!17626 matmul to mul

Merge pull request !17626 from lingyunli63/matmul_to_mul

mindspore/_extends/graph_kernel/expanders/__init__.py

@@ -53,3 +53,4 @@ from .softmax_grad_ext import SoftmaxGradExt
 from .square_sum_v1 import SquareSumV1
 from .fused_mul_add import FusedMulAdd
 from .conv2d import Conv2D
+from .matmul import MatMul, BatchMatMul

mindspore/_extends/graph_kernel/expanders/matmul.py

@@ -0,0 +1,74 @@
+# 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.
+# ===========================================================================
+"""generate json desc for BatchMatMul and MatMul"""
+from mindspore._extends.graph_kernel.model.model import DataFormat as DF
+from mindspore._extends.graph_kernel.model.model import GraphKernelUnsupportedException as GKException
+from ._utils import Expander, ExpanderInfoValidator as VLD
+
+@VLD.check_attrs('transpose_a', 'transpose_b', 'left_format', 'right_format')
+class MatMul(Expander):
+    """
+    MatMul expander
+    """
+
+    def __init__(self, expand_info):
+        super().__init__(expand_info)
+        self.transpose_a = self.attrs['transpose_a']
+        self.transpose_b = self.attrs['transpose_b']
+        self.left_format = self.attrs['left_format']
+        self.right_format = self.attrs['right_format']
+        self.shape_a = self.inputs[0]['shape']
+        self.shape_b = self.inputs[1]['shape']
+
+    def _optimize_to_mul(self):
+        """check if matmul can be replace by mul"""
+        if self.left_format != DF.DEFAULT or self.right_format != DF.DEFAULT:
+            return False
+        k_a = self.shape_a[-2] if self.transpose_a else self.shape_a[-1]
+        k_b = self.shape_b[-1] if self.transpose_b else self.shape_b[-2]
+        if k_a != 1 or k_b != 1:
+            return False
+        return True
+
+    def _check(self):
+        input_num = len(self.inputs)
+        if input_num < 2:
+            raise GKException("matul inputs number should bigger than 1, but got {}.".format(input_num))
+
+    def _trans_shape(self, shape):
+        trans_shape = list(shape)
+        trans_shape[-2] = shape[-1]
+        trans_shape[-1] = shape[-2]
+        return trans_shape
+
+    def _expand(self, graph_builder):
+        if not self._optimize_to_mul():
+            raise GKException("MatMul/BatchMatMul do not need to be replaced by Mul")
+        #Matmul is replaced by Mul([b m k], [b k n]) when k==1
+        input_a = self.inputs[0]
+        input_b = self.inputs[1]
+        if self.transpose_a:
+            shape_a_trans = self._trans_shape(self.shape_a)
+            input_a = graph_builder.emit('Reshape', [input_a], attrs={'shape': shape_a_trans})
+        if self.transpose_b:
+            shape_b_trans = self._trans_shape(self.shape_b)
+            input_b = graph_builder.emit('Reshape', [input_b], attrs={'shape': shape_b_trans})
+        result = graph_builder.emit('Mul', [input_a, input_b])
+        if 'dst_type' in self.attrs and self.inputs[0].dtype != self.attrs['dst_type']:
+            result = graph_builder.emit('Cast', [result], attrs={'dst_type': self.attrs['dst_type']})
+        return result
+
+class BatchMatMul(MatMul):
+    """BatchMatMul expander"""

mindspore/_extends/graph_kernel/model/op_infer.py

@@ -16,7 +16,6 @@
 
 import copy
 import sys
-from functools import reduce
 from .model import GraphKernelUnsupportedException as GKException
 from .model import PrimLib, DataFormat as DF
 
@@ -102,19 +101,60 @@ class OpInfer:
 class _Elemwise(OpInfer):
     """Common infer for elementwise operators"""
 
-    def _infer_shape(self):
-        """returns the input shape with largest flatten size"""
-        shape = (1,)
-        max_flatten_size = 1
-        for t in self.inputs:
-            if t.data_format != DF.DEFAULT:
-                return t.shape
-            flatten_size = reduce(lambda x, y: x * y, t.shape)
-            if flatten_size > max_flatten_size or (flatten_size == max_flatten_size and len(t.shape) > len(shape)):
-                max_flatten_size = flatten_size
-                shape = t.shape
+    def _broadcast_shape(self, shapes):
+        """deduce broadcast shape using same rules as numpy"""
+        dim_size = max([len(shape) for shape in shapes])
+        align_shapes = [[1] * (dim_size - len(shape)) + shape for shape in shapes]
+        out_shape = [1] * dim_size
+        for i in range(dim_size):
+            for align_shape in align_shapes:
+                if align_shape[i] > 1:
+                    if out_shape[i] == 1:
+                        out_shape[i] = align_shape[i]
+                    if out_shape[i] != align_shape[i]:
+                        raise GKException("shape broadcast failed!")
+        return out_shape
+
+    def _to_nz(self, default_shape):
+        """default format shape to fractal_Nz format shape"""
+        if len(default_shape) not in (1, 2):
+            raise GKException("shape is too long!")
+        # (32) or (1, 32) -> (2, 1, 1, 16)
+        if len(default_shape) == 1 or (len(default_shape) == 2 and default_shape[0] == 1):
+            shape = [default_shape[-1] // 16, 1, 1, 16]
+            if default_shape[-1] % 16 != 0:
+                raise GKException("should be multiplies of 16")
+            return shape
+        #(32, 1) -> (1, 2, 16, 1)
+        if len(default_shape) == 2 and default_shape[1] == 1:
+            shape = [1, default_shape[0] // 16, 16, 1]
+            if default_shape[0] % 16 != 0:
+                raise GKException("should be multiples of 16")
+            return shape
+        #(32, 48) -> (3, 2, 16, 16)
+        shape = [default_shape[1] // 16, default_shape[0] // 16, 16, 16]
+        if default_shape[0] % 16 != 0 or defautl_shape[1] % 16 != 0:
+            raise GKException("should be multiples of 16")
         return shape
 
+    def _infer_shape(self):
+        """returns the output shape with broadcast"""
+
+        # in case all inputs are default format/NHWC/NCHW
+        is_default = [input.data_format in (DF.DEFAULT, DF.NHWC, DF.NCHW) for input in self.inputs]
+        if all(is_default):
+            return self._broadcast_shape([input.shape for input in self.inputs])
+
+        # in case formats are fractal_nz, default_fromat/NHWC/HCHW(optional)
+        is_default_frac_nz = [input.data_format in (DF.DEFAULT, DF.NHWC, DF.NCHW, DF.FRAC_NZ) \
+            for input in self.inputs]
+        if all(is_default_frac_nz):
+            nz_shapes = [self._to_nz(input.shape) if input.data_format != DF.FRAC_NZ else input.shape \
+                for input in self.inputs]
+            return self._broadcast_shape(nz_shapes)
+
+        raise GKException("Only support default and fractal_nz")
+
     def _infer_format(self):
         for tensor in self.inputs:
             if tensor.data_format != DF.DEFAULT:

mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_expander.cc

@@ -56,6 +56,8 @@ std::vector<PrimitivePtr> GetExpandOps() {
     prim::kPrimLogSoftmax,
     prim::kPrimLogSoftmaxGrad,
     prim::kPrimTile,
+    prim::kPrimMatMul,
+    prim::kPrimBatchMatMul,
 #if ENABLE_D
     prim::kPrimSqrtGrad,
     prim::kPrimClipByNormNoDivSum,

tests/st/ops/graph_kernel/test_matmul_to_mul.py

@@ -0,0 +1,91 @@
+# 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.
+# ============================================================================
+
+import numpy as np
+import pytest
+import mindspore.context as context
+from mindspore import Tensor
+from mindspore.nn import Cell
+import mindspore.ops.operations as P
+
+class Net(Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.matmul = P.MatMul(transpose_a=False, transpose_b=False)
+
+    def construct(self, x, y):
+        return self.matmul(x, y)
+
+class Net1(Cell):
+    def __init__(self):
+        super(Net1, self).__init__()
+        self.bmm = P.BatchMatMul(transpose_a=False, transpose_b=False)
+
+    def construct(self, x, y):
+        return self.bmm(x, y)
+
+def get_output(i0, i1, net_cls, enable_graph_kernel=False):
+    context.set_context(enable_graph_kernel=enable_graph_kernel)
+    net = net_cls()
+    output = net(i0, i1)
+    return output
+
+def test_matmul():
+    i0 = Tensor(np.random.normal(1, 0.01, [96, 1]).astype(np.float32))
+    i1 = Tensor(np.random.normal(1, 0.01, [1, 128]).astype(np.float32))
+    expect = get_output(i0, i1, Net, False)
+    output = get_output(i0, i1, Net, True)
+    expect_np = expect.asnumpy().copy()
+    output_np = output.asnumpy().copy()
+    assert np.allclose(expect_np, output_np, 1.e-4, 1.e-7)
+
+def test_batchmatmul():
+    i0 = Tensor(np.random.normal(1, 0.01, [16, 96, 1]).astype(np.float32))
+    i1 = Tensor(np.random.normal(1, 0.01, [16, 1, 128]).astype(np.float32))
+    expect = get_output(i0, i1, Net1, False)
+    output = get_output(i0, i1, Net1, True)
+    expect_np = expect.asnumpy().copy()
+    output_np = output.asnumpy().copy()
+    assert np.allclose(expect_np, output_np, 6.e-4, 6.e-4)
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_matmul_ascend():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+    test_matmul()
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_batchmatmul_ascend():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+    test_batchmatmul()
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_matmul_gpu():
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+    test_matmul()
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_batchmatmul_gpu():
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+    test_batchmatmul()