| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
| |
from proto and verifying it with HloVerifier.
PiperOrigin-RevId: 216447947
|
| |
|
|
|
|
| |
Add a variant of CustomCall which specifies arbitrary layout constraints on the operands and result. The existing non-layout-constrained CustomCall is changed to have no layout preference and can now be assigned arbitrary layouts by layout assignment.
PiperOrigin-RevId: 216249615
|
| |
|
|
| |
PiperOrigin-RevId: 210998142
|
| |
|
|
|
|
|
|
| |
Unlike Printf, StrFormat does not require type-length qualifiers, e.g
%z, %ll. Nor does it require that you call c_str() to print strings.
So these are fixed up here as well.
PiperOrigin-RevId: 210435915
|
| |
|
|
| |
PiperOrigin-RevId: 210049592
|
| |
|
|
| |
PiperOrigin-RevId: 210018843
|
| |
|
|
|
|
|
| |
Unfortunately this has to be one big patch, because e.g. absl::StrCat
doesn't accept a TF StringPiece, but as soon as we switch to
absl::string_view, we have to switch away from all of the TF functions.
PiperOrigin-RevId: 209957896
|
| |
|
|
| |
PiperOrigin-RevId: 207213865
|
| |
|
|
|
|
|
|
| |
This won't catch all overflows, but will do the right thing for the "normal" flow.
Also fix layout validation to reject padded sparse layouts.
PiperOrigin-RevId: 202151215
|
| |
|
|
|
|
|
|
|
|
|
|
| |
A TOKEN primitive type was added with cl/199215963 and XLA also has an OPAQUE primitive type. However, in many places in XLA we assume either a tuple or array. This CL fixes many of those instances, but some may remain. Identified instances were discovered by searching for IsTuple or IsArray so the set of fixes is not exhaustive.
Also opportunistically addressed a couple potential points of confusion in the ShapeUtil interface:
(1) Rename ShapeUtil::HasZeroElements to ShapeUtil::IsZeroElementArray. The point of confusion here is that tuples can also have zero elements and HasZeroElements would check fail on tuple shapes. Method no longer check fails if the given shape is not an array.
(2) ShapeUtil::IsNil now returns true only for empty tuples. Previously it also returned true for zero-element array types which was confusing because ShapeUtil::MakeNil creates an empty tuple.
PiperOrigin-RevId: 200452672
|
| |
|
|
|
|
|
|
| |
The token type will be threaded through side-effecting ops to order them. Subsequent cls will add new opcodes and change side effecting operations to support this ordering.
This CL also does some cleanup in shape_util and layout_util where we have assumed that shapes are either arrays or tuples.
PiperOrigin-RevId: 199215963
|
| |
|
|
|
|
|
|
|
|
| |
Also move AlgorithmPicker after layout assignment, as now
cudnn_convolution_runner will return failures on invalid input layouts.
Also add a backend debug option to switch the layout heuristic. By default
it has the old behavior (all NCHW).
PiperOrigin-RevId: 197983747
|
| |
|
|
|
|
|
| |
These are type aliases of one another; we'd like to be consistent and
use the shorter one.
PiperOrigin-RevId: 196322955
|
| |
|
|
|
|
|
|
|
|
| |
This isn't necessary today, but it will be after an optimization change I'm
about to make.
LLVM has a constant merging pass too, but one of the motivations here is to
avoid the LLVM compile time overhead of having many large arrays in the IR.
PiperOrigin-RevId: 195032900
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Add methods for manipulating sparse literals to xla::Literal
- Make LayoutUtil::HumanString handle sparse layouts
- Make ShapeUtil::ParseShape handle sparse shapes
- Syntax for shapes has changed:
- Old way of expressing layouts still works, e.g. f32[1,2,3]{2,1,0}
- Can now make dense format explicit: f32[1,2,3]dense{2,1,0}
- Can express sparse layouts; the max_sparse_elements value is in the
braces, e.g.: f32[1,2,3]sparse{10}
- The shape should not include braces for the layout if the shape is scalar;
e.g. f32[]{} is not valid shape syntax.
- The shape should not include braces for the layout if the shape is a dense
rank-1 array; e.g. f32[10]{0} is not valid shape syntax
- Sparse literals use a dictionary-liky syntax, e.g.:
f32[2,3,4]sparse{10} {[0,1,2]: 10, [1,2,3]: 11}
- For rank-1 sparse arrays, the square brackets around indices may be omitted, e.g.:
f32[100]sparse{10} {5: 10, 20: 30}
PiperOrigin-RevId: 181813837
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Adds SparseIndexArray and support methods to Literal. SparseIndexArray manages
the array of sparse indices and is exposed by sparse Literals.
Also adds HloSupportChecker classes for CPU and GPU. This will be run as the
first HloPass during compilation, and verifies that the graph is supported by
the backend. Currently only verifies shapes, and that the layout is not sparse
since no backend supports sparse layouts yet.
PiperOrigin-RevId: 181244401
|
| |
|
|
|
|
| |
We should not use const here.
PiperOrigin-RevId: 179614367
|
| |
|
|
|
|
|
|
|
|
|
| |
Format will describe the method used to store array data in memory. Currently
only DENSE is supported, which represents the way XLA currently stores arrays.
Scalars have a DENSE format. Tuples and opaque shapes use INVALID_FORMAT.
Adds checks to code that uses minor_to_major to ensure the layout is dense.
PiperOrigin-RevId: 179475450
|
| |
|
|
|
|
|
|
|
| |
these classes by eliminating a large number of duplicated code.
Removing the old API is non-trivial because of the existing users
outside of tensorflow.
PiperOrigin-RevId: 172920837
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
column major.
The naive dot lowering to LLVM IR (already present in XLA today) is
cache efficient if the dot has LHS of shape [1,K]{1,0} and RHS of
shape [K x N]{0,1}. This change teaches the layout assignment pass to
exploit this property by converting a constant RHS matrix to a column
major layout when possible.
Couple of related things I had to touch in this change:
- In LayoutAssignmentTest.TupleLayout we used to generate a kCopy to satisfy
the conflicting constraints between the result and the constant shapes, but
with this change we change the layout of the constants themselves. So the
EXPECT_FALSE is now an EXPECT_TRUE.
- The extra instruction layout constraints added at the end of
CpuLayoutAssignment::AddBackendConstraints seemed redundant. The layout
assignment pass already tries to make all unconstrained buffers have the
default row-major layout. Moreover, they were blocking this optimization in
some cases by introducing conflicting constraints.
- The changes to literal_util.h have to be made to deal with the
Literal::Relayout calls we now get on literals of various types.
PiperOrigin-RevId: 168761204
|
| |
|
|
|
|
|
|
|
| |
* Mark the evaluator plugin as alwayslink so it doesn't get stripped out by the linker.
* Add a generic LayoutAssignment pass to the pass pipeline; otherwise the entry computation has no layout and Service::Execute CHECK-fails in the AllocationTracker.
* Register the default computation placer for the evaluator backend.
* Add an replay_computation_hlo_evaluator binary that can replay computation snapshots via the HLO evaluator.
PiperOrigin-RevId: 164364780
|
| |
|
|
|
|
|
| |
The default layout is just major-to-minor, as we don't have sufficient testing
for alternative default layouts.
PiperOrigin-RevId: 162766231
|
| |
|
|
|
|
|
|
|
|
|
| |
Define semantics of negative padding in the Pad instruction to be identical to padding inside of convolution operation ConvWithGeneralPadding. Also make negative padding work in the backends.
Specific changes:
(1) Add transformation to algebraic simplifier which replaces negative padding with slices.
(2) fix ReferenceUtil to properly handle negative padding and interior padding.
(3) Add negative padding explanation to operation semantics g3doc.
(4) Extend LayoutsInShapesEqual and CopyLayoutBetweenShapes to work with shapes which are not exactly compatible but have the same rank and tuple structure.
Change: 145355127
|
| |
|
|
|
|
| |
memory as effective reduction to vector
Change: 144353864
|
|
|
XLA is a compiler-based linear algebra execution engine that targets CPUs, GPUs and custom accelerators.
XLA is still experimental; we are releasing it early to get the community involved.
Change: 143990941
|
A. Unique TensorFlower
Mark Heffernan
Tim Shen
Justin Lebar
Chris Leary
Michael Kuperstein
Sanjoy Das
Peter Hawkins
Eli Bendersky