b_shape_;
+    if (in_a_shape.size() == 1) {
+        in_a_shape.insert(in_a_shape.begin(), 1);
+    }
+
+    if (in_b_shape.size() == 1) {
+        in_b_shape.insert(in_b_shape.end(), 1);
+    }
     auto new_a_shape = to_4d(in_a_shape);
     auto new_b_shape = to_4d(in_b_shape);
     auto a_unit_size = new_a_shape[2] * new_a_shape[3];

src/Native/src/kernels/stackvm/reference/reduce_arg.cpp

@@ -60,10 +60,8 @@ result<void> reduce_arg_impl(TReducer &&reducer, T init_value, const T *input,
             out_map[out_idx].clear();
             out_map[out_idx].push_back(index[axes[0]]);
             dst = src;
-        } else if constexpr (std::is_same_v<T, float>) {
-            if (fabs(src - dst) < epsilon) {
-                out_map[out_idx].push_back(index[axes[0]]);
-            }
+        } else if (std::fabs(src - dst) < epsilon) {
+            out_map[out_idx].push_back(index[axes[0]]);
         }
         return ok();
     }));

src/Native/src/kernels/stackvm/shape_infer.h

@@ -287,13 +287,24 @@ inline dims_t onehot_infer_shape(gsl::span<const size_t> indices_shape,
     return new_shape;
 }
 
-inline result<dims_t> matmul_infer_shape(gsl::span<const size_t> lhs_shape,
-                                         gsl::span<const size_t> rhs_shape) {
+inline result<dims_t> matmul_infer_shape(gsl::span<const size_t> lhs_shape_,
+                                         gsl::span<const size_t> rhs_shape_) {
+    dims_t lhs_shape = lhs_shape_;
+    dims_t rhs_shape = rhs_shape_;
+
     if (lhs_shape.size() == 2 && rhs_shape.size() == 2) {
         auto new_shape = dims_t{lhs_shape[0], rhs_shape[1]};
         return ok(new_shape);
     }
 
+    if (lhs_shape.size() == 1) {
+        lhs_shape.insert(lhs_shape.begin(), 1);
+    }
+
+    if (rhs_shape.size() == 1) {
+        rhs_shape.insert(rhs_shape.end(), 1);
+    }
+
     auto new_a_shape = runtime::to_4d(lhs_shape);
     auto new_b_shape = runtime::to_4d(rhs_shape);
     auto big_shape = std::max(lhs_shape.size(), rhs_shape.size());
@@ -304,6 +315,14 @@ inline result<dims_t> matmul_infer_shape(gsl::span<const size_t> lhs_shape,
     }
     new_shape.push_back(lhs_shape[lhs_shape.size() - 2]);
     new_shape.push_back(rhs_shape.back());
+    if (lhs_shape_.size() == 1) {
+        new_shape.erase(new_shape.begin() + big_shape - 2);
+        big_shape--;
+    }
+
+    if (rhs_shape_.size() == 1) {
+        new_shape.erase(new_shape.begin() + big_shape - 1);
+    }
     return ok(new_shape);
 }
 

src/Nncase.Core/IR/Tensors/Functional.cs

@@ -31,7 +31,7 @@ public static Expr NHWCToNCHW(Expr input)
         }
         else if (input.CheckedShape.Rank == 3)
         {
-            perm = new[] { 2, 0, 1 };
+            perm = new[] { 0, 2, 1 };
         }
         else
         {
@@ -50,7 +50,7 @@ public static Expr NCHWToNHWC(Expr input)
         }
         else if (input.CheckedShape.Rank == 3)
         {
-            perm = new[] { 1, 2, 0 };
+            perm = new[] { 0, 2, 1 };
         }
         else
         {

src/Nncase.Diagnostics/Diagnostics/CSharpPrintVisitor.cs

@@ -78,7 +78,7 @@ public override string VisitType(CallableType type) =>
     public override string VisitType(TensorType type) => type.DType switch
     {
         PrimType ptype => ptype.GetDisplayName() + (type.Shape.IsScalar ? string.Empty : type.Shape.ToString()),
-        PointerType { ElemType: PrimType etype } ptype => $"*{etype.GetDisplayName()}",
+        PointerType { ElemType: PrimType etype } => $"*{etype.GetDisplayName()}",
         ValueType => $"{type.DType.ToString()}",
         _ => throw new NotSupportedException(type.DType.GetType().Name),
     };

src/Nncase.Diagnostics/Diagnostics/ILPrintVisitor.cs

@@ -265,7 +265,7 @@ public override string VisitType(CallableType type) =>
     public override string VisitType(TensorType type) => type.DType switch
     {
         PrimType ptype => ptype.GetDisplayName() + (type.Shape.IsScalar ? string.Empty : type.Shape.ToString()),
-        PointerType { ElemType: PrimType etype } ptype => $"*{etype.GetDisplayName()}",
+        PointerType { ElemType: PrimType etype } => $"*{etype.GetDisplayName()}",
         ValueType => $"{type.DType.ToString()}",
         _ => throw new NotSupportedException(type.DType.GetType().Name),
     };

src/Nncase.Diagnostics/Diagnostics/ScriptPrintVisitor.cs

@@ -129,7 +129,7 @@ public ScriptPrintVisitor(TextWriter textWriter, bool display_callable)
     public override string VisitType(TensorType type) => type.DType switch
     {
         PrimType ptype => ptype.GetDisplayName() + (type.Shape.IsScalar ? string.Empty : type.Shape.ToString()),
-        PointerType { ElemType: PrimType etype } ptype => $"*{etype.GetDisplayName()}",
+        PointerType { ElemType: PrimType etype } => $"*{etype.GetDisplayName()}",
         ValueType vtype => vtype.GetDisplayName() + (type.Shape.IsScalar ? string.Empty : type.Shape.ToString()),
         _ => throw new NotSupportedException(type.DType.GetType().Name),
     };

src/Nncase.Evaluator/NN/BatchToSpace.cs

@@ -112,7 +112,7 @@ public Expr Visit(IShapeEvaluateContext context, BatchToSpace target)
 
         if (input.CheckedShape.Rank == 3)
         {
-            inShape = Stack(new IR.Tuple(inShape[1], inShape[2], inShape[0]), 0);
+            inShape = Stack(new IR.Tuple(inShape[0], inShape[2], inShape[1]), 0);
         }
 
         var blockShape = context.GetArgument(target, BatchToSpace.BlockShape);
@@ -142,7 +142,7 @@ public Expr Visit(IShapeEvaluateContext context, BatchToSpace target)
 
         if (input.CheckedShape.Rank == 3)
         {
-            return Stack(new IR.Tuple(outShapeList[2], outShapeList[0], outShapeList[1]), 0);
+            return Stack(new IR.Tuple(outShapeList[0], outShapeList[2], outShapeList[1]), 0);
         }
 
         throw new NotImplementedException();
@@ -186,7 +186,7 @@ private IRType Visit(ITypeInferenceContext context, BatchToSpace target, TensorT
     {
         var inShape = input.Shape.Rank == 4
             ? TypeInference.ApplyPerm(input.Shape, new[] { 0, 2, 3, 1 })
-            : TypeInference.ApplyPerm(input.Shape, new[] { 1, 2, 0 });
+            : TypeInference.ApplyPerm(input.Shape, new[] { 0, 2, 1 });
         var batch = inShape[0];
         if (context.GetArgument(target, BatchToSpace.BlockShape) is TensorConst blockShapeValue &&
             context.GetArgument(target, BatchToSpace.Crops) is TensorConst cropsValue)
@@ -211,7 +211,7 @@ private IRType Visit(ITypeInferenceContext context, BatchToSpace target, TensorT
             var outShape =
                 outShapeList.Length == 4
                 ? TypeInference.ApplyPerm(outShapeList, new[] { 0, 3, 1, 2 })
-                : TypeInference.ApplyPerm(outShapeList, new[] { 2, 0, 1 });
+                : TypeInference.ApplyPerm(outShapeList, new[] { 0, 2, 1 });
             return input with { Shape = outShape };
         }
         else

src/Nncase.Evaluator/NN/LayerNorm.cs

@@ -12,7 +12,8 @@ namespace Nncase.Evaluator.NN;
 /// <summary>
 /// Evaluator for <see cref="LayerNorm"/>.
 /// </summary>
-public class LayerNormEvaluator : IEvaluator<LayerNorm>, ITypeInferencer<LayerNorm>, ICostEvaluator<LayerNorm>, IShapeEvaluator<LayerNorm>, IMetricEvaluator<LayerNorm>
+public class LayerNormEvaluator : IEvaluator<LayerNorm>, ITypeInferencer<LayerNorm>, ICostEvaluator<LayerNorm>,
+    IShapeEvaluator<LayerNorm>, IMetricEvaluator<LayerNorm>
 {
     /// <inheritdoc/>
     public IValue Visit(IEvaluateContext context, LayerNorm layerNorm)
@@ -77,21 +78,19 @@ private Tensor LayerNormImpl(Tensor input, Tensor scale, Tensor bias, int axis,
         float[] inputArray = input.ToArray<float>();
         float[] outputArray = new float[inputArray.Length];
         int[] inShape = input.Shape.ToValueArray();
+        if (axis < 0)
+        {
+            axis += inShape.Length;
+        }
+
         for (int i = 0; i < axis; i++)
         {
             outputSize *= inShape[i];
         }
 
         for (int i = axis; i < inShape.Length; i++)
         {
-            if (i < 0)
-            {
-                innerSize *= inShape[^System.Math.Abs(i)];
-            }
-            else
-            {
-                innerSize *= inShape[i];
-            }
+            innerSize *= inShape[i];
         }
 
         for (int batch = 0; batch < outputSize; batch++)
@@ -131,7 +130,8 @@ private Tensor LayerNormImpl(Tensor input, Tensor scale, Tensor bias, int axis,
 
             for (int i = 0; i < innerSize; i++)
             {
-                outputArray[(i + (batch * innerSize)) % outputArray.Length] = (div[i] * scale.ToArray<float>()[i % scale.Length]) + bias.ToArray<float>()[i % bias.Length];
+                outputArray[(i + (batch * innerSize)) % outputArray.Length] =
+                    (div[i] * scale.ToArray<float>()[i % scale.Length]) + bias.ToArray<float>()[i % bias.Length];
             }
         }