add akg load_im2col

add load_im2col

add load_im2col

add load_im2col

mindspore/ccsrc/backend/kernel_compiler/kernel_query.cc

@@ -112,6 +112,12 @@ void KernelQuery(const CNodePtr &kernel_node, std::vector<std::shared_ptr<kernel
   if (IsPrimitiveCNode(kernel_node, kPrimProdForceSeA)) {
     kernel_type = KernelType::AKG_KERNEL;
   }
+
+  const PrimitivePtr kPrimLoadIm2Col = std::make_shared<Primitive>("LoadIm2Col");
+  if (IsPrimitiveCNode(kernel_node, kPrimLoadIm2Col)) {
+    kernel_type = KernelType::AKG_KERNEL;
+  }  // use LoadIm2Col only for THOR optimizer
+
   switch (kernel_type) {
     case KernelType::AKG_KERNEL:
       AkgMetadataInfo(kernel_node, kernel_info_list);

mindspore/ops/_op_impl/akg/ascend/__init__.py

@@ -44,5 +44,6 @@ from .sqrt import _sqrt_akg
 from .square import _square_akg
 from .sub import _sub_akg
 from .prod_force_se_a import _prod_force_se_a_akg
+from .load_im2col import _load_im2col_akg
 
 # Please insert op register in lexicographical order of the filename.

mindspore/ops/_op_impl/akg/ascend/load_im2col.py

@@ -0,0 +1,33 @@
+# Copyright 2021 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""LoadIm2Col op"""
+from mindspore.ops.op_info_register import op_info_register, AkgAscendRegOp, DataType as DT
+
+op_info = AkgAscendRegOp("LoadIm2Col") \
+    .fusion_type("OPAQUE") \
+    .input(0, "x") \
+    .output(0, "output") \
+    .attr("ksizes", "required", "listInt") \
+    .attr("strides", "required", "listInt") \
+    .dtype_format(DT.F16_5HD, DT.F16_Default) \
+    .dtype_format(DT.F32_5HD, DT.F32_Default) \
+    .get_op_info()
+
+
+@op_info_register(op_info)
+def _load_im2col_akg():
+    """im2col Akg register"""
+    return

mindspore/ops/operations/__init__.py

@@ -96,7 +96,8 @@ from .other_ops import (Assign, InplaceAssign, IOU, BoundingBoxDecode, BoundingB
 from ._thor_ops import (CusBatchMatMul, CusCholeskyTrsm, CusFusedAbsMax1, CusImg2Col, CusMatMulCubeDenseLeft,
                         CusMatMulCubeFraczRightMul, CusMatMulCube, CusMatrixCombine, CusTranspose02314,
                         CusMatMulCubeDenseRight,
-                        CusMatMulCubeFraczLeftCast, Im2Col, UpdateThorGradient, Cholesky, CholeskyTrsm, DetTriangle,
+                        CusMatMulCubeFraczLeftCast, Im2Col, LoadIm2Col, UpdateThorGradient, Cholesky, CholeskyTrsm,
+                        DetTriangle,
                         ProdForceSeA)
 from .sparse_ops import (SparseToDense, SparseTensorDenseMatmul)
 from ._embedding_cache_ops import (CacheSwapTable, UpdateCache, MapCacheIdx, SubAndFilter,

mindspore/ops/operations/_thor_ops.py

@@ -31,6 +31,7 @@ __all__ = ["CusBatchMatMul",
            "CusTranspose02314",
            "CusMatMulCubeDenseRight",
            "CusMatMulCubeFraczLeftCast",
+           "LoadIm2Col"
            ]
 
 
@@ -362,6 +363,7 @@ class CusTranspose02314(PrimitiveWithInfer):
 
     def get_bprop(self):
         """Get backprop for CusTranspose02314."""
+
         def bprop(x, out, dout):
             return (C.zeros_like(x),)
 
@@ -529,6 +531,55 @@ class Im2Col(PrimitiveWithInfer):
         return x_dtype
 
 
+class LoadIm2Col(PrimitiveWithInfer):
+    """
+    extracts image patches from image.
+
+    The rank of input_x1 must be `4`, data_format is "NCHW".
+    Only supports when C is divisible by 16.
+
+    Inputs:
+        - **input_x1** (Tensor) - The feature map.
+          The shape of the tensor is :math:`(N, C, H, W)`.
+    Outputs:
+        Tensor.
+    Examples:
+        >>> input_x = Tensor(np.random.rand(32, 16, 224, 224).astype(np.float16))
+        >>> img2col = ops.LoadIm2Col(kernel_size=(7,7), stride=(2,2))
+        >>> output = img2col(input_x)
+    """
+
+    @prim_attr_register
+    def __init__(self,
+                 ksizes,
+                 strides,
+                 pad_mode="same",
+                 dilates=(1, 1, 1, 1)):
+        """Initialize LoadIm2Col"""
+
+        self.init_prim_io_names(inputs=['x1'], outputs=['y'])
+        self.ksizes = ksizes
+        self.strides = strides
+        self.pad_mode = validator.check_string(pad_mode, ['same'], 'pad_mode', self.name)
+        self.dilation = dilates
+
+    def infer_shape(self, data1_shape):
+        bs, c, h, w = data1_shape
+        stride_h, stride_w = self.strides
+        k_w, k_h = self.ksizes
+        h_out = math.ceil(h / stride_h)
+        w_out = math.ceil(w / stride_w)
+        m = h_out * w_out
+        if m % 16 != 0:
+            shape = [(bs * m) // 16, (c * k_h * k_w) // 16, 16, 16]
+        else:
+            shape = [bs, m // 16, (c * k_h * k_w) // 16, 16, 16]
+        return shape
+
+    def infer_dtype(self, data1_dtype):
+        return data1_dtype
+
+
 class UpdateThorGradient(PrimitiveWithInfer):
     """
     Updates Thor Gradient with Approximate Fisher info matrix(for GPU backend).