perf(functional/matmul): reimplement matmul with subgraph

GitOrigin-RevId: 456b2a51d35852152c46d31548dac96f977d5b41

imperative/python/megengine/functional/math.py

@@ -8,17 +8,21 @@
 # "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 import collections
 import math
+from functools import lru_cache
 from typing import Optional, Sequence, Tuple, Union
 
 from ..core._imperative_rt.core2 import apply, dtype_promotion
+from ..core._imperative_rt.ops import SubgraphBuilder as _SubgraphBuilder
 from ..core._trace_option import use_symbolic_shape
 from ..core.ops import builtin
+from ..core.ops.builtin import BatchNorm, Elemwise, GetVarShape, Reduce, TypeCvt
 from ..core.ops.special import Const
 from ..core.tensor import amp
-from ..core.tensor.utils import _normalize_axis, cast_tensors, setscalar
+from ..core.tensor.utils import _normalize_axis, cast_tensors, setscalar, subgraph
+from ..jit import exclude_from_trace
 from ..tensor import Tensor
 from .debug_param import get_execution_strategy
-from .elemwise import clip
+from .elemwise import clip, minimum
 from .tensor import broadcast_to, concat, expand_dims, squeeze
 
 __all__ = [
@@ -763,6 +767,216 @@ def matinv(inp: Tensor) -> Tensor:
     return result
 
 
+@lru_cache(maxsize=None)
+def _get_extentedMatrixMulOp(
+    device, dtype, dim1, dim2, transpose_a, transpose_b, compute_mode, format, strategy,
+):
+    @subgraph("extentedMatrixMulOp", dtype, device, 2, gopt_level=3)
+    def extentedMatrixMulOp(inputs, f, c):
+        assert len(inputs) == 2
+        inp1, inp2 = inputs
+        _dim1, _dim2 = dim1, dim2
+
+        def build_shape_head(shape, idx=-1):
+            # shape[:idx]
+            return f(
+                builtin.Subtensor(items=[[0, False, True, False, False]]),
+                shape,
+                c(idx, "int32"),
+            )
+
+        def build_shape_tail(shape, idx=-1):
+            # shape[idx:]
+            return f(
+                builtin.Subtensor(items=[[0, True, False, False, False]]),
+                shape,
+                c(idx, "int32"),
+            )
+
+        remove_row, remove_col = False, False
+        if _dim1 == 1:
+            _dim1 = 2
+            remove_row = True
+        if _dim2 == 1:
+            _dim2 = 2
+            remove_col = True
+
+        if remove_row:
+            inp1 = f(builtin.AddAxis(axis=[0,]), inp1)
+        if remove_col:
+            inp2 = f(builtin.AddAxis(axis=[1,]), inp2)
+
+        shape1 = f(GetVarShape(), inp1)
+        shape2 = f(GetVarShape(), inp2)
+        if _dim1 > 2:
+            inp1 = f(
+                builtin.Reshape(),
+                inp1,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    f(builtin.Reduce(mode="product", axis=0), build_shape_head(shape1)),
+                    build_shape_tail(shape1),
+                ),
+            )
+        if _dim2 > 2:
+            inp2 = f(
+                builtin.Reshape(),
+                inp2,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    f(builtin.Reduce(mode="product", axis=0), build_shape_head(shape2)),
+                    build_shape_tail(shape2),
+                ),
+            )
+        op = builtin.MatrixMul(
+            transposeA=transpose_a,
+            transposeB=transpose_b,
+            compute_mode=compute_mode,
+            format=format,
+            strategy=strategy,
+        )
+        result = f(op, inp1, inp2)
+        result_shape = f(GetVarShape(), result)
+        if _dim1 > 2:
+            result = f(
+                builtin.Reshape(),
+                result,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    build_shape_head(shape1),
+                    build_shape_tail(result_shape),
+                ),
+            )
+        if _dim2 > 2:
+            result = f(
+                builtin.Reshape(),
+                result,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    build_shape_head(shape2),
+                    build_shape_tail(result_shape),
+                ),
+            )
+        maxdim = _dim1 if _dim1 > _dim2 else _dim2
+        if remove_row:
+            result = f(builtin.RemoveAxis(axis=[maxdim - 2]), result)
+        if remove_col:
+            result = f(builtin.RemoveAxis(axis=[maxdim - 1]), result)
+        return (result,), (True,)
+
+    return extentedMatrixMulOp
+
+
+@lru_cache(maxsize=None)
+def _get_extentedBatchedMatrixMulOp(
+    device, dtype, dim1, dim2, transpose_a, transpose_b, compute_mode, format, strategy,
+):
+    @subgraph("extentedBatchedMatrixMulOp", dtype, device, 2, gopt_level=3)
+    def extentedBatchedMatrixMulOp(inputs, f, c):
+        assert len(inputs) == 2
+        inp1, inp2 = inputs
+        _dim1, _dim2 = dim1, dim2
+
+        def build_shape_head(shape, idx=-2):
+            # shape[:idx]
+            return f(
+                builtin.Subtensor(items=[[0, False, True, False, False]]),
+                shape,
+                c(idx, "int32"),
+            )
+
+        def build_shape_tail(shape, idx=-2):
+            # shape[idx:]
+            return f(
+                builtin.Subtensor(items=[[0, True, False, False, False]]),
+                shape,
+                c(idx, "int32"),
+            )
+
+        remove_row, remove_col = False, False
+        if _dim1 == 1:
+            _dim1 = 2
+            remove_row = True
+        if _dim2 == 1:
+            _dim2 = 2
+            remove_col = True
+
+        if remove_row:
+            inp1 = f(builtin.AddAxis(axis=[0,]), inp1)
+        if remove_col:
+            inp2 = f(builtin.AddAxis(axis=[1,]), inp2)
+        shape1 = f(GetVarShape(), inp1)
+        shape2 = f(GetVarShape(), inp2)
+        maxdim = _dim1 if _dim1 > _dim2 else _dim2
+        if _dim1 > _dim2:
+            # broadcast
+            shape2 = f(
+                builtin.Concat(axis=0, comp_node=device),
+                build_shape_head(shape1, idx=-_dim2),  # shape1[:-_dim2]
+                shape2,
+            )
+            inp2 = f(builtin.Broadcast(), inp2, shape2)
+            batch_shape = build_shape_head(shape1)
+        if _dim2 > _dim1:
+            # broadcast
+            shape1 = f(
+                builtin.Concat(axis=0, comp_node=device),
+                build_shape_head(shape2, idx=-_dim1),  # shape2[:-_dim1]
+                shape1,
+            )
+            inp1 = f(builtin.Broadcast(), inp1, shape1)
+            batch_shape = build_shape_head(shape2)
+        if _dim1 == _dim2:
+            batch_shape = build_shape_head(shape1)
+
+        # compress inputs to 3d
+        if maxdim > 3:
+            inp1 = f(
+                builtin.Reshape(),
+                inp1,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    f(builtin.Reduce(mode="product", axis=0), batch_shape),
+                    build_shape_tail(shape1),
+                ),
+            )
+            inp2 = f(
+                builtin.Reshape(),
+                inp2,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    f(builtin.Reduce(mode="product", axis=0), batch_shape),
+                    build_shape_tail(shape2),
+                ),
+            )
+        op = builtin.BatchedMatrixMul(
+            transposeA=transpose_a,
+            transposeB=transpose_b,
+            compute_mode=compute_mode,
+            format=format,
+            strategy=strategy,
+        )
+        result = f(op, inp1, inp2)
+
+        if maxdim > 3:
+            result = f(
+                builtin.Reshape(),
+                result,
+                f(
+                    builtin.Concat(axis=0, comp_node=device),
+                    batch_shape,
+                    build_shape_tail(f(GetVarShape(), result)),
+                ),
+            )
+        if remove_row:
+            result = f(builtin.RemoveAxis(axis=[maxdim - 2]), result)
+        if remove_col:
+            result = f(builtin.RemoveAxis(axis=[maxdim - 1]), result)
+        return (result,), (True,)
+
+    return extentedBatchedMatrixMulOp
+
+
 def matmul(
     inp1: Tensor,
     inp2: Tensor,
@@ -822,84 +1036,38 @@ def matmul(
         if inp2.dtype != dtype:
             inp2 = inp2.astype(dtype)
 
-    remove_row, remove_col = False, False
     dim1, dim2 = inp1.ndim, inp2.ndim
-    # handle dim=1 cases, dot and matrix-vector multiplication
-    if dim1 == 1 and dim2 == 1:
-        return dot(inp1, inp2)
-    # the underlying matmul op requires input dims to be at least 2
-    if dim1 == 1:
-        inp1 = expand_dims(inp1, 0)
-        dim1 = 2
-        remove_row = True
-    if dim2 == 1:
-        inp2 = expand_dims(inp2, 1)
-        dim2 = 2
-        remove_col = True
-
-    batch_shape = None
-    shape1 = inp1.shape
-    shape2 = inp2.shape
-
+    assert dim1 > 0 and dim2 > 0
     maxdim = dim1 if dim1 > dim2 else dim2
-    if dim1 >= 3 or dim2 >= 3:
-        if use_symbolic_shape():
-            if dim1 > dim2:
-                shape2 = concat([shape1[:-2], shape2[-2:]])
-                inp2 = broadcast_to(inp2, shape2)
-            if dim1 < dim2:
-                shape1 = concat([shape2[:-2], shape1[-2:]])
-                inp1 = broadcast_to(inp1, shape1)
-            if maxdim > 3:
-                batch_shape = shape1[:-2]
-                # compress inputs to 3d
-                (inp1,) = apply(
-                    builtin.Reshape(), inp1, concat([prod(shape1[:-2]), shape1[-2:]])
-                )
-                (inp2,) = apply(
-                    builtin.Reshape(), inp2, concat([prod(shape2[:-2]), shape2[-2:]])
-                )
-        else:
-            if dim1 > dim2:
-                shape2 = shape1[:-2] + shape2[-2:]
-                inp2 = broadcast_to(inp2, shape2)
-            if dim1 < dim2:
-                shape1 = shape2[:-2] + shape1[-2:]
-                inp1 = broadcast_to(inp1, shape1)
-            if maxdim > 3:
-                batch_shape = shape1[:-2]
-                # compress inputs to 3d
-                inp1 = inp1.reshape((-1, shape1[-2], shape1[-1]))
-                inp2 = inp2.reshape((-1, shape2[-2], shape2[-1]))
-
-        op = builtin.BatchedMatrixMul(
-            transposeA=transpose_a,
-            transposeB=transpose_b,
-            compute_mode=compute_mode,
-            format=format,
+    if dim1 == 1 and dim2 == 1:  # dispatch to Dot
+        return dot(inp1, inp2)
+    elif maxdim <= 2 or dim2 <= 2:  # dispath to MatrixMul
+        extentedMatrixMulOp = _get_extentedMatrixMulOp(
+            inp1.device,
+            inp1.dtype,
+            dim1,
+            dim2,
+            transpose_a,
+            transpose_b,
+            compute_mode,
+            format,
             strategy=get_execution_strategy(),
         )
-    else:
-        op = builtin.MatrixMul(
-            transposeA=transpose_a,
-            transposeB=transpose_b,
-            compute_mode=compute_mode,
-            format=format,
+        (result,) = apply(extentedMatrixMulOp, inp1, inp2)
+        return result
+    else:  # dispath to BatchedMatrixMul
+        extentedBatchedMatrixMulOp = _get_extentedBatchedMatrixMulOp(
+            inp1.device,
+            inp1.dtype,
+            dim1,
+            dim2,
+            transpose_a,
+            transpose_b,
+            compute_mode,
+            format,
             strategy=get_execution_strategy(),
         )
-
-    (result,) = apply(op, inp1, inp2)
-    if maxdim > 3:
-        if use_symbolic_shape():
-            (result,) = apply(
-                builtin.Reshape(), result, concat([batch_shape, result.shape[-2:]])
-            )
-        else:
-            result = result.reshape(batch_shape + result.shape[-2:])
-    if remove_row:
-        result = squeeze(result, axis=-2)
-    if remove_col:
-        result = squeeze(result, axis=-1)
+        (result,) = apply(extentedBatchedMatrixMulOp, inp1, inp2)
         return result