[Mlir-commits] [mlir] [mlir][python] enable memref.subview (PR #79393)

Maksim Levental llvmlistbot at llvm.org
Thu Jan 25 16:55:25 PST 2024


https://github.com/makslevental updated https://github.com/llvm/llvm-project/pull/79393

>From eb0f298facf34cc6239e168890e6c5aff0d9e0be Mon Sep 17 00:00:00 2001
From: max <maksim.levental at gmail.com>
Date: Wed, 24 Jan 2024 18:25:54 -0600
Subject: [PATCH] [mlir][python] enable memref.subview

---
 mlir/include/mlir-c/BuiltinTypes.h   |   7 ++
 mlir/lib/Bindings/Python/IRTypes.cpp |   9 ++
 mlir/lib/CAPI/IR/BuiltinTypes.cpp    |  14 +++
 mlir/python/mlir/dialects/memref.py  | 121 +++++++++++++++++++
 mlir/test/python/dialects/memref.py  | 166 ++++++++++++++++++++++++++-
 5 files changed, 315 insertions(+), 2 deletions(-)

diff --git a/mlir/include/mlir-c/BuiltinTypes.h b/mlir/include/mlir-c/BuiltinTypes.h
index 1fd5691f41eec3..2523bddc475d82 100644
--- a/mlir/include/mlir-c/BuiltinTypes.h
+++ b/mlir/include/mlir-c/BuiltinTypes.h
@@ -408,6 +408,13 @@ MLIR_CAPI_EXPORTED MlirAffineMap mlirMemRefTypeGetAffineMap(MlirType type);
 /// Returns the memory space of the given MemRef type.
 MLIR_CAPI_EXPORTED MlirAttribute mlirMemRefTypeGetMemorySpace(MlirType type);
 
+/// Returns the strides of the MemRef if the layout map is in strided form.
+/// Both strides and offset are out params. strides must point to pre-allocated
+/// memory of length equal to the rank of the memref.
+MLIR_CAPI_EXPORTED void mlirMemRefTypeGetStridesAndOffset(MlirType type,
+                                                          int64_t *strides,
+                                                          int64_t *offset);
+
 /// Returns the memory spcae of the given Unranked MemRef type.
 MLIR_CAPI_EXPORTED MlirAttribute
 mlirUnrankedMemrefGetMemorySpace(MlirType type);
diff --git a/mlir/lib/Bindings/Python/IRTypes.cpp b/mlir/lib/Bindings/Python/IRTypes.cpp
index 56e895d3053796..86f01a6381ae4e 100644
--- a/mlir/lib/Bindings/Python/IRTypes.cpp
+++ b/mlir/lib/Bindings/Python/IRTypes.cpp
@@ -618,6 +618,15 @@ class PyMemRefType : public PyConcreteType<PyMemRefType, PyShapedType> {
               return mlirMemRefTypeGetLayout(self);
             },
             "The layout of the MemRef type.")
+        .def_property_readonly(
+            "strides_and_offset",
+            [](PyMemRefType &self) -> std::pair<std::vector<int64_t>, int64_t> {
+              std::vector<int64_t> strides(mlirShapedTypeGetRank(self));
+              int64_t offset;
+              mlirMemRefTypeGetStridesAndOffset(self, strides.data(), &offset);
+              return {strides, offset};
+            },
+            "The strides and offset of the MemRef type.")
         .def_property_readonly(
             "affine_map",
             [](PyMemRefType &self) -> PyAffineMap {
diff --git a/mlir/lib/CAPI/IR/BuiltinTypes.cpp b/mlir/lib/CAPI/IR/BuiltinTypes.cpp
index 6e645188dac861..6a3653d8baf304 100644
--- a/mlir/lib/CAPI/IR/BuiltinTypes.cpp
+++ b/mlir/lib/CAPI/IR/BuiltinTypes.cpp
@@ -16,6 +16,8 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/Types.h"
 
+#include <algorithm>
+
 using namespace mlir;
 
 //===----------------------------------------------------------------------===//
@@ -426,6 +428,18 @@ MlirAttribute mlirMemRefTypeGetMemorySpace(MlirType type) {
   return wrap(llvm::cast<MemRefType>(unwrap(type)).getMemorySpace());
 }
 
+void mlirMemRefTypeGetStridesAndOffset(MlirType type, int64_t *strides,
+                                       int64_t *offset) {
+  MemRefType memrefType = llvm::cast<MemRefType>(unwrap(type));
+  std::pair<SmallVector<int64_t>, int64_t> stridesOffsets =
+      getStridesAndOffset(memrefType);
+  assert(stridesOffsets.first.size() == memrefType.getRank() &&
+         "Strides and rank don't match for memref");
+  (void)std::copy(stridesOffsets.first.begin(), stridesOffsets.first.end(),
+                  strides);
+  *offset = stridesOffsets.second;
+}
+
 MlirTypeID mlirUnrankedMemRefTypeGetTypeID() {
   return wrap(UnrankedMemRefType::getTypeID());
 }
diff --git a/mlir/python/mlir/dialects/memref.py b/mlir/python/mlir/dialects/memref.py
index 3afb6a70cb9e0d..6ab6e0602e7a95 100644
--- a/mlir/python/mlir/dialects/memref.py
+++ b/mlir/python/mlir/dialects/memref.py
@@ -1,5 +1,126 @@
 #  Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 #  See https://llvm.org/LICENSE.txt for license information.
 #  SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+import operator
+from itertools import accumulate
+from typing import Optional
 
 from ._memref_ops_gen import *
+from .arith import ConstantOp
+from .transform.structured import _dispatch_mixed_values, MixedValues
+from ..ir import Value, MemRefType, StridedLayoutAttr, ShapedType
+
+
+def _is_constant(i):
+    return isinstance(i, Value) and isinstance(i.owner.opview, ConstantOp)
+
+
+def _is_static(i):
+    return (isinstance(i, int) and not ShapedType.is_dynamic_size(i)) or _is_constant(i)
+
+
+def _infer_memref_subview_result_type(
+    source_memref_type, offsets, static_sizes, static_strides
+):
+    source_strides, source_offset = source_memref_type.strides_and_offset
+    # "canonicalize" from tuple|list -> list
+    offsets, static_sizes, static_strides, source_strides = map(
+        list, (offsets, static_sizes, static_strides, source_strides)
+    )
+
+    assert all(
+        all(_is_static(i) for i in s)
+        for s in [
+            static_sizes,
+            static_strides,
+            source_strides,
+        ]
+    ), f"Only inferring from python or mlir integer constant is supported"
+
+    for s in [offsets, static_sizes, static_strides]:
+        for idx, i in enumerate(s):
+            if _is_constant(i):
+                s[idx] = i.owner.opview.literal_value
+
+    if any(not _is_static(i) for i in offsets + [source_offset]):
+        target_offset = ShapedType.get_dynamic_size()
+    else:
+        target_offset = source_offset
+        for offset, target_stride in zip(offsets, source_strides):
+            target_offset += offset * target_stride
+
+    target_strides = []
+    for source_stride, static_stride in zip(source_strides, static_strides):
+        target_strides.append(source_stride * static_stride)
+
+    # If default striding then no need to complicate things for downstream ops (e.g., expand_shape).
+    default_strides = list(accumulate(static_sizes[1:][::-1], operator.mul))[::-1] + [1]
+    if target_strides == default_strides and target_offset == 0:
+        layout = None
+    else:
+        layout = StridedLayoutAttr.get(target_offset, target_strides)
+    return (
+        offsets,
+        static_sizes,
+        static_strides,
+        MemRefType.get(
+            static_sizes,
+            source_memref_type.element_type,
+            layout,
+            source_memref_type.memory_space,
+        ),
+    )
+
+
+_generated_subview = subview
+
+
+def subview(
+    source: Value,
+    offsets: MixedValues,
+    sizes: MixedValues,
+    strides: MixedValues,
+    *,
+    result_type: Optional[MemRefType] = None,
+    loc=None,
+    ip=None,
+):
+    if offsets is None:
+        offsets = []
+    if sizes is None:
+        sizes = []
+    if strides is None:
+        strides = []
+    source_strides, source_offset = source.type.strides_and_offset
+    if result_type is None and all(
+        all(_is_static(i) for i in s) for s in [sizes, strides, source_strides]
+    ):
+        # If any are arith.constant results then this will canonicalize to python int
+        # (which can then be used to fully specific the subview).
+        (
+            offsets,
+            sizes,
+            strides,
+            result_type,
+        ) = _infer_memref_subview_result_type(source.type, offsets, sizes, strides)
+    else:
+        assert (
+            result_type is not None
+        ), "mixed static/dynamic offset/sizes/strides requires explicit result type"
+
+    offsets, _packed_offsets, static_offsets = _dispatch_mixed_values(offsets)
+    sizes, _packed_sizes, static_sizes = _dispatch_mixed_values(sizes)
+    strides, _packed_strides, static_strides = _dispatch_mixed_values(strides)
+
+    return _generated_subview(
+        result_type,
+        source,
+        offsets,
+        sizes,
+        strides,
+        static_offsets,
+        static_sizes,
+        static_strides,
+        loc=loc,
+        ip=ip,
+    )
diff --git a/mlir/test/python/dialects/memref.py b/mlir/test/python/dialects/memref.py
index 0c8a7ee282fe16..205acbd6d42c4d 100644
--- a/mlir/test/python/dialects/memref.py
+++ b/mlir/test/python/dialects/memref.py
@@ -1,9 +1,10 @@
 # RUN: %PYTHON %s | FileCheck %s
 
-from mlir.ir import *
-import mlir.dialects.func as func
+import mlir.dialects.arith as arith
 import mlir.dialects.memref as memref
 import mlir.extras.types as T
+from mlir.dialects.memref import _infer_memref_subview_result_type
+from mlir.ir import *
 
 
 def run(f):
@@ -88,3 +89,164 @@ def testMemRefAttr():
             memref.global_("objFifo_in0", T.memref(16, T.i32()))
         # CHECK: memref.global @objFifo_in0 : memref<16xi32>
         print(module)
+
+
+# CHECK-LABEL: TEST: testSubViewOpInferReturnType
+ at run
+def testSubViewOpInferReturnType():
+    with Context() as ctx, Location.unknown(ctx):
+        module = Module.create()
+        with InsertionPoint(module.body):
+            x = memref.alloc(T.memref(10, 10, T.i32()), [], [])
+            # CHECK: %[[ALLOC:.*]] = memref.alloc() : memref<10x10xi32>
+            print(x.owner)
+
+            y = memref.subview(x, [1, 1], [3, 3], [1, 1])
+            assert y.owner.verify()
+            # CHECK: %{{.*}} = memref.subview %[[ALLOC]][1, 1] [3, 3] [1, 1] : memref<10x10xi32> to memref<3x3xi32, strided<[10, 1], offset: 11>>
+            print(y.owner)
+
+            z = memref.subview(
+                x,
+                [arith.constant(T.index(), 1), 1],
+                [3, 3],
+                [1, 1],
+            )
+            # CHECK: %{{.*}} =  memref.subview %[[ALLOC]][1, 1] [3, 3] [1, 1] : memref<10x10xi32> to memref<3x3xi32, strided<[10, 1], offset: 11>>
+            print(z.owner)
+
+            z = memref.subview(
+                x,
+                [arith.constant(T.index(), 3), arith.constant(T.index(), 4)],
+                [3, 3],
+                [1, 1],
+            )
+            # CHECK: %{{.*}} =  memref.subview %[[ALLOC]][3, 4] [3, 3] [1, 1] : memref<10x10xi32> to memref<3x3xi32, strided<[10, 1], offset: 34>>
+            print(z.owner)
+
+            s = arith.addi(arith.constant(T.index(), 3), arith.constant(T.index(), 4))
+            z = memref.subview(
+                x,
+                [s, 0],
+                [3, 3],
+                [1, 1],
+            )
+            # CHECK: {{.*}} = memref.subview %[[ALLOC]][%0, 0] [3, 3] [1, 1] : memref<10x10xi32> to memref<3x3xi32, strided<[10, 1], offset: ?>>
+            print(z)
+
+            try:
+                _infer_memref_subview_result_type(
+                    x.type,
+                    [arith.constant(T.index(), 3), arith.constant(T.index(), 4)],
+                    [ShapedType.get_dynamic_size(), 3],
+                    [1, 1],
+                )
+            except AssertionError as e:
+                # CHECK: Only inferring from python or mlir integer constant is supported
+                print(e)
+
+            try:
+                memref.subview(
+                    x,
+                    [arith.constant(T.index(), 3), arith.constant(T.index(), 4)],
+                    [ShapedType.get_dynamic_size(), 3],
+                    [1, 1],
+                )
+            except AssertionError as e:
+                # CHECK: mixed static/dynamic offset/sizes/strides requires explicit result type
+                print(e)
+
+            layout = StridedLayoutAttr.get(ShapedType.get_dynamic_size(), [10, 1])
+            x = memref.alloc(
+                T.memref(
+                    10,
+                    10,
+                    T.i32(),
+                    layout=layout,
+                ),
+                [],
+                [arith.constant(T.index(), 42)],
+            )
+            # CHECK: %[[DYNAMICALLOC:.*]] = memref.alloc()[%c42] : memref<10x10xi32, strided<[10, 1], offset: ?>>
+            print(x.owner)
+            y = memref.subview(
+                x,
+                [1, 1],
+                [3, 3],
+                [1, 1],
+                result_type=T.memref(3, 3, T.i32(), layout=layout),
+            )
+            # CHECK: %{{.*}} = memref.subview %[[DYNAMICALLOC]][1, 1] [3, 3] [1, 1] : memref<10x10xi32, strided<[10, 1], offset: ?>> to memref<3x3xi32, strided<[10, 1], offset: ?>>
+            print(y.owner)
+
+
+# CHECK-LABEL: TEST: testSubViewOpInferReturnTypeExtensiveSlicing
+ at run
+def testSubViewOpInferReturnTypeExtensiveSlicing():
+    def check_strides_offset(memref, np_view):
+        layout = memref.type.layout
+        dtype_size_in_bytes = np_view.dtype.itemsize
+        golden_strides = (np.array(np_view.strides) // dtype_size_in_bytes).tolist()
+        golden_offset = (
+            np_view.ctypes.data - np_view.base.ctypes.data
+        ) // dtype_size_in_bytes
+
+        assert (layout.strides, layout.offset) == (golden_strides, golden_offset)
+
+    with Context() as ctx, Location.unknown(ctx):
+        module = Module.create()
+        with InsertionPoint(module.body):
+            shape = (10, 22, 333, 4444)
+            golden_mem = np.zeros(shape, dtype=np.int32)
+            mem1 = memref.alloc(T.memref(*shape, T.i32()), [], [])
+
+            # fmt: off
+            check_strides_offset(memref.subview(mem1, (1, 0, 0, 0), (1, 22, 333, 4444), (1, 1, 1, 1)), golden_mem[1:2, ...])
+            check_strides_offset(memref.subview(mem1, (0, 1, 0, 0), (10, 1, 333, 4444), (1, 1, 1, 1)), golden_mem[:, 1:2])
+            check_strides_offset(memref.subview(mem1, (0, 0, 1, 0), (10, 22, 1, 4444), (1, 1, 1, 1)), golden_mem[:, :, 1:2])
+            check_strides_offset(memref.subview(mem1, (0, 0, 0, 1), (10, 22, 333, 1), (1, 1, 1, 1)), golden_mem[:, :, :, 1:2])
+            check_strides_offset(memref.subview(mem1, (0, 1, 0, 1), (10, 1, 333, 1), (1, 1, 1, 1)), golden_mem[:, 1:2, :, 1:2])
+            check_strides_offset(memref.subview(mem1, (1, 0, 0, 1), (1, 22, 333, 1), (1, 1, 1, 1)), golden_mem[1:2, :, :, 1:2])
+            check_strides_offset(memref.subview(mem1, (1, 1, 0, 0), (1, 1, 333, 4444), (1, 1, 1, 1)), golden_mem[1:2, 1:2, :, :])
+            check_strides_offset(memref.subview(mem1, (0, 0, 1, 1), (10, 22, 1, 1), (1, 1, 1, 1)), golden_mem[:, :, 1:2, 1:2])
+            check_strides_offset(memref.subview(mem1, (0, 1, 1, 0), (10, 1, 1, 4444), (1, 1, 1, 1)), golden_mem[:, 1:2, 1:2, :])
+            check_strides_offset(memref.subview(mem1, (1, 0, 1, 0), (1, 22, 1, 4444), (1, 1, 1, 1)), golden_mem[1:2, :, 1:2, :])
+            check_strides_offset(memref.subview(mem1, (1, 1, 0, 1), (1, 1, 333, 1), (1, 1, 1, 1)), golden_mem[1:2, 1:2, :, 1:2])
+            check_strides_offset(memref.subview(mem1, (1, 0, 1, 1), (1, 22, 1, 1), (1, 1, 1, 1)), golden_mem[1:2, :, 1:2, 1:2])
+            check_strides_offset(memref.subview(mem1, (0, 1, 1, 1), (10, 1, 1, 1), (1, 1, 1, 1)), golden_mem[:, 1:2, 1:2, 1:2])
+            check_strides_offset(memref.subview(mem1, (1, 1, 1, 0), (1, 1, 1, 4444), (1, 1, 1, 1)), golden_mem[1:2, 1:2, 1:2, :])
+            # fmt: on
+
+            # default strides and offset means no stridedlayout attribute means affinemap layout
+            assert memref.subview(
+                mem1, (0, 0, 0, 0), (10, 22, 333, 4444), (1, 1, 1, 1)
+            ).type.layout == AffineMapAttr.get(
+                AffineMap.get(
+                    4,
+                    0,
+                    [
+                        AffineDimExpr.get(0),
+                        AffineDimExpr.get(1),
+                        AffineDimExpr.get(2),
+                        AffineDimExpr.get(3),
+                    ],
+                )
+            )
+
+            shape = (7, 22, 333, 4444)
+            golden_mem = np.zeros(shape, dtype=np.int32)
+            mem2 = memref.alloc(T.memref(*shape, T.i32()), [], [])
+            # fmt: off
+            check_strides_offset(memref.subview(mem2, (0, 0, 0, 0), (7, 11, 333, 4444), (1, 2, 1, 1)), golden_mem[:, 0:22:2])
+            check_strides_offset(memref.subview(mem2, (0, 0, 0, 0), (7, 11, 11, 4444), (1, 2, 30, 1)), golden_mem[:, 0:22:2, 0:330:30])
+            check_strides_offset(memref.subview(mem2, (0, 0, 0, 0), (7, 11, 11, 11), (1, 2, 30, 400)), golden_mem[:, 0:22:2, 0:330:30, 0:4400:400])
+            check_strides_offset(memref.subview(mem2, (0, 0, 100, 1000), (7, 22, 20, 20), (1, 1, 5, 50)), golden_mem[:, :, 100:200:5, 1000:2000:50])
+            # fmt: on
+
+            shape = (8, 8)
+            golden_mem = np.zeros(shape, dtype=np.int32)
+            # fmt: off
+            mem3 = memref.alloc(T.memref(*shape, T.i32()), [], [])
+            check_strides_offset(memref.subview(mem3, (0, 0), (4, 4), (1, 1)), golden_mem[0:4, 0:4])
+            check_strides_offset(memref.subview(mem3, (4, 4), (4, 4), (1, 1)), golden_mem[4:8, 4:8])
+            # fmt: on



More information about the Mlir-commits mailing list