[Mlir-commits] [mlir] [MLIR, Python] Support converting boolean numpy arrays to and from mlir attributes (unrevert) (PR #115481)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Fri Nov 8 05:26:15 PST 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
Author: Kasper Nielsen (kasper0406)
<details>
<summary>Changes</summary>
This PR re-introduces the functionality of https://github.com/llvm/llvm-project/pull/113064, which was reverted in https://github.com/llvm/llvm-project/commit/0a68171b3c67503f7143856580f1b22a93ef566e due to memory lifetime issues.
Notice that I was not able to re-produce the ASan results myself, so I have not been able to verify that this PR really fixes the issue.
---
Currently it is unsupported to:
1. Convert a MlirAttribute with type i1 to a numpy array
2. Convert a boolean numpy array to a MlirAttribute
Currently the entire Python application violently crashes with a quite poor error message https://github.com/pybind/pybind11/issues/3336
The complication handling these conversions, is that MlirAttribute represent booleans as a bit-packed i1 type, whereas numpy represents booleans as a byte array with 8 bit used per boolean.
This PR proposes the following approach:
1. When converting a i1 typed MlirAttribute to a numpy array, we can not directly use the underlying raw data backing the MlirAttribute as a buffer to Python, as done for other types. Instead, a copy of the data is generated using numpy's unpackbits function, and the result is send back to Python.
2. When constructing a MlirAttribute from a numpy array, first the python data is read as a uint8_t to get it converted to the endianess used internally in mlir. Then the booleans are bitpacked using numpy's bitpack function, and the bitpacked array is saved as the MlirAttribute representation.
---
Full diff: https://github.com/llvm/llvm-project/pull/115481.diff
2 Files Affected:
- (modified) mlir/lib/Bindings/Python/IRAttributes.cpp (+177-95)
- (modified) mlir/test/python/ir/array_attributes.py (+72)
``````````diff
diff --git a/mlir/lib/Bindings/Python/IRAttributes.cpp b/mlir/lib/Bindings/Python/IRAttributes.cpp
index ead81a76c0538d..dba8220b5543a1 100644
--- a/mlir/lib/Bindings/Python/IRAttributes.cpp
+++ b/mlir/lib/Bindings/Python/IRAttributes.cpp
@@ -13,6 +13,7 @@
#include "IRModule.h"
#include "PybindUtils.h"
+#include <pybind11/numpy.h>
#include "llvm/ADT/ScopeExit.h"
#include "llvm/Support/raw_ostream.h"
@@ -757,103 +758,10 @@ class PyDenseElementsAttribute
throw py::error_already_set();
}
auto freeBuffer = llvm::make_scope_exit([&]() { PyBuffer_Release(&view); });
- SmallVector<int64_t> shape;
- if (explicitShape) {
- shape.append(explicitShape->begin(), explicitShape->end());
- } else {
- shape.append(view.shape, view.shape + view.ndim);
- }
- MlirAttribute encodingAttr = mlirAttributeGetNull();
MlirContext context = contextWrapper->get();
-
- // Detect format codes that are suitable for bulk loading. This includes
- // all byte aligned integer and floating point types up to 8 bytes.
- // Notably, this excludes, bool (which needs to be bit-packed) and
- // other exotics which do not have a direct representation in the buffer
- // protocol (i.e. complex, etc).
- std::optional<MlirType> bulkLoadElementType;
- if (explicitType) {
- bulkLoadElementType = *explicitType;
- } else {
- std::string_view format(view.format);
- if (format == "f") {
- // f32
- assert(view.itemsize == 4 && "mismatched array itemsize");
- bulkLoadElementType = mlirF32TypeGet(context);
- } else if (format == "d") {
- // f64
- assert(view.itemsize == 8 && "mismatched array itemsize");
- bulkLoadElementType = mlirF64TypeGet(context);
- } else if (format == "e") {
- // f16
- assert(view.itemsize == 2 && "mismatched array itemsize");
- bulkLoadElementType = mlirF16TypeGet(context);
- } else if (isSignedIntegerFormat(format)) {
- if (view.itemsize == 4) {
- // i32
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 32)
- : mlirIntegerTypeSignedGet(context, 32);
- } else if (view.itemsize == 8) {
- // i64
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 64)
- : mlirIntegerTypeSignedGet(context, 64);
- } else if (view.itemsize == 1) {
- // i8
- bulkLoadElementType = signless ? mlirIntegerTypeGet(context, 8)
- : mlirIntegerTypeSignedGet(context, 8);
- } else if (view.itemsize == 2) {
- // i16
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 16)
- : mlirIntegerTypeSignedGet(context, 16);
- }
- } else if (isUnsignedIntegerFormat(format)) {
- if (view.itemsize == 4) {
- // unsigned i32
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 32)
- : mlirIntegerTypeUnsignedGet(context, 32);
- } else if (view.itemsize == 8) {
- // unsigned i64
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 64)
- : mlirIntegerTypeUnsignedGet(context, 64);
- } else if (view.itemsize == 1) {
- // i8
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 8)
- : mlirIntegerTypeUnsignedGet(context, 8);
- } else if (view.itemsize == 2) {
- // i16
- bulkLoadElementType = signless
- ? mlirIntegerTypeGet(context, 16)
- : mlirIntegerTypeUnsignedGet(context, 16);
- }
- }
- if (!bulkLoadElementType) {
- throw std::invalid_argument(
- std::string("unimplemented array format conversion from format: ") +
- std::string(format));
- }
- }
-
- MlirType shapedType;
- if (mlirTypeIsAShaped(*bulkLoadElementType)) {
- if (explicitShape) {
- throw std::invalid_argument("Shape can only be specified explicitly "
- "when the type is not a shaped type.");
- }
- shapedType = *bulkLoadElementType;
- } else {
- shapedType = mlirRankedTensorTypeGet(shape.size(), shape.data(),
- *bulkLoadElementType, encodingAttr);
- }
- size_t rawBufferSize = view.len;
- MlirAttribute attr =
- mlirDenseElementsAttrRawBufferGet(shapedType, rawBufferSize, view.buf);
+ MlirAttribute attr = getAttributeFromBuffer(view, signless, explicitType,
+ explicitShape, context);
if (mlirAttributeIsNull(attr)) {
throw std::invalid_argument(
"DenseElementsAttr could not be constructed from the given buffer. "
@@ -963,6 +871,13 @@ class PyDenseElementsAttribute
// unsigned i16
return bufferInfo<uint16_t>(shapedType);
}
+ } else if (mlirTypeIsAInteger(elementType) &&
+ mlirIntegerTypeGetWidth(elementType) == 1) {
+ // i1 / bool
+ // We can not send the buffer directly back to Python, because the i1
+ // values are bitpacked within MLIR. We call numpy's unpackbits function
+ // to convert the bytes.
+ return getBooleanBufferFromBitpackedAttribute();
}
// TODO: Currently crashes the program.
@@ -1016,6 +931,173 @@ class PyDenseElementsAttribute
code == 'q';
}
+ static MlirType
+ getShapedType(std::optional<MlirType> bulkLoadElementType,
+ std::optional<std::vector<int64_t>> explicitShape,
+ Py_buffer &view) {
+ SmallVector<int64_t> shape;
+ if (explicitShape) {
+ shape.append(explicitShape->begin(), explicitShape->end());
+ } else {
+ shape.append(view.shape, view.shape + view.ndim);
+ }
+
+ if (mlirTypeIsAShaped(*bulkLoadElementType)) {
+ if (explicitShape) {
+ throw std::invalid_argument("Shape can only be specified explicitly "
+ "when the type is not a shaped type.");
+ }
+ return *bulkLoadElementType;
+ } else {
+ MlirAttribute encodingAttr = mlirAttributeGetNull();
+ return mlirRankedTensorTypeGet(shape.size(), shape.data(),
+ *bulkLoadElementType, encodingAttr);
+ }
+ }
+
+ static MlirAttribute getAttributeFromBuffer(
+ Py_buffer &view, bool signless, std::optional<PyType> explicitType,
+ std::optional<std::vector<int64_t>> explicitShape, MlirContext &context) {
+ // Detect format codes that are suitable for bulk loading. This includes
+ // all byte aligned integer and floating point types up to 8 bytes.
+ // Notably, this excludes exotics types which do not have a direct
+ // representation in the buffer protocol (i.e. complex, etc).
+ std::optional<MlirType> bulkLoadElementType;
+ if (explicitType) {
+ bulkLoadElementType = *explicitType;
+ } else {
+ std::string_view format(view.format);
+ if (format == "f") {
+ // f32
+ assert(view.itemsize == 4 && "mismatched array itemsize");
+ bulkLoadElementType = mlirF32TypeGet(context);
+ } else if (format == "d") {
+ // f64
+ assert(view.itemsize == 8 && "mismatched array itemsize");
+ bulkLoadElementType = mlirF64TypeGet(context);
+ } else if (format == "e") {
+ // f16
+ assert(view.itemsize == 2 && "mismatched array itemsize");
+ bulkLoadElementType = mlirF16TypeGet(context);
+ } else if (format == "?") {
+ // i1
+ // The i1 type needs to be bit-packed, so we will handle it seperately
+ return getBitpackedAttributeFromBooleanBuffer(view, explicitShape,
+ context);
+ } else if (isSignedIntegerFormat(format)) {
+ if (view.itemsize == 4) {
+ // i32
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 32)
+ : mlirIntegerTypeSignedGet(context, 32);
+ } else if (view.itemsize == 8) {
+ // i64
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 64)
+ : mlirIntegerTypeSignedGet(context, 64);
+ } else if (view.itemsize == 1) {
+ // i8
+ bulkLoadElementType = signless ? mlirIntegerTypeGet(context, 8)
+ : mlirIntegerTypeSignedGet(context, 8);
+ } else if (view.itemsize == 2) {
+ // i16
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 16)
+ : mlirIntegerTypeSignedGet(context, 16);
+ }
+ } else if (isUnsignedIntegerFormat(format)) {
+ if (view.itemsize == 4) {
+ // unsigned i32
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 32)
+ : mlirIntegerTypeUnsignedGet(context, 32);
+ } else if (view.itemsize == 8) {
+ // unsigned i64
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 64)
+ : mlirIntegerTypeUnsignedGet(context, 64);
+ } else if (view.itemsize == 1) {
+ // i8
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 8)
+ : mlirIntegerTypeUnsignedGet(context, 8);
+ } else if (view.itemsize == 2) {
+ // i16
+ bulkLoadElementType = signless
+ ? mlirIntegerTypeGet(context, 16)
+ : mlirIntegerTypeUnsignedGet(context, 16);
+ }
+ }
+ if (!bulkLoadElementType) {
+ throw std::invalid_argument(
+ std::string("unimplemented array format conversion from format: ") +
+ std::string(format));
+ }
+ }
+
+ MlirType type = getShapedType(bulkLoadElementType, explicitShape, view);
+ return mlirDenseElementsAttrRawBufferGet(type, view.len, view.buf);
+ }
+
+ // There is a complication for boolean numpy arrays, as numpy represents them
+ // as 8 bits (1 byte) per boolean, whereas MLIR bitpacks them into 8 booleans
+ // per byte.
+ static MlirAttribute getBitpackedAttributeFromBooleanBuffer(
+ Py_buffer &view, std::optional<std::vector<int64_t>> explicitShape,
+ MlirContext &context) {
+ if (llvm::endianness::native != llvm::endianness::little) {
+ // Given we have no good way of testing the behavior on big-endian systems
+ // we will throw
+ throw py::type_error("Constructing a bit-packed MLIR attribute is "
+ "unsupported on big-endian systems");
+ }
+
+ py::array_t<uint8_t> unpackedArray(view.len,
+ static_cast<uint8_t *>(view.buf));
+
+ py::module numpy = py::module::import("numpy");
+ py::object packbitsFunc = numpy.attr("packbits");
+ py::object packedBooleans =
+ packbitsFunc(unpackedArray, "bitorder"_a = "little");
+ py::buffer_info pythonBuffer = packedBooleans.cast<py::buffer>().request();
+
+ MlirType bitpackedType =
+ getShapedType(mlirIntegerTypeGet(context, 1), explicitShape, view);
+ assert(pythonBuffer.itemsize == 1 && "Packbits must return uint8");
+ // Notice that `mlirDenseElementsAttrRawBufferGet` copies the memory of
+ // packedBooleans, hence the MlirAttribute will remain valid even when
+ // packedBooleans get reclaimed by the end of the function.
+ return mlirDenseElementsAttrRawBufferGet(bitpackedType, pythonBuffer.size,
+ pythonBuffer.ptr);
+ }
+
+ // This does the opposite transformation of
+ // `getBitpackedAttributeFromBooleanBuffer`
+ py::buffer_info getBooleanBufferFromBitpackedAttribute() {
+ if (llvm::endianness::native != llvm::endianness::little) {
+ // Given we have no good way of testing the behavior on big-endian systems
+ // we will throw
+ throw py::type_error("Constructing a numpy array from a MLIR attribute "
+ "is unsupported on big-endian systems");
+ }
+
+ int64_t numBooleans = mlirElementsAttrGetNumElements(*this);
+ int64_t numBitpackedBytes = llvm::divideCeil(numBooleans, 8);
+ uint8_t *bitpackedData = static_cast<uint8_t *>(
+ const_cast<void *>(mlirDenseElementsAttrGetRawData(*this)));
+ py::array_t<uint8_t> packedArray(numBitpackedBytes, bitpackedData);
+
+ py::module numpy = py::module::import("numpy");
+ py::object unpackbitsFunc = numpy.attr("unpackbits");
+ py::object unpackedBooleans =
+ unpackbitsFunc(packedArray, "bitorder"_a = "little");
+ py::buffer pythonBuffer = unpackedBooleans.cast<py::buffer>();
+
+ // Make sure the returned py::buffer_view claims ownership of the data in
+ // `pythonBuffer` so it remains valid when Python reads it
+ return pythonBuffer.request();
+ }
+
template <typename Type>
py::buffer_info bufferInfo(MlirType shapedType,
const char *explicitFormat = nullptr) {
diff --git a/mlir/test/python/ir/array_attributes.py b/mlir/test/python/ir/array_attributes.py
index 2bc403aace8348..256a69a939658d 100644
--- a/mlir/test/python/ir/array_attributes.py
+++ b/mlir/test/python/ir/array_attributes.py
@@ -326,6 +326,78 @@ def testGetDenseElementsF64():
print(np.array(attr))
+### 1 bit/boolean integer arrays
+# CHECK-LABEL: TEST: testGetDenseElementsI1Signless
+ at run
+def testGetDenseElementsI1Signless():
+ with Context():
+ array = np.array([True], dtype=np.bool_)
+ attr = DenseElementsAttr.get(array)
+ # CHECK: dense<true> : tensor<1xi1>
+ print(attr)
+ # CHECK{LITERAL}: [ True]
+ print(np.array(attr))
+
+ array = np.array([[True, False, True], [True, True, False]], dtype=np.bool_)
+ attr = DenseElementsAttr.get(array)
+ # CHECK{LITERAL}: dense<[[true, false, true], [true, true, false]]> : tensor<2x3xi1>
+ print(attr)
+ # CHECK{LITERAL}: [[ True False True]
+ # CHECK{LITERAL}: [ True True False]]
+ print(np.array(attr))
+
+ array = np.array(
+ [[True, True, False, False], [True, False, True, False]], dtype=np.bool_
+ )
+ attr = DenseElementsAttr.get(array)
+ # CHECK{LITERAL}: dense<[[true, true, false, false], [true, false, true, false]]> : tensor<2x4xi1>
+ print(attr)
+ # CHECK{LITERAL}: [[ True True False False]
+ # CHECK{LITERAL}: [ True False True False]]
+ print(np.array(attr))
+
+ array = np.array(
+ [
+ [True, True, False, False],
+ [True, False, True, False],
+ [False, False, False, False],
+ [True, True, True, True],
+ [True, False, False, True],
+ ],
+ dtype=np.bool_,
+ )
+ attr = DenseElementsAttr.get(array)
+ # CHECK{LITERAL}: dense<[[true, true, false, false], [true, false, true, false], [false, false, false, false], [true, true, true, true], [true, false, false, true]]> : tensor<5x4xi1>
+ print(attr)
+ # CHECK{LITERAL}: [[ True True False False]
+ # CHECK{LITERAL}: [ True False True False]
+ # CHECK{LITERAL}: [False False False False]
+ # CHECK{LITERAL}: [ True True True True]
+ # CHECK{LITERAL}: [ True False False True]]
+ print(np.array(attr))
+
+ array = np.array(
+ [
+ [True, True, False, False, True, True, False, False, False],
+ [False, False, False, True, False, True, True, False, True],
+ ],
+ dtype=np.bool_,
+ )
+ attr = DenseElementsAttr.get(array)
+ # CHECK{LITERAL}: dense<[[true, true, false, false, true, true, false, false, false], [false, false, false, true, false, true, true, false, true]]> : tensor<2x9xi1>
+ print(attr)
+ # CHECK{LITERAL}: [[ True True False False True True False False False]
+ # CHECK{LITERAL}: [False False False True False True True False True]]
+ print(np.array(attr))
+
+ array = np.array([], dtype=np.bool_)
+ attr = DenseElementsAttr.get(array)
+ # CHECK: dense<> : tensor<0xi1>
+ print(attr)
+ # CHECK{LITERAL}: []
+ print(np.array(attr))
+
+
### 16 bit integer arrays
# CHECK-LABEL: TEST: testGetDenseElementsI16Signless
@run
``````````
</details>
https://github.com/llvm/llvm-project/pull/115481
More information about the Mlir-commits
mailing list